Appendix A

Lyapunov Stability

Lyapunov stability theory [1] plays a central role in systems theory and engineering.
An equilibrium point is stable if all solutions starting at nearby points stay nearby;
otherwise, it is unstable. It is asymptotically stable if all solutions starting at nearby
points not only stay nearby, but also tend to the equilibrium point as time approaches
infinity. These notions are made precise as follows.

Definition A.1 Consider the autonomous system [1]

ẋ = f (x), (A.1)

where f : D → Rn is a locally Lipschitz map from a domain D into Rn .

The equilibrium point x = 0 is
• Stable if, for each  > 0, there is δ = δ() > 0 such that
   
x(0) < δ ⇒ x(t) < , ∀t ≥ 0;

• Unstable if it is not stable;

• Asymptotically stable if it is stable and δ can be chosen such that
 
x(0) < δ ⇒ lim x(t) = 0.
t→∞

Theorem A.1 Let x = 0 be an equilibrium point for system (A.1) and D ⊂ Rn be a

domain containing x = 0. Let V : D → R be a continuously differentiable function
such that

V (0) = 0 and V (x) > 0 in D − {0}, (A.2a)

V̇ (x) ≤ 0 in D. (A.2b)

Then, x = 0 is stable. Moreover, if

V̇ (x) < 0 in D − {0}, (A.3)

then x = 0 is asymptotically stable.

H. Chen, B. Gao, Nonlinear Estimation and Control of Automotive Drivetrains, 197

DOI 10.1007/978-3-642-41572-2,
© Science Press Beijing and Springer-Verlag Berlin Heidelberg 2014
198 A Lyapunov Stability

A function V (x) satisfying (A.2a) is said to be positive definite. A continuously

differentiable function V (x) satisfying (A.2a) and (A.2b) is called a Lyapunov func-
tion. A function V (x) satisfying

V (x) → ∞ as x → ∞ (A.4)

is said to be radially unbounded.

Theorem A.2 Let x = 0 be an equilibrium point for system (A.1). Let V : Rn → R

be a continuously differentiable, radially unbounded, positive definite function such
that
V̇ (x) < 0, ∀x = 0. (A.5)
Then x = 0 is globally asymptotically stable.

In some cases of V̇ (x) ≤ 0, asymptotical stability of x = 0 can be proved if we

can establish that no system trajectory stays forever at points where V̇ (x) = 0 except
at x = 0. This follows from LaSalle’s invariance principle. Before stating LaSalle’s
invariance principle, we give the notation of invariance.
A set M is said to be an invariant set with respect to system (A.1) if

x(0) ∈ M ⇒ x(t) ∈ M, ∀t ∈ R. (A.6)

That is, if a solution of (A.1) belongs M at some time instant, it belongs to M for
all future and past time.
A set M is said to be a positively invariant set with respect to system (A.1) if

x(0) ∈ M ⇒ x(t) ∈ M, ∀t ≥ 0. (A.7)

That is, any trajectory of (A.1) starting from M stays in M for all future time.

Theorem A.3 Let D be a compact (closed and bounded) set with the property that
every trajectory of system (A.1) starting from D remains in D for all future time.
Let V : D → R be a continuously differentiable positive definite function such that

V̇ (x) ≤ 0 in D. (A.8)

Let E be the set of all points in D where V̇ (x) = 0 and M be the largest invariant
set in E. Then every trajectory of (A.1) starting from D approaches M as t → ∞.

Corollary A.1 Let x = 0 be an equilibrium point for system (A.1) and D ⊂ Rn be

a domain containing x = 0. Let V : D → R be a continuously differentiable positive
definite function such that
V̇ (x) ≤ 0 in D. (A.9)
Let S = {x ∈ D | V̇ (x) = 0}, and suppose that no solution other than the trivial
solution can forever stay in S. Then x = 0 is asymptotically stable.
A Lyapunov Stability 199

Corollary A.2 Let x = 0 be an equilibrium point for system (A.1). Let V : Rn → R

be a continuously differentiable, radially unbounded, positive definite function such
that
V̇ (x) ≤ 0 for all x ∈ Rn . (A.10)
Let S = {x ∈ Rn | V̇ (x) = 0}, and suppose that no solution other than the trivial
solution can forever stay in S. Then x = 0 is globally asymptotically stable.

If V̇ (x) is negative definite, S = {0}. Then, Corollaries A.1 and A.2 coincide with
Theorems A.1 and A.2.

Example A.1 Consider the pendulum equation with friction as follows:

ẋ1 = x2 , (A.11a)
g k
ẋ2 = − sin x1 − x2 . (A.11b)
l m
We respectively apply Corollary A.1 and Theorem A.1 to discuss the stability prop-
erties of this system.
Let us first choose the energy as a Lyapunov function candidate, namely
g 1
V (x) = (1 − cos x1 ) + x22 (A.12)
l 2
which satisfies V (0) = 0 and V (x) > 0 in D − {0} with D = {x ∈ R2 | − π < x1 <
π}. Differentiating V (x) along the trajectories of the system leads to
g
V̇ (x) = ẋ1 sin x1 + x2 ẋ2
l
 
g g k
= x2 sin x1 + x2 − sin x1 − x2
l l m
k 2
=− x ,
m 2
which implies that V̇ (x) ≤ 0 in D. Let S be the set in D which contains all states
where V̇ (x) = 0 is maintained, i.e., S = {x ∈ D | V̇ (x) = 0}. For the pendulum
system, we infer from V̇ (x) ≡ 0 that

x2 (t) ≡ 0 ⇒ ẋ1 (t) ≡ 0 ⇒ x1 (t) ≡ constant (A.13)

and
x2 (t) ≡ 0 ⇒ ẋ2 (t) ≡ 0 ⇒ sin x1 = 0. (A.14)
The only point on the segment −π < x1 < π rendering sin x1 = 0 is x1 = 0. Hence,
no trajectory of the pendulum system (A.11a), (A.11b) other than the trivial solu-
tion can forever stay in S, i.e., S = {0}. Then, x = 0 is asymptotically stable by
Corollary A.1.
200 A Lyapunov Stability

We can also show the stability property by choosing other Lyapunov function
candidates. One possibility is to replace the term 12 x22 in (A.12) by the quadratic
form 12 x T P x for some positive definite matrix P . That is, we choose

g 1
V (x) = (1 − cos x1 ) + x T P x (A.15)
l 2
as a Lyapunov function candidate, where
 
p11 p12
P=
p12 p22

is to be determined. In order to ensure that 12 x T P x is positive definite, the elements

of P have to satisfy

p11 > 0, p22 > 0, p11 p22 − p12

2
> 0. (A.16)

Differentiating V (x) along the trajectories of the system leads to

g
V̇ (x) = ẋ1 sin x1 + x T P ẋ
l
g g k
= x2 sin x1 + p11 x1 x2 − p12 x1 sin x1 − p12 x1 x2
l l m
g k
+ p12 x22 − p22 x2 sin x1 − p22 x22
l m
 
g g k
= (1 − p22 )x2 sin x1 − p12 x1 sin x1 + p11 − p12 x1 x2
l l m
 
k
+ p12 − p22 x22 .
m

Now we can choose p11 , p12 and p22 to ensure V̇ (x) is negative definite. First of
all, we take
k
p22 = 1, p11 = p12 (A.17)
m
to cancel the cross-product terms x2 sin x1 and x1 x2 and arrive at
 
g k
V̇ (x) = − p12 x1 sin x1 + p12 − p22 x22 . (A.18)
l m

By combining (A.16) and (A.17), we have

k
0 < p12 < . (A.19)
m
References 201

Let us take p12 = 0.5 mk . Then, we have

gk k 2
V̇ (x) = − x1 sin x1 − x . (A.20)
2lm 2m 2
The term x1 sin x1 > 0 for all −π < x1 < π . Defining D = {x ∈ R2 | − π < x1 < π},
we can conclude that, with the chosen P , V (x) is positive definite and V̇ (x) is
negative definite on D. Hence, x = 0 is asymptotically stable by Theorem A.1.

Results for autonomous systems can be extended to non-autonomous systems [1].

