7.

Feedback Linearization

Nonlinear Control Systems

António Pedro Aguiar

pedro@isr.ist.utl.pt

7. Feedback Linearization

IST-DEEC PhD Course

http://users.isr.ist.utl.pt/%7Epedro/NCS2012/

2012

1
7. Feedback Linearization

Feedback Linearization

Given a nonlinear system of the form

ẋ = f (x) + G(x)u
y = h(x)

Does exist a state feedback control law

u = α(x) + β(x)v

and a change of variables

z = T (x)
that transforms the nonlinear system into a an equivalent linear system
(ż = Az + Bv) ?

2
7. Feedback Linearization

Feedback Linearization

Example: Consider the following system

` ´
ẋ = Ax + Bγ(x) u − α(x)

where γ(x) is nonsingular for all x in some domain D.

Then,
u = α(x) + β(x)v, with β(x) = γ −1 (x)
yields
ẋ = Ax + Bv
If we would like to stabilize the system, we design

v = −Kx such that A − BK is Hurwitz

Therefore
u = α(x) − β(x)Kx

3
7. Feedback Linearization

Feedback Linearization
Example: Consider now this example:
ẋ1 = a sin x2
ẋ2 = −x21 + u
How can we do this? We cannot simply choose u to cancel the nonlinear term a sin x2 !

However, if we first change the variables

z 1 = x1
z2 = a sin x2 = ẋ1
then
ż1 = z2
ż2 = a cos x2 ẋ2 = a cos x2 (−x21 + u)
Therefore with
1
u = x21 + v, −π/2 < x2 < π/2
a cos x2
we obtain the linear system
ż1 = z2
ż2 = v

4
7. Feedback Linearization

Feedback Linearization

• A continuously differentiable map T (x) is a diffeormorphism if T −1 (x) is

∂T
continuously differentiable. This is true if the Jacobian matrix ∂x
is nonsingular
∀x ∈ D.

• T (x) is a global diffeormorphism if and only if ∂T is nonsingular ∀x ∈ Rn and

∂x
T (x) is proper, that is, limkxk→∞ kT (x)k = ∞.

Definition
A nonlinear system

ẋ = f (x) + G(x)u (1)

where f : D → Rn and G : D → Rn×p are sufficiently smooth on a domain D ⊂ Rn

is said to be feedback linearizable (or input-state linearizable) if there exists a
diffeomorphism T : D → Rn such tat Dz = T (D) contains the origin and the change
of variables z = T (x) transforms (1) into the form
` ´
ż = Az + Bγ(x) u − α(x)

5
7. Feedback Linearization

Feedback Linearization
Suppose that we would like to solve the tracking problem for the system

ẋ1 = a sin x2
ẋ2 = −x21 + u
y = x2

If we use state feedback linearization we obtain

z1 = x1 ż1 = z2
z2 = a sin x2 = ẋ1 −→ ż2 = v
1
u = x21 + a cos x
v y = sin−1 (z2 /a)
2

which is not good!

Linearizing the state equation does not necessarily linearize the output equation.

Notice however if we set u = x21 + v we obtain

ẋ2 = v
y = x2

There is one catch: The linearizing feedback control law made x1 unobservable from
y. We have to make sure that x1 whose dynamics are given by ẋ1 = a sin x2 is well
behaved. For example, if y = yd = cte −→ x1 (t) = x1 (0) + ta sin yd . It is
unbounded!
6
7. Feedback Linearization

Input-Output Linearization
SISO system
ẋ = f (x) + g(x)u
y = h(x)
where f, g, h are sufficiently smooth in a domain D ⊂ Rn . The mappings f : D → Rn
and g : D → Rn are called vector fields on D.

Computing the first output derivative...

∂h ∂h
ẏ =
ẋ = [f (x) + g(x)u] =: Lf h(x) + Lg h(x)u
∂x ∂x
In the sequel we will use the following notation:
∂h
Lf h(x) = f (x) −→ Lie Derivative of h with respect to f
∂x
∂(Lf h)
Lg Lf h(x) = g(x)
∂x
L0f h(x) = h(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

7
7. Feedback Linearization

Input-Output Linearization

ẏ = Lf h(x) + Lg h(x)u
If Lg h(x)u = 0 then ẏ = Lf h(x) (independent of u).

Computing the second derivative...

∂(Lf h)
y (2) = [f (x) + g(x)u] = L2f h(x) + Lg Lf h(x)u
∂x

If Lg Lf h(x)u = 0 then ẏ (2) = L2f h(x) (independent of u).

