Khalil - Nonlinear Systems Slides

Nonlinear Systems and Control
Lecture # 1
Introduction
– p. 1/1
Nonlinear State Model
ẋ1 = f1 (t, x1 , . . . , xn , u1 , . . . , up )
ẋ2 = f2 (t, x1 , . . . , xn , u1 , . . . , up )
.. ..
. .
ẋn = fn (t, x1 , . . . , xn , u1 , . . . , up )
ẋi denotes the derivative of xi with respect to the time

variable t
u1 , u2 , . . ., up are input variables
x1 , x2 , . . ., xn the state variables
– p. 2/1
   
x1 f1 (t, x, u)
 
u1
   
   
   

 x2 




  f2 (t, x, u)



 
u
   
 2 
     
 ..   .. 
x= . , u =   , f (t, x, u) =  .
   
   .  



  ..  



.. ..
     
.
 



   
 . 




 u p




xn fn (t, x, u)
ẋ = f (t, x, u)
– p. 3/1
ẋ = f (t, x, u)
y = h(t, x, u)
x is the state, u is the input

y is the output (q -dimensional vector)
Special Cases:
Linear systems:
ẋ = A(t)x + B(t)u
y = C(t)x + D(t)u
Unforced state equation:
ẋ = f (t, x)
Results from ẋ = f (t, x, u) with u = γ(t, x)

– p. 4/1
Autonomous System:
ẋ = f (x)
Time-Invariant System:
ẋ = f (x, u)
y = h(x, u)
A time-invariant state model has a time-invariance property

with respect to shifting the initial time from t0 to t0 + a,
provided the input waveform is applied from t0 + a rather
than t0
– p. 5/1
Existence and Uniqueness of Solutions
ẋ = f (t, x)
f (t, x) is piecewise continuous in t and locally Lipschitz in

x over the domain of interest
f (t, x) is piecewise continuous in t on an interval J ⊂ R if

for every bounded subinterval J0 ⊂ J , f is continuous in t
for all t ∈ J0 , except, possibly, at a finite number of points
where f may have finite-jump discontinuities
f (t, x) is locally Lipschitz in x at a point x0 if there is a

neighborhood N (x0 , r) = {x ∈ Rn | kx − x0 k < r}
where f (t, x) satisfies the Lipschitz condition
kf (t, x) − f (t, y)k ≤ Lkx − yk, L > 0
– p. 6/1
A function f (t, x) is locally Lipschitz in x on a domain
(open and connected set) D ⊂ Rn if it is locally Lipschitz at
every point x0 ∈ D
When n = 1 and f depends only on x
|f (y) − f (x)|
≤L
|y − x|
On a plot of f (x) versus x, a straight line joining any two

points of f (x) cannot have a slope whose absolute value is
greater than L
Any function f (x) that has infinite slope at some point is

not locally Lipschitz at that point
– p. 7/1
A discontinuous function is not locally Lipschitz at the points
of discontinuity
The function f (x) = x1/3 is not locally Lipschitz at x = 0

since
f ′ (x) = (1/3)x−2/3 → ∞ a x → 0
On the other hand, if f ′ (x) is continuous at a point x0 then
f (x) is locally Lipschitz at the same point because
continuity of f ′ (x) ensures that |f ′ (x)| is bounded by a
constant k in a neighborhood of x0 ; which implies that
f (x) satisfies the Lipschitz condition L = k
More generally, if for t ∈ J ⊂ R and x in a domain

D ⊂ Rn , f (t, x) and its partial derivatives ∂fi /∂xj are
continuous, then f (t, x) is locally Lipschitz in x on D
– p. 8/1
Lemma: Let f (t, x) be piecewise continuous in t and
locally Lipschitz in x at x0 , for all t ∈ [t0 , t1 ]. Then, there is
δ > 0 such that the state equation ẋ = f (t, x), with
x(t0 ) = x0 , has a unique solution over [t0 , t0 + δ]
Without the local Lipschitz condition, we cannot ensure

uniqueness of the solution. For example, ẋ = x1/3 has
x(t) = (2t/3)3/2 and x(t) ≡ 0 as two different solutions
when the initial state is x(0) = 0
The lemma is a local result because it guarantees existence

and uniqueness of the solution over an interval [t0 , t0 + δ],
but this interval might not include a given interval [t0 , t1 ].
Indeed the solution may cease to exist after some time
– p. 9/1
Example:
ẋ = −x2
f (x) = −x2 is locally Lipschitz for all x
1
x(0) = −1 ⇒ x(t) =
(t − 1)
x(t) → −∞ as t → 1
the solution has a finite escape time at t = 1
In general, if f (t, x) is locally Lipschitz over a domain D

and the solution of ẋ = f (t, x) has a finite escape time te ,
then the solution x(t) must leave every compact (closed
and bounded) subset of D as t → te
– p. 10/1
Global Existence and Uniqueness
A function f (t, x) is globally Lipschitz in x if
kf (t, x) − f (t, y)k ≤ Lkx − yk
for all x, y ∈ Rn with the same Lipschitz constant L
If f (t, x) and its partial derivatives ∂fi /∂xj are continuous

for all x ∈ Rn , then f (t, x) is globally Lipschitz in x if and
only if the partial derivatives ∂fi /∂xj are globally bounded,
uniformly in t
f (x) = −x2 is locally Lipschitz for all x but not globally

Lipschitz because f ′ (x) = −2x is not globally bounded
– p. 11/1
globally Lipschitz in x for all t ∈ [t0 , t1 ]. Then, the state
equation ẋ = f (t, x), with x(t0 ) = x0 , has a unique
solution over [t0 , t1 ]
The global Lipschitz condition is satisfied for linear systems

of the form
ẋ = A(t)x + g(t)
but it is a restrictive condition for general nonlinear systems
– p. 12/1
locally Lipschitz in x for all t ≥ t0 and all x in a domain
D ⊂ Rn . Let W be a compact subset of D , and suppose
that every solution of
ẋ = f (t, x), x(t0 ) = x0
with x0 ∈ W lies entirely in W . Then, there is a unique

solution that is defined for all t ≥ t0
– p. 13/1
Example:
ẋ = −x3 = f (x)
f (x) is locally Lipschitz on R, but not globally Lipschitz
because f ′ (x) = −3x2 is not globally bounded
If, at any instant of time, x(t) is positive, the derivative ẋ(t)

will be negative. Similarly, if x(t) is negative, the derivative
ẋ(t) will be positive
Therefore, starting from any initial condition x(0) = a, the

solution cannot leave the compact set {x ∈ R | |x| ≤ |a|}
Thus, the equation has a unique solution for all t ≥ 0
– p. 14/1
Equilibrium Points
A point x = x∗ in the state space is said to be an

equilibrium point of ẋ = f (t, x) if
x(t0 ) = x∗ ⇒ x(t) ≡ x∗ , ∀ t ≥ t0
For the autonomous system ẋ = f (x), the equilibrium

points are the real solutions of the equation
f (x) = 0
An equilibrium point could be isolated; that is, there are no

other equilibrium points in its vicinity, or there could be a
continuum of equilibrium points
– p. 15/1
A linear system ẋ = Ax can have an isolated equilibrium
point at x = 0 (if A is nonsingular) or a continuum of
equilibrium points in the null space of A (if A is singular)
It cannot have multiple isolated equilibrium points , for if xa

and xb are two equilibrium points, then by linearity any point
on the line αxa + (1 − α)xb connecting xa and xb will be
an equilibrium point
A nonlinear state equation can have multiple isolated

equilibrium points .For example, the state equation
ẋ1 = x2 , ẋ2 = −a sin x1 − bx2
has equilibrium points at (x1 = nπ, x2 = 0) for

n = 0, ±1, ±2, · · ·
– p. 16/1
Linearization
A common engineering practice in analyzing a nonlinear

system is to linearize it about some nominal operating point
and analyze the resulting linear model
What are the limitations of linearization?

Since linearization is an approximation in the
neighborhood of an operating point, it can only predict
the “local” behavior of the nonlinear system in the
vicinity of that point. It cannot predict the “nonlocal” or
“global” behavior
There are “essentially nonlinear phenomena” that can

take place only in the presence of nonlinearity
– p. 17/1
Nonlinear Phenomena
Finite escape time
Multiple isolated equilibrium points
Limit cycles
Subharmonic, harmonic, or almost-periodic oscillations
Chaos
Multiple modes of behavior
– p. 18/1
Lecture # 2
Examples of Nonlinear Systems
– p. 1/1
Pendulum Equation
l
θ
•
mg
mlθ̈ = −mg sin θ − klθ̇
x1 = θ, x2 = θ̇
– p. 2/1
ẋ1 = x2
g k
ẋ2 = − sin x1 − x2
l m
Equilibrium Points:
0 = x2
g k
0 = − sin x1 − x2
l m
(nπ, 0) for n = 0, ±1, ±2, . . .
Nontrivial equilibrium points at (0, 0) and (π, 0)
– p. 3/1
Pendulum without friction:
ẋ1 = x2
g
ẋ2 = − sin x1
l
Pendulum with torque input:
ẋ1 = x2
g k 1
ẋ2 = − sin x1 − x2 + T
l m ml2
– p. 4/1
Tunnel-Diode Circuit
iL L
s i,mA
+ vL
X
X

1
CC iC CC iR i=h(v)
R
P
P

P
P
P

P
P

P

0.5
P
P
+ +
vC C J
J
J
vR 0
E
−0.5
(a) 0 0.5 1 v,V
(b)
dvC diL
iC = C , vL = L
dt dt
x1 = vC , x2 = iL , u=E
– p. 5/1
iC + iR − iL = 0 ⇒ iC = −h(x1 ) + x2
vC − E + RiL + vL = 0 ⇒ vL = −x1 − Rx2 + u
1
ẋ1 = [−h(x1 ) + x2 ]
C
1
ẋ2 = [−x1 − Rx2 + u]
L
Equilibrium Points:
0 = −h(x1 ) + x2
0 = −x1 − Rx2 + u
– p. 6/1
E 1
h(x1 ) = − x1
R R
i
R
1.2
1
0.8
Q
Q 2
0.6 1
0.4
Q
0.2 3
0
0 0.5 1 v
R
– p. 7/1
Mass–Spring System
Ff
Fsp
B B B m F-
B B B p -y
mÿ + Ff + Fsp = F
Sources of nonlinearity:
Nonlinear spring restoring force Fsp = g(y)
Static or Coulomb friction

– p. 8/1
Fsp = g(y)
g(y) = k(1 − a2 y 2 )y, |ay| < 1 (softening spring)

g(y) = k(1 + a2 y 2 )y (hardening spring)
Ff may have components due to static, Coulomb, and
viscous friction
When the mass is at rest, there is a static friction force Fs

that acts parallel to the surface and is limited to ±µs mg
(0 < µs < 1). Fs takes whatever value, between its limits,
to keep the mass at rest
Once motion has started, the resistive force Ff is modeled

as a function of the sliding velocity v = ẏ
– p. 9/1
Ff Ff
v v
(a) (b)
F Ff
f
v v
(c) (d)
(a) Coulomb friction; (b) Coulomb plus linear viscous friction; (c) static, Coulomb, and linear
viscous friction; (d) static, Coulomb, and linear viscous friction—Stribeck effect
– p. 10/1
Negative-Resistance Oscillator
i

X
X

+ i = h(v)

C L

v Resistive
Element
CC
iC CC
iL v
(a)
(b)
h(0) = 0, h′ (0) < 0
h(v) → ∞ as v → ∞, and h(v) → −∞ as v → −∞

– p. 11/1
iC + iL + i = 0
dv 1
Z t
C + v(s) ds + h(v) = 0
dt L −∞
Differentiating with respect to t and multiplying by L:
d2 v ′ dv
CL + v + Lh (v) =0
dt2 dt
√
τ = t/ CL
dv √ dv d2 v d2 v
= CL , = CL
dτ dt dτ 2 dt2
– p. 12/1
Denote the derivative of v with respect to τ by v̇
′
p
v̈ + εh (v)v̇ + v = 0, ε = L/C
Special case: Van der Pol equation
h(v) = −v + 13 v 3
v̈ − ε(1 − v 2 )v̇ + v = 0
State model: x1 = v, x2 = v̇
ẋ1 = x2
ẋ2 = −x1 − εh′ (x1 )x2
– p. 13/1
Another State Model: z1 = iL , z2 = vC
1
ż1 = z2
ε
ż2 = −ε[z1 + h(z2 )]
Change of variables: z = T (x)
x1 = v = z2
√
r
dv dv L
x2 = = CL = [−iL − h(vC )]
dτ dt C
= ε[−z1 − h(z2 )]
" # " #
−h(x1 ) − 1ε x2 −1 z2
T (x) = , T (z) =
x1 −εz1 − εh(z2 )
– p. 14/1
Adaptive Control
P lant : ẏp = ap yp + kp u
Ref erenceM odel : ẏm = am ym + km r

u(t) = θ1∗ r(t) + θ2∗ yp (t)
km am − ap
θ1∗ = and θ2∗ =
kp kp
When ap and kp are unknown, we may use
u(t) = θ1 (t)r(t) + θ2 (t)yp (t)
where θ1 (t) and θ2 (t) are adjusted on-line
– p. 15/1
Adaptive Law (gradient algorithm):
θ̇1 = −γ(yp − ym )r
θ̇2 = −γ(yp − ym )yp , γ>0
State Variables: eo = yp − ym , φ1 = θ1 − θ1∗ , φ2 = θ2 − θ2∗

ẏm = ap ym + kp (θ1∗ r + θ2∗ ym )
ẏp = ap yp + kp (θ1 r + θ2 yp )
ėo = ap eo + kp (θ1 − θ1∗ )r + kp (θ2 yp − θ2∗ ym )

= · · · · · · + kp [θ2∗ yp − θ2∗ yp ]
= (ap + kp θ2∗ )eo + kp (θ1 − θ1∗ )r + kp (θ2 − θ2∗ )yp
– p. 16/1
Closed-Loop System:
ėo = am eo + kp φ1 r(t) + kp φ2 [eo + ym (t)]

φ̇1 = −γeo r(t)
φ̇2 = −γeo [eo + ym (t)]
– p. 17/1
Lecture # 3
Second-Order Systems
– p. 1/?
ẋ1 = f1 (x1 , x2 ) = f1 (x)
ẋ2 = f2 (x1 , x2 ) = f2 (x)
Let x(t) = (x1 (t), x2 (t)) be a solution that starts at initial

state x0 = (x10 , x20 ). The locus in the x1 –x2 plane of the
solution x(t) for all t ≥ 0 is a curve that passes through the
point x0 . This curve is called a trajectory or orbit
The x1 –x2 plane is called the state plane or phase plane
The family of all trajectories is called the phase portrait
The vector field f (x) = (f1 (x), f2 (x)) is tangent to the
trajectory at point x because
dx2 f2 (x)
=
dx1 f1 (x)
– p. 2/?
Vector Field diagram
Represent f (x) as a vector based at x; that is, assign to x

the directed line segment from x to x + f (x)
x2
x + fq (x) = (3, 2)

f (x)

*
q

x = (1, 1)
x1
Repeat at every point in a grid covering the plane
– p. 3/?
6
2
2
0
x
−2
−4
−6
−5 0 5
x
1
ẋ1 = x2 , ẋ2 = −10 sin x1

– p. 4/?
Numerical Construction of the Phase Portrait:
Select a bounding box in the state plane
Select an initial point x0 and calculate the trajectory

through it by solving
ẋ = f (x), x(0) = x0
in forward time (with positive t) and in reverse time (with

negative t)
ẋ = −f (x), x(0) = x0
Repeat the process interactively

Use Simulink or pplane
– p. 5/?
Qualitative Behavior of Linear Systems
ẋ = Ax, A is a 2 × 2 real matrix
x(t) = M exp(Jr t)M −1 x0

" # " # " # " #
λ1 0 λ 0 λ 1 α −β
Jr = or or or
0 λ2 0 λ 0 λ β α
x(t) = M z(t)
ż = Jr z(t)
– p. 6/?
Case 1. Both eigenvalues are real: λ1 6= λ2 6= 0
M = [v1 , v2 ]
v1 & v2 are the real eigenvectors associated with λ1 & λ2
ż1 = λ1 z1 , ż2 = λ2 z2
z1 (t) = z10 eλ1 t , z2 (t) = z20 eλ2 t

λ /λ1
z2 = cz1 2 , c = z20 /(z10 )λ2 /λ1
The shape of the phase portrait depends on the signs of λ1
and λ2
– p. 7/?
λ2 < λ1 < 0
eλ1 t and eλ2 t tend to zero as t → ∞
eλ2 t tends to zero faster than eλ1 t
Call λ2 the fast eigenvalue (v2 the fast eigenvector) and λ1

the slow eigenvalue (v1 the slow eigenvector)
The trajectory tends to the origin along the curve

λ /λ
z2 = cz1 2 1 with λ2 /λ1 > 1
dz2 λ2 [(λ2 /λ1 )−1]

=c z1
dz1 λ1
– p. 8/?
z2
z1
Stable Node
λ2 > λ1 > 0
Reverse arrowheads
Reverse arrowheads =⇒ Unstable Node – p. 9/?
x2 v2 x2 v2
v1 v1
x1 x1
(a) (b)
Stable Node Unstable Node
– p. 10/?
λ2 < 0 < λ1
eλ1 t → ∞, while eλ2 t → 0 as t → ∞

Call λ2 the stable eigenvalue (v2 the stable eigenvector)
and λ1 the unstable eigenvalue (v1 the unstable
eigenvector)
λ /λ1
z2 = cz1 2 , λ2 /λ1 < 0
Saddle
– p. 11/?
z2 x2
v1
v2
z1 x1
(a)
(b)
Phase Portrait of a Saddle Point
– p. 12/?
Case 2. Complex eigenvalues: λ1,2 = α ± jβ
ż1 = αz1 − βz2 , ż2 = βz1 + αz2
−1 z2
q
r = z12 + z22 , θ = tan
z1
r(t) = r0 eαt and θ(t) = θ0 + βt
α < 0 ⇒ r(t) → 0 as t → ∞
α > 0 ⇒ r(t) → ∞ as t → ∞
α = 0 ⇒ r(t) ≡ r0 ∀ t
– p. 13/?
z2 (a) z2 (b) z
2 (c)
z1 z1 z1
α<0 α>0 α=0
Stable Focus Unstable Focus Center
x2 x2 x2
(a) (b) (c)
x1 x x1
1
– p. 14/?
Effect of Perturbations
A → A + δA (δA arbitrarily small)
The eigenvalues of a matrix depend continuously on its

parameters
A node (with distinct eigenvalues), a saddle or a focus is

structurally stable because the qualitative behavior remains
the same under arbitrarily small perturbations in A
A stable node with multiple eigenvalues could become a

stable node or a stable focus under arbitrarily small
perturbations in A
– p. 15/?
A center is not structurally stable
" #
µ 1
−1 µ
Eigenvalues = µ ± j
µ < 0 ⇒ Stable Focus
µ > 0 ⇒ Unstable Focus
– p. 16/?
Lecture # 4
Qualitative Behavior Near
Equilibrium Points
&
Multiple Equilibria
– p. 1/?
The qualitative behavior of a nonlinear system near an
equilibrium point can take one of the patterns we have seen
with linear systems. Correspondingly the equilibrium points
are classified as stable node, unstable node, saddle, stable
focus, unstable focus, or center
Can we determine the type of the equilibrium point of a

nonlinear system by linearization?
– p. 2/?
Let p = (p1 , p2 ) be an equilibrium point of the system
ẋ1 = f1 (x1 , x2 ), ẋ2 = f2 (x1 , x2 )
where f1 and f2 are continuously differentiable

Expand f1 and f2 in Taylor series about (p1 , p2 )
ẋ1 = f1 (p1 , p2 ) + a11 (x1 − p1 ) + a12 (x2 − p2 ) + H.O.T.

ẋ2 = f2 (p1 , p2 ) + a21 (x1 − p1 ) + a22 (x2 − p2 ) + H.O.T.
∂f1 (x1 , x2 ) ∂f1 (x1 , x2 )

a11 = , a12 =
∂x1
x=p ∂x2
x=p
∂f2 (x1 , x2 ) ∂f2 (x1 , x2 )

a21 = , a22 =
∂x1
x=p ∂x2
x=p
– p. 3/?
f1 (p1 , p2 ) = f2 (p1 , p2 ) = 0
y1 = x1 − p1 y2 = x2 − p2
ẏ1 = ẋ1 = a11 y1 + a12 y2 + H.O.T.

ẏ2 = ẋ2 = a21 y1 + a22 y2 + H.O.T.
ẏ ≈ Ay
   ∂f ∂f1

1
a11 a12 ∂x1 ∂x2

∂f
A= = =
   

∂x x=p
 
∂f2 ∂f2
a21 a22 ∂x1 ∂x2

x=p
– p. 4/?
Eigenvalues of A Type of equilibrium point
of the nonlinear system
λ2 < λ1 < 0 Stable Node
λ2 > λ1 > 0 Unstable Node
λ2 < 0 < λ1 Saddle
α ± jβ, α < 0 Stable Focus
α ± jβ, α > 0 Unstable Focus
±jβ Linearization Fails
– p. 5/?
Example
ẋ1 = −x2 − µx1 (x21 + x22 )
ẋ2 = x1 − µx2 (x21 + x22 )
x = 0 is an equilibrium point
" #
∂f −µ(3x21 + x22 ) −(1 + 2µx1 x2 )
=
∂x (1 − 2µx1 x2 ) −µ(x21 + 3x22 )
" #
∂f 0 −1
A= =
∂x x=0
1 0
x1 = r cos θ and x2 = r sin θ ⇒ ṙ = −µr 3 and θ̇ = 1

Stable focus when µ > 0 and Unstable focus when µ < 0
– p. 6/?
For a saddle point, we can use linearization to generate the
stable and unstable trajectories
Let the eigenvalues of the linearization be λ1 > 0 > λ2 and

the corresponding eigenvectors be v1 and v2
The stable and unstable trajectories will be tangent to the

stable and unstable eigenvectors, respectively, as they
approach the equilibrium point p
For the unstable trajectories use x0 = p ± αv1
For the stable trajectories use x0 = p ± αv2
α is a small positive number

– p. 7/?
Multiple Equilibria
Example: Tunnel-diode circuit
ẋ1 = 0.5[−h(x1 ) + x2 ]
ẋ2 = 0.2(−x1 − 1.5x2 + 1.2)
h(x1 ) = 17.76x1 −103.79x21 +229.62x31 −226.31x41 +83.72x51
i
R
1
0.8 Q1 = (0.063, 0.758)

Q Q
0.6 2
1
Q2 = (0.285, 0.61)
0.4
0.2 Q
3 Q3 = (0.884, 0.21)
0
0 0.5 1 v
R
– p. 8/?
" #
∂f −0.5h′ (x1 ) 0.5
=
∂x −0.2 −0.3
" #
−3.598 0.5
A1 = , Eigenvalues : − 3.57, −0.33
−0.2 −0.3
" #
1.82 0.5
A2 = , Eigenvalues : 1.77, −0.25
−0.2 −0.3
" #
−1.427 0.5
A3 = , Eigenvalues : − 1.33, −0.4
−0.2 −0.3
Q1 is a stable node; Q2 is a saddle; Q3 is a stable node

– p. 9/?
2 3 4 5
x ’ = 0.5 ( − 17.76 x + 103.79 x − 229.62 x + 226.31 x − 83.72 x + y)
y ’ = 0.2 ( − x − 1.5 y + 1.2)
1.5
1
y
0.5
−0.5
−0.5 0 0.5 1 1.5
x
– p. 10/?
Hysteresis characteristics of the tunnel-diode circuit
u = E, y = vR
i y
1.2
R
F
1 1 C
0.8 0.8
Q Q 0.6
0.6 1 2
D
0.4 0.4
0.2 Q 0.2
3 B
E A
0
0 0 1 2 3
0 0.5 1 v
R u
– p. 11/?
Lecture # 5
Limit Cycles
– p. 1/?
Oscillation: A system oscillates when it has a nontrivial
periodic solution
x(t + T ) = x(t), ∀ t ≥ 0
Linear (Harmonic) Oscillator:

" #
0 −β
ż = z
β 0
z1 (t) = r0 cos(βt + θ0 ), z2 (t) = r0 sin(βt + θ0 )
−1 z2 (0)
q
r0 = z12 (0) + z22 (0), θ0 = tan
z1 (0)
– p. 2/?
The linear oscillation is not practical because
It is not structurally stable. Infinitesimally small
perturbations may change the type of the equilibrium
point to a stable focus (decaying oscillation) or unstable
focus (growing oscillation)
The amplitude of oscillation depends on the initial

conditions
The same problems exist with oscillation of nonlinear

systems due to a center equilibrium point (e.g., pendulum
without friction)
– p. 3/?
1
Limit Cycles:
Example: Negative Resistance Oscillator
i

X
X

+ i = h(v)

C L

v Resistive
Element
CC
iC CC
iL v
(a)
(b)
– p. 4/?
ẋ1 = x2
ẋ2 = −x1 − εh′ (x1 )x2
There is a unique equilibrium point at the origin

 
0 1
∂f

A= =
 
∂x x=0

−1 −εh′ (0)
λ2 + εh′ (0)λ + 1 = 0
h′ (0) < 0 ⇒ Unstable Focus or Unstable Node
– p. 5/?
Energy Analysis:
E = 12 CvC
2
+ 12 Li2L
1
vC = x1 and iL = −h(x1 ) − x2
ε
E = 12 C{x21 + [εh(x1 ) + x2 ]2 }
Ė = C{x1 ẋ1 + [εh(x1 ) + x2 ][εh′ (x1 )ẋ1 + ẋ2 ]}

= C{x1 x2 + [εh(x1 ) + x2 ][εh′ (x1 )x2 − x1 − εh′ (x1 )x2 ]}
= C[x1 x2 − εx1 h(x1 ) − x1 x2 ]
= −εCx1 h(x1 )
– p. 6/?
−a x
b 1
Ė = −εCx1 h(x1 )
– p. 7/?
Example: Van der Pol Oscillator
ẋ1 = x2
ẋ2 = −x1 + ε(1 − x21 )x2
4 x2 x
3 2
3
2 2
1 1
0 0
x x
1 −1 1
−1
−2 −2
−3
−2 0 2 4 −2 0 2 4
(a) (b)
ε = 0.2 ε=1
– p. 8/?
1
ż1 = z2
ε
ż2 = −ε(z1 − z2 + 13 z23 )
3
10 x z
2 2
2
5 1
0
0 z
−1 1
x
1
−5 −2
−3
−5 0 5 10 −2 0 2
(a) (b)
ε=5
– p. 9/?
x x
2 2
x x
1 1
(a) (b)
Stable Limit Cycle Unstable Limit Cycle
– p. 10/?
Example: Wien-Bridge Oscillator
s R1
B B B
C1
+ v1
#
BB BB BB
"!
+ +
v2 C2 R2 gv
P
P

( 2)
P
P
P

P
P
P

P
P
Equivalent Circuit
– p. 11/?
State variables x1 = v1 and x2 = v2
1
ẋ1 = [−x1 + x2 − g(x2 )]
C1 R1
1 1
ẋ2 = − [−x1 + x2 − g(x2 )] − x2
C2 R1 C2 R2
There is a unique equilibrium point at x = 0
Numerical data: C1 = C2 = R1 = R2 = 1
g(v) = 3.234v − 2.195v 3 + 0.666v 5
– p. 12/?
x ’ = − x + y − (3.234 y − 2.195 y3 + 0.666 y5)
y ’ = − ( − x + y − (3.234 y − 2.195 y3 + 0.666 y5)) − y
0.8
0.6
0.4
0.2
0
y
−0.2
−0.4
−0.6
−0.8
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1

x
– p. 13/?
x ’ = − x + y − (3.234 y − 2.195 y3 + 0.666 y5)
y ’ = − ( − x + y − (3.234 y − 2.195 y3 + 0.666 y5)) − y
0
y
−2
−4
−6
−6 −4 −2 0 2 4 6
x
– p. 14/?
Lecture # 6
Bifurcation
– p. 1/?
Bifurcation is a change in the equilibrium points or periodic
orbits, or in their stability properties, as a parameter is
varied
Example
ẋ1 = µ − x21
ẋ2 = −x2
Find the equilibrium points and their types for different

values of µ
√
For µ > 0 there are two equilibrium points at ( µ, 0) and
√
(− µ, 0)
– p. 2/?
√
Linearization at ( µ, 0):
" √ #
−2 µ 0
0 −1
√
( µ, 0) is a stable node
√
Linearization at (− µ, 0):
" √ #
2 µ 0
0 −1
√
(− µ, 0) is a saddle
– p. 3/?
ẋ1 = µ − x21 , ẋ2 = −x2
No equilibrium points when µ < 0
As µ decreases, the saddle and node approach each other,

collide at µ = 0, and disappear for µ < 0
x x x
2 2 2
x x x
1 1 1
µ>0 µ=0 µ<0
– p. 4/?
µ is called the bifurcation parameter and µ = 0 is the
bifurcation point
Bifurcation Diagram
µ
(a) Saddle−node bifurcation
– p. 5/?
Example
ẋ1 = µx1 − x21 , ẋ2 = −x2
Two equilibrium points at (0, 0) and (µ, 0)
" #
µ 0
The Jacobian at (0, 0) is
0 −1
(0, 0) is a stable node for µ < 0 and a saddle for µ > 0

" #
−µ 0
The Jacobian at (µ, 0) is
0 −1
(µ, 0) is a saddle for µ < 0 and a stable node for µ > 0

An eigenvalue crosses the origin as µ crosses zero
– p. 6/?
While the equilibrium points persist through the bifurcation
point µ = 0, (0, 0) changes from a stable node to a saddle
and (µ, 0) changes from a saddle to a stable node
µ µ
(a) Saddle−node bifurcation (b) Transcritical bifurcation
dangerous or hard safe or soft
– p. 7/?
Example
ẋ1 = µx1 − x31 , ẋ2 = −x2
For µ < 0, there is a stable node at the origin
For µ > 0, there are three equilibrium points: a saddle at

√ √
(0, 0) and stable nodes at ( µ, 0), and (− µ, 0)
µ
(c) Supercritical pitchfork bifurcation
– p. 8/?
Example
ẋ1 = µx1 + x31 , ẋ2 = −x2
For µ < 0, there are three equilibrium points: a stable node
√
at (0, 0) and two saddles at (± −µ, 0)
For µ > 0, there is a saddle at (0, 0)
µ
(d) Subcritical pitchfork bifurcation
– p. 9/?
Notice the difference between supercritical and subcritical
pitchfork bifurcations
µ µ
(c) Supercritical pitchfork bifurcation (d) Subcritical pitchfork bifurcation
safe or soft dangerous or hard
– p. 10/?
Example: Tunnel diode Circuit
1
ẋ1 = [−h(x1 ) + x2 ]
C
1
ẋ2 = [−x1 − Rx2 + µ]
L
x2
x = h(x )
2 1
A B x1 A B µ
(a) (b)
– p. 11/?
Example
ẋ1 = x1 (µ − x21 − x22 ) − x2

ẋ2 = x2 (µ − x21 − x22 ) + x1

" #
µ −1
Linearization:
1 µ
Stable focus for µ < 0, and unstable focus for µ > 0
A pair of complex eigenvalues cross the imaginary axis as

µ crosses zero
– p. 12/?
ṙ = µr − r 3 and θ̇ = 1
√
For µ > 0, there is a stable limit cycle at r = µ
x x
2 2
x x
1 1
µ<0 µ>0
Supercritical Hopf bifurcation
– p. 13/?
Example
2 2 2 2 2

ẋ1 = x1 µ + (x1 + x2 ) − (x1 + x2 ) − x2
2 2 2 2 2

ẋ2 = x2 µ + (x1 + x2 ) − (x1 + x2 ) + x1

" #
µ −1
Linearization:
1 µ
Stable focus for µ < 0, and unstable focus for µ > 0
A pair of complex eigenvalues cross the imaginary axis as

µ crosses zero
– p. 14/?
ṙ = µr + r 3 − r 5 and θ̇ = 1
Sketch of µr + r 3 − r 5 :
r r
µ<0 µ>0
For small |µ|, the stable limit cycles are approximated by

r = 1, while the unstable limit cycle for µ < 0 is
p
approximated by r = |µ|
– p. 15/?
As µ increases from negative to positive values, the stable
focus at the origin merges with the unstable limit cycle and
bifurcates into unstable focus
Subcritical Hopf bifurcation
µ µ
(e) Supercritical Hopf bifurcation (f) Subcrtitical Hopf bifurcation
safe or soft dangerous or hard
– p. 16/?
All six types of bifurcation occur in the vicinity of an
equilibrium point. They are called local bifurcations
Example of Global Bifurcation
ẋ1 = x2
ẋ2 = µx2 + x1 − x21 + x1 x2
There are two equilibrium points at (0, 0) and (1, 0). By

linearization, we can see that (0, 0) is always a saddle,
while (1, 0) is an unstable focus for −1 < µ < 1
Limit analysis to the range −1 < µ < 1
– p. 17/?
x x
2 2
µ=−0.95 µ=−0.88
x x
1 1
x2 x2
µ=−0.8645 µ=−0.8
x x
1 1
Saddle–connection (or homoclinic) bifurcation
– p. 18/?
Lecture # 7
Stability of Equilibrium Points
Basic Concepts & Linearization
– p. 1/?
ẋ = f (x)
f is locally Lipschitz over a domain D ⊂ Rn
Suppose x̄ ∈ D is an equilibrium point; that is, f (x̄) = 0
Characterize and study the stability of x̄
For convenience, we state all definitions and theorems for

the case when the equilibrium point is at the origin of Rn ;
that is, x̄ = 0. No loss of generality
y = x − x̄
def
ẏ = ẋ = f (x) = f (y + x̄) = g(y), where g(0) = 0
– p. 2/?
Definition: The equilibrium point x = 0 of ẋ = f (x) is
stable if for each ε > 0 there is δ > 0 (dependent on ε)
such that
kx(0)k < δ ⇒ kx(t)k < ε, ∀ t ≥ 0
unstable if it is not stable
asymptotically stable if it is stable and δ can be chosen

such that
kx(0)k < δ ⇒ lim x(t) = 0

t→∞
– p. 3/?
First-Order Systems (n = 1)
The behavior of x(t) in the neighborhood of the origin can

be determined by examining the sign of f (x)
The ε–δ requirement for stability is violated if xf (x) > 0 on

either side of the origin
f(x) f(x) f(x)
x x x
Unstable Unstable Unstable
– p. 4/?
The origin is stable if and only if xf (x) ≤ 0 in some
neighborhood of the origin
f(x) f(x) f(x)
x x x
Stable Stable Stable
– p. 5/?
The origin is asymptotically stable if and only if xf (x) < 0
in some neighborhood of the origin
f(x) f(x)
−a b x x
(a) (b)
Asymptotically Stable Globally Asymptotically Stable
– p. 6/?
Definition: Let the origin be an asymptotically stable
equilibrium point of the system ẋ = f (x), where f is a
locally Lipschitz function defined over a domain D ⊂ Rn
( 0 ∈ D)
The region of attraction (also called region of
asymptotic stability, domain of attraction, or basin) is the
set of all points x0 in D such that the solution of
ẋ = f (x), x(0) = x0
is defined for all t ≥ 0 and converges to the origin as t

tends to infinity
The origin is said to be globally asymptotically stable if

the region of attraction is the whole space Rn
– p. 7/?
Second-Order Systems (n = 2)
Type of equilibrium point Stability Property

Center
Stable Node
Stable Focus
Unstable Node
Unstable Focus
Saddle
– p. 8/?
Example: Tunnel Diode Circuit
x ’ = 0.5 ( − 17.76 x + 103.79 x2 − 229.62 x3 + 226.31 x4 − 83.72 x5 + y)
y ’ = 0.2 ( − x − 1.5 y + 1.2)
1.5
1
y
0.5
−0.5
−0.5 0 0.5 1 1.5
x
– p. 9/?
Example: Pendulum Without Friction
x’=y
y ’ = − 10 sin(x)
0
y
−2
−4
−6
−8
−4 −3 −2 −1 0 1 2 3 4
x
– p. 10/?
Example: Pendulum With Friction
x’=y
y ’ = − 10 sin(x) − y
0
y
−2
−4
−6
−8
−4 −3 −2 −1 0 1 2 3 4
x
– p. 11/?
Linear Time-Invariant Systems
ẋ = Ax
x(t) = exp(At)x(0)
P −1 AP = J = block diag[J1 , J2 , . . . , Jr ]
 
λi 1 0 ... ... 0
 0 λ 1 0 . . . 0 
 i 
 . ... .. 
 .. . 
Ji =  .
 