Theorem A.4 Consider the non-autonomous system

ẋ = f (t, x), (A.21)

where f : [0, ∞) × D → Rn is piecewise continuous in t and locally Lipschitz in x

on [0, ∞) × D, and D ⊂ Rn is a domain that contains the origin x = 0.
Let x = 0 be an equilibrium point for (A.21) and D ⊂ Rn be a domain containing
x = 0. Let V : [0, ∞) × D → R be a continuously differentiable function such that

W1 (x) ≤ V (t, x) ≤ W2 (x), (A.22a)

∂V ∂V
+ f (t, x) ≤ 0 (A.22b)
∂t ∂x
for all t ≥ 0 and for all x ∈ D, where W1 (x) and W2 (x) are continuous positive
definite functions on D. Then, x = 0 is uniformly stable.

References
1. Khalil HK (2002) Nonlinear Systems. Prentice Hall, New York
Appendix B
Input-to-State Stability (ISS)

For a linear time-invariant system

ẋ = Ax + Bw (B.1)

with a Hurwitz matrix A, we can write the solution as

 t
x(t) = e(t−t0 )A x(t0 ) + e(t−τ )A Bw(τ ) dτ (B.2)
t0

and have
   
x(t) ≤ β(t)x(t0 ) + γ w ∞, (B.3)
where

  ∞ 
β(t) = eA(t−t0 )  → 0 and γ = B eA(s−t0 )  ds < ∞.
t0

Moreover, we use the bound e(t−t0 )A ≤ ke−λ(t−t0 ) to estimate the solution by

   
x(t) ≤ ke−λ(t−t0 ) x(t0 ) + k B w ∞, (B.4)
λ
where λ could be given by λ := min{|Re{eig(A)}|}. Since λ > 0, the above esti-
mate shows that the zero-input response decays exponentially, while the zero-state
response is bounded for every bounded input, that is, has a bounded-input–bounded-
state property.
For general nonlinear systems, however, it should not be surprising that these
properties may not hold [1]. ISS (Input-to-state stability) is a notation of stability
of nonlinear systems, which is suggested by Eduardo Sontag [2, 3] and merges two
different views of stability, namely state space approach usually associated with the
name of Lyapunov and the operator approach of which George Zames is one of the
main initiators [4]. The Lyapunov concept addresses the stability property of sys-
tems without external inputs, while the operator concept studies the I/O properties

H. Chen, B. Gao, Nonlinear Estimation and Control of Automotive Drivetrains, 203

DOI 10.1007/978-3-642-41572-2,
© Science Press Beijing and Springer-Verlag Berlin Heidelberg 2014
204 B Input-to-State Stability (ISS)

of systems under different external input signals including L2 and L∞ and provides
elegant results for linear systems.

B.1 Comparison Functions

Definition B.1 [1] A continuous function α : [0, a) → [0, ∞) 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 a = ∞, and α(r) → ∞ when r → ∞.

Definition B.2 A continuous function β : [0, a) × [0, ∞) → [0, ∞) is said to be-

long 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 → ∞.

Using these comparison functions, we can restate the stability definition in Ap-
pendix A in a more precise fashion [1].

Definition B.3 The equilibrium point x = 0 of (A.21) is

• Uniformly stable if there exists a class K function γ (·) and a positive constant δ,
independent of t0 , such that
   
x(t) ≤ γ x(t0 ) , ∀t ≥ t0 ≥ 0, ∀x(t0 ) ∈ {x| |x| < δ}; (B.5)

• Uniformly asymptotically stable if there exists a class KL function β(·, ·) and a

positive constant δ, independent of t0 , such that
    
x(t) ≤ β x(t0 ), t − t0 , ∀t ≥ t0 ≥ 0, ∀x(t0 ) ∈ {x| |x| < δ}; (B.6)

• Exponentially stable if (B.6) is satisfied with β(r, s) = kre−αs , k > 0, α > 0;

• Globally uniformly stable, if (B.5) is satisfied with γ ∈ K∞ for any initial state
x(t0 );
• Globally uniformly asymptotically stable if (B.6) is satisfied with β ∈ KL∞ for
any initial state x(t0 );
• Globally exponentially stable if (B.6) is satisfied with β(r, s) = kre−αs , k > 0,
α > 0 for any initial state x(t0 ).

Some properties of comparison functions are summarized as follows [1, 4]:

• If V : Rn → R is continuous, define two functions
   
α(r) := min V (x), α(r) := max V (x).
|x|≥r |x|≤r

Then, they are of class K∞ and

     
α |x| ≤ V (x) ≤ α |x| , ∀x ∈ Rn ;
B.2 Input-to-State Stability 205

• Suppose α, γ ∈ K∞ , β ∈ KL, then

– α −1 (β(·, ·)) ∈ KL;
– α −1 (γ (·)) ∈ K∞ ;
– sups∈[0,t] γ (|u(s)|) = γ ( u[0,t] ∞ ) := γ (sups∈[0,t] (|u(s)|)) due to the mono-
tonic increase of γ ;
– γ (a + b) ≤ γ (2a) + γ (2b) for a, b ∈ R≥0 ;
– If α(r) ≤ β(s, t) + γ (t), then r ≤ α −1 (β(s, t) + γ (t)) ≤ α −1 (2β(s, t)) +
α −1 (2γ (t)).

B.2 Input-to-State Stability

Consider the nonlinear system

ẋ = f (x, w), (B.7)

where x ∈ Rn and w ∈ Rm are the state and the external input, respectively, and f
is locally Lipschitz in x and w. One has the following definitions [4].

Definition B.4 The system (B.7) is said to be input-to-state stable (ISS) if there
exist a class KL function β and a class K function γ such that, for any x(0) and for
any input w(·) continuous and bounded on [0, ∞), the solution of (B.7) satisfies
    
x(t) ≤ β x(0), t + γ w ∞ , ∀t ≥ 0. (B.8)

Definition B.5 The system (B.7) is said to be integral-input-to-state stable (iISS) if

there exist a class KL function β and class K∞ functions α and γ such that, for any
x(0) and for any input w(·) continuous and bounded on [0, ∞), the solution of (B.7)
satisfies
 t
     
α x(t) ≤ β |x0 |, t + γ w(s) ds, t ≥ 0. (B.9)
0

The following results are stated in [4].

Theorem B.1 For the system (B.7), the following properties are equivalent:
• It is ISS;
• There exists a smooth positive definite function V : Rn → R+ such that for all
x ∈ Rn and w ∈ Rm ,
   
α1 |x| ≤ V (x) ≤ α2 |x| , (B.10a)
  ∂V  
|x| ≥ γ |w| ⇒ f (x, w) ≤ −α3 |x| , (B.10b)
∂x
where α1 , α2 , and γ are class K∞ functions and α3 is a class K function;
206 B Input-to-State Stability (ISS)

Fig. B.1 Cascade of systems

• There exists a smooth positive definite radially unbounded function V and class
K∞ functions α and γ such that the following dissipation inequality is satisfied
∂V    
f (x, w) ≤ −α |x| + γ |w| . (B.11)
∂x

A continuously differentiable function V (x) satisfying (B.11) is called an ISS-

Lyapunov function.

Theorem B.2 For the system (B.7), the following properties are equivalent:
• It is iISS;
• There exists a smooth positive definite radially unbounded function V : Rn →
R+ , a class K∞ function γ and a positive definite function α : R+ → R+ such
that the following inequality is satisfied:
∂V    
f (x, w) ≤ −α |x| + γ |w| . (B.12)
∂x

A continuously differentiable function V (x) satisfying (B.12) is called an iISS-

Lyapunov function.

Theorem B.3 The cascaded system in the form of

ẋ = f (x, z),
ż = g(z, w),

shown in Fig. B.1, is ISS with input w, if the x-subsystem is ISS with z being viewed
as input and the z-subsystem is ISS with input w.