Repeating this process, it follows that if

Lg Li−1
f h(x) = 0, i = 1, 2, . . . , ρ − 1
Lg Lρ−1
f h(x) 6= 0

then u does not appear in y, ẏ, . . . , y (ρ−1) and

(ρ−1)
y (ρ) = Lρf h(x) + Lg Lf h(x)u

8
7. Feedback Linearization

Input-Output Linearization

y (ρ) = Lρf h(x) + Lg Lρ−1

f h(x)u
Therefore, by setting
1
u= [−Lρf h(x) + v]
Lg Lρ−1
f h(x)
the system is input-output linearizable and reduces to

y (ρ) = v −→ chain of ρ integrators

Definition
The nonlinear system

ẋ = f (x) + g(x)u
y = h(x)

is said to have relative degree ρ, 1 ≤ ρ ≤ n, in the region D0 ⊂ D if for all x ∈ D0

Lg Li−1
f h(x) = 0, i = 1, 2, . . . , ρ − 1
Lg Lρ−1
f h(x) 6= 0

9
7. Feedback Linearization

Examples
Example 1: Van der Pol system
ẋ1 = x2
ẋ2 = −x1 + (1 − x21 )x2 + u, >0
1. y = x1
Calculating the derivatives...
ẏ = ẋ1 = x2
ÿ = ẋ2 = −x1 + (1 − x21 )x2 + u
Thus the system has relative degree ρ = 2 in R2 .

2. y = x2
Then
ẏ = ẋ2 = −x1 + (1 − x21 )x2 + u
In this case the system has relative degree ρ = 1 in R2 .

3. y = x1 + x22
Then
ẏ = x2 + 2x2 (−x1 + (1 − x21 )x2 + u)
In this case the system has relative degree ρ = 1 in D0 = {x ∈ R2 : x2 6= 0}.
10
7. Feedback Linearization

Examples

Example 2:

ẋ1 = x1
ẋ2 = x2 + u
y = x1

Calculating the derivatives...

ẏ = ẋ1 = x1 = y −→ y (n) = y = x1 , ∀n ≥ 1

The system does not have a well defined relative degree!

Why? Because the output y(t) = x1 (t) = et x1 (0) is independent of the input u.

11
7. Feedback Linearization

Examples

Example 3:
bm sm + bm−1 sm−1 + · · · + b0
H(s) =
sn + an−1 sn−1 + · · · + a0
where m < n and bm 6= 0.
A state model of the system is the following

ẋ = Ax + Bu
y = Cx

with
2 0 1 0 ··· 0
3 203
0 0 1 0 ··· :
6 . 6.7
.. ..
7
A=6 B=6 .7 C = [ b0 b1 ··· bm 0 ···0 ]1×n
4 .. 4.5
7
. . 0 5
0 ··· ··· 0 1
−a0 −a1 ··· ··· −an−1 0
n×n 1 n×1

What is the relative degree ρ ?

12
7. Feedback Linearization

Examples

ẏ = CAx + CBu
If m = n − 1 −→ CB = bm 6= 0 −→ ρ = 1

Otherwise, CB = 0
y (2) = CA2 x + CABu
Note that CA is obtained by shifting the elements of C one position to the right and
CAi by shifting i positions.

Therefore,

CAi−1 B = 0, for i = 1, 2, . . . n − m − 1
n−m−1
CA B = bm 6= 0

y (n−m) = CAn−m x + CAn−m−1 Bu −→ ρ = n − m

In this case the relative degree of the system is the difference between the degrees of
the denominator and numerator polynomials of H(s).

13
7. Feedback Linearization

Consider again the linear system given by the transfer function

8
N (s) < deg D = n
H(s) = with deg N = m<n
D(s) :
ρ = n−m

D(s) can be written as

D(s) = Q(s)N (s) + R(s)
where the degree of the quotient deg Q = n − m = ρ and the degree of the reminder
deg R < m
Thus
1
N (s) Q(s)
H(s) = =
Q(s)N (s) + R(s) 1 R(s)
1+ Q(s) N (s)

and therefore we can conclude that H(s) can be represented as a negative feedback
connection with 1/Q(s) in the forward path and R(s)/N (s) in the feedback path.

14
7. Feedback Linearization

Note that the ρ-order transfer function 1/Q(s) has no zeros and can be realized by

ξ̇ = (Ac + Bc λT )ξ + Bc bm e
y = Cc ξ

where ˜T
∈ Rρ
ˆ
ξ= y ẏ ... y (ρ−1)
and (Ac , Bc , Cc ) is a canonical form representation of a chain of ρ integrators:

0
2 3 2 3
0 1 0 ··· 0
60 0 1 ··· 07 607
6 7 6 7
6. .. .. 7 6.7
6. . 6.7 ˆ ˜
Ac = 6 . .7 Bc = 6 . 7 Cc = 1 0 ··· 0 0 1×ρ
7
6 ..
6 7 6 7
7 6 7
4. 0 15 405
0 ··· ··· 0 0 ρ×ρ 1 ρ×1

0 ··· 0
2 3
6. .. 7
6.
Bc λT = 6. .77, λ ∈ Rρ
40 ··· 05
λT

15
7. Feedback Linearization

R(s) η̇ = A0 η + B0 y
−→
N (s) w = C0 η

The eigenvalues of A0 are the zeros of the polynomial N (s), which are the zeros of
the transfer function H(s).