. . . .

 . 0 
 ..
 
 . . . . 1 

0 . . . . . . . . . 0 λi m×m
– p. 12/?
r X
X mi
exp(At) = P exp(Jt)P −1 = tk−1 exp(λi t)Rik
i=1 k=1
mi is the order of the Jordan block Ji
Re[λi ] < 0 ∀ i ⇔ Asymptotically Stable
Re[λi ] > 0 for some i ⇒ Unstable

Re[λi ] ≤ 0 ∀ i & mi > 1 for Re[λi ] = 0 ⇒ Unstable
Re[λi ] ≤ 0 ∀ i & mi = 1 for Re[λi ] = 0 ⇒ Stable
If an n × n matrix A has a repeated eigenvalue λi of
algebraic multiplicity qi , then the Jordan blocks of λi have
order one if and only if rank(A − λi I) = n − qi
– p. 13/?
Theorem: The equilibrium point x = 0 of ẋ = Ax is stable if
and only if all eigenvalues of A satisfy Re[λi ] ≤ 0 and for
every eigenvalue with Re[λi ] = 0 and algebraic multiplicity
qi ≥ 2, rank(A − λi I) = n − qi , where n is the dimension
of x. The equilibrium point x = 0 is globally asymptotically
stable if and only if all eigenvalues of A satisfy Re[λi ] < 0
When all eigenvalues of A satisfy Re[λi ] < 0, A is called a

Hurwitz matrix
When the origin of a linear system is asymptotically stable,

its solution satisfies the inequality
kx(t)k ≤ kkx(0)ke−λt , ∀t≥0
k ≥ 1, λ > 0
– p. 14/?
Exponential Stability
Definition: The equilibrium point x = 0 of ẋ = f (x) is said

to be exponentially stable if
kx(t)k ≤ kkx(0)ke−λt , ∀t≥0
k ≥ 1, λ > 0, for all kx(0)k < c
It is said to be globally exponentially stable if the inequality

is satisfied for any initial state x(0)
Exponential Stability ⇒ Asymptotic Stability
– p. 15/?
Example
ẋ = −x3
The origin is asymptotically stable
x(0)
x(t) = p
1 + 2tx2 (0)
x(t) does not satisfy |x(t)| ≤ ke−λt |x(0)| because
e2λt
|x(t)| ≤ ke−λt |x(0)| ⇒ ≤ k2
1 + 2tx2 (0)
e2λt
Impossible because lim =∞
t→∞ 1+ 2tx2 (0)
– p. 16/?
Linearization
ẋ = f (x), f (0) = 0
f is continuously differentiable over D = {kxk < r}
∂f
J(x) = (x)
∂x
h(σ) = f (σx) for 0 ≤ σ ≤ 1
h′ (σ) = J(σx)x
Z 1
h(1) − h(0) = h′ (σ) dσ, h(0) = f (0) = 0
0
Z 1
f (x) = J(σx) dσ x
0
– p. 17/?
Z 1
f (x) = J(σx) dσ x
0
Set A = J(0) and add and subtract Ax

Z 1
f (x) = [A + G(x)]x, where G(x) = [J(σx) − J(0)] dσ
0
G(x) → 0 as x → 0
This suggests that in a small neighborhood of the origin we
can approximate the nonlinear system ẋ = f (x) by its
linearization about the origin ẋ = Ax
– p. 18/?
Theorem:
The origin is exponentially stable if and only if
Re[λi ] < 0 for all eigenvalues of A
The origin is unstable if Re[λi ] > 0 for some i
Linearization fails when Re[λi ] ≤ 0 for all i, with

Re[λi ] = 0 for some i
Example
ẋ = ax3

∂f 2

A= = 3ax x=0 = 0
∂x x=0

Stable if a = 0; Asymp stable if a < 0; Unstable if a > 0

When a < 0, the origin is not exponentially stable
– p. 19/?
Lecture # 8
Lyapunov Stability
– p. 1/1
Let V (x) be a continuously differentiable function defined in
a domain D ⊂ Rn ; 0 ∈ D . The derivative of V along the
trajectories of ẋ = f (x) is
n n
X ∂V X ∂V
V̇ (x) = ẋi = fi (x)
i=1
∂xi i=1
∂xi
 
f1 (x)
i f2 (x)
h 
∂V ∂V ∂V
 
= ∂x1
, ∂x2
, ... , ∂xn
 .. 
.
 
 
fn (x)
∂V
= f (x)
∂x
– p. 2/1
If φ(t; x) is the solution of ẋ = f (x) that starts at initial
state x at time t = 0, then
d

V̇ (x) = V (φ(t; x))
dt t=0
If V̇ (x) is negative, V will decrease along the solution of

ẋ = f (x)
If V̇ (x) is positive, V will increase along the solution of

ẋ = f (x)
– p. 3/1
Lyapunov’s Theorem:
If there is V (x) such that
V (0) = 0 and V (x) > 0, ∀ x ∈ D/{0}
V̇ (x) ≤ 0, ∀x∈D
then the origin is a stable
Moreover, if
V̇ (x) < 0, ∀ x ∈ D/{0}
then the origin is asymptotically stable
– p. 4/1
Furthermore, if V (x) > 0, ∀ x 6= 0,
kxk → ∞ ⇒ V (x) → ∞
and V̇ (x) < 0, ∀ x 6= 0, then the origin is globally

asymptotically stable
Proof: D 0 < r ≤ ε, Br = {kxk ≤ r}

Br
α = min V (x) > 0
kxk=r
Bδ
Ω 0<β<α
β
Ωβ = {x ∈ Br | V (x) ≤ β}
kxk ≤ δ ⇒ V (x) < β
– p. 5/1
Solutions starting in Ωβ stay in Ωβ because V̇ (x) ≤ 0 in Ωβ
x(0) ∈ Bδ ⇒ x(0) ∈ Ωβ ⇒ x(t) ∈ Ωβ ⇒ x(t) ∈ Br
kx(0)k < δ ⇒ kx(t)k < r ≤ ε, ∀ t ≥ 0

⇒ The origin is stable
Now suppose V̇ (x) < 0 ∀ x ∈ D/{0}. V (x(t) is
monotonically decreasing and V (x(t)) ≥ 0
lim V (x(t)) = c ≥ 0
t→∞
lim V (x(t)) = c ≥ 0 Show that c = 0

t→∞
Suppose c > 0. By continuity of V (x), there is d > 0 such

that Bd ⊂ Ωc . Then, x(t) lies outside Bd for all t ≥ 0
– p. 6/1
γ = − max V̇ (x)
d≤kxk≤r
Z t
V (x(t)) = V (x(0)) + V̇ (x(τ )) dτ ≤ V (x(0)) − γt
0
This inequality contradicts the assumption c > 0
⇒ The origin is asymptotically stable
The condition kxk → ∞ ⇒ V (x) → ∞ implies that the

set Ωc = {x ∈ Rn | V (x) ≤ c} is compact for every c > 0.
This is so because for any c > 0, there is r > 0 such that
V (x) > c whenever kxk > r . Thus, Ωc ⊂ Br . All solutions
starting Ωc will converge to the origin. For any point
p ∈ Rn , choosing c = V (p) ensures that p ∈ Ωc
⇒ The origin is globally asymptotically stable
– p. 7/1
Terminology
V (0) = 0, V (x) ≥ 0 for x 6= 0 Positive semidefinite
V (0) = 0, V (x) > 0 for x 6= 0 Positive definite
V (0) = 0, V (x) ≤ 0 for x 6= 0 Negative semidefinite
V (0) = 0, V (x) < 0 for x 6= 0 Negative definite
kxk → ∞ ⇒ V (x) → ∞ Radially unbounded
Lyapunov’ Theorem: The origin is stable if there is a

continuously differentiable positive definite function V (x) so
that V̇ (x) is negative semidefinite, and it is asymptotically
stable if V̇ (x) is negative definite. It is globally
asymptotically stable if the conditions for asymptotic
stability hold globally and V (x) is radially unbounded
– p. 8/1
A continuously differentiable function V (x) satisfying the
conditions for stability is called a Lyapunov function. The
surface V (x) = c, for some c > 0, is called a Lyapunov
surface or a level surface
c3
c2
V (x) = c 1
c 1 <c 2 <c 3
– p. 9/1
Why do we need the radial unboundedness condition to
show global asymptotic stability?
It ensures that Ωc = {x ∈ Rn | V (x) ≤ c} is bounded for
every c > 0
Without it Ωc might not bounded for large c
Example
x21 2
V (x) = + x 2
1 + x21
x2 c
c
c
c
c
c
hhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhh
c
c x1
c
c
c
c
c
– p. 10/1
Lecture # 9
Lyapunov Stability
– p. 1/1
Quadratic Forms
n X
X n
V (x) = xT P x = pij xi xj , P = PT
i=1 j=1
λmin (P )kxk2 ≤ xT P x ≤ λmax (P )kxk2

P ≥ 0 (Positive semidefinite) if and only if λi (P ) ≥ 0 ∀i
P > 0 (Positive definite) if and only if λi (P ) > 0 ∀i
V (x) is positive definite if and only if P is positive definite
V (x) is positive semidefinite if and only if P is positive
semidefinite
P > 0 if and only if all the leading principal minors of P are
positive
– p. 2/1
Linear Systems
ẋ = Ax
V (x) = xT P x, P = PT > 0
def
V̇ (x) = x P ẋ + ẋ P x = x (P A + A P )x = −xT Qx
T T T T
If Q > 0, then A is Hurwitz
Or choose Q > 0 and solve the Lyapunov equation
P A + AT P = −Q
If P > 0, then A is Hurwitz
Matlab: P = lyap(A′ , Q)
– p. 3/1
Theorem A matrix A is Hurwitz if and only if for any
Q = QT > 0 there is P = P T > 0 that satisfies the
Lyapunov equation
P A + AT P = −Q
Moreover, if A is Hurwitz, then P is the unique solution
Idea of the proof: Sufficiency follows from Lyapunov’s

theorem. Necessity is shown by verifying that
Z ∞
P = exp(AT t)Q exp(At) dt
0
is positive definite and satisfies the Lyapunov equation
– p. 4/1
Linearization
ẋ = f (x) = [A + G(x)]x
G(x) → 0 as x → 0
Suppose A is Hurwitz. Choose Q = QT > 0 and solve the
Lyapunov equation P A + AT P = −Q for P . Use
V (x) = xT P x as a Lyapunov function candidate for
ẋ = f (x)
V̇ (x) = xT P f (x) + f T (x)P x

= xT P [A + G(x)]x + xT [AT + GT (x)]P x
= xT (P A + AT P )x + 2xT P G(x)x
= −xT Qx + 2xT P G(x)x
– p. 5/1
V̇ (x) ≤ −xT Qx + 2kP k kG(x)k kxk2
For any γ > 0, there exists r > 0 such that
kG(x)k < γ, ∀ kxk < r
xT Qx ≥ λmin (Q)kxk2 ⇔ −xT Qx ≤ −λmin (Q)kxk2
V̇ (x) < −[λmin (Q) − 2γkP k]kxk2 , ∀ kxk < r

Choose
λmin (Q)
γ<
2kP k
V (x) = xT P x is a Lyapunov function for ẋ = f (x)
– p. 6/1
We can use V (x) = xT P x to estimate the region of
attraction
Suppose V̇ (x) < 0, ∀ 0 < kxk < r
Take c = min xT P x = λmin (P )r 2

kxk=r
{xT P x < c} ⊂ {kxk < r}

All trajectories starting in the set {xT P x < c} approach the
origin as t tends to ∞. Hence, the set {xT P x < c} is a
subset of the region of attraction (an estimate of the region
of attraction)
– p. 7/1
Example
ẋ1 = −x2
ẋ2 = x1 + (x21 − 1)x2
" #
∂f

0 −1
A= =
∂x x=0
1 −1
√
has eigenvalues (−1 ± j 3)/2. Hence the origin is
" #
T 1.5 −0.5
Take Q = I, P A + A P = −I ⇒ P =
−0.5 1
λmin (P ) = 0.691
– p. 8/1
V (x) = xT P x = 1.5x21 − x1 x2 + x22
V̇ (x) = (3x1 − x2 )(−x2 ) + (−x1 + 2x2 )[x1 + (x21 − 1)x2 ]

= −(x21 + x22 ) − (x31 x2 − 2x21 x22 )
√
2 2 5
V̇ (x) ≤ −kxk +|x1 | |x1 x2 | |x1 −2x2 | ≤ −kxk + kxk4
2
1 2
√
where |x1 | ≤ kxk, |x1 x2 | ≤ 2 kxk , |x1 − 2x2 | ≤ 5kxk
2 2 def 2
V̇ (x) < 0 for 0 < kxk < √ = r
5
2 2
Take c = λmin (P )r = 0.691 × √ = 0.618
5
{V (x) < c} is an estimate of the region of attraction
– p. 9/1
Example:
ẋ = −g(x)
g(0) = 0; xg(x) > 0, ∀ x 6= 0 and x ∈ (−a, a)
Z x
V (x) = g(y) dy
0
∂V
V̇ (x) = [−g(x)] = −g 2 (x) < 0, ∀ x ∈ (−a, a), x 6= 0
∂x
If xg(x) > 0 for all x 6= 0, use

Z x
V (x) = 21 x2 + g(y) dy
0
– p. 10/1
Z x
V (x) = 21 x2 + g(y) dy
0
is positive definite for all x and radially unbounded since

V (x) ≥ 12 x2
V̇ (x) = −xg(x) − g 2 (x) < 0, ∀ x 6= 0
The origin is globally asymptotically stable
– p. 11/1
Example: Pendulum equation without friction
ẋ1 = x2
ẋ2 = − a sin x1
V (x) = a(1 − cos x1 ) + 12 x22

V (0) = 0 and V (x) is positive definite over the domain
−2π < x1 < 2π
V̇ (x) = aẋ1 sin x1 + x2 ẋ2 = ax2 sin x1 − ax2 sin x1 = 0
The origin is stable
Since V̇ (x) ≡ 0, the origin is not asymptotically stable
– p. 12/1
Example: Pendulum equation with friction
ẋ1 = x2
ẋ2 = − a sin x1 − bx2
1
V (x) = a(1 − cos x1 ) + x22
2
V̇ (x) = aẋ1 sin x1 + x2 ẋ2 = − bx22
The origin is stable
V̇ (x) is not negative definite because V̇ (x) = 0 for x2 = 0

irrespective of the value of x1
– p. 13/1
The conditions of Lyapunov’s theorem are only sufficient.
Failure of a Lyapunov function candidate to satisfy the
conditions for stability or asymptotic stability does not mean
that the equilibrium point is not stable or asymptotically
stable. It only means that such stability property cannot be
established by using this Lyapunov function candidate
Try
1 T
V (x) = 2x P x +" a(1 − cos#x"1 ) #
1 p11 p12 x1
= 2 [x1 x2 ] + a(1 − cos x1 )
p12 p22 x2
p11 > 0, p11 p22 − p212 > 0
– p. 14/1
V̇ (x) = (p11 x1 + p12 x2 + a sin x1 ) x2
+ (p12 x1 + p22 x2 ) (−a sin x1 − bx2 )
= a(1 − p22 )x2 sin x1 − ap12 x1 sin x1
+ (p11 − p12 b) x1 x2 + (p12 − p22 b) x22
p22 = 1, p11 = bp12 ⇒ 0 < p12 < b, Take p12 = b/2

V̇ (x) = − 21 abx1 sin x1 − 12 bx22
D = {x ∈ R2 | |x1 | < π}
V (x) is positive definite and V̇ (x) is negative definite over D
Read about the variable gradient method in the textbook
– p. 15/1
Lecture # 10
The Invariance Principle
– p. 1/1
Example: Pendulum equation with friction
ẋ1 = x2
ẋ2 = − a sin x1 − bx2
1
V (x) = a(1 − cos x1 ) + x22
2
V̇ (x) = aẋ1 sin x1 + x2 ẋ2 = − bx22
The origin is stable. V̇ (x) is not negative definite because
V̇ (x) = 0 for x2 = 0 irrespective of the value of x1
However, near the origin, the solution cannot stay

identically in the set {x2 = 0}
– p. 2/1
Definitions: Let x(t) be a solution of ẋ = f (x)
A point p is said to be a positive limit point of x(t) if there is

a sequence {tn }, with limn→∞ tn = ∞, such that
x(tn ) → p as n → ∞
The set of all positive limit points of x(t) is called the

positive limit set of x(t); denoted by L+
If x(t) approaches an asymptotically stable equilibrium

point x̄, then x̄ is the positive limit point of x(t) and L+ = x̄
A stable limit cycle is the positive limit set of every solution

starting sufficiently near the limit cycle
– p. 3/1
A set M is an invariant set with respect to ẋ = f (x) if
x(0) ∈ M ⇒ x(t) ∈ M, ∀ t ∈ R
Examples:
Equilibrium points
Limit Cycles
A set M is a positively invariant set with respect to

ẋ = f (x) if
x(0) ∈ M ⇒ x(t) ∈ M, ∀ t ≥ 0
Example: The set Ωc = {V (x) ≤ c} with V̇ (x) ≤ 0 in Ωc
– p. 4/1
The distance from a point p to a set M is defined by
dist(p, M ) = inf kp − xk
x∈M
x(t) approaches a set M as t approaches infinity, if for

each ε > 0 there is T > 0 such that
dist(x(t), M ) < ε, ∀ t > T
Example: every solution x(t) starting sufficiently near a

stable limit cycle approaches the limit cycle as t → ∞
Notice, however, that x(t) does converge to any specific

point on the limit cycle
– p. 5/1
Lemma: If a solution x(t) of ẋ = f (x) is bounded and
belongs to D for t ≥ 0, then its positive limit set L+ is a
nonempty, compact, invariant set. Moreover, x(t)
approaches L+ as t → ∞
LaSalle’s theorem: Let f (x) be a locally Lipschitz function

defined over a domain D ⊂ Rn and Ω ⊂ D be a compact
set that is positively invariant with respect to ẋ = f (x). Let
V (x) be a continuously differentiable function defined over
D such that V̇ (x) ≤ 0 in Ω. Let E be the set of all points in
Ω where V̇ (x) = 0, and M be the largest invariant set in E .
Then every solution starting in Ω approaches M as t → ∞
– p. 6/1
Proof:
V̇ (x) ≤ in Ω ⇒ V (x(t)) is a decreasing
V (x) is continuous in Ω ⇒ V (x) ≥ b = min V (x)

x∈Ω
⇒ lim V (x(t)) = a
t→∞
x(t) ∈ Ω ⇒ x(t) is bounded ⇒ L+ exists

Moreover, L+ ⊂ Ω and x(t) approaches L+ as t → ∞
For any p ∈ L+ , there is {tn } with limn→∞ tn = ∞ such
that x(tn ) → p as n → ∞
V (x) is continuous ⇒ V (p) = lim V (x(tn )) = a

n→∞
– p. 7/1
V (x) = a on L+ and L+ invariant ⇒ V̇ (x) = 0, ∀ x ∈ L+
L+ ⊂ M ⊂ E ⊂ Ω
x(t) approaches L+ ⇒ x(t) approaches M (as t → ∞)
– p. 8/1
Theorem: Let f (x) be a locally Lipschitz function defined
over a domain D ⊂ Rn ; 0 ∈ D . Let V (x) be a continuously
differentiable positive definite function defined over D such
that V̇ (x) ≤ 0 in D . Let S = {x ∈ D | V̇ (x) = 0}
If no solution can stay identically in S , other than the
trivial solution x(t) ≡ 0, then the origin is asymptotically
stable
Moreover, if Γ ⊂ D is compact and positively invariant,

then it is a subset of the region of attraction
Furthermore, if D = Rn and V (x) is radially

unbounded, then the origin is globally asymptotically
stable
– p. 9/1
Example:
ẋ1 = x2
ẋ2 = −h1 (x1 ) − h2 (x2 )
hi (0) = 0, yhi (y) > 0, for 0 < |y| < a
Z x1
V (x) = h1 (y) dy + 21 x22
0
D = {−a < x1 < a, −a < x2 < a}
V̇ (x) = h1 (x1 )x2 +x2 [−h1 (x1 )−h2 (x2 )] = −x2 h2 (x2 ) ≤ 0
V̇ (x) = 0 ⇒ x2 h2 (x2 ) = 0 ⇒ x2 = 0
S = {x ∈ D | x2 = 0}
– p. 10/1
ẋ1 = x2 , ẋ2 = −h1 (x1 ) − h2 (x2 )
x2 (t) ≡ 0 ⇒ ẋ2 (t) ≡ 0 ⇒ h1 (x1 (t)) ≡ 0 ⇒ x1 (t) ≡ 0

The only solution that can stay identically in S is x(t) ≡ 0
Thus, the origin is asymptotically stable

Ry
Suppose a = ∞ and 0 h1 (z) dz → ∞ as |y| → ∞
R x1
Then, D = R2 and V (x) = 0 h1 (y) dy + 12 x22 is radially
unbounded. S = {x ∈ R2 | x2 = 0} and the only solution
that can stay identically in S is x(t) ≡ 0

– p. 11/1
Example: m-link Robot Manipulator

H Load
H

q
2

r

q
r 1

Two-link Robot Manipulator
– p. 12/1
M (q)q̈ + C(q, q̇)q̇ + D q̇ + g(q) = u
q is an m-dimensional vector of joint positions
u is an m-dimensional control (torque) inputs
M = M T > 0 is the inertia matrix
C(q, q̇)q̇ accounts for centrifugal and Coriolis forces
(Ṁ − 2C)T = −(Ṁ − 2C)
D q̇ accounts for viscous damping; D = D T ≥ 0
g(q) accounts for gravity forces; g(q) = [∂P (q)/∂q]T
P (q) is the total potential energy of the links due to gravity

– p. 13/1
Investigate the use of the (PD plus gravity compensation)
control law
u = g(q) − Kp (q − q ∗ ) − Kd q̇
to stabilize the robot at a desired position q ∗ , where Kp and

Kd are symmetric positive definite matrices
e = q − q∗ , ė = q̇
M ë = M q̈
= −C q̇ − D q̇ − g(q) + u
= −C q̇ − D q̇ − Kp (q − q ∗ ) − Kd q̇
= −C ė − D ė − Kp e − Kd ė
– p. 14/1
M ë = −C ė − D ė − Kp e − Kd ė
V = 21 ėT M (q)ė + 21 eT Kp e
V̇ = ėT M ë + 12 ėT Ṁ ė + eT Kp ė
= −ėT C ė − ėT D ė − ėT Kp e − ėT Kd ė

+ 21 ėT Ṁ ė + eT Kp ė
1 T
= 2 ė (Ṁ − 2C)ė − ėT (Kd + D)ė
= −ėT (Kd + D)ė ≤ 0
– p. 15/1
(Kd + D) is positive definite
V̇ = −ėT (Kd + D)ė = 0 ⇒ ė = 0
M ë = −C ė − D ė − Kp e − Kd ė
ė(t) ≡ 0 ⇒ ë(t) ≡ 0 ⇒ Kp e(t) ≡ 0 ⇒ e(t) ≡ 0
By LaSalle’s theorem the origin (e = 0, ė = 0) is globally
– p. 16/1
Lecture # 11
Exponential Stability
&
Region of Attraction
– p. 1/1
Exponential Stability:
The origin of ẋ = f (x) is exponentially stable if and only if
the linearization of f (x) at the origin is Hurwitz
Theorem: Let f (x) be a locally Lipschitz function defined

over a domain D ⊂ Rn ; 0 ∈ D . Let V (x) be a
continuously differentiable function such that
k1 kxka ≤ V (x) ≤ k2 kxka
V̇ (x) ≤ −k3 kxka

for all x ∈ D , where k1 , k2 , k3 , and a are positive
constants. Then, the origin is an exponentially stable
equilibrium point of ẋ = f (x). If the assumptions hold
globally, the origin will be globally exponentially stable
– p. 2/1
Proof: Choose c > 0 small enough that
{k1 kxka ≤ c} ⊂ D
V (x) ≤ c ⇒ k1 kxka ≤ c
Ωc = {V (x) ≤ c} ⊂ { k1 kxka ≤ c} ⊂ D
Ωc is compact and positively invariant; ∀ x(0) ∈ Ωc
a k3
V̇ ≤ −k3 kxk ≤− V
k2
dV k3
≤− dt
V k2
V (x(t)) ≤ V (x(0))e−(k3 /k2 )t
– p. 3/1
1/a
V (x(t))

kx(t)k ≤
k1
" #1/a
V (x(0))e−(k3 /k2 )t
≤
k1
" #1/a
k2 kx(0)ka e−(k3 /k2 )t
≤
k1
1/a
k2

= e−γt kx(0)k, γ = k3 /(k2 a)
k1
– p. 4/1
Example
ẋ1 = x2
ẋ2 = −h(x1 ) − x2
c1 y 2 ≤ yh(y) ≤ c2 y 2 , ∀ y, c1 > 0, c2 > 0

" # Z x1
1 T 1 1
V (x) = 2 x x+2 h(y) dy
1 2 0
Z x1
c1 x21 ≤ 2 h(y) dy ≤ c2 x21
0
V̇ = [x1 + x2 + 2h(x1 )]x2 + [x1 + 2x2 ][−h(x1 ) − x2 ]

= −x1 h(x1 ) − x22 ≤ −c1 x21 − x22
The origin is globally exponentially stable
– p. 5/1
Region of Attraction
Lemma: If x = 0 is an asymptotically stable equilibrium

point for ẋ = f (x), then its region of attraction RA is an
open, connected, invariant set. Moreover, the boundary of
RA is formed by trajectories
– p. 6/1
Example
ẋ1 = −x2
ẋ2 = x1 + (x21 − 1)x2
4
x
2
2
0
x
1
−2
−4
−4 −2 0 2 4
– p. 7/1
Example
ẋ1 = x2
ẋ2 = −x1 + 13 x31 − x2
4
x
2
2
0
x
1
−2
−4
−4 −2 0 2 4
– p. 8/1
Estimates of the Region of Attraction: Find a subset of the
region of attraction
Warning: Let D be a domain with 0 ∈ D such that for all

x ∈ D , V (x) is positive definite and V̇ (x) is negative
definite
Is D a subset of the region of attraction?
NO
Why?
– p. 9/1
Example: Reconsider
ẋ1 = x2
ẋ2 = −x1 + 13 x31 − x2
" #
1 T 1 1 R x1
V (x) = 2
x x+2 0 (y − 13 y 3 ) dy
1 2
3 2
= x
2 1
− 16 x41 + x1 x2 + x22
V̇ (x) = −x21 (1 − 13 x21 ) − x22

√ √
D = {− 3 < x1 < 3}
Is D a subset of the region of attraction?
– p. 10/1
The simplest estimate is the bounded component of
{V (x) < c}, where c = minx∈∂D V (x)
For V (x) = xT P x, where P = P T > 0, the minimum of

V (x) over ∂D is given by
For D = {kxk < r}, min xT P x = λmin (P )r 2

kxk=r
r2
For D = {|bT x| < r}, min xT P x =
|bT x|=r bT P −1 b
For D = {|bTi x| < ri , i = 1, . . . , p},

ri2
Take c = min ≤ min xT P x
1≤i≤p bTi P −1 bi x∈∂D
– p. 11/1
Example (Revisited)
ẋ1 = −x2
ẋ2 = x1 + (x21 − 1)x2
V (x) = 1.5x21 − x1 x2 + x22
V̇ (x) = −(x21 + x22 ) − (x31 x2 − 2x21 x22 )

2 def 2
2
V̇ (x) < 0 for 0 < kxk < √ = r
5
2 2
Take c = λmin (P )r = 0.691 × √ = 0.618
5
{V (x) < c} is an estimate of the region of attraction
– p. 12/1
x1 = ρ cos θ, x2 = ρ sin θ
V̇ = −ρ2 + ρ4 cos2 θ sin θ(2 sin θ − cos θ)

≤ −ρ2 + ρ4 | cos2 θ sin θ| · |2 sin θ − cos θ|
≤ −ρ2 + ρ4 × 0.3849 × 2.2361
1
≤ −ρ2 + 0.861ρ4 < 0, for ρ2 < 0.861
2 0.691
Take c = λmin (P )r = = 0.803
0.861
Alternatively, choose c as the largest constant such that
{xT P x < c} is a subset of {V̇ (x) < 0}
– p. 13/1
2 x 3 x
2 2
2
1
1
0 x 0 x
1 1
−1
−1
−2
−2 −3
−2 −1 0 1 2 −2 0 2
(a) (b)
(a) Contours of V̇ (x) = 0 (dashed)

V (x) = 0.8 (dash-dot), V (x) = 2.25 (solid)
(b) comparison of the region of attraction with its estimate
– p. 14/1
If D is a domain where V (x) is positive definite and V̇ (x)
is negative definite (or V̇ (x) is negative semidefinite and no
solution can stay identically in the set V̇ (x) = 0 other than
x = 0), then according to LaSalle’s theorem any compact
positively invariant subset of D is a subset of the region of
attraction
Example
ẋ1 = x2
ẋ2 = −4(x1 + x2 ) − h(x1 + x2 )
h(0) = 0; uh(u) ≥ 0, ∀ |u| ≤ 1
– p. 15/1
" #
T T 2 1
V (x) = x P x = x x = 2x21 + 2x1 x2 + x22
1 1
V̇ (x) = (4x1 + 2x2 )ẋ1 + 2(x1 + x2 )ẋ2

= −2x21 − 6(x1 + x2 )2 − 2(x1 + x2 )h(x1 + x2 )
≤ −2x21"− 6(x1# + x2 )2 , ∀ |x1 + x2 | ≤ 1
T 8 6
= −x x
6 6
V̇ (x) is negative definite in {|x1 + x2 | ≤ 1}
T T 1
b = [1 1], c= min x Px = =1
|x1 +x2 |=1 bT P −1 b
– p. 16/1
σ = x1 + x2
d
σ 2 = 2σx2 − 8σ 2 − 2σh(σ) ≤ 2σx2 − 8σ 2 , ∀ |σ| ≤ 1
dt
d
On σ = 1, σ 2 ≤ 2x2 − 8 ≤ 0, ∀ x2 ≤ 4
dt
d
On σ = −1, σ 2 ≤ −2x2 − 8 ≤ 0, ∀ x2 ≥ −4
dt
c1 = V (x)|x1 =−3,x2 =4 = 10, c2 = V (x)|x1 =3,x2 =−4 = 10
Γ = {V (x) ≤ 10 and |x1 + x2 | ≤ 1}

is a subset of the region of attraction
– p. 17/1
5 x2
(−3,4)
V(x) = 10
0 x1
V(x) = 1
(3,−4)
−5
−5 0 5
– p. 18/1
Lecture # 12
Converse Lyapunov Functions
&
Time Varying Systems
– p. 1/1
Converse Lyapunov Theorem–Exponential Stability
Let x = 0 be an exponentially stable equilibrium point for

the system ẋ = f (x), where f is continuously differentiable
on D = {kxk < r}. Let k, λ, and r0 be positive constants
with r0 < r/k such that
kx(t)k ≤ kkx(0)ke−λt , ∀ x(0) ∈ D0 , ∀ t ≥ 0
where D0 = {kxk < r0 }. Then, there is a continuously

differentiable function V (x) that satisfies the inequalities
– p. 2/1
c1 kxk2 ≤ V (x) ≤ c2 kxk2
∂V
f (x) ≤ −c3 kxk2
∂x
∂V