Theorem B.4 For the cascaded system in the form of

ẋ = f (x, z),
ż = g(z),

one has the following:

• It is globally asymptotically stable if the x-subsystem is ISS with z being viewed
as input and the z-subsystem is globally asymptotically stable;
• It is globally asymptotically stable if the x-subsystem is affine in z and iISS with z
being viewed as input, and the z-subsystem is globally asymptotically stable and
locally exponentially stable.
B.2 Input-to-State Stability 207

Fig. B.2 Uncertain system

Some of equivalences for ISS are listed as follows:

• (Nonlinear superposition principle) A system is ISS if and only if it is zero-input
stable and satisfies the asymptotic gain property. A system satisfies the asymptotic
gain property if there is some γ ∈ K∞ such that
   
lim supx(t) ≤ γ w ∞ , ∀x(0), w(·), t ≥ 0;
t→∞

• A system is ISS if and only if it is robustly stable, where the system is perturbed
by the uncertainty as shown in Fig. B.2;
• A system is ISS if and only if it is dissipative with the supply function of
s(x, w) = γ (|w|) − α(|x|), i.e., the following dissipation inequality
 t2
     
V x(t2 ) − V x(t1 ) ≤ s x(τ ), w(τ ) dτ
t1

holds along all trajectories of the system and for some α, γ ∈ K∞ ;

• A system is ISS if and only if it satisfies the following L2 → L2 estimate
 t  t
     
α1 x(τ ) dτ ≤ α0 x(0) + γ w(τ ) dτ
0 0

along all trajectories of the system and for some α0 , α1 , γ ∈ K∞ .

B.2.1 Useful Lemmas

Lemma B.1 (Young’s Inequality [p. 75, 1]) If the constants p > 1 and q > 1 are
such that (p − 1)(q − 1) = 1, then for all  > 0 and all (x, y) ∈ R2 we have

p p 1
xy ≤ |x| + q |y|q .
p q

Choosing p = q = 2 and  2 = 2κ, the above inequality becomes

1 2
xy ≤ κx 2 + y .
4κ
208 B Input-to-State Stability (ISS)

Lemma B.2 ([pp. 495, 505, 1]) Let v and ρ be real-valued functions defined on
R+ , and let b and c be positive constants. If they satisfy the differential inequality

v̇ ≤ −cv + bρ(t)2 , v(0) ≥ 0 (B.13)

then
(a) The following integral inequality holds:
 t
v(t) ≤ v(0)e−ct + b e−c(t−τ ) ρ(τ )2 dτ ; (B.14)
0

(b) If, in addition, ρ ∈ L2 , then v ∈ L1 and

1 
v 1 ≤ v(0) + b ρ 22 ; (B.15)
c
(c) If ρ ∈ L∞ , then v ∈ L∞ and
b
v(t) ≤ v(0)e−ct + ρ ∞;
2
(B.16)
c
(d) If ρ ∈ L2 , then v ∈ L∞ and

v(t) ≤ v(0)e−ct + b ρ 22 . (B.17)

References
1. Khalil HK (2002) Nonlinear Systems. Prentice Hall, New York
2. Sontag ED (1989) Smooth stabilization implies coprime factorization. IEEE Trans Autom Con-
trol 34:435–443
3. Sontag ED, Wang Y (1996) New characterizations of input-to-state stability. IEEE Trans Autom
Control 41:1283–1294
4. Sontag ED (2008) Input to state stability: basic concepts and results. In: Cachan JM, Gronin-
gen FT, Paris BT (eds) Nonlinear and optimal control theory. Lecture notes in mathematics.
Springer, Berlin, pp 163–220
Appendix C
Backstepping

In control theory, backstepping [1–3] is a technique for designing stabilizing con-

trollers for a special class of nonlinear dynamical systems. The designer can start
the design process at the known-stable system and “back out” new controllers that
progressively stabilize each outer subsystem. The process terminates when the final
external control is reached. Hence, this process is known as backstepping. Because
in each step, Control Lyapunov Function (CLF) is constructed to obtain the virtual
control, the control law obtained by backstepping is globally asymptotically stable.

C.1 About CLF

Definition C.1 For a time-invariant nonlinear system [1]

ẋ = f (x, u), x ∈ Rn , u ∈ R, f (0, 0) = 0, (C.1)

a smooth positive definite and radially unbounded function V : Rn → R+ is called

a control Lyapunov function (CLF) if
 
∂V
inf (x)f (x, u) < 0, ∀x = 0. (C.2)
u∈R ∂x

For systems affine in the control,

ẋ = f (x) + g(x)u, f (0) = 0, (C.3)

the CLF inequality becomes

∂V ∂V
f (x) + g(x)α(x) ≤ −W (x), (C.4)
∂x ∂x
where α(x) is the control law designed for u, and W : Rn → R is positive definite
(or positive semi-definite, in this case one needs to apply Theorem A.3 to discuss
stability).

H. Chen, B. Gao, Nonlinear Estimation and Control of Automotive Drivetrains, 209

DOI 10.1007/978-3-642-41572-2,
© Science Press Beijing and Springer-Verlag Berlin Heidelberg 2014
210 C Backstepping

If V (x) is a CLF for (C.3), then a stabilizing control law α(x), smooth for all
x = 0, is given by Sontag’s formula [1, 4]
⎧ ∂V
⎨ ∂x f + ( ∂V
∂x f ) +( ∂x g)
2 ∂V 4
− ∂x g = 0,
, ∂V
α(x) = ∂V
∂x g (C.5)
⎩
0, ∂V
∂x f = 0,

witch results in
 2  4
∂V ∂V
W (x) = f + g > 0, ∀x = 0.
∂x ∂x

C.2 Backstepping Design

We first use an example to introduce the idea of backstepping.

Example C.1 Consider the system [1]

ẋ1 = x12 − x13 + x2 , (C.6a)
ẋ2 = u. (C.6b)

We start with the first equation ẋ1 = x12 − x13 + x2 , where x2 is viewed as virtual
control, and proceed to design the feedback control x2 = α(x1 ) to stabilize the origin
x1 = 0. With
x2 = −x12 − x1 (C.7)
we cancel the nonlinear term x12 to obtain

ẋ1 = −x1 − x13 (C.8)

and infer V (x1 ) = 12 x12 satisfying

V̇ = −x12 − x14 , ∀x1 ∈ R (C.9)

along the solution of (C.8). This implies that V̇ is negative definite. Hence, the origin
of (C.8) is globally exponentially stable.
Indeed, x2 is one of system states and x2 = α(x1 ). To backstep, we define the
error between x2 and the desired value α(x1 ) as follows:
z2 = x2 − α(x1 ) = x2 + x1 + x12 , (C.10)

which serves as a change of variables to transform the system into

ẋ1 = −x1 − x13 + z2 , (C.11a)
 
ż2 = u + (1 + 2x1 ) −x1 − x13 + z2 . (C.11b)
C.2 Backstepping Design 211

Taking
1 1
Vc (x) = x12 + z22 (C.12)
2 2
as a composite Lyapunov function, we obtain for the transformed system (C.11a),
(C.11b) that satisfies
   
V̇c = −x12 − x14 + z2 x1 + (1 + 2x1 ) −x1 − x13 + z2 + u . (C.13)

Choosing
 
u = −x1 − (1 + 2x1 ) −x1 − x13 + z2 − z2 (C.14)
yields
V̇c = −x12 − x14 − z22 , (C.15)
which is negative definite. Hence, the origin of the transformed system (C.11a),
(C.11b), and hence the original system (C.6a), (C.6b), is globally stable.

Now we consider the system having the following strict feedback form [1, 5]:

ẋ = f (x) + g(x)ξ1 , (C.16a)

ξ̇1 = f1 (x, ξ1 ) + g1 (x, ξ1 )ξ2 , (C.16b)
ξ̇2 = f2 (x, ξ1 , ξ2 ) + g2 (x, ξ1 , ξ2 )ξ3 , (C.16c)
..
. (C.16d)
ξ̇k = fk (x, ξ1 , . . . , ξk ) + gk (x, ξ1 , . . . , ξk )u, (C.16e)

where x ∈ Rn and ξ1 , . . . , ξk , which are scalars, are the state variables. Many phys-
ical systems can be represented as a strict feedback system, such as the system
described in Chap. 4.
The whole design process is based on the following assumption:

Assumption C.1 Consider the system

ẋ = f (x) + g(x)u, f (0) = 0, (C.17)

where x ∈ Rn is the state and u ∈ R is the scalar control input. There exists a con-
tinuously differentiable feedback control law

u = α(x), α(0) = 0, (C.18)

and a smooth, positive definite, radially unbounded function V : Rn → R such that

∂V  
(x) f (x) + g(x)α(x) ≤ −W (x) (C.19)
∂x
212 C Backstepping

with W : Rn → R positive definite (or positive semi-definite, in this case one needs
to apply Theorem A.3 to discuss stability).