Thus, the system H(s) can be realized by the state model

η̇ = A0 η + B0 Cc ξ
ξ̇ = Ac ξ + Bc (λT ξ − bm C0 η + bm u)
y = Cc ξ

Note that y = Cc ξ and

ξ̇1 = ξ2
ξ̇2 = ξ3
ξ̇ = Ac ξ + Bc (λT ξ − bm C0 η + bm u) ←→ ..
.
ξ̇ρ = λT ξ − bm C0 η + bm u

and therefore y (ρ) = λT ξ − bm C0 η + bm u

16
7. Feedback Linearization

y (ρ) = λT ξ − bm C0 η + bm u
Thus, setting
1
u= [−λT ξ + bm C0 η + v]
bm
results in
η̇ = A0 η + B0 Cc ξ −→ Internal dynamics: It is unobservable from the output y
ξ̇ = Ac ξ + Bc v −→ chain of integrators
y = Cc ξ

Suppose we would like to stabilize the output y at a constant reference r, that is,
ξ → ξ ? = (r, 0, . . . , 0)T .
Defining ζ = ξ − ξ ? we obtain
ζ̇ = Ac ζ + Bc v
Therefore, setting
v = −Kζ = −K(ξ − ξ ? )
with (Ac − Bc K) Hurwitz we obtain the closed-loop system
η̇ = A0 η + B0 Cc (ξ ? + ζ)
ζ̇ = (Ac − Bc K)ζ
y = Cc ξ
where the eigenvalues of A0 are the zeros of H(s). If H(s) is minmum phase (zeros in
the open left-half complex plan) then A0 is Hurwitz.
17
7. Feedback Linearization

Feedback Linearization
Can we extend this result
η̇ = A0 η + B0 `Cc η
ξ̇ = Ac ξ + Bc λT ξ − bm C0 η + bm u
´
y = Cc x
for the nonlinear system (SISO)
ẋ = f (x) + g(x)u
y = h(x)
that is find a z = T (x), where
φ1 (x)
2 3
6 .. 7
6
6 . 7
7
» –
η
6 φn−ρ (x) 7
z= =6
6 7
ξ 6 h(x) 7
7
6
6 .. 7
7
4 . 5
Lρ−1
f h(x)
∂φi
such that T (x) is a diffeomorphism on D0 ⊂ D and ∂x
g(x) = 0, for
1 ≤ i ≤ n − ρ, ∀x ∈ D. Note that
∂φi ∂φi ∂φi
η̇ = ẋ = f (x) + g(x)u
∂x ∂x ∂x
Does exist such T (x)?
18
7. Feedback Linearization

Normal form

Theorem (13.1)
Consider the SISO system
ẋ = f (x) + g(x)u
y = h(x)
and suppose that it has relative degree ρ ≤ n in D. Then, for every x0 ∈ D, there
exists a such diffeomorphism T (x) on a neighborhood of x0 .

Using this transformation we obtain the system re-written in normal form:

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

where ξ ∈ Rρ , η ∈ Rn−ρ and (Ac , Bc , Cc ) is the canonical form representation of a

chain of integrators, and

Lρf h(x)
˛
∂φ
γ(x) = Lg Lρ−1
˛
f0 (η, ξ) := f (x)˛˛ f h(x) α(x) = −
∂x x=T −1 (z) Lg Lfρ−1 h(x)

19
7. Feedback Linearization

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

The external part can be linearized by

u = α(x) + β(x)v

with β(x) = γ −1 (x). The internal part is described by

η̇ = f0 (η, ξ)

Setting ξ = 0 result

η̇ = f0 (η, 0) −→ This is called the zero-dynamics

Note that for the linear case we have η̇ = A0 η, where the eigenvalues of A0 are the
zeros of H(s).

20
7. Feedback Linearization

Definition
The system is said to be minimum phase if η̇ = f0 (η, 0) has an asymptotically stable
equilibrium point in the domain of interest.