≤ c4 kxk
∂x
for all x ∈ D0 , with positive constants c1 , c2 , c3 , and c4

Moreover, if f is continuously differentiable for all x, globally
Lipschitz, and the origin is globally exponentially stable,
then V (x) is defined and satisfies the aforementioned
inequalities for all x ∈ Rn
– p. 3/1
Idea of the proof: Let ψ(t; x) be the solution of
ẏ = f (y), y(0) = x
Take
Z δ
V (x) = ψ T (t; x) ψ(t; x) dt, δ>0
0
– p. 4/1
Example: Consider the system ẋ = f (x) where f is
continuously differentiable in the neighborhood of the origin
and f (0) = 0. Show that the origin is exponentially stable
only if A = [∂f /∂x](0) is Hurwitz
f (x) = Ax + G(x)x, G(x) → 0 as x → 0
Given any L > 0, there is r1 > 0 such that
kG(x)k ≤ L, ∀ kxk < r1
Because the origin of ẋ = f (x) is exponentially stable, let

V (x) be the function provided by the converse Lyapunov
theorem over the domain {kxk < r0 }. Use V (x) as a
Lyapunov function candidate for ẋ = Ax
– p. 5/1
∂V ∂V ∂V
Ax = f (x) − G(x)x
∂x ∂x ∂x
≤ −c3 kxk2 + c4 Lkxk2
= −(c3 − c4 L)kxk2
def
Take L < c3 /c4 , γ = (c3 − c4 L) > 0 ⇒
∂V
Ax ≤ −γkxk2 , ∀ kxk < min{r0 , r1 }
∂x
The origin of ẋ = Ax is exponentially stable
– p. 6/1
Converse Lyapunov Theorem–Asymptotic Stability
Let x = 0 be an asymptotically stable equilibrium point for

ẋ = f (x), where f is locally Lipschitz on a domain
D ⊂ Rn that contains the origin. Let RA ⊂ D be the region
of attraction of x = 0. Then, there is a smooth, positive
definite function V (x) and a continuous, positive definite
function W (x), both defined for all x ∈ RA , such that
V (x) → ∞ as x → ∂RA
∂V
f (x) ≤ −W (x), ∀ x ∈ RA
∂x
and for any c > 0, {V (x) ≤ c} is a compact subset of RA
When RA = Rn , V (x) is radially unbounded
– p. 7/1
Time-varying Systems
ẋ = f (t, x)
f (t, x) is piecewise continuous in t and locally Lipschitz in

x for all t ≥ 0 and all x ∈ D . The origin is an equilibrium
point at t = 0 if
f (t, 0) = 0, ∀ t ≥ 0
While the solution of the autonomous system

ẋ = f (x), x(t0 ) = x0
depends only on (t − t0 ), the solution of

ẋ = f (t, x), x(t0 ) = x0
may depend on both t and t0

– p. 8/1
Comparison Functions
A scalar continuous function α(r), defined for r ∈ [0, a)
is said to belong to class K if it is strictly increasing and
α(0) = 0. It is said to belong to class K∞ if it defined
for all r ≥ 0 and α(r) → ∞ as r → ∞
A scalar continuous function β(r, s), defined for

r ∈ [0, a) and s ∈ [0, ∞) is said to belong to class KL
if, for each fixed s, the mapping β(r, s) belongs to class
K with respect to r and, for each fixed r , the mapping
β(r, s) is decreasing with respect to s and β(r, s) → 0
as s → ∞
– p. 9/1
Example
α(r) = tan−1 (r) is strictly increasing since
α′ (r) = 1/(1 + r 2 ) > 0. It belongs to class K, but not
to class K∞ since limr→∞ α(r) = π/2 < ∞
α(r) = r c , for any positive real number c, is strictly

increasing since α′ (r) = cr c−1 > 0. Moreover,
limr→∞ α(r) = ∞; thus, it belongs to class K∞
α(r) = min{r, r 2 } is continuous, strictly increasing,

and limr→∞ α(r) = ∞. Hence, it belongs to class K∞
– p. 10/1
β(r, s) = r/(ksr + 1), for any positive real number k,
is strictly increasing in r since
∂β 1
= >0
∂r (ksr + 1)2
and strictly decreasing in s since
∂β −kr 2
= <0
∂s (ksr + 1)2
Moreover, β(r, s) → 0 as s → ∞. Therefore, it belongs

to class KL
β(r, s) = r c e−s , for any positive real number c, belongs

to class KL
– p. 11/1
Definition: The equilibrium point x = 0 of ẋ = f (t, x) is
uniformly stable if there exist a class K function α and a
positive constant c, independent of t0 , such that
kx(t)k ≤ α(kx(t0 )k), ∀ t ≥ t0 ≥ 0, ∀ kx(t0 )k < c
uniformly asymptotically stable if there exist a class KL

function β and a positive constant c, independent of t0 ,
such that
kx(t)k ≤ β(kx(t0 )k, t−t0 ), ∀ t ≥ t0 ≥ 0, ∀ kx(t0 )k < c
globally uniformly asymptotically stable if the foregoing

inequality is satisfied for any initial state x(t0 )
– p. 12/1
exponentially stable if there exist positive constants c,
k, and λ such that
kx(t)k ≤ kkx(t0 )ke−λ(t−t0 ) , ∀ kx(t0 )k < c
globally exponentially stable if the foregoing inequality

is satisfied for any initial state x(t0 )
– p. 13/1
Theorem: Let the origin x = 0 be an equilibrium point for
ẋ = f (t, x) and D ⊂ Rn be a domain containing x = 0.
Suppose f (t, x) is piecewise continuous in t and locally
Lipschitz in x for all t ≥ 0 and x ∈ D . Let V (t, x) be a
(1) W1 (x) ≤ V (t, x) ≤ W2 (x)
∂V ∂V
(2) + f (t, x) ≤ 0
∂t ∂x
for all t ≥ 0 and x ∈ D , where W1 (x) and W2 (x) are
continuous positive definite functions on D . Then, the origin
is uniformly stable
– p. 14/1
Theorem: Suppose the assumptions of the previous
theorem are satisfied with
∂V ∂V
+ f (t, x) ≤ −W3 (x)
∂t ∂x
for all t ≥ 0 and x ∈ D , where W3 (x) is a continuous
positive definite function on D . Then, the origin is uniformly
asymptotically stable. Moreover, if r and c are chosen such
that Br = {kxk ≤ r} ⊂ D and c < minkxk=r W1 (x), then
every trajectory starting in {x ∈ Br | W2 (x) ≤ c} satisfies
kx(t)k ≤ β(kx(t0 )k, t − t0 ), ∀ t ≥ t0 ≥ 0
for some class KL function β . Finally, if D = Rn and

W1 (x) is radially unbounded, then the origin is globally
uniformly asymptotically stable
– p. 15/1
Theorem: Suppose the assumptions of the previous
theorem are satisfied with
k1 kxka ≤ V (t, x) ≤ k2 kxka
∂V ∂V
+ f (t, x) ≤ −k3 kxka
∂t ∂x
for all t ≥ 0 and x ∈ D , where k1 , k2 , k3 , and a are
positive constants. Then, the origin is exponentially stable.
If the assumptions hold globally, the origin will be globally
exponentially stable.
– p. 16/1
Example:
ẋ = −[1 + g(t)]x3 , g(t) ≥ 0, ∀ t ≥ 0
V (x) = 12 x2
V̇ (t, x) = −[1 + g(t)]x4 ≤ −x4 , ∀ x ∈ R, ∀ t ≥ 0

The origin is globally uniformly asymptotically stable
Example:
ẋ1 = −x1 − g(t)x2
ẋ2 = x1 − x2
0 ≤ g(t) ≤ k and ġ(t) ≤ g(t), ∀ t ≥ 0
– p. 17/1
V (t, x) = x21 + [1 + g(t)]x22
x21 + x22 ≤ V (t, x) ≤ x21 + (1 + k)x22 , ∀ x ∈ R2
V̇ (t, x) = −2x21 + 2x1 x2 − [2 + 2g(t) − ġ(t)]x22

2 + 2g(t) − ġ(t) ≥ 2 + 2g(t) − g(t) ≥ 2
" #
2 2 T 2 −1
V̇ (t, x) ≤ −2x1 + 2x1 x2 − 2x2 = − x x
−1 2
– p. 18/1
Lecture # 13
Perturbed Systems
– p. 1/?
Nominal System:
ẋ = f (x), f (0) = 0
Perturbed System:
ẋ = f (x) + g(t, x), g(t, 0) = 0
Case 1: The origin of the nominal system is exponentially

stable
∂V
f (x) ≤ −c3 kxk2
∂x
∂V

≤ c4 kxk
∂x
– p. 2/?
Use V (x) as a Lyapunov function candidate for the
perturbed system
∂V ∂V
V̇ (t, x) = f (x) + g(t, x)
∂x ∂x
Assume that
kg(t, x)k ≤ γkxk, γ≥0
∂V

2
V̇ (t, x) ≤ −c3 kxk +
∂x kg(t, x)k

≤ −c3 kxk2 + c4 γkxk2
– p. 3/?
c3
γ<
c4
V̇ (t, x) ≤ −(c3 − γc4 )kxk2

The origin is an exponentially stable equilibrium point of the
perturbed system
– p. 4/?
Example
ẋ1 = x2
ẋ2 = −4x1 − 2x2 + βx32 , β≥0
ẋ = Ax + g(x)
" # " #
0 1 0
A= , g(x) =
−4 −2 βx32
√
The eigenvalues of A are −1 ± j 3
 
3 1
2 8
P A + AT P = −I ⇒ P = 
 

1 5
8 16
– p. 5/?
T ∂V
V (x) = x P x, Ax = −xT x
∂x
c3 = 1, c4 = 2 kP k = 2λmax (P ) = 2 × 1.513 = 3.026
kg(x)k = β|x2 |3
g(x) satisfies the bound kg(x)k ≤ γkxk over compact sets
of x. Consider the compact set
Ωc = {V (x) ≤ c} = {xT P x ≤ c}, c>0
k2 = max |x2 | = max |[0 1]x|

xT P x≤c xT P x≤c
– p. 6/?
Fact:
√
max kLxk = c kLP −1/2 k
xT P x≤c
Proof
T 1 T 1
x Px ≤ c ⇔ x Px ≤ 1 ⇔ xT P 1/2 P 1/2 x ≤ 1
c c
1
y = √ P 1/2 x
c
√ −1/2 √
max kLxk = max kL c P yk = c kLP −1/2 k
xT P x≤c y T y≤1
– p. 7/?
√ −1/2 √
k2 = max |[0 1]x| = c k[0 1]P k = 1.8194 c
xT P x≤c
kg(x)k ≤ β c (1.8194)2 kxk, ∀ x ∈ Ωc

kg(x)k ≤ γkxk, ∀ x ∈ Ωc , γ = β c (1.8194)2
c3 1 0.1
γ< ⇔ β< ≈
c4 3.026 × (1.8194)2 c c
β < 0.1/c ⇒ V̇ (x) ≤ −(1 − 10βc)kxk2

Hence, the origin is exponentially stable and Ωc is an
estimate of the region of attraction
– p. 8/?
Alternative Bound on β
V̇ (x) = −kxk2 + 2xT P g(x)

≤ −kxk2 + 18 βx32 ([2 5]x)
√
29
≤ −kxk2 + 8 βx2
2 kxk2
Over Ωc , x22 ≤ (1.8194)2 c

√
V̇ (x) ≤ − 1 − 829 β(1.8194)2 c kxk2
βc

= − 1− kxk2
0.448
If β < 0.448/c, the origin will be exponentially stable and

Ωc will be an estimate of the region of attraction
– p. 9/?
Remark: The inequality β < 0.448/c shows a tradeoff
between the estimate of the region of attraction and the
estimate of the upper bound on β
– p. 10/?
Case 2: The origin of the nominal system is asymptotically
stable
∂V ∂V ∂V

V̇ (t, x) = f (x)+ g(t, x) ≤ −W3 (x)+ g(t, x)
∂x ∂x ∂x
Under what condition will the following inequality hold?

∂V

∂x g(t, x) < W3 (x)

Special Case: Quadratic-Type Lyapunov function

∂V ∂V

2
f (x) ≤ −c3 φ (x), ≤ c4 φ(x)
∂x ∂x
– p. 11/?
V̇ (t, x) ≤ −c3 φ2 (x) + c4 φ(x)kg(t, x)k
c3
If kg(t, x)k ≤ γφ(x), with γ <
c4
V̇ (t, x) ≤ −(c3 − c4 γ)φ2 (x)
– p. 12/?
Example
ẋ = −x3 + g(t, x)
V (x) = x4 is a quadratic-type Lyapunov function for the
nominal system ẋ = −x3
∂V ∂V

3 6 = 4|x|3
(−x ) = −4x ,
∂x ∂x
φ(x) = |x|3 , c3 = 4, c4 = 4
Suppose |g(t, x)| ≤ γ|x|3 , ∀ x, with γ < 1
V̇ (t, x) ≤ −4(1 − γ)φ2 (x)
Hence, the origin is a globally uniformly asymptotically
stable
– p. 13/?
Remark: A nominal system with asymptotically, but not
exponentially, stable origin is not robust to smooth
perturbations with arbitrarily small linear growth bounds
Example
ẋ = −x3 + γx
The origin is unstable for any γ > 0
– p. 14/?
Lecture # 14
Passivity
Memoryless Functions
&
State Models
– p. 1/1
Memoryless Functions
+ y -
y
PPP
u P P
P
u
(a)
(b)
power inflow = uy
Resistor is passive if uy ≥ 0
– p. 2/1
y y y
u u u
(a) (b) (c)

Passive Passive Not passive
y = h(t, u), h ∈ [0, ∞]

Vector case:
h i
y = h(t, u), hT = h1 , h2 , · · · , hp
power inflow = Σpi=1 ui yi = uT y

– p. 3/1
Definition: y = h(t, u) is
passive if uT y ≥ 0
lossless if uT y = 0
input strictly passive if uT y ≥ uT ϕ(u) for some

function ϕ where uT ϕ(u) > 0, ∀ u 6= 0
output strictly passive if uT y ≥ y T ρ(y) for some

function ρ where y T ρ(y) > 0, ∀ y 6= 0
– p. 4/1
Sector Nonlinearity: h belongs to the sector [α, β]
(h ∈ [α, β]) if
αu2 ≤ uh(t, u) ≤ βu2
y=βu
y=β u
y y
y=αu
u u
y=αu
(a) α>0 (b) α<0
Also, h ∈ (α, β], h ∈ [α, β), h ∈ (α, β)

– p. 5/1
αu2 ≤ uh(t, u) ≤ βu2 ⇔ [h(t, u) − αu][h(t, u) − βu] ≤ 0
Definition: A memoryless function h(t, u) is said to belong

to the sector
[0, ∞] if uT h(t, u) ≥ 0
[K1 , ∞] if uT [h(t, u) − K1 u] ≥ 0
[0, K2 ] with K2 = K2T > 0 if

hT (t, u)[h(t, u) − K2 u] ≤ 0
[K1 , K2 ] with K = K2 − K1 = K T > 0 if
[h(t, u) − K1 u]T [h(t, u) − K2 u] ≤ 0

– p. 6/1
Example
" #
h1 (u1 )
h(u) = , hi ∈ [αi , βi ], βi > αi i = 1, 2
h2 (u2 )
" # " #
α1 0 β1 0
K1 = , K2 =
0 α2 0 β2
h ∈ [K1 , K2 ]
" #
β1 − α1 0
K = K2 − K1 =
0 β2 − α2
– p. 7/1
Example
kh(u) − Luk ≤ γkuk
K1 = L − γI, K2 = L + γI
[h(u) − K1 u]T [h(u) − K2 u] =

kh(u) − Luk2 − γ 2 kuk2 ≤ 0
K = K2 − K1 = 2γI
– p. 8/1
A function in the sector [K1 , K2 ] can be transformed into a
function in the sector [0, ∞] by input feedforward followed
by output feedback
+ +
- - - y = h(t, u) - -
K −1
+6 −6
- K1
Feedforward K −1 Feedback
[K1 , K2 ] [0, K] [0, I] [0, ∞]
−→ −→ −→
– p. 9/1
State Models
v2 = h2(i2) L
y
B B B
i2 + v2 iL
- - BB BB BB -
+ + + +
u v1 i1 = h1(v1) vC C v3 i3 = h3(v3)
P
P P
P

P
P
PP PP

P
P
PP PP

P
P
P P
i1 ?
i3
?
Lẋ1 = u − h2 (x1 ) − x2
C ẋ2 = x1 − h3 (x2 )
y = x1 + h1 (u)
– p. 10/1
V (x) = 12 Lx21 + 12 Cx22
Z t
u(s)y(s) ds ≥ V (x(t)) − V (x(0))
0
u(t)y(t) ≥ V̇ (x(t), u(t))
V̇ = Lx1 ẋ1 + Cx2 ẋ2

= x1 [u − h2 (x1 ) − x2 ] + x2 [x1 − h3 (x2 )]
= x1 [u − h2 (x1 )] − x2 h3 (x2 )
= [x1 + h1 (u)]u − uh1 (u) − x1 h2 (x1 ) − x2 h3 (x2 )
= uy − uh1 (u) − x1 h2 (x1 ) − x2 h3 (x2 )
– p. 11/1
uy = V̇ + uh1 (u) + x1 h2 (x1 ) + x2 h3 (x2 )
If h1 , h2 , and h3 are passive, uy ≥ V̇ and the system is

passive
Case 1: If h1 = h2 = h3 = 0, then uy = V̇ ; no energy

dissipation; the system is lossless
Case 2: If h1 ∈ (0, ∞] (uh1 (u) > 0 for u 6= 0), then
uy ≥ V̇ + uh1 (u)
The energy absorbed over [0, t] will be greater than the

increase in the stored energy, unless the input u(t) is
identically zero. This is a case of input strict passivity
– p. 12/1
Case 3: If h1 = 0 and h2 ∈ (0, ∞], then
y = x1 and uy ≥ V̇ + yh2 (y)
The energy absorbed over [0, t] will be greater than the

increase in the stored energy, unless the output y is
identically zero. This is a case of output strict passivity
Case 4: If h2 ∈ (0, ∞) and h3 ∈ (0, ∞), then
uy ≥ V̇ + x1 h2 (x1 ) + x2 h3 (x2 )
x1 h2 (x1 ) + x2 h3 (x2 ) is a positive definite function of x.

This is a case of state strict passivity because the energy
absorbed over [0, t] will be greater than the increase in the
stored energy, unless the state x is identically zero
– p. 13/1
Definition: The system
ẋ = f (x, u), y = h(x, u)
is passive if there is a continuously differentiable positive

semidefinite function V (x) (the storage function) such that
T ∂V
u y ≥ V̇ = f (x, u), ∀ (x, u)
∂x
Moreover, it is said to be
lossless if uT y = V̇
input strictly passive if uT y ≥ V̇ + uT ϕ(u) for some

function ϕ such that uT ϕ(u) > 0, ∀ u 6= 0
– p. 14/1
output strictly passive if uT y ≥ V̇ + y T ρ(y) for some
function ρ such that y T ρ(y) > 0, ∀ y = 6 0
strictly passive if uT y ≥ V̇ + ψ(x) for some positive

definite function ψ
Example
ẋ = u, y=x
V (x) = 12 x2 ⇒ uy = V̇ ⇒ Lossless
– p. 15/1
Example
ẋ = u, y = x + h(u), h ∈ [0, ∞]
V (x) = 21 x2 ⇒ uy = V̇ + uh(u) ⇒ Passive

h ∈ (0, ∞] ⇒ uh(u) > 0 ∀ u 6= 0
⇒ Input strictly passive
Example
ẋ = −h(x) + u, y = x, h ∈ [0, ∞]
V (x) = 21 x2 ⇒ uy = V̇ + yh(y) ⇒ Passive

h ∈ (0, ∞] ⇒ Output strictly passive
– p. 16/1
Example
ẋ = u, y = h(x), h ∈ [0, ∞]
Z x
V (x) = h(σ) dσ ⇒ V̇ = h(x)ẋ = yu ⇒ Lossless
0
Example
aẋ = −x + u, y = h(x), h ∈ [0, ∞]

Z x
V (x) = a h(σ) dσ ⇒ V̇ = h(x)(−x+u) = yu−xh(x)
0
yu = V̇ + xh(x) ⇒ Passive
h ∈ (0, ∞] ⇒ Strictly passive
– p. 17/1
Lecture # 15
Positive Real Transfer Functions
&
Connection with Lyapunov Stability
– p. 1/?
Definition: A p × p proper rational transfer function matrix
G(s) is positive real if
poles of all elements of G(s) are in Re[s] ≤ 0
for all real ω for which jω is not a pole of any element of
G(s), the matrix G(jω) + GT (−jω) is positive
semidefinite
any pure imaginary pole jω of any element of G(s) is a
simple pole and the residue matrix
lims→jω (s − jω)G(s) is positive semidefinite Hermitian
G(s) is called strictly positive real if G(s − ε) is positive real
for some ε > 0
– p. 2/?
Scalar Case (p = 1):
G(jω) + GT (−jω) = 2Re[G(jω)]
Re[G(jω)] is an even function of ω . The second condition

of the definition reduces to
Re[G(jω)] ≥ 0, ∀ ω ∈ [0, ∞)
which holds when the Nyquist plot of of G(jω) lies in the

closed right-half complex plane
This is true only if the relative degree of the transfer function

is zero or one
– p. 3/?
Lemma: A p × p proper rational transfer function matrix
G(s) is strictly positive real if and only if
G(s) is Hurwitz
G(jω) + GT (−jω) > 0, ∀ ω ∈ R
G(∞) + GT (∞) > 0 or
lim ω 2(p−q) det[G(jω) + GT (−jω)] > 0

ω→∞
where q = rank[G(∞) + GT (∞)]
– p. 4/?
Scalar Case (p = 1): G(s) is strictly positive real if and only
if
G(s) is Hurwitz
Re[G(jω)] > 0, ∀ ω ∈ [0, ∞)
G(∞) > 0 or
lim ω 2 Re[G(jω)] > 0

ω→∞
– p. 5/?
Example:
1
G(s) =
s
has a simple pole at s = 0 whose residue is 1
1

Re[G(jω)] = Re = 0, ∀ ω 6= 0
jω
Hence, G is positive real. It is not strictly positive real since

1
(s − ε)
has a pole in Re[s] > 0 for any ε > 0
– p. 6/?
Example:
1
G(s) = , a > 0, is Hurwitz
s+a
a
Re[G(jω)] = > 0, ∀ ω ∈ [0, ∞)
ω2 + a2
ω2a
lim ω 2 Re[G(jω)] = lim = a > 0 ⇒ G is SPR
ω→∞ ω→∞ ω 2 + a2
Example:
1 1 − ω2
G(s) = , Re[G(jω)] =
s2 +s+1 (1 − ω 2 )2 + ω 2
G is not PR
– p. 7/?
Example:
s+2 1
 
s+1 s+2
G(s) =   is Hurwitz
 
−1 2
s+2 s+1
2(2+ω 2 ) −2jω
 
1+ω 2 4+ω 2
G(jω) + GT (−jω) =   > 0, ∀ ω ∈ R
 
2jω 4
4+ω 2 1+ω 2
" #
T 2 0
G(∞) + G (∞) = , q=1
0 0
lim ω 2 det[G(jω) + GT (−jω)] = 4 ⇒ G is SPR

ω→∞
– p. 8/?
Positive Real Lemma: Let
G(s) = C(sI − A)−1 B + D
where (A, B) is controllable and (A, C) is observable.

G(s) is positive real if and only if there exist matrices
P = P T > 0, L, and W such that
P A + AT P = −LT L
P B = C T − LT W
W T W = D + DT
– p. 9/?
Kalman–Yakubovich–Popov Lemma: Let
G(s) = C(sI − A)−1 B + D
where (A, B) is controllable and (A, C) is observable.

G(s) is strictly positive real if and only if there exist matrices
P = P T > 0, L, and W , and a positive constant ε such
that
P A + AT P = −LT L − εP
P B = C T − LT W
W T W = D + DT
– p. 10/?
Lemma: The linear time-invariant minimal realization
ẋ = Ax + Bu
y = Cx + Du
with
G(s) = C(sI − A)−1 B + D
is
passive if G(s) is positive real
strictly passive if G(s) is strictly positive real
Proof: Apply the PR and KYP Lemmas, respectively, and

use V (x) = 12 xT P x as the storage function
– p. 11/?
T ∂V
u y− (Ax + Bu)
∂x
= uT (Cx + Du) − xT P (Ax + Bu)
= uT Cx + 12 uT (D + D T )u
− 12 xT (P A + AT P )x − xT P Bu
= uT (B T P + W T L)x + 12 uT W T W u
+ 12 xT LT Lx + 12 εxT P x − xT P Bu
1
= 2 (Lx + W u)T (Lx + W u) + 12 εxT P x ≥ 12 εxT P x
In the case of the PR Lemma, ε = 0, and we conclude that

the system is passive; in the case of the KYP Lemma,
ε > 0, and we conclude that the system is strictly passive
– p. 12/?
Connection with Lyapunov Stability
Lemma: If the system
ẋ = f (x, u), y = h(x, u)
is passive with a positive definite storage function V (x),

then the origin of ẋ = f (x, 0) is stable
Proof:
T ∂V ∂V
u y≥ f (x, u) ⇒ f (x, 0) ≤ 0
∂x ∂x
– p. 13/?
ẋ = f (x, u), y = h(x, u)
is strictly passive, then the origin of ẋ = f (x, 0) is

asymptotically stable. Furthermore, if the storage function
is radially unbounded, the origin will be globally
Proof: The storage function V (x) is positive definite
T ∂V ∂V
u y≥ f (x, u) + ψ(x) ⇒ f (x, 0) ≤ −ψ(x)
∂x ∂x
Why is V (x) positive definite? Let φ(t; x) be the solution
of ż = f (z, 0), z(0) = x
– p. 14/?
V̇ ≤ −ψ(x)
Z τ
V (φ(τ, x)) − V (x) ≤ − ψ(φ(t; x)) dt, ∀ τ ∈ [0, δ]
0
Z τ
V (φ(τ, x)) ≥ 0 ⇒ V (x) ≥ ψ(φ(t; x)) dt
0
Z τ
V (x̄) = 0 ⇒ ψ(φ(t; x̄)) dt = 0, ∀ τ ∈ [0, δ]
0
⇒ ψ(φ(t; x̄)) ≡ 0 ⇒ φ(t; x̄) ≡ 0 ⇒ x̄ = 0
– p. 15/?
ẋ = f (x, u), y = h(x, u)
is zero-state observable if no solution of ẋ = f (x, 0) can

stay identically in S = {h(x, 0) = 0}, other than the zero
solution x(t) ≡ 0
Linear Systems
ẋ = Ax, y = Cx
Observability of (A, C) is equivalent to
y(t) = CeAt x(0) ≡ 0 ⇔ x(0) = 0 ⇔ x(t) ≡ 0
– p. 16/?
ẋ = f (x, u), y = h(x, u)
is output strictly passive and zero-state observable, then

the origin of ẋ = f (x, 0) is asymptotically stable.
Furthermore, if the storage function is radially unbounded,
the origin will be globally asymptotically stable
Proof: The storage function V (x) is positive definite
T ∂V T ∂V
u y≥ f (x, u) + y ρ(y) ⇒ f (x, 0) ≤ −y T ρ(y)
∂x ∂x
V̇ (x(t)) ≡ 0 ⇒ y(t) ≡ 0 ⇒ x(t) ≡ 0

Apply the invariance principle
– p. 17/?
Example
ẋ1 = x2 , ẋ2 = −ax31 − kx2 + u, y = x2 , a, k > 0
V (x) = 14 ax41 + 12 x22
V̇ = ax31 x2 + x2 (−ax31 − kx2 + u) = −ky 2 + yu

The system is output strictly passive
y(t) ≡ 0 ⇔ x2 (t) ≡ 0 ⇒ ax31 (t) ≡ 0 ⇒ x1 (t) ≡ 0
The system is zero-state observable. V is radially

unbounded. Hence, the origin of the unforced system is
globally asymptotically stable
– p. 18/?
Lecture # 16
Feedback Systems: Passivity

Theorems
– p. 1/2
e1 -
u1- y1 -
+ H1
−6
y2 e + u
?

2 + 2
H2
ẋi = fi (xi , ei ), yi = hi (xi , ei )
yi = hi (t, ei )
– p. 2/2
Passivity Theorems
Theorem 6.1: The feedback connection of two passive

systems is passive
Theorem 6.3: Consider the feedback connection of two

dynamical systems. When u = 0, the origin of the
closed-loop system is asymptotically stable if each
feedback component is either
strictly passive, or
output strictly passive and zero-state observable
Furthermore, if the storage function for each component is
radially unbounded, the origin is globally asymptotically
stable
– p. 3/2
Theorem 6.4: Consider the feedback connection of a
strictly passive dynamical system with a passive
memoryless function. When u = 0, the origin of the
closed-loop system is uniformly asymptotically stable. if the
storage function for the dynamical system is radially
unbounded, the origin will be globally uniformly
Prove using V = V1 + V2 as a Lyapunov function candidate
Proof of Theorem 6.3: H1 is SP; H2 is OSP & ZSO
eT1 y1 ≥ V̇1 + ψ1 (x1 ), ψ1 (x1 ) > 0, ∀ x1 6= 0
eT2 y2 ≥ V̇2 + y2T ρ2 (y2 ), y2T ρ(y2 ) > 0, ∀y2 6= 0

– p. 4/2
eT1 y1 +eT2 y2 = (u1 −y2 )T y1 +(u2 +y1 )T y2 = uT1 y1 +uT2 y2
V (x) = V1 (x1 ) + V2 (x2 )
V̇ ≤ uT y − ψ1 (x1 ) − y2T ρ2 (y2 )
u = 0 ⇒ V̇ ≤ −ψ1 (x1 ) − y2T ρ2 (y2 )
V̇ = 0 ⇒ x1 = 0 and y2 = 0
y2 (t) ≡ 0 ⇒ e1 (t) ≡ 0 ( & x1 (t) ≡ 0) ⇒ y1 (t) ≡ 0
y1 (t) ≡ 0 ⇒ e2 (t) ≡ 0
By zero-state observability of H2 : y2 (t) ≡ 0 ⇒ x2 (t) ≡ 0
– p. 5/2

Example

ẋ1 = x2 ẋ3 = x4

ẋ2 = 3
−ax1 − kx2 + e1 ẋ4 = −bx3 − x34 + e2
y1 = x2 + e1 y = x
2 4
| {z } | {z }
H
H
1 2
a, b, k > 0
V1 = 14 ax41 + 21 x22
V̇1 = ax31 x2 − ax31 x2 − kx22 + x2 e1 = −ky12 + y1 e1

With e1 = 0, y1 (t) ≡ 0 ⇔ x2 (t) ≡ 0 ⇒ x1 (t) ≡ 0
H1 is output strictly passive and zero-state observable
– p. 6/2
V2 = 12 bx23 + 12 x24
V̇2 = bx3 x4 − bx3 x4 − x44 + x4 e2 = −y24 + y2 e2

With e2 = 0, y2 (t) ≡ 0 ⇔ x4 (t) ≡ 0 ⇒ x3 (t) ≡ 0
H2 is output strictly passive and zero-state observable
V1 and V2 are radially unbounded
– p. 7/2
Loop Transformations
Recall that a memoryless function in the sector [K1 , K2 ]

can be transformed into a function in the sector [0, ∞] by
input feedforward followed by output feedback
+ +
- - - y = h(t, u) - -
K −1
+6 −6
- K
1
– p. 8/2
+h
- -H -
1
−6
H2
H1 is a dynamical system
H2 is a memoryless function in the sector [K1 , K2 ]
– p. 9/2
+h
- +h -
- -
H1
−6 −6
K1
+

h H2
−
6
K1
– p. 10/2
+h
- +h -
- - K -
H1
−6 −6
K1
+

h H2 K −1
−
6
K1
– p. 11/2
H̃1
+h
- +h -
- - K ?
+
- h -
H1
−6 −6 +
K1
+
+
h H2 K −1
h
−
6 +6
K1
H̃2
– p. 12/2
Example

ẋ1 = x2

ẋ2 = −h(x1 ) + bx2 + e1 y2 = σ(e2 )
| {z }
y1 = x2
H2
| {z }
H1
σ ∈ [α, β], h ∈ [α1 , ∞], b > 0, α1 > 0, k = β − α > 0

ẋ1 = x2

ẋ2 = −h(x1 ) − ax2 + ẽ1 ỹ2 = σ̃(ẽ2 )
| {z }
ỹ1 = kx2 + ẽ1 H̃2
| {z }
H̃1
σ̃ ∈ [0, ∞], a=α−b

– p. 13/2
Assume a = α − b > 0 and show that H̃1 is strictly passive
Z x1
V1 = k h(s) ds + xT P x
0
Z x1
V1 = k h(s) ds + p11 x21 + 2p12 x1 x2 + p22 x22
0
V̇ = kh(x1 )x2 + 2(p11 x1 + p12 x2 )x2

2(p12 x1 + p22 x2 )[−h(x1 ) − ax2 + ẽ1 ]
Take p22 = k/2, p11 = ap12
– p. 14/2
V̇ = −2p12 x1 h(x1 ) − (ka − 2p12 )x22
+ kx2 ẽ1 + 2p12 x1 ẽ1
= −2p12 x1 h(x1 ) − (ka − 2p12 )x22
+ (kx2 + ẽ1 )ẽ1 − ẽ21 + 2p12 x1 ẽ1
ỹ1 ẽ1 = V̇ + 2p12 x1 h(x1 ) + (ka − 2p12 )x22
+ (ẽ1 − p12 x1 )2 − p212 x21
≥ V̇ + p12 (2α1 − p12 )x21 + (ka − 2p12 )x22

ak k
Take 0 < p12 < min , 2α1 ⇒ p212 < 2p12 = p11 p22
2 2
H̃1 is strictly passive. By Theorem 6.4 the origin is globally
asymptotically stable (when u = 0)
– p. 15/2
+
- - H1 -