A special case is when the x-state has dimension 1, i.e., n = 1. Then, by con-
structing a Lyapunov function
1
V (x) = x 2 (C.20)
2
for (C.16a), where ξ1 is regarded as the virtual control, the control law α(x) is
determined to satisfy (C.4), i.e.,
 
x f (x) + g(x)α(x) ≤ −W (x) (C.21)

with W : Rn → R positive definite (or positive semi-definite). Hence, one choice of

α(x) is

− W (x)+xf (x)
, x = 0,
α(x) = xg(x) (C.22)
0, x = 0.

More specially, one can choose W (x) = k1 x 2 with k1 > 0 for simplicity to get

− k1 x+f (x)
g(x) , x = 0,
α(x) = (C.23)
0, x = 0,

if g(x) = 0 for all x. Note that the x-subsystem is uncontrollable at the points of
g(x) = 0.
In the following, we assume that Assumption C.1 is satisfied in general.
Since ξ1 is just a state variable and not the control, we define e1 as the deviation
of ξ1 from its desired value α(x):

e1 = ξ1 − α(x) (C.24)

and infer
∂V ∂V   ∂V
V̇ = ẋ = f (x) + g(x)α(x) + g(x)e1 ≤ −W (x) + g(x)e1 . (C.25)
∂x ∂x ∂x
Then, the second step begins, and the second Lyapunov function is defined as

1
V1 (x, ξ1 ) = V (x) + e12 . (C.26)
2
Let the desired value of ξ2 be α1 (x, ξ1 ), and introduce the second error

e2 = ξ2 − α1 (x, ξ1 ). (C.27)

Then we have
C.2 Backstepping Design 213

Fig. C.1 Backstepping

design procedure [1]

V̇1 = V̇ + e1 ė1
 
∂V ∂α(x)
≤ −W (x) + g(x)e1 + e1 ξ̇1 − ẋ
∂x ∂x

∂V
≤ −W (x) + e1 g(x) + f1 (x, ξ1 ) + g1 (x, ξ1 )α1 (x, ξ1 )
∂x

∂α  
+ g1 (x, ξ1 )e2 − f (x) + g(x)ξ1 . (C.28)
∂x

If g1 (x, ξ1 ) = 0 for all x and ξ1 , the choice of

 
1 ∂V ∂α  
α1 (x, ξ1 ) = −c1 e1 − g(x) − f1 (x, ξ1 ) + f (x) + g(x)ξ1
g1 (x, ξ1 ) ∂x ∂x
(C.29)
with c1 > 0 leads to

∂V1
V̇1 ≤ −W1 (x, ξ1 ) + g1 (x, ξ1 )e2 , (C.30)
∂ξ1

where W1 (x, e1 ) = W (x) + c1 e12 is positive definite.

Repeat the procedure step by step for the other subsystems (see Fig. C.1) until
the final external control is reached. Lyapunov functions are defined as

1 2
l
Vj (x, ξ1 , . . . , ξj ) = V (x) + ei , j = 1, 2, . . . , k (C.31)
2
i=1

with ej = ξj − αj −1 (x, ξ1 , . . . , ξj −1 ). If the non-singularity conditions

gj (x, ξ1 , . . . , ξj ) = 0, ∀x ∈ Rn , ∀ξi ∈ Rn , i = 1, 2, . . . , j, (C.32)

214 C Backstepping

hold then a final control law can be given by

 
1 ∂Vk−1 ∂αk−1
u= −ck ek − gk−1 − fk + (Fk−1 + Gk−1 ξk ) , (C.33)
gk ∂ξk−1 ∂Xk−1

where
 
Xj −1
Xj = , j = 2, 3, . . . , k − 1, (C.34a)
ξj
 
Fj −1 (Xj −1 ) + Gj −1 (Xj −1 )ξj
Fj (Xj ) = (C.34b)
fj (Xj −1 , ξj )
 
0
Gj (Xj ) = , (C.34c)
gj (Xj −1 , ξj )

and
     
x f (x) + g(x)ξ1 0
X1 = , F1 (X1 ) = , G1 (X1 ) = .
ξ1 f1 (x, ξ1 ) g1 (x, ξ1 )
(C.35)
Under the above control law, the system is globally asymptotically stable, since

V̇k (x, e1 , . . . , ek ) ≤ −Wk (x, e1 , . . . , ek ) (C.36)

with
Wj (x, e1 , . . . , ej ) = Wj −1 + cj ej2 , j = 2, 3, . . . , k, (C.37)
which are positive definite.
Applying the backstepping technique, results for some special forms of systems
are listed as follows.
(a) Integrator Backstepping. Consider the system

ẋ = f (x) + g(x)ξ, (C.38a)

ξ̇ = u (C.38b)

and suppose that the first equation satisfies Assumption C.1 with ξ ∈ R as con-
trol. If W (x) is positive definite, then

1 2
Va (x, ξ ) = V (x) + ξ − α(x) (C.39)
2
is a CLF for the whole system, and one of controls rendering the system asymp-
totically stable is given by
  ∂α   ∂V
u = −c ξ − α(x) + f (x) + g(x)ξ − g(x), c > 0. (C.40)
∂x ∂x
C.3 Adaptive Backstepping 215

(b) Linear Block Backstepping. Consider the cascade system

ẋ = f (x) + g(x)y, f (0) = 0, x ∈ Rn , y ∈ R, (C.41a)

ξ̇ = Aξ + bu, y = hξ, ξ ∈ Rq , u ∈ R, (C.41b)

where the linear subsystem is a minimum phase system of relative degree one
(hb = 0). If the x-subsystem satisfies Assumption C.1 with y ∈ R as control
and W (x) is positive definite, then there exists a feedback control guaranteeing
that the equilibrium x = 0, ξ = 0 is globally asymptotically stable. One of such
controls is
 
1   ∂α   ∂V
u= −c y − α(x) − hAξ + f (x) + g(x)y − g(x) , c > 0.
hb ∂x ∂x
(C.42)
(c) Nonlinear Block Backstepping. Consider the cascade system

ẋ = fx (x) + gx (x)y, f (0) = 0, x ∈ Rn , y ∈ R, (C.43a)

ξ̇ = fξ (x, ξ ) + gξ (x, ξ )u, y = h(ξ ), h(0) = 0, ξ ∈ R , u ∈ R,
q
(C.43b)

where the ξ -subsystem has globally defined and relative degree 1 uniformly
in x and its zero dynamics is input-to-state stable with respect to x and y as
its inputs. If the x-subsystem satisfies Assumption C.1 with y ∈ R as control
and W (x) is positive definite, then there exists a feedback control guaranteeing
that the equilibrium x = 0, ξ = 0 is globally asymptotically stable. One of such
controls is
 −1
∂h
u= gξ (x, ξ )
∂ξ
 
  ∂h ∂α   ∂V
× −c y − α(x) − fξ (x, ξ ) + f (x) + g(x)y − g(x) ,
∂ξ ∂x ∂x
c > 0. (C.44)

C.3 Adaptive Backstepping

Some systems consist of unknown constant parameters which appear linearly in
the system equations. In the presence of such parametric uncertainties, we will be
able to achieve both boundedness of the closed-loop states and convergence of the
tracking error to zero.
Consider the nonlinear system

ẋ1 = x2 + θ ϕ(x1 ), (C.45a)

ẋ2 = u. (C.45b)
216 C Backstepping

If θ were known, we would apply the backstepping technique to design a stabilizing

controller. First, we view x2 as virtual control and design

α1 (x1 , θ ) = −c1 x1 − θ ϕ(x1 ) (C.46)

to achieve the derivative of the Lyapunov function

1
V0 (x1 ) = x12 (C.47)
2
negative definite as follows:
 
V̇0 = x1 ẋ1 = x1 −c1 x1 − θ ϕ(x1 ) + θ ϕ(x1 ) = −c1 x12 , (C.48)

where c1 > 0. Then, we define the difference between x2 and α1 (x1 , θ ) as

z1 = x2 − α1 (x1 , θ ) (C.49)

to reformulate the system as

ẋ1 = z1 − c1 x1 , (C.50a)
∂α1 ∂α1 ∂α1
ż1 = ẋ2 − ẋ1 − θ̇ = u − (z1 − c1 x1 ), (C.50b)
∂x1 ∂θ ∂x1
where θ̇ = 0 is used (θ is assumed to be constant). By defining the Lyapunov func-
tion
1 1
V1 (x1 , z1 ) = x12 + z12 (C.51)
2 2
and differentiating it along the above system, we arrive at