The zero dynamics can be characterized in the original coordinates by notting that

y(t) = 0, ∀t ≥ 0 ⇒ ξ(t) = 0 ⇒ u(t) = α(x(t))

where the first implication is due to the fact that ξ = [y, ẏ, ...]T and the second due
to ξ̇ = A0 ξ + B0 γ(x)[u − α(x)].

Thus, when y(t) = 0, the solution of the state equation is confined to the set
n o
Z ∗ = x ∈ D0 : h(x) = Lf h(x) = ... = Lρ−1 f h(x) = 0

and the input

u = u∗ (x) := α(x)|x∈Z ∗
that is
ẋ = f ∗ (x) := [f (x) + g(x)α(x)]x∈Z ∗

In the special case that ρ = n ⇒ η does not exist. In that case the system has no zero
dynamics and by default is said to be minimum phase.

21
7. Feedback Linearization

Example
Example 1

ẋ1 = x2
ẋ2 = −x1 + (1 − x21 )x2 + u
y = x2

It is in the normal form (ξ = y, η = x1 )

Zero-dynamics?
ẋ1 = 0, which does not have an asymptotic stable equilibrium point. Hence, the
system is not minimum phase.

Example 2
2 + x3
ẋ1 = −x1 + u
1 + x23
ẋ2 = x3
ẋ3 = x2 x3 + u
y = x2
What is the relative degree and the zero dynamics?

22
7. Feedback Linearization

2 + x3
ẋ1 = −x1 + u
1 + x23
ẋ2 = x3
ẋ3 = x2 x3 + u
y = x2

Computing the time-derivative...

ẏ = ẋ2 = x3
ÿ = ẋ3 = x1 x3 + u

Thus, the relative degree is ρ = 2. Analyzing the zero-dynamics we have

y=0
ẏ = 0
ÿ = 0

we have x2 = x3 = 0 and from the last we have u = −x1 x3 = 0. Therefore,

ẋ1 = −x1 and the system is minimum phase.

23
7. Feedback Linearization

Full-State Linearization
The single-input system
ẋ = f (x) + g(x)u
with f , g sufficiently smooth in a domain D ⊂ Rn is feedback linearizable if there
exists a sufficiently smooth h : D → R such that the system

ẋ = f (x) + g(x)u
y = h(x)

has relative degree n in a region D0 ⊂ D.

This implies that the normal form reduces to

ż = Ac z + Bc γ(x)[u − α(x)]
y = Cc z

Note that
z = T (x)
Thus
∂T
ż = ẋ
∂x
which is equivalent to
∂T
Ac z + Bc γ(x)[u − α(x)] = [f (x) + g(x)u]
∂x

24
7. Feedback Linearization

Splinting in two we obtain

∂T
f (x) = Ac T (x) − Bc γ(x)α(x) (2)
∂x
∂T
g(x) = Bc γ(x) (3)
∂x

Equation (2) is equivalent to

∂T1
f (x) = T2 (x)
∂x
∂T2
f (x) = T3 (x)
∂x
..
.
∂Tn−1
f (x) = Tn (x)
∂x
∂Tn
f (x) = −α(x)γ(x)
∂x

25
7. Feedback Linearization

and (3) is equivalent to

∂T1
g(x) = 0
∂x
∂T2
g(x) = 0
∂x
..
.
∂Tn−1
g(x) = 0
∂x
∂Tn
g(x) = γ(x) 6= 0
∂x
Setting h(x) = T1 , we see that

Ti+1 (x) = Lf Ti (x) = Lif h(x), i = 1, 2, ..., n − 1

and
Lg Li−1
f h(x) = 0, i = 1, 2, ..., n − 1
(4)
Lg Ln−1
f 6= 0

Therefore we can conclude that if h(.) satisfies (4) the system is feedback linearizable.

26
7. Feedback Linearization

The existence of h(.) can be characterized by necessary and sufficient conditions on

the vector fields f and g. First we need some terminology.

Definition
Given two vector fields f and g on D ⊂ Rn , the Lie Bracket [f, g] is the vector field

∂g ∂f
[f, g](x) = f (x) − g(x)
∂x ∂x

Note that
[f, g] = −[g, f ]
f = g = cte ⇒ [f, g] = 0
Adjoint representation

ad0f g(x) = g(x)

ad1f g(x) = [f, g](x)
adkf g(x) = [f, adk−1
f g](x), k ≥ 1

27
7. Feedback Linearization

Example 1
» – » –
x2 0
f (x) = , g(x) =
− sin x1 − x2 x1
Then,

" ∂g ∂g1
# " ∂f ∂f1
#
1 1
∂g ∂f ∂x1 ∂x2 ∂x1 ∂x2
[f, g](x) = f (x) − g(x) = ∂g2 ∂g2 f (x) − ∂f2 ∂f2 g(x)
∂x ∂x ∂x1 ∂x2 ∂x1 ∂x2
» –» – » –» –
0 0 x2 0 1 0
= −
1 0 − sin x1 − x2 − cos x1 −1 x1
» – » – » –
0 x1 −x1
= − =
x2 −x1 x1 + x2