−6
H2
– p. 16/2
H̃1
+
- - H1 - W −1 (s) -

−6
H2 W (s)
H̃2
– p. 17/2
Example
H1 :
ẋ = Ax + Be1 , y1 = Cx
" # " #
0 1 0 h i
A= , B= , C= 1 0
−1 −1 1
H2 : y2 = h(e2 ), h ∈ [0, ∞]
−1 1
C(sI − A) B= Not PR
(s2 + s + 1)
1 (as + 1)
W (s) = ⇒
as + 1 (s2 + s + 1)
h i
H̃1 : ẋ = Ax + Be1 , ỹ1 = C̃x = 1 a x
– p. 18/2
(as + 1)
(s2 + s + 1)

1 + jωa 1 + (a − 1)ω 2
Re =
1− ω2
+ jω (1 − ω 2 )2 + ω 2

2 1 + jωa
lim ω Re 2
=a−1
ω→∞ 1 − ω + jω
(as + 1)
a>1 ⇒ is SPR
(s2 + s + 1)
V1 = 12 xT P x, P A + AT P = −LT L − εP, P B = C̃ T
– p. 19/2
H̃2 : aė2 = −e2 + ẽ2 , y2 = h(e2 ), h ∈ [0, ∞]
R e2
H̃2 is strictly passive with V2 = a 0 h(s) ds. Use
Z e2
V = V1 + V2 = 21 xT P x + a h(s) ds
0
as a Lyapunov function candidate for the original feedback

connection
– p. 20/2
1 T 1 T
V̇ = 2 x P ẋ + 2 ẋ P x + ah(e2 )ė2
1 T
= 2 x P [Ax − Bh(e 2 )] + 1
2 [Ax − Bh(e2 )]T P x
+ ah(e2 )C[Ax − Bh(e2 )]
= − 21 xT LT Lx − (ε/2)xT P x − xT C̃ T h(e2 )
+ ah(e2 )CAx
= − 12 xT LT Lx − (ε/2)xT P x
− xT [C + aCA]T h(e2 ) + ah(e2 )CAx
= − 12 xT LT Lx − (ε/2)xT P x − eT2 h(e2 )
≤ −(ε/2)xT P x

– p. 21/2
Lecture # 17
Circle & Popov Criteria
– p. 1/2
Absolute Stability
r + u
- -
y -
G(s)
−6

ψ(·)
The system is absolutely stable if (when r = 0) the origin is

globally asymptotically stable for all memoryless
time-invariant nonlinearities in a given sector
– p. 2/2
Circle Criterion
Suppose G(s) = C(sI − A)−1 B + D is SPR and
ψ ∈ [0, ∞]
ẋ = Ax + Bu
y = Cx + Du
u = −ψ(y)
By the KYP Lemma, ∃ P = P T > 0, L, W, ε > 0
P B = C T − LT W
W T W = D + DT
V (x) = 12 xT P x
– p. 3/2
1 T 1 T
V̇ = 2
x P ẋ + 2
ẋ P x
1 T T T
= 2 x (P A + A P )x + x P Bu
= − 12 xT LT Lx − 12 εxT P x + xT (C T − LT W )u
= − 12 xT LT Lx − 12 εxT P x + (Cx + Du)T u
T T T
− u Du − x L W u
uT Du = 21 uT (D + D T )u = 12 uT W T W u
V̇ = − 12 εxT P x − 1
2
(Lx + W u)T (Lx + W u) − y T ψ(y)
y T ψ(y) ≥ 0 ⇒ V̇ ≤ − 12 εxT P x
– p. 4/2
What if ψ ∈ [K1 , ∞]?
-
+f - G(s) - -
+f -+f -
G(s) -
− − −
6 6 6
K1
ψ(·) ψ(·)
+
f
−
6
K1 ψ̃(·)
ψ̃ ∈ [0, ∞]; hence the origin is globally exponentially stable

if G(s)[I + K1 G(s)]−1 is SPR
– p. 5/2
What if ψ ∈ [K1 , K2 ]?
-
+f - G(s) - -
+f -+f -
G(s) - K -?+ -
f
− − − +
6 6 6
K1
ψ(·) ψ(·) −1
+
f +
f
− K +
6 6
K1 ψ̃(·)
ψ̃ ∈ [0, ∞]; hence the origin is globally exponentially stable

if I + KG(s)[I + K1 G(s)]−1 is SPR
– p. 6/2
I + KG(s)[I + K1 G(s)]−1 = [I + K2 G(s)][I + K1 G(s)]−1
Theorem (Circle Criterion): The system is absolutely stable

if
ψ ∈ [K1 , ∞] and G(s)[I + K1 G(s)]−1 is SPR, or
ψ ∈ [K1 , K2 ] and [I + K2 G(s)][I + K1 G(s)]−1 is SPR
Scalar Case: ψ ∈ [α, β], β > α
The system is absolutely stable if
1 + βG(s)
is Hurwitz and
1 + αG(s)
1 + βG(jω)

Re > 0, ∀ ω ∈ [0, ∞]
1 + αG(jω)
– p. 7/2
Case 1: α > 0
By the Nyquist criterion
1 + βG(s) 1 βG(s)
= +
1 + αG(s) 1 + αG(s) 1 + αG(s)
is Hurwitz if the Nyquist plot of G(jω) does not intersect the

point −(1/α) + j0 and encircles it m times in the
counterclockwise direction, where m is the number of poles
of G(s) in the open right-half complex plane
1
1 + βG(jω) β + G(jω)
>0 ⇔ 1 >0
1 + αG(jω) α + G(jω)
– p. 8/2
"1 #
β + G(jω)
Re 1 > 0, ∀ ω ∈ [0, ∞]
α + G(jω)
D(α,β) q
θ θ
2 1
−1/α −1/β
The system is absolutely stable if the Nyquist plot of G(jω)

does not enter the disk D(α, β) and encircles it m times in
the counterclockwise direction
– p. 9/2
Case 2: α = 0
1 + βG(s)
Re[1 + βG(jω)] > 0, ∀ ω ∈ [0, ∞]

1
Re[G(jω)] > − , ∀ ω ∈ [0, ∞]
β
The system is absolutely stable if G(s) is Hurwitz and the
Nyquist plot of G(jω) lies to the right of the vertical line
defined by Re[s] = −1/β
– p. 10/2
Case 3: α < 0 < β
"1 #

1 + βG(jω)

β + G(jω)
Re > 0 ⇔ Re 1 <0
1 + αG(jω) α + G(jω)
The Nyquist plot of G(jω) must lie inside the disk D(α, β).
The Nyquist plot cannot encircle the point −(1/α) + j0.
From the Nyquist criterion, G(s) must be Hurwitz
The system is absolutely stable if G(s) is Hurwitz and the

Nyquist plot of G(jω) lies in the interior of the disk D(α, β)
– p. 11/2
Example
4
G(s) =
(s + 1)( 12 s + 1)( 13 s + 1)
6
Im G
4
0
Re G
−2
−4
−5 0 5
– p. 12/2
Apply Case 3 with center (0, 0) and radius = 4
Sector is (−0.25, 0.25)

Apply Case 3 with center (1.5, 0) and radius = 2.834
Sector is [−0.227, 0.714]
Apply Case 2
The Nyquist plot is to the right of Re[s] = −0.857
Sector is [0, 1.166]
[0, 1.166] includes the saturation nonlinearity
– p. 13/2
Example
4
G(s) =
(s − 1)( 12 s + 1)( 13 s + 1)
0.4 Im G
0.2
G is not Hurwitz
0
Re G
−0.2
Apply Case 1
−0.4
−4 −2 0
Center = (−3.2, 0), Radius = 0.168 ⇒ [0.2969, 0.3298]
– p. 14/2
Popov Criterion
-
+f -
G(s) -
−
6
ψ(·)
ẋ = Ax + Bu
y = Cx
ui = −ψi (yi ), 1 ≤ i ≤ p
ψi ∈ [0, ki ], 1 ≤ i ≤ p, (0 < ki ≤ ∞)
G(s) = C(sI − A)−1 B
Γ = diag(γ1 , . . . , γp ), M = diag(1/k1 , · · · , 1/kp )
– p. 15/2
- M
H̃1
- ?
+ +
- i - G(s) - (I + sΓ) i -
−6 +
+
ψ(·) (I + sΓ)−1
i
+6
H̃2
M
Show that H̃1 and H̃2 are passive

– p. 16/2
M + (I + sΓ)G(s)
= M + (I + sΓ)C(sI − A)−1 B
= M + C(sI − A)−1 B + ΓCs(sI − A)−1 B
= M + C(sI − A)−1 B + ΓC(sI − A + A)(sI − A)−1 B
= (C + ΓCA)(sI − A)−1 B + M + ΓCB
If M + (I + sΓ)G(s) is SPR, then H̃1 is strictly passive

with the storage function V1 = 12 xT P x, where P is given by
the KYP equations
P B = (C + ΓCA)T − LT W
W T W = 2M + ΓCB + B T C T Γ
– p. 17/2
H̃2 consists of p decoupled components:
1
γi żi = −zi + ψi (zi ) + ẽ2i , ỹ2i = ψi (zi )
ki
Z zi
V2i = γi ψi (σ) dσ
0
h i
1
V̇2i = γi ψi (zi )żi = ψi (zi ) −zi + ki ψi (zi ) + ẽ2i
1
= y2i e2i + ki ψi (zi ) [ψi (zi ) − ki z i ]
ψi ∈ [0, ki ] ⇒ ψi (ψi − ki zi ) ≤ 0 ⇒ V̇2i ≤ y2i e2i

H̃2 is passive Rwith the storage function
Pp z
V2 = i=1 γi 0 i ψi (σ) dσ
– p. 18/2
p
X Z yi
Use V = 21 xT P x + γi ψi (σ) dσ
i=1 0
as a Lyapunov function candidate for the original feedback

connection
ẋ = Ax + Bu, y = Cx, u = −ψ(y)
1 T 1 T T (y)Γẏ
V̇ = 2
x P ẋ + 2
ẋ P x + ψ
1 T T P )x + xT P Bu
= 2
x (P A + A
+ ψ T (y)ΓC(Ax + Bu)
= − 12 xT LT Lx − 12 εxT P x
+ xT (C T + AT C T Γ − LT W )u
+ ψ T (y)ΓCAx + ψ T (y)ΓCBu
– p. 19/2
V̇ = − 21 εxT P x − 12 (Lx + W u)T (Lx + W u)
− ψ(y)T [y − M ψ(y)]
≤ − 12 εxT P x − ψ(y)T [y − M ψ(y)]
ψi ∈ [0, ki ] ⇒ ψ(y)T [y−M ψ(y)] ≥ 0 ⇒ V̇ ≤ − 21 εxT P x

Popov Criterion: The system is absolutely stable if, for
1 ≤ i ≤ p, ψi ∈ [0, ki ] and there exists a constant γi ≥ 0,
with (1 + λk γi ) 6= 0 for every eigenvalue λk of A, such that
M + (I + sΓ)G(s) is strictly positive real
– p. 20/2
Scalar case
1
+ (1 + sγ)G(s)
k
is SPR if G(s) is Hurwitz and
1
+ Re[G(jω)] − γωIm[G(jω)] > 0, ∀ ω ∈ [0, ∞)
k
If
1

lim + Re[G(jω)] − γωIm[G(jω)] =0
ω→∞ k
we also need
2 1

lim ω + Re[G(jω)] − γωIm[G(jω)] > 0
ω→∞ k
– p. 21/2
1
+ Re[G(jω)] − γωIm[G(jω)] > 0, ∀ ω ∈ [0, ∞)
k
slope = 1/γ ωIm[G(jω)]
−1/k Re[G(jω)]
Popov Plot
– p. 22/2
Example
ẋ1 = x2 , ẋ2 = −x2 − h(y), y = x1
ẋ2 = −αx1 − x2 − h(y) + αx1 , α>0

1
G(s) = , ψ(y) = h(y) − αy
s2 +s+α
h ∈ [α, β] ⇒ ψ ∈ [0, k] (k = β − α > 0)
α − ω 2 + γω 2
γ>1 ⇒ > 0, ∀ ω ∈ [0, ∞)
(α − ω 2 )2 + ω2
ω 2 (α − ω 2 + γω 2 )
and lim =γ−1>0
ω→∞ (α − ω 2 )2 + ω2
– p. 23/2
The system is absolutely stable for ψ ∈ [0, ∞] (h ∈ [α, ∞])
ω Im G
0.2 slope=1
0
Re G
−0.2
−0.4
−0.6
−0.8
−1
−0.5 0 0.5 1
Compare with the circle criterion (γ = 0)
1 α − ω2 √
+ > 0, ∀ ω ∈ [0, ∞], for k < 1 + 2 α
k (α − ω 2 )2 + ω 2
– p. 24/2
Lecture # 18
Boundedness
&
Ultimate Boundedness
– p. 1/1
Definition: The solutions of ẋ = f (t, x) are
uniformly bounded if ∃ c > 0 and for every
0 < a < c, ∃ β = β(a) > 0 such that
kx(t0 )k ≤ a ⇒ kx(t)k ≤ β, ∀ t ≥ t0 ≥ 0
uniformly ultimately bounded with ultimate bound b if

∃ b and c and for every 0 < a < c, ∃ T = T (a, b) ≥ 0
such that
kx(t0 )k ≤ a ⇒ kx(t)k ≤ b, ∀ t ≥ t0 + T
“Globally” if a can be arbitrarily large
Drop “uniformly” if ẋ = f (x)

– p. 2/1
Lyapunov Analysis: Let V (x) be a cont. diff. positive
definite function and suppose that the sets
Ωc = {V (x) ≤ c}, Ωε = {V (x) ≤ ε}, Λ = {ε ≤ V (x) ≤ c}
are compact for some c > ε > 0
Ωε
Ωc
– p. 3/1
Suppose
∂V
V̇ (t, x) = f (t, x) ≤ −W3 (x), ∀ x ∈ Λ, ∀ t ≥ 0
∂x
W3 (x) is continuous and positive definite
Ωc and Ωε are positively invariant
k = min W3 (x) > 0

x∈Λ
V̇ (t, x) ≤ −k, ∀ x ∈ Λ, ∀ t ≥ t0 ≥ 0
V (x(t)) ≤ V (x(t0 )) − k(t − t0 ) ≤ c − k(t − t0 )
x(t) enters the set Ωε within the interval [t0 , t0 + (c − ε)/k]
– p. 4/1
Suppose
V̇ (t, x) ≤ −W3 (x), ∀ µ ≤ kxk ≤ r, ∀ t ≥ 0
Choose c and ε such that Λ ⊂ {µ ≤ kxk ≤ r}
Bµ
Ωε
Ωc
Br
– p. 5/1
Let α1 and α2 be class K functions such that
α1 (kxk) ≤ V (x) ≤ α2 (kxk)
V (x) ≤ c ⇒ α1 (kxk) ≤ c ⇔ kxk ≤ α−1

1 (c)
c = α1 (r) ⇒ Ωc ⊂ Br
kxk ≤ µ ⇒ V (x) ≤ α2 (µ)
ε = α2 (µ) ⇒ Bµ ⊂ Ωε
What is the ultimate bound?
V (x) ≤ ε ⇒ α1 (kxk) ≤ ε ⇔ kxk ≤ α−1

1 (ε) = α−1
1 (α2 (µ))
– p. 6/1
Theorem (special case of Thm 4.18): Suppose
α1 (kxk) ≤ V (x) ≤ α2 (kxk)
∂V
f (t, x) ≤ −W3 (x), ∀ kxk ≥ µ > 0
∂x
∀ t ≥ 0 and kxk ≤ r , where α1 , α2 ∈ K, W3 (x) is
continuous & positive definite, and µ < α−12 (α1 (r)). Then,
for every initial state x(t0 ) ∈ {kxk ≤ α−1
2 (α1 (r))}, there is
T ≥ 0 (dependent on x(t0 ) and µ) such that
kx(t)k ≤ α−1
1 (α2 (µ)), ∀ t ≥ t0 + T
If the assumptions hold globally and α1 ∈ K∞ , then the

conclusion holds for any initial state x(t0 )
– p. 7/1
Remarks:
The ultimate bound is independent of the initial state
The ultimate bound is a class K function of µ; hence,

the smaller the value of µ, the smaller the ultimate
bound. As µ → 0, the ultimate bound approaches zero
– p. 8/1
Example
ẋ1 = x2 , ẋ2 = −(1 + x21 )x1 − x2 + M cos ωt, M ≥0
With M = 0, ẋ2 = −(1 + x21 )x1 − x2 = −h(x1 ) − x2

 
1 1
2 2 Z x1
V (x) = xT  x + 2 (y + y 3
) dy (Example 4.5)
 

1 0
2 1
 
3 1
2 2
1 4 def T
V (x) = xT  x + x = x P x + 1 4
2 x1
 
 2 1
1
2
1
– p. 9/1
λmin (P )kxk2 ≤ V (x) ≤ λmax (P )kxk2 + 12 kxk4
α1 (r) = λmin (P )r 2 , α2 (r) = λmax (P )r 2 + 12 r 4
V̇ = −x21 − x41 − x22 + (x1 + 2x2 )M cos ω

2 4
√
≤ −kxk − x1 + M 5kxk
2 4 2
√
= −(1 − θ)kxk − x1 − θkxk + M 5kxk
(0 < θ < 1)
2 4
√ def
≤ −(1 − θ)kxk − x1 , ∀ kxk ≥ M 5/θ = µ
The solutions are GUUB by

s
−1 λmax (P )µ2 + µ4 /2
b = α1 (α2 (µ)) =
λmin (P )
– p. 10/1
Lecture # 19
Perturbed Systems
&
Input-to-State Stability
– p. 1/?
Perturbed Systems: Nonvanishing Perturbation
Nominal System:
ẋ = f (x), f (0) = 0
Perturbed System:
ẋ = f (x) + g(t, x), g(t, 0) 6= 0
Case 1: The origin of ẋ = f (x) is exponentially stable
∂V ∂V

2

f (x) ≤ −c3 kxk , ≤ c4 kxk
∂x ∂x
∀ x ∈ Br = {kxk ≤ r}
– p. 2/?
Use V (x) to investigate ultimate boundedness of the
perturbed system
∂V ∂V
V̇ (t, x) = f (x) + g(t, x)
∂x ∂x
Assume
kg(t, x)k ≤ δ, ∀ t ≥ 0, x ∈ Br

V̇ (t, x) ≤ −c3 kxk2 + ∂V kg(t, x)k

∂x
≤ −c3 kxk2 + c4 δkxk
= −(1 − θ)c3 kxk2 − θc3 kxk2 + c4 δkxk
0<θ<1
def
≤ −(1 − θ)c3 kxk2 , ∀ kxk ≥ δc4 /(θc3 ) = µ
– p. 3/?
Apply Theorem 4.18
c1
r
kx(t0 )k ≤ α−1
2 (α1 (r)) ⇔ kx(t0 )k ≤ r
c2
δc4 c1 c3 c1
r r
µ< α−1
2 (α1 (r)) ⇔ <r ⇔ δ< θr
θc3 c2 c4 c2
c2 δc4 c2
r r
−1
b = α1 (α2 (µ)) ⇔ b = µ ⇔ b=
c1 θc3 c1
p
For all kx(t0 )k ≤ r c1 /c2 , the solutions of the perturbed
system are ultimately bounded by b
– p. 4/?
Example
ẋ1 = x2 , ẋ2 = −4x1 − 2x2 + βx32 + d(t)
β ≥ 0, |d(t)| ≤ δ, ∀ t ≥ 0
 
3 1
2 8
T T
V (x) = x P x = x   x (Lecture 13)
 
1 5
8 16
2 1 5 2
2βx22

V̇ (t, x) = −kxk + 8 x1 x2 + 16 x2
1 5

+ 2d(t) x +
8 1
x
16 2
√ √
2 29 29δ
≤ −kxk + βk22 kxk2 + kxk
8 8
– p. 5/?
√
k2 = max |x2 | = 1.8194 c
xT P x≤c
√
Suppose β ≤ 8(1 − ζ)/( 29k22 ) (0 < ζ < 1)
√
29δ
V̇ (t, x) ≤ −ζkxk2 + 8
kxk
√
29δ def
≤ −(1 − θ)ζkxk2 , ∀ kxk ≥ 8ζθ
= µ
(0 < θ < 1)
If µ2 λmax (P ) < c, then all solutions of the perturbed
system, starting in Ωc , are uniformly ultimately bounded by
√ s
29δ λmax (P )
b=
8ζθ λmin (P )
– p. 6/?
Case 2: The origin of ẋ = f (x) is asymptotically stable
α1 (kxk) ≤ V (x) ≤ α2 (kxk)
∂V ∂V

f (x) ≤ −α3 (kxk), ≤k
∂x ∂x
∀ x ∈ Br = {kxk ≤ r}, αi ∈ K, i = 1, 2, 3

V̇ (t, x) ≤ −α3 (kxk) + ∂V
∂x kg(t, x)k

≤ −α3 (kxk) + δk
≤ −(1 − θ)α3 (kxk) − θα3 (kxk) + δk
0<θ<1

−1 δk def
≤ −(1 − θ)α3 (kxk), ∀ kxk ≥ α3 θ
= µ
– p. 7/?
Apply Theorem 4.18
δk

µ< α−1
2 (α1 (r)) ⇔ α−1
3 < α−1
2 (α1 (r))
θ
θα3 (α−1
2 (α1 (r)))
⇔ δ<
k
−1 δk

−1
µ < α2 (α1 (r)) ⇔ α3 < α−12 (α1 (r))
θ
−1
θα3 (α2 (α1 (r))) c3 c1
r
⇔ δ< Compare with δ < θr
k c4 c2
Example
x
ẋ = −
1 + x2
∂V x 4x4

V (x) = x4 ⇒ − 2
=− 2
– p. 8/?
θα3 (α−1
2 (α1 (r))) θα3 (r) rθ
= =
k k 1 + r2
rθ
→ 0 as r → ∞
1+ r2
x
ẋ = − + δ, δ>0
1+ x2
1
δ> 2 ⇒ lim x(t) = ∞
t→∞
– p. 9/?
Input-to-State Stability (ISS)
Definition: The system ẋ = f (x, u) is input-to-state stable if

there exist β ∈ KL and γ ∈ K such that for any initial state
x(t0 ) and any bounded input u(t)
!
kx(t)k ≤ β(kx(t0 )k, t − t0 ) + γ sup ku(τ )k
t0 ≤τ ≤t
ISS of ẋ = f (x, u) implies

BIBS stability
x(t) is ultimately bounded by a class K function of

supt≥t0 ku(t)k
limt→∞ u(t) = 0 ⇒ limt→∞ x(t) = 0
The origin of ẋ = f (x, 0) is GAS
– p. 10/?
Theorem (Special case of Thm 4.19): Let V (x) be a
α1 (kxk) ≤ V (x) ≤ α2 (kxk)
∂V
f (x, u) ≤ −W3 (x), ∀ kxk ≥ ρ(kuk) > 0
∂x
∀ x ∈ Rn , u ∈ Rm , where α1 , α2 ∈ K∞ , ρ ∈ K, and
W3 (x) is a continuous positive definite function. Then, the
system ẋ = f (x, u) is ISS with γ = α−11 ◦ α2 ◦ ρ
Proof: Let µ = ρ(supτ ≥t0 ku(τ )k); then
∂V
f (x, u) ≤ −W3 (x), ∀ kxk ≥ µ
∂x
– p. 11/?
Choose ε and c such that
∂V
f (x, u) ≤ −W3 (x), ∀ x ∈ Λ = {ε ≤ V (x) ≤ c}
∂x
Suppose x(t0 ) ∈ Λ and x(t) reaches Ωε at t = t0 + T . For
t0 ≤ t ≤ t0 + T , V satisfies the conditions for the uniform
asymptotic stability. Therefore, the trajectory behaves as if
the origin was uniformly asymptotically stable and satisfies
kx(t)k ≤ β(kx(t0 )k, t − t0 ), for some β ∈ KL
For t ≥ t0 + T ,
kx(t)k ≤ α−1
1 (α2 (µ))
– p. 12/?
kx(t)k ≤ β(kx(t0 )k, t − t0 ) + α−1
1 (α2 (µ)), ∀ t ≥ t0
!
kx(t)k ≤ β(kx(t0 )k, t − t0 ) + γ sup ku(τ )k , ∀ t ≥ t0
τ ≥t0
Since x(t) depends only on u(τ ) for t0 ≤ τ ≤ t, the

supremum on the right-hand side can be taken over [t0 , t]
– p. 13/?
Example
ẋ = −x3 + u
The origin of ẋ = −x3 is globally asymptotically stable
V = 12 x2
V̇ = −x4 + xu
= −(1 − θ)x4 − θx4 + xu
1/3
≤ −(1 − θ)x4 , ∀ |x| ≥ |u|
θ
0<θ<1
The system is ISS with
γ(r) = (r/θ)1/3
– p. 14/?
Example
ẋ = −x − 2x3 + (1 + x2 )u2
The origin of ẋ = −x − 2x3 is globally exponentially stable
V = 12 x2
V̇ = −x2 − 2x4 + x(1 + x2 )u2

= −x4 − x2 (1 + x2 ) + x(1 + x2 )u2
≤ −x4 , ∀ |x| ≥ u2
The system is ISS with γ(r) = r 2
– p. 15/?
Example
ẋ1 = −x1 + x22 , ẋ2 = −x2 + u
Investigate GAS of ẋ1 = −x1 + x22 , ẋ2 = −x2
V (x) = 12 x21 + 14 x42

1 2 2 1
−x21 x1 x22 x42 x42

V̇ = + − = −(x1 − 2 x2 ) − 1− 4
Now u 6= 0, V̇ = − 12 (x1 − x22 )2 − 12 (x21 + x42 ) + x32 u

≤ − 12 (x21 + x42 ) + |x2 |3 |u|
V̇ ≤ − 12 (1 − θ)(x21 + x42 ) − 12 θ(x21 + x42 ) + |x2 |3 |u|

(0 < θ < 1)
– p. 16/?
− 12 θ(x21 + x42 ) + |x2 |3 |u| ≤ 0
2
2|u| 2|u| 2|u|

if |x2 | ≥ or |x2 | ≤ and |x1 | ≥
θ θ θ
s
2
2|u| 2|u|

if kxk ≥ 1+
θ θ
s
2
2r 2r

ρ(r) = 1+
θ θ
V̇ ≤ − 12 (1 − θ)(x21 + x42 ), ∀ kxk ≥ ρ(|u|)

The system is ISS
– p. 17/?
Find γ
V (x) = 12 x21 + 14 x42
1
For |x2 | ≤ |x1 |, 4
(x2
1 + x22 ) ≤ 14 x21 + 14 x21 = 12 x21 ≤ V (x)
1 1
For |x2 | ≥ |x1 |, 16 (x21 +x22 )2 ≤ 16 (x22 +x22 )2 = 14 x42 ≤ V (x)
min 4 kxk , 16 kxk ≤ V (x) ≤ 21 kxk2 + 14 kxk4

1 2 1 4

1
α1 (r) = 4 min r , 4 r , α2 (r) = 12 r 2 + 14 r 4
2 1 4
γ = α−1
1 ◦ α2 ◦ ρ
( 1
α−1 (s) = 2(s) 4 , if s ≤ 1
1 √
2 s, if s ≥ 1
– p. 18/?
Lecture # 20
Input-Output Stability
– p. 1/1
Input-Output Models
y = Hu
u(t) is a piecewise continuous function of t and belongs to

a linear space of signals
The space of bounded functions: supt≥0 ku(t)k < ∞
The
R∞ T
space of square-integrable functions:
0 u (t)u(t) dt < ∞
Norm of a signal kuk:

kuk ≥ 0 and kuk = 0 ⇔ u = 0
kauk = akuk for any a > 0
Triangle Inequality: ku1 + u2 k ≤ ku1 k + ku2 k

– p. 2/1
Lp spaces:
L∞ : kukL∞ = sup ku(t)k < ∞

t≥0
sZ
∞
L2 ; kukL2 = uT (t)u(t) dt < ∞
0
Z ∞ 1/p
Lp ; kukLp = ku(t)kp dt < ∞, 1 ≤ p < ∞
0
Notation Lm
p : p is the type of p-norm used to define the
space and m is the dimension of u
– p. 3/1
Extended Space: Le = {u | uτ ∈ L, ( ∀ τ ∈ [0, ∞)}
u(t), 0 ≤ t ≤ τ
uτ is a truncation of u: uτ (t) =
0, t>τ
Le is a linear space and L ⊂ Le

(
t, 0 ≤ t ≤ τ
Example: u(t) = t, uτ (t) =
0, t>τ
/ L∞ but uτ ∈ L∞e
u∈
Causality: A mapping H : Lm e → L q is causal if the value
e
of the output (Hu)(t) at any time t depends only on the
values of the input up to time t
(Hu)τ = (Huτ )τ
– p. 4/1
Definition: A mapping H : Lm
e → L q is L stable if ∃ α ∈ K
e
β ≥ 0 such that
k(Hu)τ kL ≤ α (kuτ kL ) + β, ∀ u ∈ Lm
e and τ ∈ [0, ∞)
It is finite-gain L stable if ∃ γ ≥ 0 and β ≥ 0 such that
k(Hu)τ kL ≤ γkuτ kL + β, ∀ u ∈ Lm
e and τ ∈ [0, ∞)
It is small-signal L stable (respectively, finite-gain L stable)

if ∃ r > 0 such that the inequality is satisfied for all u ∈ Lme
with sup0≤t≤τ ku(t)k ≤ r
– p. 5/1
Example: Memoryless function y = h(u)
ecu − e−cu
h(u) = a + b tanh cu = a + b , a, b, c > 0
ecu + e−cu
′ 4bc
h (u) = 2 ≤ bc ⇒ |h(u)| ≤ a+bc|u|, ∀ u ∈ R
(ecu + e−cu )
Finite-gain L∞ stable with β = a and γ = bc
h(u) = b tanh cu, |h(u)| ≤ bc|u|, ∀ u ∈ R
Z ∞ Z ∞
|h(u(t))|p dt ≤ (bc)p |u(t)|p dt, for p ∈ [1, ∞)
0 0
Finite-gain Lp stable with β = 0 and γ = bc
– p. 6/1
h(u) = u2
!2
sup |h(u(t))| ≤ sup |u(t)|
t≥0 t≥0
L∞ stable with β = 0 and α(r) = r 2

It is not finite-gain L∞ stable. Why?
h(u) = tan u
π tan r

|u| ≤ r < ⇒ |h(u)| ≤ |u|
2 r
Small-signal finite-gain Lp stable with β = 0 and γ = tan r/r
– p. 7/1
Example: SISO causal convolution operator
Z t
y(t) = h(t − σ)u(σ) dσ, h(t) = 0 for t < 0
0
Z ∞
Suppose h ∈ L1 ⇔ khkL1 = |h(σ)| dσ < ∞
0
Rt
|y(t)| ≤ 0 |h(t − σ)| |u(σ)| dσ
Rt
≤ 0 |h(t − σ)| dσ sup0≤σ≤τ |u(σ)|
Rt
= 0 |h(s)| ds sup0≤σ≤τ |u(σ)|
kyτ kL∞ ≤ khkL1 kuτ kL∞ , ∀ τ ∈ [0, ∞)
Finite-gain L∞ stable
Also, finite-gain Lp stable for p ∈ [1, ∞) (see textbook)
– p. 8/1
L Stability of State Models
ẋ = f (x, u), y = h(x, u)
0 = f (0, 0), 0 = h(0, 0)

Case 1: The origin of ẋ = f (x, 0) is exponentially stable
∂V ∂V

2
f (x, 0) ≤ −c3 kxk , ≤ c4 kxk
∂x ∂x
kf (x, u)−f (x, 0)k ≤ Lkuk, kh(x, u)k ≤ η1 kxk+η2 kuk
∀ kxk ≤ r and kuk ≤ ru
– p. 9/1
V̇ = ∂V
∂x f (x, 0) + ∂V
∂x [f (x, u) − f (x, 0)] √
≤ −c3 kxk2 + c4 Lkxk kuk ≤ − cc32 V + c4 L
√
c1
kuk V
c3 c4 L

p
W (t) = V (x(t)) ⇒ Ẇ ≤ − W + √ ku(t)k
2c2 2 c1
tc c4 L
Z t (t−τ )c 3
− 2c3 −
W (t) ≤ e 2 W (0) + √ e 2c 2
ku(τ )k dτ
2 c1 0
r
c2 tc c4 L
Z t (t−τ )c 3
− 2c3 −
kx(t)k ≤ kx(0)ke 2 + e 2c 2
ku(τ )k dτ
c1 2c1 0
Z t
ky(t)k ≤ k0 kx(0)ke−at +k2 e−a(t−τ ) ku(τ )k dτ +k3 ku(t)k
0
– p. 10/1
p
Theorem 5.1: For each x(0) with kx(0)k ≤ r c1 /c2 , the
system is small-signal finite-gain Lp stable for each
p ∈ [1, ∞]
If the assumptions hold globally, then, for each x(0) ∈ Rn ,

the system is finite-gain Lp stable for each p ∈ [1, ∞]
Example
ẋ = −x − x3 + u, y = tanh x + u
V = 12 x2 ⇒ x(−x − x3 ) ≤ −x2
c1 = c2 = 12 , c3 = c4 = 1, L = η1 = η2 = 1
Finite-gain Lp stable for each x(0) ∈ R and each p ∈ [1, ∞]
– p. 11/1
Case 2: The origin of ẋ = f (x, 0) is asymptotically stable
Theorem 5.3: Suppose that, for all (x, u), f is locally

Lipschitz and h is continuous and satisfies
kh(x, u)k ≤ α1 (kxk) + α2 (kuk) + η, α1 , α2 ∈ K, η ≥ 0
If ẋ = f (x, u) is ISS, then, for each x(0) ∈ Rn , the system
ẋ = f (x, u), y = h(x, u)
is L∞ stable
– p. 12/1
Proof
!
kx(t)k ≤ β(kx(0)k, t)+γ sup ku(t)k , β ∈ KL, γ ∈ K
0≤t≤τ

ky(t)k ≤ α1 β(kx(0)k, t) + γ sup0≤t≤τ ku(t)k
+ α2 (ku(t)k) + η
α1 (a + b) ≤ α1 (2a) + α1 (2b)

ky(t)k ≤ α1 (2β(kx(0)k, t)) + α1 2γ sup0≤t≤τ ku(t)k
+ α2 (ku(t)k) + η
kyτ kL∞ ≤ γ0 (kuτ kL∞ ) + β0
γ0 = α1 ◦ 2γ + α2 and β0 = α1 (2β(kx(0)k, 0)) + η
– p. 13/1
Theorem (Rephrasing of Thm 5.2): Suppose f is locally
Lipschitz and h is continuous in some neighborhood of
(x = 0, u = 0). If the origin of ẋ = f (x, 0) is
asymptotically stable, then there is a constant k1 > 0 such
that for each x(0) with kx(0)k < k1 , the system
ẋ = f (x, u), y = h(x, u)
is small-signal L∞ stable
– p. 14/1
Example
ẋ1 = −x31 + x2 , ẋ2 = −x1 − x32 + u, y = x1 + x2
V = (x21 + x22 ) ⇒ V̇ = −2x41 − 2x42 + 2x2 u
x41 + x42 ≥ 21 kxk4
V̇ ≤ −kxk4 + 2kxk|u|
= −(1 − θ)kxk4 − θkxk4 + 2kxk|u|, 0 < θ < 1
1/3
≤ −(1 − θ)kxk4 , ∀ kxk ≥ 2|u|
θ
ISS
L∞ stable
– p. 15/1
Lecture # 21
L2 Gain
&
The Small-Gain theorem
– p. 1/1
Theorem 5.4: Consider the linear time-invariant system
ẋ = Ax + Bu, y = Cx + Du
where A is Hurwitz. Let G(s) = C(sI − A)−1 B + D .

