Introduction to Nonlinear Control

Lecture # 5

State Feedback Stabilization

– 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 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/?
Definition: The system

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 = 12 x2 ⇒ V̇ = −kx2 + xd(t)
 
V̇ ≤ − k2 x2 − |x| k2 |x| − δ

2δ
V̇ ≤ − k2 x2 , ∀ |x| ≥
k
By choosing k large enough we can achieve global practical
stabilization

– p. 17/?
 Normal Form

Relative Degree: Consider the system

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. 18/?
 ∂(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. 19/?
 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
Definition: The system

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. 20/?
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. 21/?
Definition: a map z = T (x) is a diffeomorphism on a
domain D if there is an inverse map T −1 (·) such that
x = T −1 (z) for all z ∈ T (D), and both T (·) and T −1 (·)
are continuously differentiable

If the Jacobian [∂T /∂x] is nonsingular at a point x0 , then

there is a neighborhood N of x0 such that T restricted to N
is a diffeomorphism on N

A map T is a global diffeomorphism if it is a diffeomorphsim

on Rn and T (Rn ) = Rn , which is equivalent to
∂f
is nonsingular for all x and lim kT (x)k = ∞
∂x kxk→∞

– p. 22/?
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. 23/?
 ∂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. 24/?
 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 1 
  0
 


0 ... ... 0 0 1

h i −Lρf h
Cc = 1 0 ... 0 0 , α= ρ−1 , γ = Lg Lρ−1
f h
Lg Lf h
– p. 25/?
Normal Form:
η̇ = f0 (η, ξ)
ξ̇ = Ac ξ + Bc γ(x)[u − α(x)], γ(x) 6= 0
y = Cc ξ

Zero Dynamics

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

– p. 26/?
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)

– p. 27/?
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. 28/?
 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. 29/?
Normal Form (ρ < n):

η̇ = f0 (η, ξ)
ξ̇ = Ac ξ + Bc γ(x)[u − α(x)]
y = Cc ξ

Normal Form (ρ = n):

ξ̇ = Ac ξ + Bc γ(x)[u − α(x)]
y = Cc ξ

– p. 30/?
Given the system

ẋ = f (x) + g(x)u

the existence of h(x) such that the system

ẋ = f (x) + g(x)u
y = h(x)

has relative degree n is characterized by geometric

conditions on the vector fields f and g ( See Theorem 13.2
of the textbook)

– p. 31/?
 Feedback Linearization

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 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. 32/?
 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. 33/?
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

– p. 34/?
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. 35/?
 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. 36/?
Example (Pendulum Equation):

θ̈ = −a sin θ − bθ̇ + cT
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. 37/?
 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. 38/?
 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. 39/?
 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. 40/?
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 ⇒ V̇ = h(x1 )ẋ1 + x2 ẋ2 = yu
0

– p. 41/?
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. 42/?
 Partial Feedback linearization

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. 43/?
 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

asymptotically stable
p
Proof: V (η, ξ) = V1 (η)+k ξT P ξ

– p. 44/?
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. 45/?
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. 46/?
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. 47/?
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. 48/?
 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. 49/?
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. 50/?
Proof–First Part:

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. 51/?
 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. 52/?