∂adf g ∂f
ad2f g = [f, adf g] = f (x) − adf g(x)
∂x ∂x
» –» – » –» –
−1 0 x2 0 1 −x1
= −
1 1 − sin x1 − x2 − cos x1 −1 x1 + x2
» –
−x1 − 2x2
=
x1 + x2 − sin x1 − x1 cos x1

28
7. Feedback Linearization

Example 2: f (x) = Ax and g(x) = g is a constant vector field.

Then,
∂g ∂f
adf g(x) = [f, g](x) = f (x) − g(x) = −Ag
∂x ∂x
∂adf g ∂f
ad2f g(x) = [f, adf g](x) = f− adf g = −A(−Ag) = A2 g
∂x ∂x

adkf g = (−1)k Ak g

Definition
For vector fields f1 , f2 , ..., fk on D ⊂ Rn , a distribution ∆ is a collection of all vector
spaces ∆(x) = span {f1 (x), ..., fk (x)}, where for each fixed x ∈ D, ∆(x) is the
subspace of Rn spanned by the vectors f1 (x), ..., fk (x).
The dimension of ∆(x) is defined by

dim(∆(x)) = rank[f1 (x), f2 (x), ..., fk (x)]

which may depend on x.

If f1 , f2 , ..., fk are linearly independent, then dim(∆(x)) = k, ∀x ∈ D. In this case,
we say that ∆ is a nonsingular distribution on D. A distribution ∆ is involutive if
g1 ∈ ∆, g2 ∈ ∆ ⇒ [g1 , g2 ] ∈ ∆.
If ∆ is a nonsingular distribution on D, generated by f1 , ..., fk then it is involutive if
and only if [fi , fj ] ∈ ∆, ∀1 ≤ i, j ≤ k
29
7. Feedback Linearization

Example 3
Let D = R3 , ∆ = span {f1 , f2 } and
2 3 2 3
2x2 1
f1 = 4 1 5, f2 = 4 0 5
0 x2
dim(∆(x)) = rank[f1 , f2 ] = 2, ∀x ∈ D

is ∆ involutive?

2 32 3 2 32 3 2 3
0 0 0 2x2 0 2 0 1 0
∂f2 ∂f1
[f1 , f2 ] = f1 − f2 = 4 0 0 0 5 4 1 5−4 0 0 0 54 0 5 = 4 0 5
∂x ∂x 0 1 0 0 0 0 0 x2 1

Checking that [f1 , f2 ] ∈ ∆ is the same to see if [f1 , f2 ] can be generated by f1 , f2 ,

that is if rank[f1 (x), f2 (x), [f1 , f2 ](x)] = 2, ∀x ∈ D. But
2 3
2x2 1 0
rank 4 1 0 0 5 = 3, ∀x ∈ D
0 x2 1

Hence, ∆ is not involutive.

30
7. Feedback Linearization

Theorem 13.2

Theorem
The system ẋ = f (x) + g(x)u, with x ∈ Rn , u ∈ R is feedback linearizable if and only
if there is a domain D0 ⊂ D such that
1. The matrix G(x) = [g(x), adf g(x), ..., adn−1
f g] has rank n ∀x ∈ D0 .
2. The distribution D = span{g, adf g(x), ..., adn−2
f g} is involutive in D0 .

Example » – » –
a sin x2 0
ẋ = f (x) + gu, f (x) = , g=
−x21 1
we have seen that
» –» – » –
∂g ∂f 0 a cos x2 0 −a cos x2
adf g = [f, g] = f− g=− =
∂x ∂x −2x1 0 1 0
» –
0 −a cos x2
The matrix G = [g, adf g] = has rank G = 2, ∀ cos x2 6= 0. The
1 0
= span {g} is involutive.
distribution D ˘ Thus, we can conclude that there exists a
T (x) in D0 = x ∈ R2 : cos x2 6= 0 that allow us to do feedback linearization.
¯

31
7. Feedback Linearization

Now we have to find h(x) that satisfies

∂h ∂(Lf h)
g = 0; g 6= 0; h(0) = 0
∂x ∂x
h i 0 –
»
∂h
∂h ∂h
∂x1 ∂x2 = =0
1 ∂x2
Thus, h(.) must be independent of x2 .