Then, the L2 gain of the system is supω∈R kG(jω)k
– p. 2/1
Lemma: Consider the time-invariant system
ẋ = f (x, u), y = h(x, u)
where f is locally Lipschitz and h is continuous for all

x ∈ Rn and u ∈ Rm . Let V (x) be a positive semidefinite
function such that
∂V
V̇ = f (x, u) ≤ a(γ 2 kuk2 − kyk2 ), a, γ > 0
∂x
Then, for each x(0) ∈ Rn , the system is finite-gain L2
stable and its L2 gain is less than or equal to γ . In particular
s
V (x(0))
kyτ kL2 ≤ γkuτ kL2 +
a
– p. 3/1
Proof
Z τ Z τ
V (x(τ ))−V (x(0)) ≤ aγ 2 ku(t)k2 dt− a ky(t)k2 dt
0 0
– p. 4/1
Proof
Z τ Z τ
0 0
V (x) ≥ 0
Z τ Z τ
V (x(0))
2 2 2
ky(t)k dt ≤ γ ku(t)k dt +
0 0 a
– p. 4/1
Proof
Z τ Z τ
0 0
V (x) ≥ 0
Z τ Z τ
V (x(0))
2 2 2
ky(t)k dt ≤ γ ku(t)k dt +
0 0 a
s
V (x(0))
kyτ kL2 ≤ γkuτ kL2 +
a
– p. 4/1
Lemma 6.5: If the system
ẋ = f (x, u), y = h(x, u)
is output strictly passive with
uT y ≥ V̇ + δy T y, δ>0
then it is finite-gain L2 stable and its L2 gain is less than or

equal to 1/δ
– p. 5/1
Lemma 6.5: If the system
ẋ = f (x, u), y = h(x, u)
is output strictly passive with
uT y ≥ V̇ + δy T y, δ>0
then it is finite-gain L2 stable and its L2 gain is less than or

equal to 1/δ
Proof
V̇ ≤ uT y − δy T y
1
= − 2δ (u − δy)T (u − δy) + 1 T
2δ u u − 2δ y T y
δ 1 T T

≤ 2 δ2 u u − y y
– p. 5/1
Example
ẋ1 = x2 , ẋ2 = −ax31 − kx2 + u, y = x2 , a, k > 0
– p. 6/1
Example
ẋ1 = x2 , ẋ2 = −ax31 − kx2 + u, y = x2 , a, k > 0

a 4
V (x) = 4 x1 + 12 x22
– p. 6/1
Example
ẋ1 = x2 , ẋ2 = −ax31 − kx2 + u, y = x2 , a, k > 0

a 4
V (x) = 4 x1 + 12 x22
V̇ = ax31 x2 + x2 (−ax31 − kx2 + u)

= −kx22 + x2 u = −ky 2 + yu
– p. 6/1
Example
ẋ1 = x2 , ẋ2 = −ax31 − kx2 + u, y = x2 , a, k > 0

a 4
V (x) = 4 x1 + 12 x22
V̇ = ax31 x2 + x2 (−ax31 − kx2 + u)

= −kx22 + x2 u = −ky 2 + yu
The system is finite-gain L2 stable and its L2 gain is less

than or equal to 1/k
– p. 6/1
Theorem 5.5: Consider the time-invariant system
ẋ = f (x) + G(x)u, y = h(x)
f (0) = 0, h(0) = 0
where f and G are locally Lipschitz and h is continuous
over Rn . Suppose ∃ γ > 0 and a continuously
differentiable, positive semidefinite function V (x) that
satisfies the Hamilton–Jacobi inequality
T
∂V 1 ∂V ∂V 1

T
f (x)+ G(x)G (x) + hT (x)h(x) ≤ 0
∂x 2γ 2 ∂x ∂x 2
∀ x ∈ Rn . Then, for each x(0) ∈ Rn , the system is

finite-gain L2 stable and its L2 gain ≤ γ
– p. 7/1
Proof
∂V ∂V
f (x) + G(x)u =
∂x ∂x
2
T
1 2 1 ∂V ∂V

T
− γ u − 2 G (x) + f (x)

2 γ ∂x ∂x
T
1 ∂V ∂V 1 2

T 2
+ G(x)G (x) + γ kuk
2γ 2 ∂x ∂x 2
– p. 8/1
Proof
∂V ∂V
f (x) + G(x)u =
∂x ∂x
2
T
1 2 1 ∂V ∂V

T
− γ u − 2 G (x) + f (x)

2 γ ∂x ∂x
T
1 ∂V ∂V 1 2

T 2
+ G(x)G (x) + γ kuk
2γ 2 ∂x ∂x 2
1 2 2 1
V̇ ≤ γ kuk − kyk2
2 2
– p. 8/1
Example
ẋ = Ax + Bu, y = Cx
– p. 9/1
Example
Suppose there is P = P T ≥ 0 that satisfies the Riccati
equation
T 1 T T
PA + A P + 2
P BB P + C C=0
γ
for some γ > 0.
– p. 9/1
Example
equation
T 1 T T
PA + A P + 2
P BB P + C C=0
γ
for some γ > 0. Verify that V (x) = 21 xT P x satisfies the

Hamilton-Jacobi equation
– p. 9/1
Example
equation
T 1 T T
PA + A P + 2
P BB P + C C=0
γ
for some γ > 0. Verify that V (x) = 21 xT P x satisfies the

Hamilton-Jacobi equation
The system is finite-gain L2 stable and its L2 gain is less

than or equal to γ
– p. 9/1
The Small-Gain Theorem
u1-e1- y1

H1 -
+−
6
y2 e + u
?

H2 2 + 2

– p. 10/1
u1-e1- y1

H1 -
+−
6
y2 e + u
?

H2 2 + 2

ky1τ kL ≤ γ1 ke1τ kL + β1 , ∀ e1 ∈ Lm
e , ∀ τ ∈ [0, ∞)
– p. 10/1
u1-e1- y1

H1 -
+−
6
y2 e + u
?

H2 2 + 2

ky1τ kL ≤ γ1 ke1τ kL + β1 , ∀ e1 ∈ Lm
e , ∀ τ ∈ [0, ∞)
ky2τ kL ≤ γ2 ke2τ kL + β2 , ∀ e2 ∈ Lqe , ∀ τ ∈ [0, ∞)
– p. 10/1
" # " # " #
u1 y1 e1
u= , y= , e=
u2 y2 e2
– p. 11/1
" # " # " #
u1 y1 e1
u= , y= , e=
u2 y2 e2
Theorem: The feedback connection is finite-gain L stable if

γ1 γ2 < 1
– p. 11/1
" # " # " #
u1 y1 e1
u= , y= , e=
u2 y2 e2

γ1 γ2 < 1
Proof
e1τ = u1τ − (H2 e2 )τ , e2τ = u2τ + (H1 e1 )τ
– p. 11/1
" # " # " #
u1 y1 e1
u= , y= , e=
u2 y2 e2

γ1 γ2 < 1
Proof
e1τ = u1τ − (H2 e2 )τ , e2τ = u2τ + (H1 e1 )τ
ke1τ kL ≤ ku1τ kL + k(H2 e2 )τ kL

≤ ku1τ kL + γ2 ke2τ kL + β2
– p. 11/1
ke1τ kL ≤ ku1τ kL + γ2 (ku2τ kL + γ1 ke1τ kL + β1 ) + β2
= γ1 γ2 ke1τ kL
+ (ku1τ kL + γ2 ku2τ kL + β2 + γ2 β1 )
– p. 12/1
= γ1 γ2 ke1τ kL
+ (ku1τ kL + γ2 ku2τ kL + β2 + γ2 β1 )
1
ke1τ kL ≤ (ku1τ kL + γ2 ku2τ kL + β2 + γ2 β1 )
1 − γ1 γ2
– p. 12/1
= γ1 γ2 ke1τ kL
+ (ku1τ kL + γ2 ku2τ kL + β2 + γ2 β1 )
1
1 − γ1 γ2
1
1 − γ1 γ2
– p. 12/1
= γ1 γ2 ke1τ kL
+ (ku1τ kL + γ2 ku2τ kL + β2 + γ2 β1 )
1
1 − γ1 γ2
1
1 − γ1 γ2
keτ kL ≤ ke1τ kL + ke2τ kL
– p. 12/1
Lecture # 22
Normal Form
– p. 1/1
Relative Degree
ẋ = f (x) + g(x)u, y = h(x)
where f , g , and h are sufficiently smooth in a domain D

f : D → Rn and g : D → Rn are called vector fields on D
∂h def
ẏ = [f (x) + g(x)u] = Lf h(x) + Lg h(x) u
∂x
∂h
Lf h(x) = f (x)
∂x
is the Lie Derivative of h with respect to f or along f
– p. 2/1
∂(Lf h)
Lg Lf h(x) = g(x)
∂x
∂(Lf h)
L2f h(x) = Lf Lf h(x) = f (x)
∂x
∂(Lk−1
f h)
Lkf h(x) = Lf Lk−1
f h(x) = f (x)
∂x
L0f h(x) = h(x)
ẏ = Lf h(x) + Lg h(x) u
Lg h(x) = 0 ⇒ ẏ = Lf h(x)
(2) ∂(Lf h)
y = [f (x) + g(x)u] = L2f h(x) + Lg Lf h(x) u
∂x
– p. 3/1
Lg Lf h(x) = 0 ⇒ y (2) = L2f h(x)
y (3) = L3f h(x) + Lg L2f h(x) u

ρ−1
Lg Li−1
f h(x) = 0, i = 1, 2, . . . , ρ − 1; Lg Lf h(x) 6= 0
ρ ρ−1
y (ρ) = Lf h(x) + Lg Lf h(x) u
ẋ = f (x) + g(x)u, y = h(x)
has relative degree ρ, 1 ≤ ρ ≤ n, in D0 ⊂ D if ∀ x ∈ D0

ρ−1
Lg Li−1
f h(x) = 0, i = 1, 2, . . . , ρ − 1; Lg Lf h(x) 6= 0
– p. 4/1
Example
ẋ1 = x2 , ẋ2 = −x1 + ε(1 − x21 )x2 + u, y = x1 , ε>0
ẏ = ẋ1 = x2
ÿ = ẋ2 = −x1 + ε(1 − x21 )x2 + u
Relative degree = 2 over R2
Example
ẋ1 = x2 , ẋ2 = −x1 + ε(1 − x21 )x2 + u, y = x2 , ε>0
ẏ = ẋ2 = −x1 + ε(1 − x21 )x2 + u

Relative degree = 1 over R2
– p. 5/1
Example
ẋ1 = x2 , ẋ2 = −x1 +ε(1−x21 )x2 +u, y = x1 +x22 , ε>0
ẏ = x2 + 2x2 [−x1 + ε(1 − x21 )x2 + u]

Relative degree = 1 over {x2 6= 0}
Example: Field-controlled DC motor
ẋ1 = −ax1 +u, ẋ2 = −bx2 +k−cx1 x3 , ẋ3 = θx1 x2 , y = x3
a, b, c, k, and θ are positive constants

ẏ = ẋ3 = θx1 x2
ÿ = θx1 ẋ2 + θ ẋ1 x2 = (·) + θx2 u
Relative degree = 2 over {x2 6= 0}
– p. 6/1
Normal Form
Change of variables:
 
φ1 (x)
 .. 

 . 
    
 φn−ρ (x)  φ(x) η
 
 def   def 
z = T (x) =  − − −  =  −−−  =  −−− 
 
 
 h(x)  ψ(x) ξ
..
 
.
 
 
ρ−1
Lf h(x)
φ1 to φn−ρ are chosen such that T (x) is a diffeomorphism

on a domain D0 ⊂ D
– p. 7/1
∂φ
η̇ = [f (x) + g(x)u] = f0 (η, ξ) + g0 (η, ξ)u
∂x
ξ̇i = ξi+1 , 1≤i≤ρ−1
ξ̇ρ = Lρf h(x) + Lg Lρ−1
f h(x) u
y = ξ1
Choose φ(x) such that T (x) is a diffeomorphism and

∂φi
g(x) = 0, for 1 ≤ i ≤ n − ρ, ∀ x ∈ D0
∂x
Always possible (at least locally)
η̇ = f0 (η, ξ)
– p. 8/1
Theorem 13.1: Suppose the system
ẋ = f (x) + g(x)u, y = h(x)
has relative degree ρ (≤ n) in D . If ρ = n, then for every

x0 ∈ D , a neighborhood N of x0 exists such that the map
T (x) = ψ(x), restricted to N , is a diffeomorphism on N . If
ρ < n, then, for every x0 ∈ D , a neighborhood N of x0
and smooth functions φ1 (x), . . . , φn−ρ (x) exist such that
∂φi
g(x) = 0, for 1 ≤ i ≤ n − ρ
∂x
" #
φ(x)
is satisfied for all x ∈ N and the map T (x) = ,
ψ(x)
restricted to N , is a diffeomorphism on N
– p. 9/1
Normal Form: η̇ = f0 (η, ξ)
ξ̇i = ξi+1 , 1≤i≤ρ−1
ρ ρ−1
ξ̇ρ = Lf h(x) + Lg Lf h(x) u
y = ξ1
   
0 1 0 ... 0 0
0 0 1 ... 0
 
0
   
 
.. ... .. ..
   
Ac =  . .  , Bc = 
 
 . 

 .. 
 0
 

 . 0 1 
 
0 ... ... 0 0 1
h i
Cc = 1 0 . . . 0 0
– p. 10/1
η̇ = f0 (η, ξ)
h i
ρ ρ−1
ξ̇ = Ac ξ + Bc Lf h(x) + Lg Lf h(x) u
y = Cc ξ
ρ
ρ−1 Lf h(x)
γ(x) = Lg Lf h(x), α(x) = −
Lg Lρ−1
f h(x)
ξ̇ = Ac ξ + Bc γ(x)[u − α(x)]
If x∗ is an open-loop equilibrium point at which y = 0; i.e.,
f (x∗ ) = 0 and h(x∗ ) = 0, then ψ(x∗) = 0. Take
φ(x∗ ) = 0 so that z = 0 is an open-loop equilibrium point.
– p. 11/1
Zero Dynamics
η̇ = f0 (η, ξ)
ξ̇ = Ac ξ + Bc γ(x)[u − α(x)]
y = Cc ξ
y(t) ≡ 0 ⇒ ξ(t) ≡ 0 ⇒ u(t) ≡ α(x(t)) ⇒ η̇ = f0 (η, 0)

Definition: The equation η̇ = f0 (η, 0) is called the zero
dynamics of the system. The system is said to be minimum
phase if zero dynamics have an asymptotically stable
equilibrium point in the domain of interest (at the origin if
T (0) = 0)
The zero dynamics can be characterized in the

x-coordinates
– p. 12/1
ρ−1
Z ∗ = {x ∈ D0 | h(x) = Lf h(x) = · · · = Lf h(x) = 0}
y(t) ≡ 0 ⇒ x(t) ∈ Z ∗
def
⇒ u = u∗ (x) = α(x)|x∈Z ∗
The restricted motion of the system is described by
def
ẋ = f ∗ (x) = [f (x) + g(x)α(x)]x∈Z ∗
– p. 13/1
Example
ẋ1 = x2 , ẋ2 = −x1 + ε(1 − x21 )x2 + u, y = x2
ẏ = ẋ2 = −x1 + ε(1 − x21 )x2 + u ⇒ ρ = 1

y(t) ≡ 0 ⇒ x2 (t) ≡ 0 ⇒ ẋ1 = 0
Non-minimum phase
– p. 14/1
Example
2 + x23
ẋ1 = −x1 + u, ẋ2 = x3 , ẋ3 = x1 x3 + u, y = x2
1+ x23
ẏ = ẋ2 = x3
ÿ = ẋ3 = x1 x3 + u ⇒ ρ = 2
L2f h(x)
γ = Lg Lf h(x) = 1, α = − = −x1 x3
Lg Lf h(x)
Z ∗ = {x2 = x3 = 0}
u = u∗ (x) = 0 ⇒ ẋ1 = −x1
Minimum phase
– p. 15/1
Find φ(x) such that
2+x23
 
∂φ h i 1+x23
∂φ ∂φ ∂φ
φ(0) = 0, g(x) = , , 0 =0
 
∂x ∂x1 ∂x2 ∂x3 
1
and h iT
T (x) = φ(x) x2 x3
is a diffeomorphism
∂φ 2 + x23 ∂φ
· + =0
∂x1 1+ x23 ∂x3
φ(x) = −x1 + x3 + tan−1 x3

– p. 16/1
h iT
T (x) = −x1 + x3 + tan−1 x3 , x2 , x3
is a global diffeomorphism
η = −x1 + x3 + tan−1 x3 , ξ1 = x2 , ξ2 = x3
!
2 + ξ22
−η + ξ2 + tan−1 ξ2

η̇ = 1+ 2
ξ2
1+ ξ2
ξ̇1 = ξ2
−1

ξ̇2 = −η + ξ2 + tan ξ2 ξ2 + u
y = ξ1
– p. 17/1
Lecture # 23
Controller Form
– p. 1/1
Definition: A nonlinear system is in the controller form if
ẋ = Ax + Bγ(x)[u − α(x)]
where (A, B) is controllable and γ(x) is a nonsingular
u = α(x) + γ −1 (x)v ⇒ ẋ = Ax + Bv
The n-dimensional single-input (SI) system
ẋ = f (x) + g(x)u
can be transformed into the controller form if ∃ h(x) s.t.
ẋ = f (x) + g(x)u, y = h(x)
has relative degree n. Why?

– p. 2/1
Transform the system into the normal form
ż = Ac z + Bc γ(z)[u − α(z)], y = Cc z
On the other hand, if there is a change of variables

ζ = S(x) that transforms the SI system
ẋ = f (x) + g(x)u
into the controller form

ζ̇ = Aζ + Bγ(ζ)[u − α(ζ)]
then there is a function h(x) such that the system

ẋ = f (x) + g(x)u, y = h(x)
has relative degree n. Why?

– p. 3/1
For any controllable pair (A, B), we can find a nonsingular
matrix M that transforms (A, B) into a controllable
canonical form:
M AM −1 = Ac + Bc λT , M B = Bc
def
z = M ζ = M S(x) = T (x)
ż = Ac z + Bc γ(·)[u − α(·)]
h(x) = T1 (x)
– p. 4/1
In summary, the n-dimensional SI system
ẋ = f (x) + g(x)u
is transformable into the controller form if and only if ∃ h(x)

such that
ẋ = f (x) + g(x)u, y = h(x)
has relative degree n

Search for a smooth function h(x) such that
Lg Li−1
f h(x) = 0, i = 1, 2, . . . , n−1, and L Ln−1
g f h(x) 6= 0
h i
T (x) = h(x), Lf h(x), ··· Ln−1
f h(x)
– p. 5/1
The Lie Bracket: For two vector fields f and g , the Lie
bracket [f, g] is a third vector field defined by
∂g ∂f
[f, g](x) = f (x) − g(x)
∂x ∂x
Notation:
ad0f g(x) = g(x), adf g(x) = [f, g](x)
adkf g(x) = [f, adfk−1 g](x), k ≥ 1

Properties:
[f, g] = −[g, f ]
For constant vector fields f and g , [f, g] = 0

– p. 6/1
Example
" # " #
x2 0
f = , g=
− sin x1 − x2 x1
" # " # " # " #
0 0 x2 0 1 0
[f, g] = −
1 0 − sin x1 − x2 − cos x1 −1 x1
" #
−x1
adf g = [f, g] =
x1 + x2
– p. 7/1
" # " #
x2 −x1
f = , adf g =
− sin x1 − x2 x1 + x2
ad2f g = [f, adf g] =

" # " #
−1 0 x2
1 1 − sin x1 − x2
" # " #
0 1 −x1
−
− cos x1 −1 x1 + x2
" #
−x1 − 2x2
=
x1 + x2 − sin x1 − x1 cos x1
– p. 8/1
Distribution: For vector fields f1 , f2 , . . . , fk on D ⊂ Rn , let
∆(x) = span{f1 (x), f2 (x), . . . , fk (x)}
The collection of all vector spaces ∆(x) for x ∈ D is called

a distribution and referred to by
∆ = span{f1 , f2 , . . . , fk }
If dim(∆(x)) = k for all x ∈ D , we say that ∆ is a

nonsingular distribution on D , generated by f1 , . . . , fk
A distribution ∆ is involutive if
g1 ∈ ∆ and g2 ∈ ∆ ⇒ [g1 , g2 ] ∈ ∆
– p. 9/1
Lemma: If ∆ is a nonsingular distribution, generated by
f1 , . . . , fk , then it is involutive if and only if
[fi , fj ] ∈ ∆, ∀ 1 ≤ i, j ≤ k
Example: D = R3 ; ∆ = span{f1 , f2 }
   
2x2 1
f1 =  1  , f2 =  0  , dim(∆(x)) = 2, ∀ x ∈ D
   
0 x2
 
0
∂f2 ∂f1
[f1 , f2 ] = f1 − f2 =  0 
 
∂x ∂x
1
– p. 10/1
rank [f1 (x), f2 (x), [f1 , f2 ](x)] =
 
2x2 1 0
rank  1 0 0  = 3, ∀ x ∈ D
 
0 x2 1
∆ is not involutive
– p. 11/1
Example: D = {x ∈ R3 | x21 + x23 6= 0}; ∆ = span{f1 , f2 }
   
2x3 −x1
f1 =  −1  , f2 =  −2x2  , dim(∆(x)) = 2, ∀ x ∈ D
   
0 x3
 
−4x3
∂f2 ∂f1
[f1 , f2 ] = f1 − f2 =  2 
 
∂x ∂x
0
 
2x3 −x1 −4x3
rank  −1 −2x2 2  = 2, ∀ x ∈ D
 
0 x3 0
∆ is involutive
– p. 12/1
Theorem: The n-dimensional SI system
ẋ = f (x) + g(x)u
is transformable into the controller form if and only if there is

a domain D0 such that
rank[g(x), adf g(x), . . . , adn−1

f g(x)] = n, ∀ x ∈ D0
and
span {g, adf g, . . . , adn−2

f g} is involutive in D0
– p. 13/1
Example
" # " #
a sin x2 0
ẋ = + u
−x21 1
" #
∂f −a cos x2
adf g = [f, g] = − g=
∂x 0
" #
0 −a cos x2
[g(x), adf g(x)] =
1 0
rank[g(x), adf g(x)] = 2, ∀ x such that cos x2 6= 0

span{g} is involutive
Find h such that Lg h(x) = 0, and Lg Lf h(x) 6= 0
– p. 14/1
∂h ∂h
g= = 0 ⇒ h is independent of x2
∂x ∂x2
∂h
Lf h(x) = a sin x2
∂x1
∂(Lf h) ∂(Lf h) ∂h
Lg Lf h(x) = g= = a cos x2
∂x ∂x2 ∂x1
∂h
Lg Lf h(x) 6= 0 in D0 = {x ∈ R2 | cos x2 6= 0} if ∂x1 6= 0
" # " #
h x1
Take h(x) = x1 ⇒ T (x) = =
Lf h a sin x2
– p. 15/1
Example (Field-Controlled DC Motor)
   
−ax1 1
ẋ =  −bx2 + k − cx1 x3  +  0  u
   
θx1 x2 0
   
a a2
adf g =  cx3  ; ad2f g =  (a + b)cx3
   

−θx2 (b − a)θx2 − θk
 
1 a a2
2
[g(x), adf g(x), adf g(x)] =  0 cx3 (a + b)cx3
 

0 −θx2 (b − a)θx2 − θk
– p. 16/1
det[·] = cθ(−k + 2bx2 )x3
rank [·] = 3 for x2 6= k/2b and x3 6= 0

span{g, adf g} is involutive if [g, adf g] ∈ span{g, adf g}
     
0 0 0 1 0
∂(adf g)
[g, adf g] = g= 0 0 c   0 = 0 
     
∂x
0 −θ 0 0 0
⇒ span{g, adf g} is involutive
3 k
D0 = {x ∈ R | x2 > and x3 > 0}
2b
Find h such that Lg h(x) = Lg Lf h(x) = 0; Lg L2f h(x) 6= 0
– p. 17/1
x∗ = [0, k/b, ω0 ]T , h(x∗ ) = 0
∂h ∂h
g= = 0 ⇒ h is independent of x1
∂x ∂x1
∂h ∂h
Lf h(x) = [−bx2 + k − cx1 x3 ] + θx1 x2
∂x2 ∂x3
∂h ∂h
[∂(Lf h)/∂x]g = 0 ⇒ cx3 = θx2
∂x2 ∂x3
h = c1 [θx22 + cx23 ] + c2 , Lg L2f h(x) = −2c1 cθ(k − 2bx2 )x3
h(x∗ ) = c1 [θ(k/b)2 + cω02 ] + c2
c1 = 1, c2 = −θ(k/b)2 − cω02
– p. 18/1
Lecture # 24
Observer, Output Feedback

&
Strict Feedback Forms
– p. 1/1
Definition: A nonlinear system is in the observer form if
ẋ = Ax + γ(y, u), y = Cx
where (A, C) is observable
Observer:
x̂˙ = Ax̂ + γ(y, u) + H(y − C x̂)
x̃ = x − x̂
x̃˙ = (A − HC)x̃
Design H such that (A − HC) is Hurwitz
– p. 2/1
Theorem: An n-dimensional single-output (SO) system
ẋ = f (x) + g(x)u, y = h(x)
is transformable into the observer form if and only if there is

a domain D0 such that
∂φ

rank (x) = n, ∀ x ∈ D0
∂x
h iT
where φ= h, Lf h, ··· Ln−1
f h
and the unique vector field solution τ of

∂φ h iT
τ = b, where b = 0, ··· 0, 1
∂x
– p. 3/1
satisfies
j
[adif τ, adf τ ] = 0, 0 ≤ i, j ≤ n − 1
and
j
[g, adf τ ] = 0, 0≤j ≤n−2
The change of variables z = T (x) is given by
∂T h i
τ1 , τ2 , ··· τn =I
∂x
where
τi = (−1)i−1 adi−1
f τ, 1≤i≤n
– p. 4/1
Example
" # " #
β1 (x1 ) + x2 0
ẋ = + u, y = x1
f2 (x) 1
" # " #
h(x) x1
φ(x) = =
Lf h(x) β1 (x1 ) + x2
" #
∂φ ∂φ

1 0
= ∂β1 ; rank (x) = 2, ∀ x
∂x ∂x1 1 ∂x
" # " #
∂φ 0 0
τ = ⇒ τ =
∂x 1 1
– p. 5/1
" #" # " #
∂f ∗ 1 0 1
adf τ = [f, τ ] = − τ =− ∂f2 =− ∂f2
∂x ∗ ∂x2 1 ∂x2
" #" #
∂(adf τ ) 0 0 0
[τ, adf τ ] = τ =− ∂ 2 f2 ∂ 2 f2
∂x ∂x1 ∂x2 ∂x22 1
∂ 2 f2
[τ, adf τ ] = 0 ⇔ = 0 ⇔ f2 (x) = β2 (x1 )+x2 β3 (x1 )
∂x22
[g, τ ] = 0 (g and τ are constant vector fields)
All the conditions are satisfied
– p. 6/1
" #
0 0
τ1 = (−1) ad0f τ =τ =
1
" #
1 1
τ2 = (−1) ad1f τ = −adf τ =
β3 (x1 )
∂T h i
τ1 , τ2 =I
∂x
" #" # " #
∂T1 ∂T1
∂x1 ∂x2 0 1 1 0
∂T2 ∂T2 =
∂x1 ∂x2
1 β3 (x1 ) 0 1
∂T1 ∂T1 ∂T1
∂x2
= 1, ∂x1
+ β3 (x1 ) ∂x 2
=0
∂T2 ∂T2 ∂T2

∂x2 = 0, ∂x1 + β3 (x1 ) ∂x 2
=1
– p. 7/1
∂T1 ∂T1
∂x2
=1 ⇒ ∂x1
+ β3 (x1 ) = 0
R x1
T1 (x) = x2 − 0 β3 (σ) dσ
∂T2 ∂T2
∂x2
=0 ⇒ ∂x1
= 1, T2 (x) = x1
Z x1
z1 = x2 − β3 (σ) dσ, z2 = x1
0
y = z2
" # " #
0 0 β2 (y) − β1 (y)β3 (y) + u
ż = z+ Ry
1 0 0 β3 (σ) dσ + β1 (y)
h i
y = 0 1 z
– p. 8/1
Definition: A nonlinear system is in the output feedback
form if
ẋ1 = x2 + γ1 (y)
ẋ2 = x3 + γ2 (y)
..
.
ẋρ−1 = xρ + γρ−1 (y)
ẋρ = xρ+1 + γρ (y) + bm u, bm > 0
..
.
ẋn−1 = xn + γn−1 (y) + b1 u
ẋn = γn (y) + b0 u
y = x1
– p. 9/1
Show that
The output feedback form is a special case of the
observer form
It has relative degree ρ
It is minimum phase if the polynomial
bm sm + · · · + b1 s + b0
is Hurwitz
– p. 10/1
Definition: A nonlinear system is in the strict feedback form
if
ẋ = f0 (x) + g0 (x)z1
ż1 = f1 (x, z1 ) + g1 (x, z1 )z2
ż2 = f2 (x, z1 , z2 ) + g2 (x, z1 , z2 )z3
..
.
żk−1 = fk−1 (x, z1 , . . . , zk−1 ) + gk−1 (x, z1 , . . . , zk−1 )zk
żk = fk (x, z1 , . . . , zk ) + gk (x, z1 , . . . , zk )u
x ∈ Rn , z1 to zk are scalars
gi (x, z1 , . . . , zi ) 6= 0 for 1 ≤ i ≤ k
– p. 11/1
Find the relative degree if y = z1
Find the zero dynamics if y = z1 and
fi (x, 0) = 0, ∀1≤i≤k
– p. 12/1
Lecture # 25
Stabilization
Basic Concepts & Linearization
– p. 1/?
We want to stabilize the system
ẋ = f (x, u)
at the equilibrium point x = xss
Steady-State Problem: Find steady-state control uss s.t.
0 = f (xss , uss )
xδ = x − xss , uδ = u − uss
def
ẋδ = f (xss + xδ , uss + uδ ) = fδ (xδ , uδ )
fδ (0, 0) = 0
uδ = γ(xδ ) ⇒ u = uss + γ(x − xss )
– p. 2/?
State Feedback Stabilization: Given
ẋ = f (x, u) [f (0, 0) = 0]
find
u = γ(x) [γ(0) = 0]
s.t. the origin is an asymptotically stable equilibrium point of
ẋ = f (x, γ(x))
f and γ are locally Lipschitz functions
– p. 3/?
Linear Systems
ẋ = Ax + Bu
(A, B) is stabilizable (controllable or every uncontrollable
eigenvalue has a negative real part)
Find K such that (A − BK) is Hurwitz
u = −Kx
Typical methods:
Eigenvalue Placement
Eigenvalue-Eigenvector Placement
LQR
– p. 4/?
Linearization
ẋ = f (x, u)
f (0, 0) = 0 and f is continuously differentiable in a domain
Dx × Du that contains the origin (x = 0, u = 0)
(Dx ⊂ Rn , Du ⊂ Rp )
ẋ = Ax + Bu

∂f ∂f
A= (x, u) ; B= (x, u)
∂x x=0,u=0 ∂u x=0,u=0
Assume (A, B) is stabilizable. Design a matrix K such that

(A − BK) is Hurwitz
u = −Kx
– p. 5/?
Closed-loop system:
ẋ = f (x, −Kx)
∂f ∂f

ẋ = (x, −Kx) + (x, −Kx) (−K) x
∂x ∂u x=0
= (A − BK)x
Since (A − BK) is Hurwitz, the origin is an exponentially

stable equilibrium point of the closed-loop system
– p. 6/?
Example (Pendulum Equation):
θ̈ = −a sin θ − bθ̇ + cT
Stabilize the pendulum at θ = δ
0 = −a sin δ + cTss
x1 = θ − δ, x2 = θ̇, u = T − Tss
ẋ1 = x2
ẋ2 = −a[sin(x1 + δ) − sin δ] − bx2 + cu
" # " #
0 1 0 1
A= =
−a cos(x1 + δ) −b −a cos δ −b
x1 =0
– p. 7/?
" # " #
0 1 0
A= ; B=
−a cos δ −b c
h i
K= k1 k2
" #
0 1
A − BK =
−(a cos δ + ck1 ) −(b + ck2 )
a cos δ b
k1 > − , k2 > −
c c
a sin δ a sin δ
T = − Kx = − k1 (θ − δ) − k2 θ̇
c c
– p. 8/?
Notions of Stabilization
ẋ = f (x, u), u = γ(x)
Local Stabilization: The origin of ẋ = f (x, γ(x)) is

asymptotically stable (e.g., linearization)
Regional Stabilization: The origin of ẋ = f (x, γ(x)) is

asymptotically stable and a given region G is a subset of
the region of attraction (for all x(0) ∈ G, limt→∞ x(t) = 0)
(e.g., G ⊂ Ωc = {V (x) ≤ c} where Ωc is an estimate of
the region of attraction)
Global Stabilization: The origin of ẋ = f (x, γ(x)) is

globally asymptotically stable
– p. 9/?
Semiglobal Stabilization: The origin of ẋ = f (x, γ(x)) is
asymptotically stable and γ(x) can be designed such that
any given compact set (no matter how large) can be
included in the region of attraction (Typically u = γp (x) is
dependent on a parameter p such that for any compact set
G, p can be chosen to ensure that G is a subset of the
region of attraction )
What is the difference between global stabilization and

semiglobal stabilization?
– p. 10/?
Example
ẋ = x2 + u
Linearization:
ẋ = u, u = −kx, k > 0
Closed-loop system:
ẋ = −kx + x2
Linearization of the closed-loop system yields ẋ = −kx.