V̇1 (x1 , z1 ) = x1 ẋ1 + z1 ż1

 
∂α1
= z1 x1 − c1 x12 + z1 u − (z1 − c1 x1 )
∂x1
 
∂α1
= −c1 x12 + z1 u + x1 − (z1 − c1 x1 ) . (C.52)
∂x1
If we choose u such that
∂α1
u + x1 − (z1 − c1 x1 ) = −c2 z1 (C.53)
∂x1
with c2 > 0, then
V̇1 (x1 , z1 ) = −c1 x12 − c2 z12 , (C.54)
which implies V̇1 is negative definite. Hence, the control law is given as
  ∂α1  
u = c2 x2 − α1 (x1 , θ ) − x1 + x2 + θ ϕ(x1 ) . (C.55)
∂x1
C.3 Adaptive Backstepping 217

Indeed, θ is unknown, we cannot implement this control law. However, we can apply
the idea of backstepping to handle this issue.
We start with x2 being a virtual control to design an adaptive control law, i.e., we
start with
ẋ1 = v + θ ϕ(x1 ). (C.56)
If θ were known, the control

v = −c1 x1 − θ ϕ(x1 ) (C.57)

would render the derivative of

1
V0 (x1 ) = x1 (C.58)
2
negative definite as V̇0 = −c1 x12 . Since θ is unknown, we apply the certainty-
equivalence principle to modify the control law as

v = −c1 x1 − θ̂1 ϕ(x1 ), (C.59)

where θ̂1 could be an estimate of θ . Then, we obtain

ẋ1 = −c1 x1 + θ̃1 ϕ(x1 ) (C.60)

with
θ̃1 = θ − θ̂1 (C.61)
being the parameter estimation error. We extend the Lyapunov function V0 as
1 1 2
V1 (x1 , θ̃1 ) = x1 + θ̃ , (C.62)
2 2γ 1
where γ > 0. Its derivative becomes
1 ˙
V̇1 = x1 ẋ1 + θ̃1 θ̃1
γ
1 ˙
= −c1 x12 + x1 θ̃1 ϕ(x1 ) + θ̃1 θ̃1
γ
 
1˙
= −c1 x1 + θ̃1 x1 ϕ(x1 ) + θ̃1 .
2
(C.63)
γ
If we choose
1˙
x1 ϕ(x1 ) + θ̃1 = 0, (C.64)
γ
then we have the seminegative definite property of V1 as

V̇1 = −c1 x12 . (C.65)

218 C Backstepping

With the assumption of θ̇ = 0, the adaptive law is then given by

θ̂˙1 = γ x1 ϕ(x1 ). (C.66)

Hence, the adaptive virtual control for x2 is given as

α1 (x1 , θ̂1 ) = −c1 x1 − θ̂1 ϕ(x1 ), (C.67a)

θ̂˙1 = γ x1 ϕ(x1 ), (C.67b)

and the x1 -equation in (C.45a), (C.45b) becomes

ẋ1 = −c1 x1 + (θ − θ̂1 )ϕ(x1 ). (C.68)

Now we define the difference between x2 and α1 (x1 , θ̂1 ) as

z2 = x2 − α1 (x1 , θ̂1 ), (C.69)

then
ẋ1 = z2 + α1 (x1 , θ̂1 ) + θ ϕ(x1 ) = z2 − c1 x1 + θ̃1 ϕ(x1 ) (C.70)
and
∂α1 ∂α1 ˙
ż2 = ẋ2 − ẋ1 − θ̂1
∂x1 ∂ θ̂1
∂α1   ∂α1
=u− z2 − c1 x1 + θ̃1 ϕ(x1 ) − γ x1 ϕ(x1 ). (C.71)
∂x1 ∂ θ̂1

We augment the Lyapunov function V1 as

1 1 2 1 2
V2 (x1 , θ̃1 , z2 ) = x12 + θ̃ + z (C.72)
2 2γ 1 2 2

and infer
1 ˙
V̇2 = x1 ẋ1 + θ̃1 θ̃1 + z2 ż2 = x1 z2 − c1 x12 + x1 θ̃1 ϕ(x1 ) − θ̃1 x1 ϕ(x1 )
γ
 
∂α1   ∂α1
+ z2 u − z2 − c1 x1 + θ̃1 ϕ(x1 ) − γ x1 ϕ(x1 )
∂x1 ∂ θ̂1
 
∂α1   ∂α1
= −c1 x12 + z2 u + x1 − z2 − c1 x1 + θ̃1 ϕ(x1 ) − γ x1 ϕ(x1 ) ,
∂x1 ∂ θ̂1

which is rendered negative semidefinite as

V̇2 = −c1 x12 − c2 z22 (C.73)

C.3 Adaptive Backstepping 219

if we choose the control law such that

∂α1   ∂α1
u + x1 − z2 − c1 x1 + θ̃1 ϕ(x1 ) − γ x1 ϕ(x1 ) = −c2 z2 . (C.74)
∂x1 ∂ θ̂1
This leads to the following control law
∂α1   ∂α1
u = −c2 z2 − x1 + z2 − c1 x1 + θ̃1 ϕ(x1 ) + γ x1 ϕ(x1 )
∂x1 ∂ θ̂1
∂α1   ∂α1
= −c2 z2 − x1 + x2 + θ ϕ(x1 ) + γ x1 ϕ(x1 ) (C.75)
∂x1 ∂ θ̂1
which is still not implementable due to the unknown θ . Hence, we need a new
estimate θ̂2 to build
∂α1   ∂α1
u = −c2 z2 − x1 + x2 + θ̂2 ϕ(x1 ) + γ x1 ϕ(x1 ). (C.76)
∂x1 ∂ θ̂1
With this choice, ż2 becomes
∂α1
ż2 = −x1 − c2 z2 − (θ − θ̂2 ) ϕ(x1 ). (C.77)
∂x1
Defining an augmented Lyapunov function V3 as
  1 1  2  1
V3 x1 , θ̃1 , z2 , (θ − θ̂2 ) = x12 + θ̃1 + (θ − θ̂2 )2 + z22 , (C.78)
2 2γ 2
we derive
1 ˙ 
V̇2 = x1 ẋ1 + θ̃1 θ̃1 + (θ − θ̂2 )(θ̇ − θ̂˙2 ) + z2 ż2
γ

(θ − θ̂2 )θ̂˙2
1
= x1 z2 − c1 x12 + x1 θ̃1 ϕ(x1 ) − θ̃1 x1 ϕ(x1 ) −
γ
 
∂α1
+ z2 −x1 − c2 z2 − (θ − θ̂2 ) ϕ(x1 )
∂x1
 
1˙ ∂α1
= −c1 x12 − c2 z22 − (θ − θ̂2 ) θ̂2 + z2 ϕ(x1 ) .
γ ∂x1
By choosing the second update law as

θ̂˙2 = −γ z2
∂α1
ϕ(x1 ), (C.79)
∂x1
we arrive at
V̇3 = −c1 x12 − c2 z22 , (C.80)
which is negative semidefinite. Hence, (C.66), (C.67a), (C.76), and (C.79) construct
the final adaptive controller for the system (C.45a), (C.45b).
220 C Backstepping

References
1. Krstić M, Kanellakopoulos I, Kokotović P (1995) Nonlinear and adaptive control design. Wiley,
New York
2. Kokotovic PV (1992) The joy of feedback: nonlinear and adaptive. In: IEEE control systems
3. Kokotovic Krstic M PV, Kanellakopoulos I (1992) Backstepping to passivity: recursive design
of adaptive systems. In: Proc 31st IEEE conf decision contr. IEEE Press, New Orleans, pp 3276–
3280
4. Sontag ED (1989) A ‘universal’ construction of Artstein’s theorem on nonlinear stabilization.
Syst Control Lett 13:117–123
5. Khalil HK (2002) Nonlinear Systems. Prentice Hall, New York
Appendix D
Model Predictive Control (MPC)