∂h h i » a sin x – ∂h
∂h 2
Lf h(x) = f (x) = 0 = a sin x2
∂x ∂x1 −x21 ∂x1
∂Lf h h “ ” “ ” i» 0 – ∂h
∂ ∂h ∂ ∂h
g= ∂x1 ∂x1
a sin x2 ∂x2 ∂x1
a sin x2 = a cos x2 6= 0
∂x 1 ∂x1

∂h ∂h
In conclusion, ∂x 6= 0 and ∂x = 0.
1 2
Examples of such h(x) include h(x) = x1 or h(x) = x1 + x31 . Given h(x) we can now
perform input-output linearization.

32
7. Feedback Linearization

Example 2
A single link manipulator with flexible points

ẋ = f (x) + gu
2 3 2 3
x2 0
6 −a sin x1 − b(x1 − x3 ) 7 6 0 7
f (x) = 6 7, g=6 7 a, b, c, d > 0
4 x4 5 4 0 5
c(x1 − x3 ) d

G = [g, adf g, ad2f g, ad3f g]

2 32 3 2 3
0 1 0 0 0 0
∂g ∂f 6 −a cos x1 − b 0 b 0 7
76 0
6 7 6 0 7
adf g = [f, g] = f− g =0−4
6 7=6
5 4 −d
7
∂x ∂x 0 0 0 1 54 0 5
c 0 −c 0 d 0
∂(adf g) ∂f
ad2f g = [f, adf g] = f − adf g
∂x ∂x
2 32 3 2 3
0 1 0 0 0 0
6 −a cos x1 − b 0 b 0 7 6 0 7 6 bd 7
=0−4
6 76 7=6
5 4 −d 5 4 0
7
0 0 0 1 5
c 0 −c 0 0 −cd

33
7. Feedback Linearization

∂ad2f g ∂f 2
ad3f g = [f, ad2f g] = f− ad g
∂x ∂x f
2 32 3 2 3
0 1 0 0 0 −bd
6 −a cos x1 − b 0 b 0 7
7 6 bd
6 7 6 0 7
=0−6 7=6 7
4 0 0 0 1 54 0 5 4 cd 5
c 0 −c 0 −cd 0
Thus,
2 3
0 0 0 −bd
6 0 0 bd 0 7
G=4
6 7 → rank G = 4, ∀x ∈ R4
0−d 0 cd 5
d 0 −cd 0
n o
The distribution ∆ = span g, adf g, ad2f is involutive since g, adf g, ad2f are
constant vector fields ([g, adf g] = 0).
Therefore, we can conclude that there exists an h(x) : R4 → R and a T (x) that make
the system full state feedback linearizable. In particular, h(x) must satisfy

∂Li−1
f h
g = 0, i = 1, 2, 3.
∂x

∂L3f h
g 6= 0, h(0) = 0
∂x
34
7. Feedback Linearization

∂h ∂h
For i = 1, we have ∂x
g =0⇒ ∂x4
= 0 and so h(x) must be independent of x4 .

∂h ∂h ∂h ∂h
Lf h(x) = ∂x
f (x) = x
∂x1 2
+ ∂x2
[−a sin x1 − b(x1 − x3 )] + x .
∂x3 4
From
∂Lf h ∂Lf h ∂h
g=0⇒ =0⇒ =0
∂x ∂x4 ∂x3
Thus, h(x) is independent of x3 .
Thus,
∂h ∂h
Lf h(x) = x2 + [−a sin x1 − b(x1 − x3 )]
∂x1 ∂x2
.

∂ ( L h) ∂ ( L h) ∂ (L h)
L2f h(x) = Lf Lf h(x) = ∂xf f (x) = ∂xf x2 + ∂xf [−a sin x1 − b(x1 − x3 )]
1 2
∂ (L h) ∂ ( L h)
+ ∂xf x4 + ∂xf c(x1 − x3 )
3 4

For i = 2, “ ”
∂ L2f h
` ´
∂ Lf h ∂h
g=0⇒ =0⇒ =0
∂x ∂x3 ∂x2
Thus h is independent of x2 .

35
7. Feedback Linearization

Hence,
“ ” “ ” “ ”
∂ L2
fh ∂ L2
fh ∂ L2
fh
L3f h(x) = Lf L2f h = ∂x
f (x) = ∂x1
x2 + ∂x1
[−a sin x1 − b(x1 − x3 )]
“ ” “ ”
∂ L2
fh ∂ L2fh
+ ∂x x4 + ∂x4
c(x1 − x3 )
3