Thus, u = −kx achieves local stabilization
The region of attraction is {x < k}. Thus, for any set

{x ≤ a} with a < k, the control u = −kx achieves
regional stabilization
– p. 11/?
The control u = −kx does not achieve global stabilization
But it achieves semiglobal stabilization because any

compact set {|x| ≤ r} can be included in the region of
attraction by choosing k > r
The control
u = −x2 − kx
achieves global stabilization because it yields the linear
closed-loop system ẋ = −kx whose origin is globally
exponentially stable
– p. 12/?
Practical Stabilization
ẋ = f (x, u) + g(x, u, t)
f (0, 0) = 0, g(0, 0, t) 6= 0
kg(x, u, t)k ≤ δ, ∀ x ∈ Dx , u ∈ Du , t ≥ 0
There is no control u = γ(x), with γ(0) = 0, that can make
the origin of
ẋ = f (x, γ(x)) + g(x, γ(x), t)
uniformly asymptotically stable because the origin is not an

equilibrium point
– p. 13/?
ẋ = f (x, u) + g(x, u, t)
is practically stabilizable if for any ε > 0 there is a control

law u = γ(x) such that the solutions of
ẋ = f (x, γ(x)) + g(x, γ(x), t)
are uniformly ultimately bounded by ε; i.e.,
kx(t)k ≤ ε, ∀t≥T
Typically, u = γp (x) is dependent on a parameter p such

that for any ε > 0, p can be chosen to ensure that ε is an
ultimate bound
– p. 14/?
With practical stabilization, we may have
local practical stabilization
regional practical stabilization
global practical stabilization, or
semiglobal practical stabilization
depending on the region of initial states
– p. 15/?
Example
ẋ = x2 + u + d(t), |d(t)| ≤ δ, ∀ t ≥ 0
u = −kx, k > 0, ⇒ ẋ = x2 − kx + d(t)
V = 12 x2 ⇒ V̇ = x3 − kx2 + xd(t)

V̇ ≤ − k3 x2 − x2 k3 − |x| − |x| k3 |x| − δ
V̇ ≤ − k3 x2 , for 3δ
k
≤ |x| ≤ k
3
3δ 3δ
Take k <ε ⇔ k≥ ε
By choosing k large enough we can achieve semiglobal
practical stabilization
– p. 16/?
ẋ = x2 + u + d(t)
u = −x2 − kx, k > 0, ⇒ ẋ = −kx + d(t)
V = 21 x2 ⇒ V̇ = −kx2 + xd(t)

V̇ ≤ − k2 x2 − |x| k2 |x| − δ
By choosing k large enough we can achieve global practical

stabilization
– p. 17/?
Lecture # 25
Stabilization
Feedback Lineaization
– p. 1/?
Consider the nonlinear system
ẋ = f (x) + G(x)u
f (0) = 0, x ∈ Rn , u ∈ Rm
Suppose there is a change of variables z = T (x), defined
for all x ∈ D ⊂ Rn , that transforms the system into the
controller form
ż = Az + Bγ(x)[u − α(x)]
where (A, B) is controllable and γ(x) is nonsingular for all

x∈D
u = α(x) + γ −1 (x)v ⇒ ż = Az + Bv
– p. 2/?
v = −Kz
Design K such that (A − BK) is Hurwitz
The origin z = 0 of the closed-loop system
ż = (A − BK)z
is globally exponentially stable
u = α(x) − γ −1 (x)KT (x)
Closed-loop system in the x-coordinates:

−1

ẋ = f (x) + G(x) α(x) − γ (x)KT (x)
– p. 3/?
What can we say about the stability of x = 0 as an
equilibrium point of
−1

ẋ = f (x) + G(x) α(x) − γ (x)KT (x)
x = 0 is asymptotically stable because T (x) is a

diffeomorphism. Show it!
Is x = 0 globally asymptotically stable? In general No
It is globally asymptotically stable if T (x) is a global

diffeomorphism (See page 508)
– p. 4/?
What information do we need to implement the control
u = α(x) − γ −1 (x)KT (x) ?
What is the effect of uncertainty in α, γ , and T ?
Let α̂(x), γ̂(x), and T̂ (x) be nominal models of α(x),

γ(x), and T (x)
u = α̂(x) − γ̂ −1 (x)K T̂ (x)
Closed-loop system:
ż = (A − BK)z + Bδ(z)
δ = γ[α̂ − α + γ −1 KT − γ̂ −1 K T̂ ]
– p. 5/?
ż = (A − BK)z + Bδ(z) (∗)
V (z) = z T P z, P (A − BK) + (A − BK)T P = −I

Lemma 13.3
If kδ(z)k ≤ kkzk for all z , where
1
0≤k<
2kP Bk
then the origin of (*) is globally exponentially stable
If kδ(z)k ≤ kkzk + ε for all z , then the state z is

globally ultimately bounded by εc for some c > 0
– p. 6/?
Example (Pendulum Equation):
a
x1 = θ − δ, x2 = θ̇, u = T − Tss = T − sin δ
c
ẋ1 = x2
ẋ2 = −a[sin(x1 + δ) − sin δ] − bx2 + cu
1
u= {a[sin(x1 + δ) − sin δ] − k1 x1 − k2 x2 }
c
" #
0 1
A − BK = is Hurwitz
−k1 −(k2 + b)
– p. 7/?
a 1
T =u+ sin δ = [a sin(x1 + δ) − k1 x1 − k2 x2 ]
c c
Let â and ĉ be nominal models of a and c
1
T = [â sin(x1 + δ) − k1 x1 − k2 x2 ]
ĉ
ẋ = (A − BK)x + Bδ(x)
âc − aĉ c − ĉ

δ(x) = sin(x1 + δ1 ) − (k1 x1 + k2 x2 )
ĉ ĉ
– p. 8/?
âc − aĉ c − ĉ

δ(x) = sin(x1 + δ1 ) − (k1 x1 + k2 x2 )
ĉ ĉ
|δ(x)| ≤ kkxk + ε
âc − aĉ c − ĉ âc − aĉ
q
k = + 2 2
k +k , ε= | sin δ1 |
1 2
ĉ ĉ ĉ
" # " #
p11 p12 p12
P = , PB =
p12 p22 p22
1
k< q
2 p212 + p222
sin δ1 = 0 ⇒ ε = 0
– p. 9/?
Is feedback linearization a good idea?
Example
ẋ = ax − bx3 + u, a, b > 0
u = −(k + a)x + bx3 , k > 0, ⇒ ẋ = −kx
−bx3 is a damping term. Why cancel it?
u = −(k + a)x, k > 0, ⇒ ẋ = −kx − bx3
Which design is better?
– p. 10/?
Example
ẋ1 = x2
ẋ2 = −h(x1 ) + u
h(0) = 0 and x1 h(x1 ) > 0, ∀ x1 6= 0
Feedback Linearization:
u = h(x1 ) − (k1 x1 + k2 x2 )
With y = x2 , the system is passive with

Z x1
V = h(z) dz + 12 x22
0
V̇ = h(x1 )ẋ1 + x2 ẋ2 = yu

– p. 11/?
The control
u = −σ(x2 ), σ(0) = 0, x2 σ(x2 ) > 0 ∀ x2 6= 0
creates a feedback connection of two passive systems with

storage function V
V̇ = −x2 σ(x2 )
x2 (t) ≡ 0 ⇒ ẋ2 (t) ≡ 0 ⇒ h(x1 (t)) ≡ 0 ⇒ x1 (t) ≡ 0

Asymptotic stability of the origin follows from the invariance
principle
Which design is better? (Read Example 13.20)
– p. 12/?
Lecture # 27
Stabilization
Partial Feedback Linearization
– p. 1/1
Consider the nonlinear system
ẋ = f (x) + G(x)u [f (0) = 0]
Suppose there is a change of variables

" # " #
η T1 (x)
z= = T (x) =
ξ T2 (x)
defined for all x ∈ D ⊂ Rn , that transforms the system into
η̇ = f0 (η, ξ)
ξ̇ = Aξ + Bγ(x)[u − α(x)]
(A, B) is controllable and γ(x) is nonsingular for all x ∈ D
– p. 2/1
u = α(x) + γ −1 (x)v
η̇ = f0 (η, ξ), ξ̇ = Aξ + Bv
Suppose the origin of η̇ = f0 (η, 0) is asymptotically stable
v = −Kξ, where (A − BK) is Hurwitz
Lemma 13.1: The origin of
η̇ = f0 (η, ξ), ξ̇ = (A − BK)ξ
is asymptotically stable if the origin of η̇ = f0 (η, 0) is

p
Proof: V (η, ξ) = V1 (η)+k ξT P ξ
– p. 3/1
If the origin of η̇ = f0 (η, 0) is globally asymptotically stable,
will the origin of
η̇ = f0 (η, ξ), ξ̇ = (A − BK)ξ
be globally asymptotically stable? In general No

Example
η̇ = −η + η 2 ξ, ξ̇ = v
The origin of η̇ = −η is globally exponentially stable, but
the origin of
η̇ = −η + η 2 ξ, ξ̇ = −kξ, k>0
is not globally asymptotically stable. The region of

attraction is {ηξ < 1 + k}
– p. 4/1
Example
η̇ = − 12 (1 + ξ2 )η 3 , ξ̇1 = ξ2 , ξ̇2 = v
The origin of η̇ = − 12 η 3 is globally asymptotically stable

" #
2 def 0 1
v = −k ξ1 −2kξ2 = −Kξ ⇒ A−BK =
−k2 −2k
The eigenvalues of (A − BK) are −k and −k

 
(1 + kt)e−kt te−kt
e(A−BK)t = 
 

−k2 te−kt (1 − kt)e−kt
– p. 5/1
Peaking Phenomenon:
2 −kt k
max{k te }= → ∞ as k → ∞
t e
ξ1 (0) = 1, ξ2 (0) = 0 ⇒ ξ2 (t) = −k2 te−kt

η̇ = − 12 1 − k2 te−kt η 3 , η(0) = η0
η02
η 2 (t) =
1 + η02 [t + (1 + kt)e−kt − 1]
If η02 > 1, the system will have a finite escape time if k is
chosen large enough
– p. 6/1
Lemma 13.2: The origin of
η̇ = f0 (η, ξ), ξ̇ = (A − BK)ξ
is globally asymptotically stable if the system η̇ = f0 (η, ξ)

is input-to-state stable
Proof: Use
Lemma 4.7: If ẋ1 = f1 (x1 , x2 ) is ISS and the origin of
ẋ2 = f2 (x2 ) is globally asymptotically stable, then the
origin of
ẋ1 = f1 (x1 , x2 ), ẋ2 = f2 (x2 )
is globally asymptotically stable
– p. 7/1
u = α(x) − γ −1 (x)KT2 (x)
What is the effect of uncertainty in α, γ , and T2 ?
Let α̂(x), γ̂(x), and T̂2 (x) be nominal models of α(x),

γ(x), and T2 (x)
u = α̂(x) − γ̂ −1 (x)K T̂2 (x)
η̇ = f0 (η, ξ), ξ̇ = (A − BK)ξ + Bδ(z)
δ = γ[α̂ − α + γ −1 KT2 − γ̂ −1 K T̂2 ]
– p. 8/1
Lemma 13.4
If kδ(z)k ≤ ε for all z and η̇ = f0 (η, ξ) is input-to-state
stable, then the state z is globally ultimately bounded by
a class K function of ε
If kδ(z)k ≤ kkzk in some neighborhood of z = 0, with

sufficiently small k, and the origin of η̇ = f0 (η, 0) is
exponentially stable, then z = 0 is an exponentially
stable equilibrium point of the system
η̇ = f0 (η, ξ), ξ̇ = (A − BK)ξ + Bδ(z)
– p. 9/1
Proof–First Part: As in Lemma 13.3
kξ(t)k≤ cε, ∀ t ≥ t0
kη(t)k ≤ β0 (kη(t0 )k, t − t0 ) + γ0 (sup kξ(t)k)

t≥t0
kη(t)k ≤ β0 (kη(t0 )k, t − t0 ) + γ0 (cε)

Proof–Second Part:
c1 kηk2 ≤ V1 (η) ≤ c2 kηk2
∂V1
f0 (η, 0) ≤ −c3 kηk2
∂η
∂V1

≤ c4 kηk
∂η
– p. 10/1
V (z) = bV1 (η) + ξT P ξ
" #T " #
kηk kηk
V̇ ≤ − Q
kξk kξk
" #
bc3 −(kkP Bk + bc4 L/2)
Q=
−(kkP Bk + bc4 L/2) 1 − 2kkP Bk
b=k
Q is positive definite for sufficiently small k
– p. 11/1
Lecture # 28
Stabilization
Backstepping
– p. 1/?
η̇ = f (η) + g(η)ξ
ξ̇ = u, η ∈ Rn , ξ, u ∈ R
Stabilize the origin using state feedback
View ξ as “virtual” control input to
η̇ = f (η) + g(η)ξ
Suppose there is ξ = φ(η) that stabilizes the origin of
η̇ = f (η) + g(η)φ(η)
∂V
[f (η) + g(η)φ(η)] ≤ −W (η), ∀η∈D
∂η
– p. 2/?
z = ξ − φ(η)
η̇ = [f (η) + g(η)φ(η)] + g(η)z

∂φ
ż = u − [f (η) + g(η)ξ]
∂η
∂φ
u= [f (η) + g(η)ξ] + v
∂η
η̇ = [f (η) + g(η)φ(η)] + g(η)z

ż = v
– p. 3/?
Vc (η, ξ) = V (η) + 12 z 2
∂V ∂V
V̇c = [f (η) + g(η)φ(η)] + g(η)z + zv
∂η ∂η
∂V
≤ −W (η) + g(η)z + zv
∂η
∂V
v=− g(η) − kz, k > 0
∂η
V̇c ≤ −W (η) − kz 2
– p. 4/?
Example
ẋ1 = x21 − x31 + x2 , ẋ2 = u
ẋ1 = x21 − x31 + x2
x2 = φ(x1 ) = −x21 − x1 ⇒ ẋ1 = −x1 − x31
V (x1 ) = 12 x21 ⇒ V̇ = −x21 − x41 , ∀ x1 ∈ R
z2 = x2 − φ(x1 ) = x2 + x1 + x21
ẋ1 = −x1 − x31 + z2

ż2 = u + (1 + 2x1 )(−x1 − x31 + z2 )
– p. 5/?
Vc (x) = 12 x21 + 21 z22
V̇c = x1 (−x1 − x31 + z2 )

+ z2 [u + (1 + 2x1 )(−x1 − x31 + z2 )]
V̇c = −x21 − x41

+ z2 [x1 + (1 + 2x1 )(−x1 − x31 + z2 ) + u]
u = −x1 − (1 + 2x1 )(−x1 − x31 + z2 ) − z2
V̇c = −x21 − x41 − z22
V̇c = −x21 − x41 − (x2 + x1 + x21 )2

– p. 6/?
Example
ẋ1 = x21 − x31 + x2 , ẋ2 = x3 , ẋ3 = u
ẋ1 = x21 − x31 + x2 , ẋ2 = x3

def
x3 = −x1 − (1 + 2x1 )(−x1 − x31 + z2 ) − z2 = φ(x1 , x2 )
V (x) = 12 x21 + 21 z22 , V̇ = −x21 − x41 − z22

z3 = x3 − φ(x1 , x2 )
ẋ1 = x21 − x31 + x2 , ẋ2 = φ(x1 , x2 ) + z3

∂φ 2 3 ∂φ
ż3 = u − (x1 − x1 + x2 ) − (φ + z3 )
∂x1 ∂x2
– p. 7/?
Vc = V + 12 z32
∂V ∂V
V̇c = (x21 − x31 + x2 ) + (z3 + φ)
∂x1 ∂x2
∂φ 2 ∂φ

3
+ z3 u − (x1 − x1 + x2 ) − (z3 + φ)
∂x1 ∂x2
V̇c = −x21 − x41 − (x2 + x1 + x21 )2

∂V ∂φ 2 ∂φ

3
+z3 − (x1 − x1 + x2 ) − (z3 + φ) + u
∂x2 ∂x1 ∂x2
∂V ∂φ ∂φ
u=− + (x21 − x31 + x2 ) + (z3 + φ) − z3
∂x2 ∂x1 ∂x2
– p. 8/?
η̇ = f (η) + g(η)ξ
ξ̇ = fa (η, ξ) + ga (η, ξ)u, ga (η, ξ) 6= 0
1
u= [v − fa (η, ξ)]
ga (η, ξ)
η̇ = f (η) + g(η)ξ
ξ̇ = v
– p. 9/?
Strict-Feedback Form
ẋ = f0 (x) + g0 (x)z1
ż1 = f1 (x, z1 ) + g1 (x, z1 )z2
ż2 = f2 (x, z1 , z2 ) + g2 (x, z1 , z2 )z3
..
.
żk−1 = fk−1 (x, z1 , . . . , zk−1 ) + gk−1 (x, z1 , . . . , zk−1 )zk
żk = fk (x, z1 , . . . , zk ) + gk (x, z1 , . . . , zk )u
gi (x, z1 , . . . , zi ) 6= 0 for 1 ≤ i ≤ k
– p. 10/?
Example
η̇ = −η + η 2 ξ, ξ̇ = u
η̇ = −η + η 2 ξ
ξ = 0 ⇒ η̇ = −η
V0 = 12 η 2 ⇒ V̇0 = −η 2 , ∀ η ∈ R
V = 12 (η 2 + ξ2 )
V̇ = η(−η + η 2 ξ) + ξu = −η 2 + ξ(η 3 + u)
u = −η 3 − kξ, k>0
V̇ = −η 2 − kξ2 Global stabilization
– p. 11/?
Lecture # 29
Stabilization
Passivity-Based Control
– p. 1/?
ẋ = f (x, u), y = h(x)
f (0, 0) = 0
T ∂V
u y ≥ V̇ = f (x, u)
∂x
Theorem 14.4: If the system is
(1) passive with a radially unbounded positive definite
storage function and
(2) zero-state observable,
then the origin can be globally stabilized by
u = −φ(y), φ(0) = 0, y T φ(y) > 0 ∀ y 6= 0
– p. 2/?
Proof:
∂V
V̇ = f (x, −φ(y)) ≤ −y T φ(y) ≤ 0
∂x
V̇ (x(t)) ≡ 0 ⇒ y(t) ≡ 0 ⇒ u(t) ≡ 0 ⇒ x(t) ≡ 0

A given system may be made passive by

(1) Choice of output,
(2) Feedback,
or both
– p. 3/?
Choice of Output
∂V
ẋ = f (x) + G(x)u, f (x) ≤ 0, ∀x
∂x
No output is defined. Choose the output as
T
∂V

def
y = h(x) = G(x)
∂x
∂V ∂V
V̇ = f (x) + G(x)u ≤ y T u
∂x ∂x
Check zero-state observability
– p. 4/?
Example
ẋ1 = x2 , ẋ2 = −x31 + u
V (x) = 41 x41 + 12 x22
With u = 0 V̇ = x31 x2 − x2 x31 = 0

∂V ∂V
Take y = G= = x2
∂x ∂x2
Is it zero-state observable?
with u = 0, y(t) ≡ 0 ⇒ x(t) ≡ 0
u = −kx2 or u = −(2k/π) tan−1 (x2 ) (k > 0)
– p. 5/?
Feedback Passivation
ẋ = f (x) + G(x)u, y = h(x)
is equivalent to a passive system if there is
u = α(x) + β(x)v
such that
ẋ = f (x) + G(x)α(x) + G(x)β(x)v, y = h(x)
is passive
– p. 6/?
Theorem [31]: The system
ẋ = f (x) + G(x)u, y = h(x)
is locally equivalent to a passive system (with a positive

definite storage function) if it has relative degree one at
x = 0 and the zero dynamics have a stable equilibrium
point at the origin with a positive definite Lyapunov function
Example: m-link Robot Manipulator
M (q)q̈ + C(q, q̇)q̇ + D q̇ + g(q) = u
M = M T > 0, (Ṁ − 2C)T = −(Ṁ − 2C), D = D T ≥ 0
– p. 7/?
Stabilize the system at q = qr
e = q − qr , ė = q̇
M (q)ë + C(q, q̇)ė + D ė + g(q) = u

(e = 0, ė = 0) is not an open-loop equilibrium point
u = g(q) − φp (e) + v, [φp (0) = 0, eT φp (e) > 0 ∀e 6= 0]
M (q)ë + C(q, q̇)ė + D ė + φp (e) = v
Z e
V = 21 ėT M (q)ė + φTp (σ) dσ
0
V̇ = 12 ėT (Ṁ −2C)ė−ėT D ė−ėT φp (e)+ėT v+φTp (e)ė ≤ ėT v

y = ė
– p. 8/?
Is it zero-state observable? Set v = 0
ė(t) ≡ 0 ⇒ ë(t) ≡ 0 ⇒ φp (e(t)) ≡ 0 ⇒ e(t) ≡ 0
v = −φd (ė), [φd (0) = 0, ėT φd (ė) > 0 ∀ė 6= 0]

u = g(q) − φp (e) − φd (ė)
Special case:
u = g(q) − Kp e − Kd ė, Kp = KpT > 0, Kd = KdT > 0
– p. 9/?
How does passivity-based control compare with feedback
linearization?
Example 13.20
ẋ1 = x2 , ẋ2 = −h(x1 ) + u
h(0) = 0, x1 h(x1 ) > 0, ∀ x1 6= 0

Feedback linearization:
u = h(x1 ) − (k1 x1 + k2 x2 )
" #
0 1
ẋ = x
−k1 −k2
– p. 10/?
Passivity-based control:
Z x1
V = h(z) dz + 12 x22
0
V̇ = x2 h(x1 ) − x2 h(x1 ) + x2 u = x2 u
Take y = x2
With u = 0, y(t) ≡ 0 ⇒ h(x1 (t)) ≡ 0 ⇒ x1 (t) ≡ 0
u = −σ(x2 ), [σ(0) = 0, yσ(y) > 0 ∀ y 6= 0]
ẋ1 = x2 , ẋ2 = −h(x1 ) − σ(x2 )
– p. 11/?
Linearization:
" #
0 1
, k = σ ′ (0)
−h′ (0) −k
s2 + ks + h′ (0) = 0
Sketch the root locus as k varies from zero to infinity
One of the p
two roots cannot be moved to the left of
Re[s] = − h′ (0)
– p. 12/?
Cascade Connection:
ż = fa (z) + F (z, y)y, ẋ = f (x) + G(x)u, y = h(x)
fa (0) = 0, f (0) = 0, h(0) = 0

∂V ∂V
f (x) + G(x)u ≤ y T u
∂x ∂x
∂W
fa (z) ≤ 0
∂z
U (z, x) = W (z) + V (x)
" T #
∂W ∂W

T T
U̇ ≤ F (z, y)y + y u = y u + F (z, y)
∂z ∂z
– p. 13/?
T
∂W

u=− F (z, y) + v ⇒ U̇ ≤ y T v
∂z
The system
ż = fa (z) + F (z, y)y

T
∂W

ẋ = f (x) − G(x) F (z, y) + G(x)v
∂z
y = h(x)
with input v and output y is passive with U as the storage

function
Read Examples 14.17 and 14.18
– p. 14/?
Lecture # 30
Stabilization
Control Lyapunov Functions
– p. 1/1
ẋ = f (x) + g(x)u, f (0) = 0, x ∈ Rn , u ∈ R
Suppose there is a continuous stabilizing state feedback

control u = ψ(x) such that the origin of
ẋ = f (x) + g(x)ψ(x)
is asymptotically stable
By the converse Lyapunov theorem, there is V (x) such that

∂V
[f (x) + g(x)ψ(x)] < 0, ∀ x ∈ D, x 6= 0
∂x
If u = ψ(x) is globally stabilizing, then D = Rn and V (x)
is radially unbounded
– p. 2/1
∂V
[f (x) + g(x)ψ(x)] < 0, ∀ x ∈ D, x 6= 0
∂x
∂V ∂V
g(x) = 0 for x ∈ D, x 6= 0 ⇒ f (x) < 0
∂x ∂x
Since ψ(x) is continuous and ψ(0) = 0, given any ε > 0,
∃ δ > 0 such that if x 6= 0 and kxk < δ , there is u with
kuk < ε such that
∂V
[f (x) + g(x)u] < 0
∂x
Small Control Property
– p. 3/1
Definition: A continuously differentiable positive definite
function V (x) is a Control Lyapunov Function (CLF) for the
system ẋ = f (x) + g(x)u if
∂V ∂V
g(x) = 0 for x ∈ D, x 6= 0 ⇒ f (x) < 0 (∗)
∂x ∂x
it satisfies the small control property
It is a Global Control Lyapunov Function if it is radially
unbounded and (∗) holds with D = Rn
The system ẋ = f (x) + g(x)u is stabilizable by a

continuous state feedback control only if it has a CLF
Is it sufficient?
– p. 4/1
Theorem: Let V (x) be a CLF for ẋ = f (x) + g(x)u, then
origin is stabilizable by u = ψ(x), where
 q
2 4
 ∂V
∂x
f + ( ∂V
∂x
f ) + ( ∂V
∂x
g ) ∂V
− , if g 6= 0


 ( ∂x )
∂V
g ∂x
ψ(x) =


if ∂V g = 0

 0,
∂x
The function ψ(x) is continuous for all x ∈ D0 (a

neighborhood of the origin) including x = 0. If f and g are
smooth, then ψ is smooth for x 6= 0. If V is a global CLF,
then the control u = ψ(x) is globally stabilizing
Sontag’s Formula
– p. 5/1
Proof: For properties of ψ , see Section 9.4 of [88]
∂V
[f (x) + g(x)ψ(x)]
∂x
∂V ∂V
If g(x) = 0, V̇ = f (x) < 0 for x 6= 0
∂x ∂x
∂V
If g(x) 6= 0
∂x
" r #
2 4
∂V ∂V ∂V ∂V
V̇ = ∂x f − ∂x f + ∂x f + ∂x g
r 2 4
∂V ∂V
= − ∂x f + ∂x g < 0 for x 6= 0
– p. 6/1
How can we find a CLF?
If we know of any stabilizing control with a corresponding

Lyapunov function V , then V is a CLF
Feedback Linearization
ẋ = f (x) + G(x)u, z = T (x), ż = (A − BK)z
P (A − BK) + (A − BK)T P = −Q, Q = QT > 0

V = z T P z = T T (x)P T (x) is a CLF
Backstepping
– p. 7/1
Example:
ẋ = ax − bx3 + u, a, b > 0
u = −ax + bx3 − kx (k > 0)
ẋ = −kx
V (x) = 12 x2 is a CLF
∂V ∂V
g = x, f = x(ax − bx3 )
∂x ∂x
– p. 8/1
r 2 4
∂V ∂V ∂V
∂x f + ∂x f + ∂x g
−
∂V
∂x g
p
x(ax − bx3 ) + x2 (ax − bx3 )2 + x4
= −
x
q
= −ax + bx3 − x (a − bx2 )2 + 1
q
ψ(x) = −ax + bx3 − x (a − bx2 )2 + 1
Compare with
−ax + bx3 − kx
– p. 9/1
Method Expression
FL-u −ax + bx3 − kx
FL-CLS ẋ = −kx
p
CLF-u 3
−ax + bx − x (a − bx2 )2 + 1
p
CLF-CLS −x (a − bx2 )2 + 1
Method Small |x| Large |x|

FL-u (−a + k)x bx3
FL-CLS ẋ = −kx ẋ = −kx
√
CLF-u −(a + a2 + 1)x −ax
√
CLF-CLS − a2 + 1x −bx3
– p. 10/1
Lemma: Let V (x) be a CLF for ẋ = f (x) + g(x)u and
suppose ∂V ∂x (0) = 0. Then, Sontag’s formula has a gain
margin [ 12 , ∞); that is, u = kψ(x) is stabilizing for all k ≥ 1
2
Proof: Let
 s 
2 4
∂V ∂V ∂V

1
q(x) = 2
− f+ f + g 
∂x ∂x ∂x
∂V
q(0) = 0, g 6= 0 ⇒ q > 0
∂x
∂V ∂V
g=0 ⇒ q=− f > 0 for x 6= 0
∂x ∂x
q(x) is positive definite
– p. 11/1
u = kψ(x) ⇒ ẋ = f (x) + g(x)kψ(x)
∂V ∂V
V̇ = f+ gkψ
∂x ∂x
∂V ∂V
g = 0 ⇒ V̇ = f < 0 for x 6= 0
∂x ∂x
∂V ∂V ∂V
g 6= 0, V̇ = −q + q + f+ gkψ
∂x ∂x ∂x
∂V ∂V
q+ f+
gkψ
∂x ∂x 
s
2 4
1  ∂V ∂V ∂V

= − k−2 f+ f + g 
∂x ∂x ∂x
– p. 12/1
Lecture # 31
Stabilization
Output Feedback
– p. 1/1
In general, output feedback stabilization requires the use of
observers. In this lecture we deal with three simple cases
where an observer is not needed
Minimum Phase Relative Degree One Systems
Passive systems
System with Passive maps from the input to the

derivative of the output
– p. 2/1
Minimum Phase Relative Degree One Systems
ẋ = f (x) + g(x)u, y = h(x)
f (0) = 0, h(0) = 0, Lg h(x) 6= 0, ∀x∈D

Normal Form:
" # " #
η φ
φ(0) = 0, Lg φ(x) = 0, =
y h(x)
η̇ = f0 (η, y), ẏ = γ(x)[u − α(x)], γ(x) 6= 0
– p. 3/1
Assumptions:
The origin of η̇ = f0 (η, 0) is exponentially stable
c1 kηk2 ≤ V1 (η) ≤ c2 kηk2
∂V1
f0 (η, 0) ≤ −c3 kηk2
∂η

∂V1
≤ c4 kηk
∂η
kf0 (η, y) − f0 (η, 0)k ≤ L1 |y|

|α(x)γ(x)| ≤ L2 kηk + L3 |y|
γ(x) ≥ γ0 > 0
– p. 4/1
High-Gain Feedback:
u = −ky, k>0
η̇ = f0 (η, y), ẏ = γ(x)[−ky − α(x)]

V (η, y) = V1 (η) + 12 y 2
∂V1
V̇ = f0 (η, y) − kγ(x)y 2 − α(x)γ(x)y
∂η
∂V1 ∂V1
V̇ = f0 (η, 0) + [f0 (η, y) − f0 (η, 0)]
∂η ∂η
− kγ(x)y 2 − α(x)γ(x)y
V̇ ≤ −c3 kηk2 + c4 L1 kηk |y| − kγ0 y 2 + L2 kηk |y| + L3 y 2
– p. 5/1
V̇ ≤ −c3 kηk2 + c4 L1 kηk |y| − kγ0 y 2 + L2 kηk |y| + L3 y 2
" #T " #" #
kηk c3 − 21 (L2 + c4 L1 ) kηk
V̇ ≤ −
|y| − 21 (L2 + c4 L1 ) (kγ0 − L3 ) |y|
| {z }
Q
det(Q) = c3 (kγ0 − L3 ) − 41 (L2 + c4 L1 )2

1 1 2

det(Q) > 0 for k > c3 L3 + 4 (L2 + c4 L1 )
c3 γ 0
The origin of the closed-loop system is exponentially stable
If the assumptions hold globally, it is globally exp. stable
– p. 6/1
Passive Systems
Suppose the system
ẋ = f (x, u), y = h(x, u)
is passive (with a positive definite storage function) and

zero-state observable
Then, it can be stabilized by
u = −ψ(y), ψ(0) = 0, y T ψ(y) > 0, ∀ y 6= 0
V̇ ≤ uT y = −y T ψ(y) ≤ 0
V̇ = 0 ⇒ y = 0 ⇒ u = 0
y(t) ≡ 0 ⇒ x(t) ≡ 0
– p. 7/1
Systems with Passive Maps from u to ẏ
ẋ = f (x, u), y = h(x)
f (0, 0) = 0, h(0) = 0, h is cont. diff.