In general, model predictive control is formulated as solving online a finite horizon

(open-loop) optimal control problem subject to system dynamics and constraints
involving states and controls [1–4]. The methodology of all the controllers belong-
ing to the MPC family is characterized by the following strategy, represented in
Fig. D.1 [5].
Based on the current measurement, say at time t, the controller uses a dynamic
model (called a prediction model) to predict the future dynamic behavior of the sys-
tem over a prediction horizon Tp , and determines (over a control horizon Tc ≤ Tp )
the control input ū such that a pre-specified performance objective is optimized (for
example, an integrated square error between the predicted output and the setpoint).
Note that the control input between Tc and Tp may be assumed constant and equal
to the control at the end of the control horizon in the case of Tc < Tp . If there are
no disturbances and model–plant mismatch, and/or if the optimization problem can
be solved for infinite prediction and control horizons, then, we can apply the input
function found at time t = 0 to the system for all the time t ≥ 0. However, this is
not possible in general. Due to the existence of disturbances and/or model–plant
mismatch, the real system behavior is different from the predicted behavior. Since
finding a solution over the infinite horizon to the optimization problem is also im-
possible in general, we do not have a control input being available forever. Thus,
the control input obtained by solving the optimization problem will be implemented
only until the next measurement becomes available. We assume that this will be the
case every δ time-units, where δ denotes the “sampling time”. Updated with the new
measurement, at time t + δ, the whole procedure—prediction and optimization—is
repeated to find a new control input, with the control and prediction horizons mov-
ing forward (for this reason, MPC is also referred to as moving horizon control or
receding horizon control). This results in a discrete feedback control with an implicit
control law because closed-loop control inputs are calculated by solving online the
optimization problem at each sampling time. Hence, MPC is characterized by the
following points:
• Model-based prediction. In contrast to other feedback controllers that calculate
the control action based on the present or past state information, model predictive

H. Chen, B. Gao, Nonlinear Estimation and Control of Automotive Drivetrains, 221

DOI 10.1007/978-3-642-41572-2,
© Science Press Beijing and Springer-Verlag Berlin Heidelberg 2014
222 D Model Predictive Control (MPC)

Fig. D.1 Principle of model

predictive control

controllers determine the control action based on the predicted future dynamics
of the to be controlled system starting from the current state. The model used to
complete the prediction can be linear or nonlinear, time-continuous or discrete-
time, deterministic or stochastic, etc. We emphasize here the functionality of the
model, namely it is able to predict the future dynamics of the to be controlled
system, while we do not care for the form of the model. Hence, any kind of
models, based on which the system dynamics can be computed, can be used as a
prediction model. Some of them are listed as follows:
– Convolution models, including step response and impulse response models;
– First principle models (state space model);
– Fuzzy models;
– Neural network models;
– Data-based models;
– ...
• Handling of constraints. In practice, most systems have to satisfy time-domain
constraints on inputs and states. For example, an actuator reaches saturation and
some states such as temperature and pressure are not allowed to exceed their lim-
itations for the reason of safe operation, or some variables have to be held under
certain threshold values to meet environmental regulations. Moreover, when mod-
eling chemical processes from mass, momentum and energy conservation laws,
algebraic equations may arise from phase equilibrium calculations and other phe-
nomenological and thermodynamic correlations [6]. These algebraic equations
may also be considered as constraints on the dynamics of the process. It is clear
that time-domain constraints impose limitations on the achievable control perfor-
mance, even if the system to be controlled is linear [7].
In MPC, time-domain constraints can be placed directly in the optimization
problem in their original form, without doing any transformation. Such a direct
and explicit handling of time-domain constraints leads to non-conservative or at
least less conservative solutions. Moreover, because the future response of the
system is predicted, early control action can be taken so as to avoid the violation
of time-domain constraints (e.g., actuator saturation, safety constraints, emission
regulation) while tracking, for example, a given reference trajectory with mini-
D.1 Linear MPC 223

mal tracking error. This performs somehow an active handling of time-domain

constraints;
• Online optimization. An objective functional that specifies mathematically the
desired control performance is minimized online at each sampling instance.
A commonly used objective functional is an integrated weighted square error
between predicted controlled variables and their desired references. There may,
however, be different objectives to describe economic requirements. In consid-
eration of time-domain constraints, a constrained dynamic optimization problem
will be repeatedly solved online. The main reasons for an online repeated solution
are listed as follows:
– In general, we cannot find an analytic solution to the involved optimization
problem. Numerical methods are used. A time-continuous input parameteriza-
tion and/or the use of an infinite horizon may lead to an infinite-dimensional
optimization problem that is numerically extremely demanding and often in-
tractable. In order to get around that, the optimization problem is formulated
with finite horizons. Through the moving horizon implementation, we can ob-
tain the control action as the time goes;
– Due to the existence of model uncertainties, the real dynamics is different from
the predicted dynamics. The measurement available at each sampling time con-
tains the information reflecting various uncertainties. Through repeating the
whole procedure, prediction and optimization, the information is used to im-
prove control performance;
– Real systems suffer in general disturbances. If we want to achieve high per-
formance for disturbance attenuation, strong control action is required, which
may lead to the violation of time-domain constraints. Hence, we need a trade-
off between satisfying constraints and achieving high performance. Through
the online solution of the optimization problem, a performance adaptation is
possible [8];
– A detailed derivation shows that MPC admits a feed-forward and feedback
structure [9]. The feed-forward information includes the measurable distur-
bance and the given reference over the prediction horizon, while the feedback
information is the measured state/output.

D.1 Linear MPC

The terminology of linear MPC refers to MPC based on linear models, even if the
existence of time-domain constraints renders the dynamics nonlinear. Since we can
reformulate the step response model and impulse response model in the state space
form [9], in the following we take the general form of state space model as an
example. Given a basic form of linear discrete state space equations
x(k + 1) = Ax(k) + Bu u(k) + Bd d(k), (D.1a)
yc (k) = Cc x(k), (D.1b)
yb (k) = Cb x(k), (D.1c)
224 D Model Predictive Control (MPC)

where x ∈ Rnx is the system state, u ∈ Rnu is the control input, d ∈ Rnd is the
measurable disturbance, yc ∈ Rnc is the controlled output, and yb ∈ Rnb is the con-
strained output.
It is well-known that the difference equation (D.1a) can be exactly obtained from
the differential equation

ẋ(t) = Ac x(t) + Bcu u(t) + Bcd d(t) (D.2)

by computing

A = eAc δ , (D.3a)
 δ
Bu = eAc τ dτ · Bcu , (D.3b)
0
 δ
Bd = eAc τ dτ · Bcd , (D.3c)
0

with δ being the sampling time.

In order to introduce the integral action to reduce offset, we rewrite (D.1a)–(D.1c)
in the incremental form

x(k + 1) = Ax(k) + Bu u(k) + Bd d(k), (D.4a)

yc (k) = Cc x(k) + yc (k − 1), (D.4b)
yb (k) = Cb x(k) + yb (k − 1), (D.4c)

where

x(k) = x(k) − x(k − 1),

u(k) = u(k) − u(k − 1),
d(k) = d(k) − d(k − 1).