Also, from condition

∂L3f h(x) ∂L2f h ∂Lf h ∂h

g 6= 0 ⇒ 6= 0 ⇒ 6= 0 ⇒ 6= 0
∂x ∂x3 ∂x2 ∂x1

Therefore, let h(x) = x1 . Then, the change of variables

z1 = h(x) = x1
z2 = Lf h(x) = x2
z3 = L2f h(x) = −a sin x1 − b(x1 − x3 )
z4 = L3f h(x) = −a cos x1 ẋ1 − bẋ1 + bẋ3 = −ax2 cos x1 − bx2 + bx4

transforms the state equation into

ż1 = z2
ż2 = z3
ż3 = z4
ż4 = −(a cos z1 + b + c)z3 + a(z22 − c) sin z1 + bd u

36
7. Feedback Linearization

Example - field controlled DC motor

Consider the system

ẋ = f (x) + gu
with 2 3 23
−ax1 1
f (x) = 4 −bx2 + k − cx1 x3 5 , g= 4 0 5
θx1 x2 0
Computing the

a2
2 3 2 3
a
adf g = [f, g] = 4 cx3 5 , ad2f g = [f, adf g] = 4 (a + b) cx3 5
−θx2 (b − a) θx2 − θk

a2
2 3
1 a
G= [g, adf g, ad2f g] =4 0 cx3 (a + b) cx3 5
0 −θx2 (b − a) θx2 − θk
Rank of G?

37
7. Feedback Linearization

det G = cθ(−k + 2bx2 )x3 .

k
Hence G has rank 3 for x2 6= 2b
and x3 6= 0. Let’s check the distribution
D = span{g, adf g}
2 32 3 2 3
ˆ ˜ ∂(adf g) 0 0 0 1 0
g, adf g = g=4 0 0 c 54 0 5 = 4 0 5
∂x 0 −θ 0 0 0
Hence D is involutive because [g, adf g] ∈ D.

Therefore, the conditions of Theorem 13.2 are satisfied in particular for the domain
 ff
k
D0 = x ∈ R3 : x2 > , x3 > 0
2b

h(.) =?

38
7. Feedback Linearization

ẋ = f (x) + gu
with 2 3 32
−ax1 1
f (x) = 4 −bx2 + k − cx1 x3 5 , g=4 0 5
θx1 x2 0

k
Equilibrium points: x1 = 0, x2 = b
, x3 = cte.

k
Suppose we are interested in the desired operating point x∗ = [0 b
w0 ]T .

Then, h(x) must satisfy (n = 3)

∂h ∂(Lf h) ∂(L2f h)
g = 0; g = 0; g 6= 0
∂x ∂x ∂x
and h(x∗ ) = 0.

∂h ∂h
g=0⇒ =0
∂x ∂x1
that is, h must be independent of x1 .

Lf h(x) = ...

39
7. Feedback Linearization

State feedback control

Consider the system

ẋ = f (x) + G(x)u
and let z = T (x) = [T1 (x), T2 (x)]T such that

η̇ = f0 (η, ξ)
ξ̇ = Aξ + Bγ(x)[u − α(x)]

Suppose that (A, B) is controllable, γ(x) is nonsingular ∀x ∈ D, f0 (0, 0) = 0 and

f0 (η, ξ), α(x), γ(x) ∈ C 1 .

Goal: Design a state feedback control law to stabilize the origin z = 0.

Setting u = α(x) + β(x)v, β(x) = γ −1 (x), we obtain the triangular system

η̇ = f0 (η, ξ)
ξ̇ = Aξ + Bv

Let v = −Kξ with (A − BK) Hurwitz then we can conclude the following:

40
7. Feedback Linearization

State feedback control

Lemma (13.1)
The origin z = 0 is asymptotically stable (AS) if the origin of η̇ = f0 (η, 0) is AS (that
is, if the system is minimum phase).

Proof.
∂V1
By the converse theorem ∃V1 (η) : ∂η 0
f (η, 0) ≤ −α3 (kxk) in some neighborhood of
η = 0, where α3 ∈ K. Let P = PT
> 0 be the solution of the Lyapunov equation
P (A − BK) + (A − BK)T P = −I.
p
V = V1 + k ξ T P ξ, k > 0

∂V1 k
V̇ ≤ f0 (η, ξ) + p ξ t [P (A − BK) + (A − BK)T P ]ξ
∂η 2 ξT P ξ
∂V1 ∂V1 kξ T ξ
≤ f0 (η, 0) + [f0 (η, ξ) − f0 (η, 0)] − p
∂η ∂η 2 ξT P ξ
≤ −α3 (kηk) + k1 kξk − kk2 kξk, k1 , k2 > 0

where the last inequality follows from restricting the state to be in any bounded
neighborhood of the origin and using the continuous differentiability property of V1
and f0 . Thus, choosing k > k1 /k2 ensures that V̇ < 0, which follows that the origin
is AS.
41
7. Feedback Linearization

State feedback control

Lemma (13.2)
The origin z = 0 is GAS if η̇ = f0 (η, ξ) is ISS.