Suppose the system
∂h def
ẋ = f (x, u), ẏ = f (x, u) = h̃(x, u)
∂x
is passive (with a positive definite storage function) and
zero-state observable
V̇ ≤ uT ẏ
With u = 0, ẏ(t) ≡ 0 ⇒ x(t) ≡ 0
– p. 8/1
+ h u- y
- Plant -
−6
z
ψ(·) bs
s+a

+ h u- y ẏ
- Plant - s -
−6
z
ψ(·) b
s+a

– p. 9/1
For 1 ≤ i ≤ m
bi s
zi is the output of driven by yi
s + ai
ui = −ψi (zi )
ai , bi > 0, ψi (0) = 0, zi ψi (zi ) > 0 ∀ zi 6= 0
żi = −ai zi + bi ẏi
Use
m
X Z zi
1
Vc (x, z) = V (x) + bi ψi (σ) dσ
i=1 0
to prove asymptotic stability of the origin of the closed-loop

system
– p. 10/1
Example: Stabilize the pendulum
mℓθ̈ + mg sin θ = u
at θ = δ1 using feedback from θ
x1 = θ − δ1 , x2 = θ̇
1
ẋ1 = x2 , ẋ2 = mℓ [−mg sin θ + u]
u = mg sin θ − kp x1 + v, kp > 0
kp 1
ẋ1 = x2 , ẋ2 = − mℓ x1 + mℓ v
y = x1 , ẏ = x2
– p. 11/1
V = 12 kp x21 + 12 mℓx22
h i
kp 1
V̇ = kp x1 x2 + mℓx2 − mℓ x1 + mℓ
v
V̇ = x2 v = ẏv
With v = 0, x2 (t) ≡ 0 ⇒ x1 (t) ≡ 0
y z
- bs -
s+a
ξ̇ = −aξ + y, z = b(−aξ + y)
v = −kd z, kd > 0
u = mg sin θ − kp (θ − δ1 ) − kd z
– p. 12/1
Lecture # 32
Robust Stabilization
Sliding Mode Control
– p. 1/1
Example
ẋ1 = x2 ẋ2 = h(x) + g(x)u, g(x) ≥ g0 > 0
Sliding Manifold (Surface):
s = a1 x1 + x2 = 0
s(t) ≡ 0 ⇒ ẋ1 = −a1 x1

a1 > 0 ⇒ lim x1 (t) = 0
t→∞
How can we bring the trajectory to the manifold s = 0?
How can we maintain it there?
– p. 2/1
ṡ = a1 ẋ1 + ẋ2 = a1 x2 + h(x) + g(x)u
Suppose

a1 x2 + h(x)
≤ ̺(x)
g(x)
V = 12 s2
V̇ = sṡ = s[a1 x2 +h(x)]+g(x)su ≤ g(x)|s|̺(x)+g(x)su

β(x) ≥ ̺(x) + β0 , β0 > 0
s > 0, u = −β(x)
V̇ ≤ g(x)|s|̺(x) − g(x)β(x)|s|
V̇ ≤ g(x)|s|̺(x) − g(x)(̺(x) + β0 )|s| = −g(x)β0 |s|
– p. 3/1
s < 0, u = β(x)
V̇ ≤ g(x)|s|̺(x) + g(x)su = g(x)|s|̺(x) − g(x)β(x)|s|
V̇ ≤ g(x)|s|̺(x) − g(x)(̺(x) + β0 )|s| = −g(x)β0 |s|

(
1, s>0
sgn(s) =
−1, s<0
u = −β(x) sgn(s)
V̇ ≤ −g(x)β0 |s| ≤ −g0 β0 |s|

√
V̇ ≤ −g0 β0 2V
– p. 4/1
√
V̇ ≤ −g0 β0 2V
dV √
√ ≤ −g0 β0 2 dt
V

√ V (s(t)) √
2 V ≤ −g0 β0 2 t
V (s(0))
p p 1
V (s(t)) ≤ V (s(0)) − g0 β0 √ t
2
|s(t)| ≤ |s(0)| − g0 β0 t
s(t) reaches zero in finite time
Once on the surface s = 0, the trajectory cannot leave it
– p. 5/1
s=0
What is the region of validity?
– p. 6/1
ẋ1 = x2 ẋ2 = h(x) − g(x)β(x)sgn(s)
ẋ1 = −a1 x1 + s ṡ = a1 x2 + h(x) − g(x)β(x)sgn(s)

sṡ ≤ −g0 β0 |s|, if β(x) ≥ ̺(x) + β0
V1 = 12 x21
V̇1 = x1 ẋ1 = −a1 x21 + x1 s ≤ −a1 x21 + |x1 |c ≤ 0

c
∀ |s| ≤ c and |x1 | ≥
a1

c
Ω= |x1 | ≤ , |s| ≤ c
a1

a1 x2 + h(x)
Ω is positively invariant if ≤ ̺(x) over Ω
g(x)
– p. 7/1

c
Ω= |x1 | ≤ , |s| ≤ c
a1
x2 6
s=0 H
HHH HH
HH HHc
HH HH
HH c/a1 -
HH
HH HH x1
HH HH
HH HH
HH HH
H H

a1 x2 + h(x)
≤ k1 < k, ∀ x ∈ Ω
g(x)
u = −k sgn(s)
– p. 8/1
Chattering
B Sliding manifold
B
B
B
B
B
H
HH B
s < 0 BH Y s>0
B
B
B
H
HH B a
B HH
B YH
H
How can we reduce or eliminate chattering?
– p. 9/1
Reduce the amplitude of the signum function
ṡ = a1 x2 + h(x) + g(x)u
[a1 x2 + ĥ(x)]
u=− +v
ĝ(x)
ṡ = δ(x) + g(x)v

g(x) g(x)
δ(x) = a1 1 − x2 + h(x) − ĥ(x)
ĝ(x) ĝ(x)

δ(x)
≤ ̺(x), β(x) ≥ ̺(x) + β0
g(x)
v = −β(x) sgn(s)
– p. 10/1
Replace the signum function by a high-slope saturation
function
s
u = −β(x) sat
ε
(
y, if |y| ≤ 1
sat(y) =
sgn(y), if |y| > 1
y
sgn(y) 6 sat ε 6
1 1

- -
y ε y

−1 −1
– p. 11/1
How can we analyze the system?
For |s| ≥ ε, u = −β(x) sgn(s)
With c ≥ ε
n o
c
Ω = |x1 | ≤ a1
, |s| ≤ c is positively invariant
The trajectory reaches the boundary layer {|s| ≤ ε}in

finite time
The boundary layer is positively invariant
– p. 12/1
Inside the boundary layer:
s
ẋ1 = −a1 x1 + s ṡ = a1 x2 + h(x) − g(x)β(x)
ε
x1 ẋ1 ≤ −a1 x21 + |x1 |ε

0<θ<1
ε
x1 ẋ1 ≤ −(1 − θ)a1 x21 , ∀ |x1 | ≥
θa1
The trajectories reach the positively invariant set
ε
Ωε = {|x1 | ≤ , |s| ≤ ε}
θa1
in finite time
– p. 13/1
What happens inside Ωε ?
Find the equilibrium points

s
0 = −a1 x1 + s = x2 , 0 = a1 x2 + h(x) − g(x)β(x)
ε

h(x)
φ(x1 ) =
a1 g(x)β(x) x2 =0
x1 = εφ(x1 )
Suppose x1 = εφ(x1 ) has an isolated root x̄1 = εk1
h(0) = 0 ⇒ x̄1 = 0
– p. 14/1
z1 = x1 − x̄1 , z2 = s − a1 x̄1
x2 = −a1 x1 + s = −a1 (x1 − x̄1 ) + s − a1 x̄1 = −a1 z1 + z2

ż1 = −a1 x1 + s = −a1 z1 + z2
s
ż2 = a1 x2 + h(x) − g(x)β(x)
ε
z2 + a1 x̄1
= a1 (z2 − a1 z1 ) + h(x) − g(x)β(x)
ε
z2
ż2 = ℓ(z) − g(x)β(x)
ε

h(x) x̄1
ℓ(z) = a1 (z2 − a1 z1 ) + a1 g(x)β(x) −
a1 g(x)β(x) ε
– p. 15/1
z2
ż1 = −a1 z1 + z2 , ż2 = ℓ(z) − g(x)β(x)
ε
ℓ(0) = 0, |ℓ(z)| ≤ ℓ1 |z1 | + ℓ2 |z2 |
g(x)β(x) ≥ g0 β0
V = 21 z12 + 12 z22

z2
V̇ = z1 (−a1 z1 + z2 ) + z2 ℓ(z) − g(x)β(x)
ε
g0 β0
V̇ ≤ −a1 z12 + (1 + ℓ1 )|z1 | |z2 | + ℓ2 z22 − z22
ε
– p. 16/1
g0 β0
V̇ ≤ −a1 z12 + (1 + ℓ1 )|z1 | |z2 | + ℓ2 z22
− z22
" #T " # ε
" #
1
|z1 | a1 − 2 (1 + ℓ1
) |z1 |
V̇ ≤ −
g0 β0
|z2 | − 12 (1 + ℓ1 ) ε − ℓ2 |z2 |
| {z }
Q

g0 β0
det(Q) = a1 − ℓ2 − 41 (1 + ℓ1 )2
ε
h(0) = 0 ⇒ lim x(t) = 0
"t→∞ #
x̄1
h(0) =
6 0 ⇒ lim x(t) =
t→∞ 0
Read Section 14.1.1

– p. 17/1
Lecture # 33
– p. 1/1
Regular Form:
η̇ = fa (η, ξ)
ξ̇ = fb (η, ξ) + g(η, ξ)u + δ(t, η, ξ, u)
η ∈ Rn−1 , ξ ∈ R, u ∈ R
fa (0, 0) = 0, fb (0, 0) = 0, g(η, ξ) ≥ g0 > 0
Sliding Manifold:
s = ξ − φ(η) = 0, φ(0) = 0
s(t) ≡ 0 ⇒ η̇ = fa (η, φ(η))

Design φ s.t. the origin of η̇ = fa (η, φ(η)) is asymp. stable
– p. 2/1
∂φ
ṡ = fb (η, ξ) − fa (η, ξ) + g(η, ξ)u + δ(t, η, ξ, u)
∂η
1 ˆ ∂φ ˆ

u=− fb − fa + v or u = v
ĝ ∂η
∂φ ˆ 1

ˆ
u = −L fb − fa + v, L = or L = 0
∂η ĝ
ṡ = g(η, ξ)v + ∆(t, η, ξ, v)
∂φ ∂φ ˆ

∆ = fb − ˆ
fa + δ − gL fb − fa
∂η ∂η
∆(t, η, ξ, v)

g(η, ξ) ≤ ̺(η, ξ) + κ0 |v|

– p. 3/1
∆(t, η, ξ, v)

g(η, ξ) ≤ ̺(η, ξ) + κ0 |v|

̺(η, ξ) ≥ 0, 0 ≤ κ0 < 1 (Known)

sṡ = sgv + s∆ ≤ sgv + |s| |∆|
sṡ ≤ g[sv + |s|(̺ + κ0 |v|)]
v = −β(η, ξ) sgn(s)
̺(η, ξ)
β(η, ξ) ≥ + β0 , β0 > 0
1 − κ0
sṡ ≤ g[−β|s| + ̺|s| + κ0 β|s|] = g[−β(1 − κ0 )|s| + ̺|s| ]
sṡ ≤ g[−̺|s| − (1 − κ0 )β0 |s| + ̺|s| ]
– p. 4/1
sṡ ≤ −g(η, ξ)(1 − κ0 )β0 |s| ≤ −g0 β0 (1 − κ0 )|s|
s

v = −β(x) sat , ε>0
ε
sṡ ≤ −g0 β0 (1 − κ0 )|s|, for |s| ≥ ε
The trajectory reaches the boundary layer {|s| ≤ ε} in finite
time and remains inside thereafter
Study the behavior of η
η̇ = fa (η, φ(η) + s)
What do we know about this system and what do we need?
– p. 5/1
α1 (kηk) ≤ V (η) ≤ α2 (kηk)
∂V
fa (η, φ(η) + s) ≤ −α3 (kηk), ∀ kηk ≥ γ(|s|)
∂η
|s| ≤ c ⇒ V̇ ≤ −α3 (kηk), for kηk ≥ γ(c)
α(r) = α2 (γ(r))
V (η) ≥ α(c) ⇔ V (η) ≥ α2 (γ(c)) ⇒ α2 (kηk) ≥ α2 (γ(c))

⇒ kηk ≥ γ(c) ⇒ V̇ ≤ −α3 (kηk) ≤ −α3 (γ(c))
The set {V (η) ≤ c0 } with c0 ≥ α(c) is positively invariant
Ω = {V (η) ≤ c0 } × {|s| ≤ c}, with c0 ≥ α(c)
– p. 6/1
V
c α(.)
0
α(c)
α(ε)
ε c |s|
Ω = {V (η) ≤ c0 } × {|s| ≤ c}, with c0 ≥ α(c)
is positively invariant and all trajectories starting in Ω reach

Ωε = {V (η) ≤ α(ε)} × {|s| ≤ ε} in finite time
– p. 7/1
Theorem 14.1: Suppose all the assumptions hold over Ω.
Then, for all (η(0), ξ(0)) ∈ Ω, the trajectory (η(t), ξ(t)) is
bounded for all t ≥ 0 and reaches the positively invariant
set Ωε in finite time. If the assumptions hold globally and
V (η) is radially unbounded, the foregoing conclusion holds
for any initial state
Theorem 14.2: Suppose all the assumptions hold over Ω
̺(0) = 0, κ0 = 0
The origin of η̇ = fa (η, φ(η)) is exponentially stale
Then there exits ε∗ > 0 such that for all 0 < ε < ε∗ , the
origin of the closed-loop system is exponentially stable and
Ω is a subset of its region of attraction. If the assumptions
hold globally, the origin will be globally uniformly
– p. 8/1
Example
ẋ1 = x2 + θ1 x1 sin x2 , ẋ2 = θ2 x22 + x1 + u
|θ1 | ≤ a, |θ2 | ≤ b
x2 = −kx1 ⇒ ẋ1 = −kx1 + θ1 x1 sin x2
V1 = 12 x21 ⇒ x1 ẋ1 ≤ −kx21 + ax21
s = x2 + kx1 , k>a
ṡ = θ2 x22 + x1 + u + k(x2 + θ1 x1 sin x2 )
u = −x1 − kx2 + v ⇒ ṡ = v + ∆(x)
∆(x) = θ2 x22 + kθ1 x1 sin x2
– p. 9/1
∆(x) = θ2 x22 + kθ1 x1 sin x2
|∆(x)| ≤ ak|x1 | + bx22
β(x) = ak|x1 | + bx22 + β0 , β0 > 0

u = −x1 − kx2 − β(x) sgn(s)
Will
s

u = −x1 − kx2 − β(x) sat
ε
stabilize the origin?
– p. 10/1
Example: Normal Form
η̇ = f0 (η, ξ)
ξ̇i = ξi+1 , 1≤i≤ρ−1
ρ ρ−1
y = ξ1
View ξρ as input to the system
η̇ = f0 (η, ξ1 , · · · , ξρ−1 , ξρ )
ξ̇i = ξi+1 , 1≤i≤ρ−2
ξ̇ρ−1 = ξρ
Design ξρ = φ(η, ξ1 , · · · , ξρ−1 ) to stabilize the origin

– p. 11/1
s = ξρ − φ(η, ξ1 , · · · , ξρ−1 )
Minimum Phase Systems: The origin of η̇ = f0 (η, 0) is

s = ξρ + k1 ξ1 + · · · + kρ−1 ξρ−1
η̇ = f0 (η, ξ1 , · · · , ξρ−1 , −k1 ξ1 − · · · − kρ−1 ξρ−1 )

    
ξ̇1 1 ξ1
 .  .  . 
 ..  ..   .. 


 =  
1
    
    
ξ̇ρ−1 −k1 −kρ−1 ξρ−1
– p. 12/1
Multi-Input Systems
η̇ = fa (η, ξ)
ξ̇ = fb (η, ξ) + G(η, ξ)E(η, ξ)u + δ(t, η, ξ, u)
η ∈ Rn−p , ξ ∈ Rp , u ∈ Rp
fa (0, 0) = 0, fb (0, 0) = 0, det(G) 6= 0, det(E) 6= 0
G = diag[g1 , g2 , · · · , gm ], gi (η, ξ) ≥ g0 > 0
Design φ s.t. the origin of η̇ = fa (η, φ(η)) is asymp. stable
s = ξ − φ(η)
∂φ
ṡ = fb (η, ξ) − fa (η, ξ) + G(η, ξ)E(η, ξ)u + δ(t, η, ξ, u)
∂η
– p. 13/1
∂φ
ṡ = fb (η, ξ) − fa (η, ξ) + G(η, ξ)E(η, ξ)u + δ(t, η, ξ, u)
∂η
∂φ ˆ

u=E −1 ˆ
−L fb − fa + v , L = Ĝ−1 or L = 0
∂η
ṡi = gi (η, ξ)vi + ∆i (t, η, ξ, v), 1 ≤ i ≤ p
∆i (t, η, ξ, v)

≤ ̺(η, ξ) + κ0 max |vi |, ∀ 1 ≤ i ≤ p
g (η, ξ) 1≤i≤p
i
̺(η, ξ) ≥ 0, 0 ≤ κ0 < 1 (Known)

̺(x)
β(x) ≥ + β0 , β0 > 0
1 − κ0
– p. 14/1
si ṡi = si gi vi + si ∆i ≤ gi {si vi + |si |[̺ + κ0 max |vi |]}
1≤i≤p
vi = −β sgn(si ), 1≤i≤p
si ṡi ≤ gi [−β + ̺ + κ0 β]|si |

= gi [−(1 − κ0 )β + ̺]|si |
≤ gi [−̺ − (1 − κ0 )β0 + ̺]|si |
≤ −g0 β0 (1 − κ0 )|si |
Now use
si

vi = −β sat , 1≤i≤p
ε
Read Theorem 14.1 and 14.2 in the textbook
– p. 15/1
Lecture # 34
Lyapunov Redesign &

Backstepping
– p. 1/?
Lyapunov Redesign (Min-max control)
ẋ = f (x) + G(x)[u + δ(t, x, u)], x ∈ Rn , u ∈ Rp
Nominal Model: ẋ = f (x) + G(x)u

Stabilizing Control: u = ψ(x)
∂V
[f (x) + G(x)ψ(x)] ≤ −W (x), ∀ x ∈ D, W is p.d.
∂x
u = ψ(x) + v
kδ(t, x, ψ(x) + v)k ≤ ρ(x) + κ0 kvk, 0 ≤ κ0 < 1
ẋ = f (x) + G(x)ψ(x) + G(x)[v + δ(t, x, ψ(x) + v)]
∂V ∂V
V̇ = (f + Gψ) + G(v + δ)
∂x ∂x
– p. 2/?
T ∂V
w = G
∂x
V̇ ≤ −W (x) + w T v + w T δ
w T v +w T δ ≤ w T v +kwk kδk ≤ w T v +kwk[ρ(x)+κ0 kvk]

w w
v = −η(x) = sgn(w) for p = 1
kwk kwk
w T v + w T δ ≤ −ηkwk + ρkwk + κ0 ηkwk

= −η(1 − κ0 )kwk + ρkwk
ρ(x)
η(x) ≥ ⇒ w T v + w T δ ≤ 0 ⇒ V̇ ≤ −W (x)
(1 − κ0 )
– p. 3/?

 −η(x) w , if η(x)kwk ≥ ε
kwk
v=
 −η 2 (x) w , if η(x)kwk < ε
ε
η(x)kwk ≥ ε ⇒ V̇ ≤ −W (x)
For η(x)kwk < ε

T 2 w
V̇ ≤ −W (x) + w −η · +δ
ε
η2
≤ −W (x) − kwk2 + ρkwk + κ0 kwkkvk
ε
η2 κ0 η 2
= −W (x) − kwk2 + ρkwk + kwk2
ε ε
– p. 4/?
!
η2
V̇ ≤ −W (x) + (1 − κ0 ) − kwk2 + ηkwk
ε
y2 ε
− +y ≤ , for y ≥ 0
ε 4
(1 − κ0 )
V̇ ≤ −W (x) + ε , ∀x∈D
4
Theorem 14.3: x(t) is uniformly ultimately bounded by a
class K function of ε. If the assumptions hold globally and
V is radially unbounded, then x(t) globally uniformly
ultimately bounded
Corollary 14.1: If ρ(0) = 0 and η(x) ≥ η0 > 0 we can
recover uniform asymptotic stability
– p. 5/?
Example: Pendulum with horizontal acceleration of
suspension point
h i
m ℓθ̈ + A(t) cos θ = T /ℓ − mg sin θ
Stabilize the pendulum at θ = π
g 1 A(t)
x1 = θ − π, x2 = θ̇, a = , c= , h(t) =
ℓ mℓ2 ℓ
ẋ1 = x2 , ẋ2 = a sin x1 + cu + h(t) cos x1
Nominal Model: ẋ1 = x2 , ẋ2 = â sin x1 + ĉu

â 1
ψ(x) = − sin x1 − (k1 x1 + k2 x2 )
ĉ ĉ
– p. 6/?
" #
0 1
, V (x) = xT P x
−k1 −k2
| {z }
Hurwitz

1 aĉ − âc
δ = sin x1 + h(t) cos x1
ĉ ĉ

c − ĉ c − ĉ
− (k1 x1 + k2 x2 ) + v
ĉ ĉ
q
c − ĉ âc − aĉ c − ĉ
≤ κ0 , + k2 + k2 ≤ k, |h(t)| ≤ H
ĉ 1 2
ĉ ĉ
(kkxk + H) def
|δ| ≤ + κ0 |v| = ρ(x) + κ0 |v|, (κ0 < 1)
ĉ
– p. 7/?
ρ(x) H
η(x) = , η(x) ≥
(1 − κ0 ) ĉ(1 − κ0 )
" #
∂V T 0
w= G = 2x P = 2(p12 x1 + p22 x2 )
∂x 1

 −η(x)sgn(w), if η(x)|w| ≥ ε
v=
 −η 2 (x) w , if η(x)|w| < ε
ε

â 1
u=− sin x1 − (k1 x1 + k2 x2 ) + v
ĉ ĉ
Will this control stabilize the origin x = 0?
– p. 8/?
Backstepping
ż1 = f1 (z1 ) + g1 (z1 )z2

ż2 = f2 (z1 , z2 ) + g2 (z1 , z2 )z3
..
.
żk−1 = fk−1 (x, z1 , . . . , zk−1 ) + gk−1 (z1 , . . . , zk−1 )zk
żk = fk (z1 , . . . , zk ) + gk (z1 , . . . , zk )u
gi 6= 0, 1≤i≤k
– p. 9/?
ż1 = f1 + g1 z2 + δ1 (z)
ż2 = f2 + g2 z3 + δ2 (z)
..
.
żk−1 = fk−1 + gk−1 zk + δk−1 (z)
żk = fk + gk u + δk (z)
|δ1 (z)| ≤ ρ1 (z1 )

|δ2 (z)| ≤ ρ2 (z1 , z2 )
..
.
|δk−1 | ≤ ρk−1 (z1 , . . . , zk−1 )
|δk | ≤ ρk (z1 , . . . , zk )
The virtual control zi = φi (z1 , . . . , zi−1 ) should be smooth

– p. 10/?
Example:
ẋ1 = x2 + θ1 x1 sin x2 , ẋ2 = θ2 x22 + x1 + u
|θ1 | ≤ a, |θ2 | ≤ b
δ1 = θ1 x1 sin x2 , δ2 = θ2 x22
ẋ1 = x2 + θ1 x1 sin x2 , |θ1 x1 sin x2 | ≤ a|x1 |
x2 = −k1 x1
V1 = 12 x21 , V̇1 ≤ −(k1 − a)x21 ; Take k1 = 1 + a
z2 = x2 + (1 + a)x1
– p. 11/?
ẋ1 = −(1 + a)x1 + θ1 x1 sin x2 + z2
ż2 = ψ1 (x) + ψ2 (x, θ) + u
ψ1 = x1 + (1 + a)x2 , ψ2 = (1 + a)θ1 x1 sin x2 + θ2 x22
Vc = 21 x21 + 12 z22
V̇c ≤ −x21 + z2 [x1 + ψ1 (x) + ψ2 (x, θ) + u]

First Approach (Example 14.13):
u = −x1 − ψ1 (x) − kz2 , k>0
V̇c ≤ −x21 − kz22 + z2 ψ2 (x, θ)

Restrict analysis to the compact set Ωc = {Vc (x) ≤ c}
– p. 12/?
ψ2 = (1 + a)θ1 x1 sin x2 + θ2 x22
|ψ2 | ≤ a(1 + a)|x1 | + bρ|x2 |, ρ = max |x2 |

x∈Ωc
x2 = z2 − (1 + a)x1
|ψ2 | ≤ (1 + a)(a + bρ)|x1 | + bρ|z2 |
V̇c ≤ −x21 − kz22 + (1 + a)(a + bρ)|x1 | |z2 | + bρz22

We can make V̇ neg. def. by choosing k large enough
Can this control achieve global stabilization?
Can it achieve semiglobal stabilization?
– p. 13/?
Second Approach (Example 14.14):
u = −x1 − ψ1 (x) − kz2 + v
V̇c ≤ −x21 − kz22 + z2 [ψ2 + v]
|ψ2 | ≤ a(1 + a)|x1 | + bx22

(
−η(x) sgn(z2 ), if η(x)|z2 | ≥ ε
v=
−η 2 (x)z2 /ε, if η(x)|z2 | < ε
η(x) = η0 + a(1 + a)|x1 | + bx22 , η0 > 0, ε>0

ε
V̇c ≤ −x21 − kz22 +
4
Show that this control is globally stabilizing
– p. 14/?
Lecture # 35
Tracking
Feedback Linearization &

– p. 1/1
SISO relative-degree ρ system:
ẋ = f (x) + g(x)u, y = h(x)
f (0) = 0, h(0) = 0
ρ−1
Lg Li−1
f h(x) = 0, for 1 ≤ i ≤ ρ − 1, Lg Lf h(x) 6= 0
Normal form:
η̇ = f0 (η, ξ)
ξ̇i = ξi+1 , 1≤i≤ρ−1
ξ̇ρ = Lρf h(x) + Lg Lρ−1
f h(x)u
y = ξ1
f0 (0, 0) = 0
– p. 2/1
Reference signal r(t)
r(t) and its derivatives up to r (ρ) (t) are bounded for all
t ≥ 0 and the ρth derivative r (ρ) (t) is a piecewise
continuous function of t;
the signals r ,. . . ,r (ρ) are available on-line.
Goal: lim [y(t) − r(t)] = 0

t→∞
   
r ξ1 − r

R= ..  
e= .. 
. , . =ξ−R
r (ρ−1) ξρ − r (ρ−1)
– p. 3/1
η̇ = f0 (η, e + R)
h i
ρ ρ−1
ė = Ac e + Bc Lf h(x) + Lg Lf h(x)u − r (ρ)
   
0 1 0 ... 0 0
 
 0 0 1 ... 0  
 0


 
 .. ... ..  .. 
Ac =  . . , Bc = 
 . 

 ..   

 . 0 
1   0 
0 ... ... 0 0 1
1 h
ρ
i
u= ρ−1 −Lf h(x) + r (ρ) + v
Lg Lf h(x)
ė = Ac e + Bc v
– p. 4/1
v = −Ke ⇒ ė = (Ac − Bc K) e
| {z }
H urwitz
lim e(t) = 0 ⇒ lim [y(t) − r(t)] = 0

t→∞ t→∞
e(t) is bounded ⇒ ξ(t) = e(t) + R(t) is bounded

What about η(t)?
η̇ = f0 (η, ξ)
Local Tracking (small kη(0)k, ke(0)k, kR(t)k):
Minimum Phase ⇒ The origin of η̇ = f0 (η, 0) is

⇒ η is bounded for sufficiently small
kη(0)k, ke(0)k, and kR(t)k
– p. 5/1
Global Tracking (large kη(0)k, ke(0)k, kR(t)k):
What condition on η̇ = f0 (η, ξ) is needed?
Example 13.21
ẋ1 = x2 , ẋ2 = −a sin x1 − bx2 + cu, y = x1
e1 = x1 − r, e2 = x2 − ṙ
ė1 = e2 , ė2 = −a sin x1 − bx2 + cu − r̈
1
u= [a sin x1 + bx2 + r̈ − k1 e1 − k2 e2 ]
c
ė1 = e2 , ė2 = −k1 e1 − k2 e2
See simulation in the textbook
– p. 6/1
ẋ = f (x) + g(x)[u + δ(t, x, u)], y = h(x)
Lg h(x) = · · · = Lg Lfρ−2 h(x) = 0, Lg Lρ−1

f h(x) ≥ a > 0
η̇ = f0 (η, ξ)
ξ̇1 = ξ2
.. ..
. .
ξ̇ρ−1 = ξρ
ρ ρ−1
ξ̇ρ = Lf h(x) + Lg Lf h(x)[u + δ(t, x, u)]
y = ξ1
e=ξ−R
– p. 7/1
η̇ = f0 (η, ξ)
ė1 = e2
.. ..
. .
ėρ−1 = eρ
ėρ = Lρf h(x) + Lg Lρ−1
f h(x)[u + δ(t, x, u)] − r (ρ)
(t)
Sliding surface:
s = (k1 e1 + · · · + kρ−1 eρ−1 ) + eρ
s(t) ≡ 0 ⇒ eρ = −(k1 e1 + · · · + kρ−1 eρ−1 )
– p. 8/1
η̇ = f0 (η, ξ)
ė1 = e2
.. ..
. .
ėρ−1 = −(k1 e1 + · · · + kρ−1 eρ−1 )
Design k1 to kρ−1 such that the matrix

 
1

 ... 

  is Hurwitz
 
 1 
−k1 −kρ−1
Assumption: The system f0 (η, ξ) is BIBS stable
– p. 9/1
ρ−1
X
s = (k1 e1 + · · · + kρ−1 eρ−1 ) + eρ = ki ei + eρ
i=1
ρ−1
X ρ ρ−1
ṡ = ki ei+1 +Lf h(x)+Lg Lf h(x)[u+δ(t, x, u)]−r (ρ) (t)
i=1
"ρ−1 #
1 X
u=− ρ−1 ki ei+1 + Lρf h(x) − r (ρ) (t) + v
Lg Lf h(x) i=1
ṡ = Lg Lρ−1
f h(x)v + ∆(t, x, v)

∆(t, x, v)

ρ−1 ≤ ̺(x) + κ0 |v|, 0 ≤ κ0 < 1
Lg L h(x)
f
– p. 10/1

s
v = −β(x) sat , ε>0
ε
̺(x)
β(x) ≥ + β0 , β0 >
(1 − κ0 )
What properties can we prove for this control?
– p. 11/1
Lecture # 36
Tracking
Equilibrium-to-Equilibrium
Transition
– p. 1/1
η̇ = f0 (η, ξ)
ξ̇i = ξi+1 , 1≤i≤ρ−1
ρ ρ−1
| {z } | {z }
fb (η,ξ) gb (η,ξ)
y = ξ1
Equilibrium point:
0 = f0 (η̄, ξ̄)
0 = ξ̄i+1 , 1≤i≤ρ−1
0 = fb (η̄, ξ̄) + gb (η̄, ξ̄)ū
ȳ = ξ̄1
– p. 2/1
ξ̄1 = ȳ, ξ̄i = 0 for 2 ≤ i ≤ ρ − 1
fb (η̄, ȳ, 0, · · · , 0)
0 = f0 (η̄, ȳ, 0, · · · , 0), ū = −
gb (η̄, ȳ, 0, · · · , 0)
Assume f0 (η̄, ȳ, 0, · · · , 0) has a unique solution η̄ in the
domain of interest
η̄ = φη (ȳ), ū = φu (ȳ)
By assumption (and without loss of generality)
φη (0) = 0, φu (0) = 0
– p. 3/1
Goal: Move the system from equilibrium at y = 0 to
equilibrium at y = ȳ , either asymptotically or over a finite
time period
First Approach: Apply a step command

(
∗ 0, for t < 0
y (t) =
ȳ for t ≥ 0
Is this allowed ?
Take r = ȳ, for t ≥ 0
r (i) = 0 for i ≥ 2
– p. 4/1
η(0) = 0, e1 (0) = −ȳ, ei (0) = 0 for i ≥ 2
The shape of the transient response depends on the

solution of
ė = (Ac − Bc K)e
in feedback linearization
or the solution of
      
ė1 1 e1

 ė2 
 
 ... 
 e2
 
  
 . = + 
 .  

 ..   s
 .   1  .  
ėρ−1 −k1 −kρ−1 eρ−1 1
in sliding mode control

What is the impact of the reaching phase?
– p. 5/1
Second Approach: Take r(t) as the zero-state response of
a Hurwitz transfer function driven by y∗
Typical Choice:
aρ
sρ + a1 sρ−1 + · · · + aρ−1 s + aρ
Choose the parameters a1 to aρ to shape the response of r
r(0) = 0 ⇒ e1 (0) = 0 ⇒ e(0) = 0
Sliding Mode Control:
– p. 6/1
The derivatives of r are generated by the pre-filter
   
1

 . ..

  
z +   ∗
ż =     y
 1   
−aρ −a1 aρ
h i
r = 1 z
r = z1 , ṙ = z2 , . . . . . . r (ρ−1) = zρ
ρ
X
r (ρ) = − aρ−i+1 zi + aρ y ∗
i=1
Does r(t) satisfy the assumptions imposed last lecture?