Assume that the state is measurable. If it is not the case, we can use an observer
to estimate the state. Then, at time k, with the measured/estimated state x(k), the
optimization problem of linear MPC is formulated as

Problem D.1
 
min J x(k), U (k), Nc , Np (D.5)
U (k)

subject to (D.4a)–(D.4c) and time-domain constraints

umin ≤ u(k + i|k) ≤ umax , i = 0, 1, . . . , Nc − 1, (D.6a)

umin ≤ u(k + i|k) ≤ umax , i = 0, 1, . . . , Nc − 1, (D.6b)
D.1 Linear MPC 225

ymin (k + i) ≤ yb (k + i|k) ≤ ymax (k + i), i = 1, . . . , Np , (D.6c)

u(k + i|k) = 0, Nc ≤ i ≤ N p (D.6d)

where the objective functional is defined as

Np
     
J x(k), U (k), Nc , Np = Γy,i yc (k + i|k) − r(k + i) 2
i=1

c −1
N
 
+ Γu,i u(k + i|k)2 , (D.7)
i=0

or in the vector form,

    2  2
J x(k), U (k), Nc , Np = Γy Yc (k +1|k)−R(k +1)  + Γu U (k) . (D.8)

In the above, Np and Nc are prediction and control horizons, respectively, satis-
fying Nc ≤ Np , Γy and Γu are weights given as

Γy = diag{Γy,1 , Γy,2 , . . . , Γy,p }, Γy,i ∈ Rnc ×nc , i = 1, 2, . . . , Np ,

Γu = diag{Γu,1 , Γu,2 , . . . , Γu,m }, Γu,j ∈ Rnu ×nu , j = 1, 2, . . . , Nc ,

R(k + 1) is the vector of the reference

⎡ ⎤
r(k + 1)
⎢ r(k + 2) ⎥
⎢ ⎥
R(k + 1) = ⎢ .. ⎥
⎣ . ⎦
r(k + Np ) Np ×1

and U (k) is the vector form of the incremental control sequences defined as
⎡ ⎤
u(k|k)
⎢ u(k + 1|k) ⎥
⎢ ⎥
U (k) = ⎢ .. ⎥ , (D.9)
⎣ . ⎦
u(k + Nc − 1|k) Nc ×1

which is the independent variable of the optimization problem. The constraints on

u and u come from actuator saturations. Note that although they are regarded as
constant here, time-varying constraints can also be dealt with if only minor revision
are added. Moreover, yc (k + i|k) and yb (k + i|k) are the controlled and constrained
outputs predicted at time k, on the basis of the prediction model (D.4a)–(D.4c). They
226 D Model Predictive Control (MPC)

can be represented in the vector form of

⎡ ⎤ ⎡ ⎤
yc (k + 1|k) yb (k + 1|k)
⎢ yc (k + 2|k) ⎥ ⎢ yb (k + 2|k) ⎥
⎢ ⎥ ⎢ ⎥
Yc (k + 1|k) = ⎢ .. ⎥ , Yb (k + 1|k) = ⎢ .. ⎥
⎣ . ⎦ ⎣ . ⎦
yc (k + Np |k) Np ×1
yb (k + Np |k) Np ×1

for a clear formulation. By iterating the difference equation in (D.4a)–(D.4c), we

get the prediction equations as follows:

Yc (k + 1|k) = Sx,c x(k) + Ic yc (k) + Sd,c d(k) + Su,c U (k), (D.10a)

Yb (k + 1|k) = Sx,b x(k) + Ib yb (k) + Sd,b d(k) + Su,b U (k), (D.10b)

where

⎡ ⎤ ⎡ ⎤
Cc A Cb A
⎢ Cc A2 + Cc A ⎥ ⎢ Cb A2 + Cb A ⎥
⎢ ⎥ ⎢ ⎥
Sx,c = ⎢ .. ⎥ , S x,b = ⎢ .. ⎥ ,
⎣ . ⎦ ⎣ . ⎦
 Np i
Np i
i=1 Cc A Np ×1 i=1 Cb A Np ×1
⎡ ⎤ ⎡ ⎤
Inc ×nc Inb ×nb
⎢ Inc ×nc ⎥ ⎢ Inb ×nb ⎥
⎢ ⎥ ⎢ ⎥
Ic = ⎢ . ⎥ , Ib = ⎢ . ⎥ ,
⎣ . ⎦ . ⎣ . ⎦ .
Inc ×nc N ×1 Inb ×nb N ×1
p p
⎡ ⎤ ⎡ ⎤
Cc Bd Cb Bd
⎢ Cc ABd + Cc Bd ⎥ ⎢ Cb ABd + Cb Bd ⎥
⎢ ⎥ ⎢ ⎥
Sd,c = ⎢ .. ⎥ , Sd,b = ⎢ .. ⎥ ,
⎣ . ⎦ ⎣ . ⎦
Np i−1
 Np i−1
i=1 Cc A Bd N
p ×1 i=1 Cb A Bd N
p ×1
⎡ Cc Bu 0 0 ... 0 ⎤
2
⎢ i=1 Cc Ai−1 Bu Cc Bu 0 ... 0 ⎥
⎢ ⎥
⎢ .. .. .. .. .. ⎥
⎢ . . . . . ⎥
⎢ ⎥
Su,c = ⎢ Nc i−1 B Nc −1 C Ai−1 B ⎥ ,
⎢ i=1 c C A u c u ... ... C c Bu ⎥
⎢ i=1 ⎥
⎢ .. .. .. .. .. ⎥
⎣ . . . . . ⎦
 Np i−1 Np −1 Np −Nc +1
i=1 Cc A Bu i=1 Cc Ai−1 Bu ... ... i=1 Cc Ai−1 Bu N ×N
p c
D.1 Linear MPC 227
⎡ Cb Bu 0 0 ... 0 ⎤
2
⎢ i=1 Cb Ai−1 Bu Cb Bu 0 ... 0 ⎥
⎢ ⎥
⎢ .. .. .. .. ⎥
⎢ . . . ... . ⎥
⎢  ⎥
Su,b = ⎢ Nc Nc −1 ⎥ .
⎢ i=1 Cb Ai−1 Bu i=1 Cb Ai−1 Bu ... ... Cb Bu ⎥
⎢ ⎥
⎢ .. .. .. .. ⎥
⎣ . . . ... . ⎦
 Np i−1 Np −1 Np −Nc +1
i=1 Cb A Bu i=1 Cb Ai−1 Bu ... ... i=1 Cb Ai−1 Bu N ×N
p c

According to the basic of MPC, the optimization problem (Problem D.1) will be
solved at each sampling time, updated with the new measurement. If we can find a
solution of Problem D.1, denoted as U ∗ (k),
the closed-loop control at time k is then defined as

u(k) := u∗ (k|k) + u(k − 1). (D.12)

If we do not consider the time-domain constraints, then the optimization problem

(Problem D.1) becomes
  2  2
min Γy Yc (k + 1|k) − R(k + 1)  + Γu U (k) (D.13)
U (k)

with Yc (k +1|k) given by (D.10a). We can then obtain the solution by calculating the
gradient of the objective function over the independent variable U (k) and setting
it to zero. The result reads
 T T −1 T T
U ∗ (k) = Su,c Γy Γy Su,c + ΓuT Γu Su,c Γy Γy Ep (k + 1|k), (D.14)

with Ep (k + 1|k) being calculated by

Ep (k + 1|k) = R(k + 1) − Sx,c x(k) − Ic yc (k) − Sd,c d(k). (D.15)

According to the basic of MPC, we pick up the first element of U ∗ (k) to build the
closed-loop control as follows

u(k) = Kmpc Ep (k + 1|k), (D.16)

where Kmpc is calculated by

   T T −1 T T
Kmpc = Inu ×nu 0 ... 0 1×Nc
Su,c Γy Γy Su,c + ΓuT Γu Su,c Γy Γy .
(D.17)
Substituting (D.15) into (D.16) leads to

u(k) = Kmpc R(k + 1) − Kmpc (Sx,c + Ic Cc )x(k)

− Kmpc Sd,c d(k) + Kmpc Sx,c x(k − 1).

It is clear that
228 D Model Predictive Control (MPC)

• Kmpc R(k + 1) represents a feed-forward depending on the future reference over

the prediction horizon;
• −Kmpc Sd,c d(k) represents a feed-forward depending on the measurable distur-
bance;
• −Kmpc (Sx,c + Ic Cc )x(k) + Kmpc Sx,c x(k − 1) represents a state feedback de-
pending on the measurement.

Hence, MPC admits a feed-forward and feedback structure.