Note that if η̇ = f0 (η, 0) is GAS or GES does NOT imply ISS. But, if it is GES +
globally Lipschitz then it is ISS.
Otherwise, we have to prove ISS by further analysis.

Example:
η̇ = −η + η 2 ξ
ξ̇ = v
Zero dynamics: η̇ = −η −→ η = 0 is GES
but η̇ = −η + η 2 ξ is not ISS, e.g., if ξ(t) = 1 and η(0) ≥ 2 then η̇(t) ≥ 2, ∀t ≥ 0,
which implies that η grows unbounded.
However with v = −Kξ, K > 0 we achieve AS.

42
7. Feedback Linearization

To view this, let ν = ηξ, then

ν̇ = η ξ̇ + η̇ξ
= ηv − ηξ + η 2 ξ 2
= −Kηξ − ηξ + η 2 ξ 2
= −(1 + K)ν + υ 2 = −[(1 + K) − ν]ν 2

Thus, with ν(0) < 1 + K ⇒ υ → 0 and therefore we can also conclude that
∃T ≥ t0 : υ(t) ≤ 12 , ∀t ≥ T .

Consider now V = 1/2η 2 . Then

V̇ = η η̇
= −η 2 + η 3 ξ
= −η 2 (1 − ηξ) = −η 2 (1 − ν) < 0, ∀t ≥ T

Thus, η → 0 and note also that ξ̇ = −Kξ. Therefore, the control law v = −Kξ can
achieve semiglobal stabilization.

43
7. Feedback Linearization

One may think that we can assign the eigenvalues of (A − BK) to the left
half-complex plane to make ξ̇ = (A − BK)ξ decay to zero arbitrarily fast. BUT this
may have consequences: the zero-dynamics may go unstable! This is due to the
peaking phenomenon.

Example

η̇ = − 12 (1 + ξ2 )η 3
ξ̇1 = ξ2
ξ̇2 = v
» –
0 1
Setting v = −Kξ = −k2 ξ1 − 2kξ2 −→ A − BK =
−k2 −2k
and the eigenvalues are −k, −k. Note that

(1 + kt)e−kt te−kt
» –
e(A−BK)t =
−k2 te−kt (1 − kt)e−kt

which shows that as k → ∞, ξ(t) will decay to zero arbitrary fast.

However, the element (2, 1) reaches a maximum value k/e at t = k1 . There is a peak
of order k! Furthermore, the interaction of peaking with nonlinear growth could
destabilize the system.

44
7. Feedback Linearization

E.g., consider the initial conditions

η(0) = η0
ξ1 (0) = 1
ξ2 (0) = 0

Then,
ξ2 (t) = −k2 te−kt
and
1
η̇ = − (1 + ξ2 )η 3
2
1
= − (1 − k2 te−kt )η 3
2
In this case the solution η(t) is given by

η02
η 2 (t) = .
1 + η02 [t + (1 + kt)e−kt − 1]

which has a finite escape if η0 > 1.

45
7. Feedback Linearization

Tracking
η̇ = f0 (η, ξ)
ξ̇ = A0 ξ + B0 γ(x)[u − α(x)]
y = C0 ξ
Goal: Design u such that y asymptotically tracks a reference signal r(t). Assume that
r(t), ṙ(t), ..., r(ρ) are bounded and available on-line. Note that the reference signal
could be the output of a pre-filter.
Example: If ρ = 2, the pre-filter could be
2
wn
G(s) =
s2 + 2ξwn s + wn
Then
ẏ1 = y2
ẏ2 = −wn2 y − 2ξw y + w 2 y
1 n 2 n d
r = y1
Note that in this case ṙ = ẏ1 = y2 and r̈ = ẏ2 .

Consider a system with relative degree ρ. Let

r ξ1 − r
2 3 2 3
.. ..
r=4 5 , and e = 4 5=ξ−r
6 7 6 7
. .
r (ρ−1) ξρ − r (ρ−1)

46
7. Feedback Linearization

Then, the error system is given by

η̇ = f0 (η, e + r)
ė = Ac e + Bc (γ(x)[u − α(x)] − r(ρ) )
1
Setting u = α(x) + β(x)[v − r(ρ) ] with β = γ(x)
it follows that

η̇ = f0 (η, e + r)
ė = Ac e + Bc v

Thus, selecting v = −Ke with (Ac − Bc K) Hurwitz we can conclude that the states
of the closed-loop system
η̇ = f0 (η, e + r)
ė = (A0 − B0 K)e
are bounded if η̇ = f0 (η, e + R) is ISS and that e → 0 as t → ∞.

47