– p. 7/1
Example 13.22
ẋ1 = x2 , ẋ2 = −10 sin x1 − x2 + 10u, y = x1
Move the pendulum from equilibrium at x1 = 0 to

equilibrium at x1 = π2
Constraint : |u(t)| ≤ 2
(
∗ 0, for t < 0
y = π
2 for t ≥ 0
Pre-Filter:
1
(τ s + 1)2
– p. 8/1
" # " #
0 1 0
ż = −1 −2 z+ 1 y∗
τ2 τ τ2
h i
r = 1 0 z
1 ∗ 2
r = z1 , ṙ = z2 , r̈ = (y − r) − z2
τ2 τ
h t i
r(t) = π
2
1 − e− τ 1 + τt
u = 0.1(10 sin x1 + x2 + r̈ − k1 e1 − k2 e2 )
– p. 9/1
τ = 0.05 τ = 0.25
2 2
1.5 1.5
Output
Output
1 1
output output
reference reference
0.5 0.5
0 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
τ = 0.05 τ = 0.25
2 2
1 1
Control
0 Control 0
−1 −1
−2 −2
0 0.5 1 1.5 2 0 0.5 1 1.5 2

Time Time
– p. 10/1
Third Approach: Plan a trajectory (r(t), ṙ(t), . . . , r (ρ) (t))
to move from (0, 0, . . . , 0) to (ȳ, 0, . . . , 0) in finite time T
Example: ρ = 2
a
r(2)
0 T
T/2
−a
r(1)
at a(T−t)
aT2/4
r
−aT2/4+aTt−at2/2
at2/2
– p. 11/1

at2

 2
for 0 ≤ t ≤ T2
aT 2 at2
r(t) = − 4 + aT t − 2 for T2 ≤ t ≤ T

 aT 2
4 for t ≥ T
4ȳ
a= ⇒ r(t) = ȳ for t ≥ T
T2
– p. 12/1
Lecture # 37
Observers
Linearization
and
Extended Kalman Filter (EKF)
– p. 1/1
Linear Observer via Linearization
ẋ = f (x, u), y = h(x)
0 = f (xss , uss ), yss = h(xss )

Linearize about the equilibrium point:
ẋδ = Axδ + Buδ , yδ = Cxδ
xδ = x − xss , uδ = u − uss , yδ = y − yss

What are A, B , C ?
x̂˙ δ = Ax̂δ + Buδ + H(yδ − C x̂δ ), x̂ = xss + x̂δ
(A − HC) is Hurwitz
It will work locally for sufficiently small kxδ (0)k, kx̂δ (0)k,
and kuδ (t)k
– p. 2/1
Feedback Control:
x̂˙ δ = Ax̂δ + Buδ + H(yδ − C x̂δ )
uδ = −K x̂δ , u = uss − K x̂δ

Verify that the closed-loop system has an equilibrium point
at
x = xss , x̃ = 0
and linearization at the equilibrium point yields
" # " #" #
ẋδ (A − BK) BK xδ
=
x̃˙ 0 (A − HC) x̃
Which theorem would justify this controller locally?

– p. 3/1
Nonlinear Observer via Linearization
ẋ = f (x, u), y = h(x)
0 = f (xss , uss ), yss = h(xss )
x̂˙ = f (x̂, u) + H[y − h(x̂)]

Equilibrium Point: x = xss , u = uss , x̂ = xss
x̃ = x − x̂
x̃˙ = g(x, x̃, u), g(xss , 0, uss ) = 0

Verify that linearization at the equilibrium point yields
x̃˙ = (A − HC)x̃
Investigate the design of H and the use in feedback control
– p. 4/1
Extended Kalman Filter (EKF)
ẋ = f (x, u) + w, y = h(x, u) + v
x̂˙ = f (x̂, u) + H(t)[y − h(x̂, u)]

x̃ = x − x̂
x̃˙ = f (x, u) + w − f (x̂, u) − H[h(x, u) + v − h(x̂, u)]
Substitute x = x̂ + x̃ and expand the RHS in a Taylor
series about x̃ = 0
x̃˙ = [A(t) − H(t)C(t)]x̃ + η(x̃, t) + ξ(t)
∂f ∂h
A(t) = (x̂(t), u(t)), C(t) = (x̂(t), u(t))
∂x ∂x
η(0, t) = 0, ξ(t) = w(t) − H(t)v(t)
– p. 5/1
Assuming that x(t), u(t), w(t), v(t), and H(t) are
bounded and f and h are twice continuously differentiable,
show that
kη(x̃, t)k ≤ k1 kx̃k2 , kξ(t)k ≤ k2
Hint:
∂f
f (x, u) − f (x̂, u) − (x̂, u)x̃
∂x
Z 1
∂f ∂f
= (σ x̃ + x̂, u) dσ x̃ − (x̂, u)x̃ (Exercise 3.23)
0 ∂x ∂x
Z 1
∂f ∂f

= (σ x̃ + x̂, u) − (x̂, u) dσ x̃
0 ∂x ∂x
– p. 6/1
Kalman Filter Design: Let Q(t) and R(t) be symmetric
positive definite matrices that satisfy
0 < q1 I ≤ Q(t) ≤ q2 I, 0 < r1 I ≤ R(t) ≤ r2 I
Let P (t) be the solution of the Riccati equation
Ṗ = AP + P AT + Q − P C T R−1 CP, P (t0 ) = P0 > 0
If (A(t), C(t) is uniformly observable, then P (t) exists for

all t ≥ t0 and satisfies
0 < p1 I ≤ P (t) ≤ p2 I ⇒ 0 < p3 I ≤ P −1 (t) ≤ p4 I

See a texbook on optimal control or optimal estimation
H(t) = P (t)C(t)T R−1 (t)
– p. 7/1
Compute A(t) and C(t)
∂f ∂h
A(t) = (x̂(t), u(t)), C(t) = (x̂(t), u(t))
∂x ∂x
Solve the Riccati equation
Compute H(t)
H(t) = P (t)C(t)T R−1 (t)
Remark: The Riccati equation and the observer equation

have to be solved simultaneously in real time because A(t)
and C(t) depend on x̂(t) and u(t)
– p. 8/1
Lemma: The origin of
x̃˙ = [A(t) − H(t)C(t)]x̃ + η(x̃, t)
is exponentially stable and the solutions of
x̃˙ = [A(t) − H(t)C(t)]x̃ + η(x̃, t) + ξ(t)
are uniformly ultimately bounded by an ultimate bound

proportional to k2
Proof:
V = x̃T P −1 x̃
−1 ˙ ˙ T −1 T d
T
V̇ = x̃ P x̃ + x̃ P x̃ + x̃ P −1 x̃
dt
– p. 9/1
d
P −1 = −P −1 Ṗ P −1
dt
V̇ = x̃T P −1 (A − P C T R−1 C)x̃
+ x̃T (AT − C T R−1 CP )P −1 x̃
− x̃T P −1 Ṗ P −1 x̃ + 2x̃T P −1 (η + ξ)
V̇ = x̃T P −1 (AP + P AT − P C T R−1 CP − Ṗ )P −1 x̃

− x̃T C T R−1 C x̃ + 2x̃T P −1 (η + ξ)
= −x̃T (P −1 QP −1 + C T R−1 C)x̃ + 2x̃T P −1 (η + ξ)
≤ −c1 kx̃k2 + c2 kx̃k3 + c3 kx̃k (c3 ∝ k2 )
– p. 10/1
Stochastic Interpretation: When w(t) and v(t) are
zero-mean, white noise stochastic processes,
uncorrelated, i.e., E{w(t)v T (τ )} = 0, ∀t, τ , and
E{w(t)w T (τ )} = Q(t)δ(t − τ )
E{v(t)v T (τ )} = R(t)δ(t − τ )
then x̂(t) is an approximation of the minimum variance
estimate that minimizes
n o
E [y(t) − h(x̂(t), u(t))]T [y(t) − h(x̂(t), u(t))]
and P (t) is an approximation of the covariance matrix

n o
E [x̂(t) − x(t)][x̂(t) − x(t)]T
– p. 11/1
Feedback Control: What can you say about the closed-loop
system when x̂ is used in feedback control?
– p. 12/1
Lecture # 38
Observers
Exact Observers
– p. 1/1
Observer with Linear Error Dynamics
Observer Form:
ẋ = Ax + γ(y, u), y = Cx
where (A, C) is observable, x ∈ Rn , u ∈ Rm , y ∈ Rp

From Lecture # 24: An n-dimensional SO system
ẋ = f (x) + g(x)u, y = h(x)
is transformable into the observer form if and only if

∂φ
h iT
n−1
φ = h, Lf h, · · · Lf h , rank (x) = n
∂x
h iT ∂φ
b= 0, ··· 0, 1 , τ =b
∂x
– p. 2/1
[adif τ, adjf τ ] = 0, 0 ≤ i, j ≤ n − 1
[g, adjf τ ] = 0, 0≤j ≤n−2

Change of variables:
τi = (−1)i−1 adi−1
f τ, 1≤i≤n
∂T h i
τ1 , τ2 , ··· τn =I
∂x
z = T (x)
– p. 3/1
ẋ = Ax + γ(y, u), y = Cx
x̂˙ = Ax̂ + γ(y, u) + H(y − C x̂)

x̃ = x − x̂
x̃˙ = (A − HC)x̃
Design H such that (A − HC) is Hurwitz
What about feedback control?
Let u = ψ(x) be a globally stabilizing state feedback
control
u = ψ(x̂)
x̂˙ = Ax̂ + γ(y, u) + H(y − C x̂)
– p. 4/1
How would you analyze the closed-loop system?
ẋ = Ax + γ(Cx, ψ(x − x̃))

x̃˙ = (A − HC)x̃
We know that
the origin of ẋ = Ax + γ(Cx, ψ(x)) is globally
the origin of x̃˙ = (A − HC)x̃ is globally exponentially
stable
What additional assumptions do we need to show that the
origin of the closed-loop system is globally asymptotically
stable?
– p. 5/1
Circle Criterion Design
ẋ = Ax + γ(y, u) − Lβ(M x), y = Cx
where (A, C) is observable, x ∈ Rn , u ∈ Rm , y ∈ Rp ,

M x ∈ Rℓ , β(η) = [ β1 (η1 ), . . . , βℓ (ηℓ ) ]T
x̂˙ = Ax̂ + γ(y, u) − Lβ(M x̂ − N (y − C x̂)) + H(y − C x̂)
x̃ = x − x̂
x̃˙ = (A − HC)x̃ − L[β(M x) − β(M x̂ − N (y − C x̂))]
x̃˙ = (A − HC)x̃ − L[β(M x) − β(M x − (M + N C)x̃)]

Define
z = (M + N C)x̃
ψ(t, z) = β(M x(t)) − β(M x(t) − z)
– p. 6/1
x̃˙ = (A − HC)x̃ − Lψ(t, z)
z = (M + N C)x̃
def
G(s) = (M + N C)[sI − (A − HC)]−1 L
0 +
- n -
z -
G(s)
−6
ψ(·)
h iT
ψ(t, z) = ψ1 (t, z1 ), . . . , ψℓ (t, zℓ )
– p. 7/1
Main Assumption: βi (·) is a nondecreasing function
(a − b)[βi (a) − βi (b)] ≥ 0, ∀ a, b ∈ R
If βi (ηi ) is continuously differentiable

dβi
≥ 0, ∀ηi ∈ R
dηi
zi ψi (t, zi ) = zi [βi ((M x)i ) − βi ((M x)i − zi )] ≥ 0

z T ψ(t, z) ≥ 0
– p. 8/1
By the circle criterion (Theorem 7.1) the origin of
x̃˙ = (A − HC)x̃ − Lψ(t, z)

z = (M + N C)x̃
is globally exponentially stable if

def
G(s) = (M + N C)[sI − (A − HC)]−1 L
is strictly positive real
Design Problem: Design H and N such that G(s) is

strictly positive real
Feasibility can be investigated using LMI (Arcak &
Kokotovic, Automatica, 2001)
– p. 9/1
Example:
ẋ1 = x2 , ẋ2 = −x31 − x32 + u, y = x1

" # " #
0 1 h i 0
A= , C= 1 0 , γ= 3
,
0 0 −y + u
" #
0 h i
3 dβ
L= , M = 0 1 , β(η) = η , = 3η 2 ≥ 0
1 dη
" #
h1
h= , N
h2
−1 s + N + h1
G(s) = (M + N C)[sI − (A − HC)] L=
s2 + h1 s + h2
– p. 10/1
From Exercise 6.7, G(s) is SPR if and only if
h1 > 0, h2 > 0, 0 < N + h1 < h1

1
h1 = 2, h2 = 1, N =− 2
3
s+ 2
G(s) =
(s + 1)2
x̂˙ 1 = x̂2 + 2(y − x̂1 )

3
x̂˙ 2 = −y + u − x̂2 +
3 1
2 (y − x̂1 ) + (y − x̂1 )
– p. 11/1
What about feedback control?
Let u = φ(x) be a globally stabilizing state feedback control
Closed-loop system under output feedback:
ẋ = Ax + γ(y, φ(x − x̃)) − Lβ(M x)

x̃˙ = (A − HC)x̃ − Lψ(t, z)
z = (M + N C)x̃
How would you analyze the closed-loop system?
ψ(t, z) depends on x(t). How would you show that ψ is

well defined?
What about the effect of uncertainty?

– p. 12/1
Lecture # 38
Observers
High-Gain Observers
Motivating Example
– p. 1/1
ẋ1 = x2 , ẋ2 = φ(x, u), y = x1
Let u = γ(x) stabilize the origin of
ẋ1 = x2 , ẋ2 = φ(x, γ(x))
Observer:
x̂˙ 1 = x̂2 + h1 (y − x̂1 ), x̂˙ 2 = φ0 (x̂, u) + h2 (y − x̂1 )
φ0 (x, u) is a nominal model φ(x, u)

x̃1 = x1 − x̂1 , x̃2 = x2 − x̂2
x̃˙ 1 = −h1 x̃1 + x̃2 , x̃˙ 2 = −h2 x̃1 + δ(x, x̃)

δ(x, x̃) = φ(x, γ(x̂)) − φ0 (x̂, γ(x̂))
– p. 2/1
" # " #
h1 −h1 1
Design H = such that Ao = is Hurwitz
h2 −h2 0
Transfer function from δ to x̃:
" #
1 1
Go (s) =
s2 + h1 s + h2 s + h1
Design H to make supω∈R kGo (jω)k as small as possible

α1 α2
h1 = , h2 = , ε>0
ε ε2
" #
ε ε
Go (s) =
(εs)2 + α1 εs + α2 εs + α1
– p. 3/1
" #
ε ε
Go (s) =
(εs)2 + α1 εs + α2 εs + α1
Observer eigenvalues are (λ1 /ε) and (λ2 /ε) where λ1 and
λ2 are the roots of
λ2 + α1 λ + α2 = 0
sup kGo (jω)k = O(ε)

ω∈R
– p. 4/1
x̃1
η1 = , η2 = x̃2
ε
εη̇1 = −α1 η1 + η2 , εη̇2 = −α2 η1 + εδ(x, x̃)
Ultimate bound of η is O(ε)
η decays faster than an exponential mode e−at/ε , a>0
Peaking Phenomenon:
x1 (0) 6= x̂1 (0) ⇒ η1 (0) = O(1/ε)
1
The solution contains a term of the form e−at/ε
ε
1
e−at/ε approaches an impulse function as ε → 0
ε
– p. 5/1
Example
ẋ1 = x2 , ẋ2 = x32 + u, y = x1
State feedback control:
u = −x32 − x1 − x2
Output feedback control:
u = −x̂32 − x̂1 − x̂2

x̂˙ 1 = x̂2 + (2/ε)(y − x̂1 )
x̂˙ 2 = (1/ε2 )(y − x̂1 )
– p. 6/1
0.5
0 SFB
−0.5 OFB ε = 0.1
1
OFB ε = 0.01
x
−1
−1.5
OFB ε = 0.005
−2
0 1 2 3 4 5 6 7 8 9 10
−1
2
x
−2
−3
0 1 2 3 4 5 6 7 8 9 10
−100
u
−200
−300
−400
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
t
– p. 7/1
ε = 0.004
0.2
0
1
−0.2
x
−0.4
−0.6
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
−200
2
x
−400
−600
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
2000
1000
u
−1000
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
t
– p. 8/1
u = sat(−x̂32 − x̂1 − x̂2 )
SFB
0.15
OFB ε = 0.1
0.1 OFB ε = 0.01
OFB ε = 0.001
x1
0.05
−0.05
0 1 2 3 4 5 6 7 8 9 10
0.05
0
x2
−0.05
−0.1
0 1 2 3 4 5 6 7 8 9 10
0
u
−0.5
−1
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
t
– p. 9/1
Region of attraction under state feedback:
2
x2
0
−1
−2
−3 −2 −1 0 1 2 3
x1
– p. 10/1
Region of attraction under outputfeedback:
0.5
x2
−0.5
−1
−1.5 −1 −0.5 0 0.5 1 1.5

x1
ε = 0.1 (dashed) and ε = 0.05 (dash-dot)

– p. 11/1
Analysis of the closed-loop system:
ẋ1 = x2 ẋ2 = φ(x, γ(x − x̃))

εη̇1 = −α1 η1 + η2 εη̇2 = −α2 η1 + εδ(x, x̃)
η6
O(1/ε)
q q
D D
D D
D D
D D
D D
D D
O(ε) D D
W W
-
Ωb - x
Ωc -
– p. 12/1
What is the effect of measurement noise?
The high-gain observer is an approximate differentiator
Transfer function from y to x̂ (with φ0 = 0):

" # " #
α2 1 + (εα1 /α2 )s 1
2
→ as ε → 0
(εs) + α1 εs + α2 s s
Differentiation amplifies the effect of measurement noise
y = x1 + v, kn = sup |v(t)| < ∞

t≥0
– p. 13/1
1.1
0.9
0.8
the error bound 0.7
0.6
0.5
0.4
0.3
0.2
0.1
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
the high−gain parameter epsilon
s !
kn
εopt = O , kd = sup |ẍ1 (t)|, kn = sup |v(t)|
kd t≥0 t≥0
– p. 14/1
Lecture # 40
Observers
High-Gain Observers
Stabilization
– p. 1/1
ẋ = Ax + Bφ(x, z, u)
ż = ψ(x, z, u)
y = Cx
ζ = q(x, z)
u ∈ Rp , y ∈ Rm , ζ ∈ Rs , x ∈ Rρ , z ∈ Rℓ
A, B, C are block diagonal matrices
   
0 1 ··· ··· 0 0
 0 0 1 ··· 0   0
   

 . .  .
. .  ..
 
Ai = 
 . . 
 , B i = 


 0 ··· ··· 0 1   0
   

0 · · · · · · · · · 0 ρ ×ρ 1 ρi ×1
i i
– p. 2/1
h i m
X
Ci = 1 0 ··· ··· 0 , ρ= ρi
1×ρi
i=1
Normal form
Mechanical and electromechanical systems
Example: Magnetic Suspension
ẋ1 = x2
k L0 ax23
ẋ2 = g − x2 −
m 2m(a + x1 )2
1 L0 ax2 x3

ẋ3 = −Rx3 + 2
+u
L(x1 ) (a + x1 )
– p. 3/1
Stabilizing (partial) state feedback controller:
u = γ(x, ζ)
ϑ̇ = Γ(ϑ, x, ζ), u = γ(ϑ, x, ζ)

Closed-loop system under state feedback:
Ẋ = f (X ), X = (x, z, ϑ)
The origin of Ẋ = f (X ) is asymptotically stable
Observer:
x̂˙ = Ax̂ + Bφ0 (x̂, ζ, u) + H(y − C x̂)
– p. 4/1
H is block diagonal
αi1 /ε
 

 αi2 /ε2 

 .. 
Hi = 
 . 

 i
 αρi −1 /ερi −1


αiρi /ερi ρi ×1
sρi + αi1 sρi −1 + · · · + αiρi −1 s + αiρi

is Hurwitz and ε > 0 (small)
φ0 (x, ζ, u) is a nominal model of φ(x, z, u), which is
globally bounded in x
– p. 5/1
Theorem 14.6 (Nonlinear Separation Principle:
Suppose the origin of Ẋ = f (X ) is asymptotically stable
and R is its region of attraction. Let S be any compact set
in the interior of R and Q be any compact subset of Rρ .
Then,
∃ ε∗1 > 0 such that, for every 0 < ε ≤ ε∗1 , the solutions
(X (t), x̂(t)) of the closed-loop system, starting in
S × Q, are bounded for all t ≥ 0
given any µ > 0, ∃ ε∗2 > 0 and T2 > 0, dependent on

µ, such that, for every 0 < ε ≤ ε∗2 , the solutions of the
closed-loop system, starting in S × Q, satisfy
kX (t)k ≤ µ and kx̂(t)k ≤ µ, ∀ t ≥ T2
– p. 6/1
given any µ > 0, ∃ ε∗3 > 0, dependent on µ, such that,
for every 0 < ε ≤ ε∗3 , the solutions of the closed-loop
system, starting in S × Q, satisfy
kX (t) − Xr (t)k ≤ µ, ∀t≥0
where Xr is the solution of Ẋ = f (X ), starting at X (0)
if the origin of Ẋ = f (X ) is exponentially stable, then ∃

ε∗4 > 0 such that, for every 0 < ε ≤ ε∗4 , the origin of the
closed-loop system is exponentially stable and S × Q is
a subset of its region of attraction.
– p. 7/1
Key ideas of the proof:
Representation of the closed-loop system as a
singularly perturbed one with X as the slow and η
(scaled estimation error) as the fast
Use of a converse Lyapunov theorem to construct

positively invariant sets
Use of global boundedness in x̂ to show that η reaches

O(ε) while X is inside a positively invariant set
Nonlocal versus local analysis
Novel Feature: Performance recovery
– p. 8/1
Example 14.19:
mℓ2 θ̈ + mg0 ℓ sin θ + k0 ℓ2 θ̇ = u
Stabilization at (θ = π, θ̇ = 0). From Section 14.1

!
a1 (θ − π) + θ̇
u = −k sat
µ
Suppose we only measure θ
θ̂˙ = ω̂ + (2/ε)(θ − θ̂)

ω̂˙ = φ0 (θ̂, u) + (1/ε2 )(θ − θ̂)
φ0 = −â sin θ̂ + ĉu
– p. 9/1
!
a1 (θ̂ − π) + ω̂
u = −k sat
µ
or
a1 (θ − π) + ω̂

u = −k sat
µ
– p. 10/1
(a) (b)
3.5 2
3
1
2.5
ω
θ
2 0
1.5 SFB
OFB ε = 0.05 −1
1 OFB ε = 0.01
0.5 −2
0 2 4 6 8 10 0 2 4 6 8 10
(c) (d)
3.5 3.5
3 3
2.5
2.5
2
θ
θ
2
1.5
1.5
1
1 0.5
0.5 0
0 2 4 6 8 10 0 2 4 6 8 10
Time Time
– p. 11/1
Lecture # 41
Integral Control
– p. 1/1
ẋ = f (x, u, w)
y = h(x, w)
ym = hm (x, w)
x ∈ Rn state, u ∈ Rp control input

y ∈ Rp controlled output, ym ∈ Rm measured output
w ∈ Rl unknown constant parameters and disturbances
Goal:
y(t) → r as t → ∞
r ∈ Rp constant reference, v = (r, w)
e(t) = y(t) − r
– p. 2/1
Assumption: e can be measured
Steady-state condition: There is a unique pair (xss , uss )

that satisfies the equations
0 = f (xss , uss , w)
0 = h(xss , w) − r
Stabilize the system at the equilibrium point x = xss
Can we reduce this to a stabilization problem by shifting the

equilibrium point to the origin via the change of variables
xδ = x − xss , uδ = u − uss ?
– p. 3/1
Integral Action:
σ̇ = e
Augmented System:
ẋ = f (x, u, w)
σ̇ = h(x, w) − r
Task: Stabilize the augmented system at (xss , σss ) where

σss produces uss
r - l - R −σ
- Stabilizing u- Plant y-
+ Controller
−6
6 6
Measured
Signals
– p. 4/1
Integral Control via Linearization
State Feedback:
u = −K1 x − K2 σ − K3 e
Closed-loop system:
ẋ = f (x, −K1 x − K2 σ − K3 (h(x, w) − r), w)

σ̇ = h(x, w) − r
Equilibrium points:
0 = f (x̄, ū, w)
0 = h(x̄, w) − r
ū = −K1 x̄ − K2 σ̄
Unique equilibrium point at x = xss , σ = σss , u = uss

– p. 5/1
Linearization about (xss , σss ):
" #
x − xss
ξδ =
σ − σss
ξ̇δ = (A − BK)ξδ
" #
A 0 ∂f ∂h
A= , A= (x, u, w) , C =
(x, w)
C 0 ∂x eq ∂x eq
" #
B ∂f
B= , B= (x, u, w)
0 ∂u eq
h i
K = K1 + K3 C K2
– p. 6/1
(A, B) is controllable if and only if (A, B) is controllable
and " #
A B
rank =n+p
C 0
Task: Design K, independent of v , such that (A − BK) is
Hurwitz for all v
(xss , σss ) is an exponentially stable equilibrium point of the

closed-loop system. All solutions starting in its region of
attraction approach it as t tends to infinity
e(t) → 0 as t → ∞
– p. 7/1
Pendulum Example:
Regulate θ to δ
x1 = θ − δ, x2 = θ̇, u=T
ẋ1 = x2
ẋ2 = −a sin(x1 + δ) − bx2 + cu
" #
0 a
xss = , uss = sin δ
0 c
σ̇ = x1
– p. 8/1
   
0 1 0 0
A =  −a cos δ −b 0  , B =  c 
   
1 0 0 0
K1 = [k1 k2 ], K 2 = k3 , K3 = 0
(A − BK) will be Hurwitz if
b+k2 c > 0, (b+k2 c)(a cos δ +k1 c)−k3 c > 0, k3 c > 0
a 1
Suppose ≤ ρ1 , ≤ ρ2
c c
k3
k2 > 0, k3 > 0, k1 > ρ1 + ρ2
k2
– p. 9/1
Output Feedback: We only measure e and ym
σ̇ = e = y − r
ż = F z + G1 σ + G2 ym
u = Lz + M1 σ + M2 ym + M3 e
Task: Design F , G1 , G2 , L, M1 , M2 , and M3 ,

independent of v , such that Ac is Hurwitz for all v
 
A + BM2 Cm + BM3 C BM1 BL
Ac =  C 0 0 
 
G2 Cm G1 F

∂hm
Cm = (x, w)
∂x eq
– p. 10/1
Integral Control via Sliding Mode Design
η̇ = f0 (η, ξ, w)
ξ̇1 = ξ2
.. ..
. .
ξ̇ρ−1 = ξρ
ξ̇ρ = b(η, ξ, u, w) + a(η, ξ, w)u
y = ξ1
a(η, ξ, w) ≥ a0 > 0
Goal:
y(t) → r as t → ∞
ξss = [r, 0, . . . , 0]T
– p. 11/1
Steady-state condition: There is a unique pair (ηss , uss )
that satisfies the equations
0 = f0 (ηss , ξss , w)
0 = b(ηss , ξss , uss , w) + a(ηss , ξss , w)uss
ė0 = y − r
   
e1 ξ1 − r
 e2   ξ2
   

z = η − ηss , e=  ..  = 
  .. 
 .  .

 
eρ ξρ
– p. 12/1
def
ż = f0 (η, ξ, w) = f˜0 (z, e, w, r)
ė0 = e1
ė1 = e2
.. ..
. .
ėρ−1 = eρ
ėρ = b(η, ξ, u, w) + a(η, ξ, w)u
Partial State Feedback: {e1 , . . . , eρ } are measured
s = k0 e0 + k1 e1 + · · · + kρ−1 eρ−1 + eρ
k0 to kρ−1 are chosen such that the polynomial
λρ + kρ−1 λρ−1 + · · · + k1 λ + k0 is Hurwitz

– p. 13/1
ṡ = k0 e1 + · · · + kρ−1 eρ + b(η, ξ, u, w) + a(η, ξ, w)u
ṡ = ∆(η, ξ, u, w, r) + a(η, ξ, w)u

∆(η, ξ, u, w, r)

≤ ̺(e) + κ0 |u|, 0 ≤ κ0 < 1
a(η, ξ, w)
s

u = −β(e) sat
µ
̺(e)
β(e) ≥ + β0 , β0 > 0
(1 − κ0 )
For |s| ≥ µ, sṡ ≤ −a0 (1 − κ0 )β0
What about the other state variables?
– p. 14/1
ż = f˜0 (z, e, w, r)
ζ̇ = Aζ + Bs (A is Hurwitz)
s

ṡ = −a(·)β(e) sat + ∆(·)
µ
ζ = [e0 , . . . , eρ−1 ]T
α̃1 (kzk) ≤ V1 (z, w, r) ≤ α̃2 (kzk)

∂V1 ˜
f0 (z, e, w, r) ≤ −α̃3 (kzk), ∀ kzk ≥ γ̃(kek)
∂z
V2 (ζ) = ζ T P ζ, P A + AT P = −I
– p. 15/1
Ω = {|s| ≤ c} ∩ {V2 ≤ c2 ρ1 } ∩ {V1 ≤ c0 }
Ωµ = {|s| ≤ µ} ∩ {V2 ≤ µ2 ρ1 } ∩ {V1 ≤ α̃2 (γ̃(µρ2 ))}

All trajectories starting in Ω enter Ωµ in finite time and stay
in thereafter
Inside Ωµ there is a unique equilibrium point at
s̄
(z = 0, e = 0, e0 = ē0 ), s̄ = k0 ē0 , uss = −β(0)
µ
Under additional conditions (the origin of ż = f˜0 (z, 0, w, r)

is exponentially stable), local analysis inside Ωµ shows that
for sufficiently small µ all trajectories converge to the
equilibrium point as time tends to infinity
– p. 16/1
Output Feedback: Only e1 is measured
High-gain Observer:
ė0 = e1
k0 e0 + k1 e1 + k2 ê2 + · · · + êρ

u = −β sat
µ
˙êi = êi+1 + αi (e1 − ê1 ),

1≤i≤ρ−1
εi
αρ

˙êρ = (e1 − ê1 )
ερ
β = β(e1 , ê2 , . . . , êρ )
– p. 17/1

Khalil - Nonlinear Systems Slides

Uploaded by

Document Information

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

Khalil - Nonlinear Systems Slides

Uploaded by

Copyright:

Available Formats

Nonlinear Systems and Control

ẋi denotes the derivative of xi with respect to the time

u1 , u2 , . . ., up are input variables

x1 , x2 , . . ., xn the state variables

x is the state, u is the input

Unforced state equation:

Results from ẋ = f (t, x, u) with u = γ(t, x)

A time-invariant state model has a time-invariance property

f (t, x) is piecewise continuous in t and locally Lipschitz in

f (t, x) is piecewise continuous in t on an interval J ⊂ R if

f (t, x) is locally Lipschitz in x at a point x0 if there is a

When n = 1 and f depends only on x

On a plot of f (x) versus x, a straight line joining any two

Any function f (x) that has infinite slope at some point is

The function f (x) = x1/3 is not locally Lipschitz at x = 0

More generally, if for t ∈ J ⊂ R and x in a domain

Without the local Lipschitz condition, we cannot ensure

The lemma is a local result because it guarantees existence

In general, if f (t, x) is locally Lipschitz over a domain D

A function f (t, x) is globally Lipschitz in x if

kf (t, x) − f (t, y)k ≤ Lkx − yk

for all x, y ∈ Rn with the same Lipschitz constant L

If f (t, x) and its partial derivatives ∂fi /∂xj are continuous

f (x) = −x2 is locally Lipschitz for all x but not globally

The global Lipschitz condition is satisfied for linear systems

ẋ = f (t, x), x(t0 ) = x0

with x0 ∈ W lies entirely in W . Then, there is a unique

If, at any instant of time, x(t) is positive, the derivative ẋ(t)

Therefore, starting from any initial condition x(0) = a, the

Thus, the equation has a unique solution for all t ≥ 0

A point x = x∗ in the state space is said to be an

For the autonomous system ẋ = f (x), the equilibrium

An equilibrium point could be isolated; that is, there are no

It cannot have multiple isolated equilibrium points , for if xa

A nonlinear state equation can have multiple isolated

ẋ1 = x2 , ẋ2 = −a sin x1 − bx2

has equilibrium points at (x1 = nπ, x2 = 0) for

A common engineering practice in analyzing a nonlinear

What are the limitations of linearization?

There are “essentially nonlinear phenomena” that can

Multiple isolated equilibrium points

Subharmonic, harmonic, or almost-periodic oscillations

Multiple modes of behavior

mlθ̈ = −mg sin θ − klθ̇

vC − E + RiL + vL = 0 ⇒ vL = −x1 − Rx2 + u

Static or Coulomb friction

g(y) = k(1 − a2 y 2 )y, |ay| < 1 (softening spring)

When the mass is at rest, there is a static friction force Fs

Once motion has started, the resistive force Ff is modeled

h(0) = 0, h′ (0) < 0

h(v) → ∞ as v → ∞, and h(v) → −∞ as v → −∞

Differentiating with respect to t and multiplying by L:

Special case: Van der Pol equation

Change of variables: z = T (x)

Ref erenceM odel : ẏm = am ym + km r

u(t) = θ1 (t)r(t) + θ2 (t)yp (t)

where θ1 (t) and θ2 (t) are adjusted on-line

State Variables: eo = yp − ym , φ1 = θ1 − θ1∗ , φ2 = θ2 − θ2∗

ėo = ap eo + kp (θ1 − θ1∗ )r + kp (θ2 yp − θ2∗ ym )

ėo = am eo + kp φ1 r(t) + kp φ2 [eo + ym (t)]

Let x(t) = (x1 (t), x2 (t)) be a solution that starts at initial

Represent f (x) as a vector based at x; that is, assign to x

ẋ1 = x2 , ẋ2 = −10 sin x1

Select a bounding box in the state plane