In the case of considering the time-domain constraints, the optimization problem
can be formulated as a standard quadratic programming (QP) problem as follows:

min U (k)T H U (k) − G(k + 1|k)T U (k) (D.18a)

U (k)

subject to Cu U (k) ≥ b(k + 1|k), (D.18b)

where

H = Su,c
T
ΓyT Γy Su,c + ΓuT Γu ,

G(k + 1|k) = 2Su,c

T
ΓyT Γy Ep (k + 1|k),
⎡ ⎤
−T
⎢ T ⎥
⎢ ⎥
⎢ −L ⎥
Cu = ⎢⎢ ⎥,
⎥
⎢ L ⎥
⎣ −Su,b ⎦
Su,b
⎡ ⎤
−umax
⎢ .. ⎥
⎢ . ⎥
⎢ ⎥
⎢ −u ⎥
⎢ max ⎥
⎢ u ⎥
⎢ min ⎥
⎢ .. ⎥
⎢ . ⎥
⎢ ⎥
⎢ umin ⎥
⎢ ⎥
⎢ −umax + u(k − 1) ⎥
⎢
b(k + 1|k) = ⎢ ⎥
.. ⎥
⎢ . ⎥
⎢ ⎥
⎢ −umax + u(k − 1) ⎥
⎢ ⎥
⎢ umin − u(k − 1) ⎥
⎢ ⎥
⎢ . ⎥
⎢ .. ⎥
⎢ ⎥
⎢ umin − u(k − 1) ⎥
⎢ ⎥
⎣ −Ymax (k + 1) + Sx,b x(k) + Ib yb (k) + Sd,b d(k) ⎦
Ymin (k + 1) − Sx,b x(k) − Ib yb (k) − Sd,b d(k)
D.2 Nonlinear MPC (NMPC) 229

with
⎡ ⎤
Inu ×nu 0 ... 0 0
⎢ 0 Inu ×nu ... 0 0 ⎥
⎢ ⎥
⎢ ⎥
T = ⎢ ... ..
.
..
.
..
.
..
. ⎥ ,
⎢ ⎥
⎣ 0 0 ... Inu ×nu 0 ⎦
0 0 ... 0 Inu ×nu Nc ×Nc
⎡ ⎤
Inu ×nu 0 ... 0 0
⎢ Inu ×nu Inu ×nu ... 0 0 ⎥
⎢ ⎥
⎢ ⎥
L = ⎢ ... ..
.
..
.
..
.
..
. ⎥ ,
⎢ ⎥
⎣ Inu ×nu Inu ×nu . . . Inu ×nu 0 ⎦
Inu ×nu Inu ×nu . . . Inu ×nu Inu ×nu Nc ×Nc
⎡ ⎤ ⎡ ⎤
ymin (k + 1) ymax (k + 1)
⎢ ymin (k + 2) ⎥ ⎢ ymax (k + 2) ⎥
⎢ ⎥ ⎢ ⎥
Ymin (k + 1) = ⎢ .. ⎥ , Ymax (k + 1) = ⎢ .. ⎥ .
⎣ . ⎦ ⎣ . ⎦
ymin (k + Np ) Np ×1
ymax (k + Np ) Np ×1

D.2 Nonlinear MPC (NMPC)

D.2.1 NMPC Based on Discrete-Time Model

Consider the following discrete nonlinear state-space equations:

 
x(k + 1) = f x(k), u(k) , k ≥ 0, (D.20a)
 
yc (k) = gc x(k), u(k) , (D.20b)
 
yb (k) = gb x(k), u(k) , (D.20c)

where x(k) ∈ Rnx is the system state, u(k) ∈ Rnu is the control input, yc (k) ∈ Rnc
is the controlled output, yb (k) ∈ Rnb is the constrained output. The constraints on
input and output are represented as

umin ≤ u(k) ≤ umax , ∀k ≥ 0, (D.21a)

umin ≤ u(k) ≤ umax , ∀k ≥ 0, (D.21b)
ymin (k) ≤ yb (k) ≤ ymax (k), ∀k ≥ 0. (D.21c)

It is assumed that all states are measurable. If not all states are measurable, an
observer has to be designed to estimate the state. Then at time k, based on the
measured/estimated state x(k), the optimization problem of discrete nonlinear MPC
is formulated as
230 D Model Predictive Control (MPC)

Problem D.2
 
min J x(k), Uk (D.22)
Uk

subject to (D.20a)–(D.20c) and time-domain constraints

umin ≤ ū(k + i) ≤ umax , 0 ≤ i < Nc , (D.23a)

umin ≤ ū(k + i) ≤ umax , (D.23b)
ū(k + i) = ū(k + i) − ū(k + i − 1), (D.23c)
ymin (k + i) ≤ ȳb (k + i) ≤ ymax (k + i), 0 < i ≤ Np , (D.23d)
ū(k + i) = 0, Nc ≤ i ≤ N p , (D.23e)

where the objective functional is defined as

N
  p
 
J x(k), Uk = ȳc (k + i) − r(k + i)2
Q
i=1

c −1
N
    
+ ū(k + i) − ur (k + i)2 + ū(k + i)2 , (D.24)
R S
i=0

and ȳc (·) and ȳb (·) as predicted controlled and constrained outputs, respectively,
can be calculated through the following dynamic equations:
 
x̄(i + 1) = f x̄(i), ū(i) , k ≤ i ≤ k + Np , x̄(k) = x(k), (D.25a)
 
ȳc (i) = gc x̄(i), ū(i) , (D.25b)
 
ȳb (i) = gb x̄(i), ū(i) . (D.25c)

In the above description, Np and Nc are the prediction and control horizons,
respectively, satisfying Nc ≤ Np ; (r(·), ur (·)) are the references of the controlled
output and corresponding control input; (Q, R, S) are weighting matrices, allowed
to be time varying; ū(·) is the predicted control input, defined as

ū(k + i) = ūi , i = 0, 1, . . . , Nc − 1, (D.26)

where ū0 , . . . , ūNc −1 constitute the independent variables of the optimization prob-
lem, denoted as Uk
⎡ ⎤
ū0
⎢ ū1 ⎥
⎢ ⎥
Uk  ⎢ . ⎥ . (D.27)
⎣ .. ⎦
ūNc −1
Note that x(k), the system state, is also the initial condition of the prediction
model (D.25a)–(D.25c), which is the key of MPC being a feedback strategy.
D.2 Nonlinear MPC (NMPC) 231

Assume that optimization problem D.2 is feasible at each sampling time and the
solution is
⎡ ∗ ⎤
ū0
⎢ ū∗1 ⎥
⎢ ⎥
Uk∗  ⎢ .. ⎥ , (D.28)
⎣ . ⎦
ū∗Nc −1
then, according to the basics of MPC, the control input is chosen as

u(k) = ū∗0 . (D.29)

Because Uk∗ depends on the values of x(k), (Q, S, R) and (Nc , Np ), u(k) is an
implicit function of these variables. Ignoring the dependence of (Q, S, R) and
(Nc , Np ), denote u(k) as
 
u(k) = κ x(k) , k ≥ 0. (D.30)

Substituting it into the controlled system (D.20a)–(D.20c), we have the closed sys-
tem
  
x(k + 1) = f x(k), κ x(k) , k ≥ 0. (D.31)
If not all states are measurable, only need to replace x(k) with x̂(k).

D.2.2 NMPC Based on Continuous-Time Model

In general, it is difficult to achieve precise enough discretization of a nonlinear sys-

tem. Hence nonlinear MPC on time continuous model is also investigated.
Consider the following continuous nonlinear system:
 
ẋ(t) = f x(t), u(t) , t ≥ 0, (D.32a)
 
yc (t) = gc x(t), u(t) , (D.32b)
 
yb (t) = gb x(t), u(t) , (D.32c)

where x(t) ∈ Rnx is the state, u(t) ∈ Rnu is the control input, yc (t) ∈ Rnc is the
controlled output, yb (t) ∈ Rnb is the constrained output. The constraints on input
and output are

umin ≤ u(t) ≤ umax , ∀t ≥ 0, (D.33a)

dumin ≤ u̇(t) ≤ dumax , ∀t ≥ 0, (D.33b)
ymin (t) ≤ yb (t) ≤ ymax (t), ∀t ≥ 0. (D.33c)

At the present time t, based on the measured/estimated state x(t) and ignoring
the constraint on the change rate of the control input, the optimization problem is
formulated as
232 D Model Predictive Control (MPC)

Problem D.3
 
min J x(t), Ut (D.34)
Ut

subject to

umin ≤ ū(τ ) ≤ umax , t ≤ τ < t + Tc , (D.35a)

ymin (τ ) ≤ ȳb (τ ) ≤ ymax (τ ), t < τ ≤ t + Tp , (D.35b)
ū(τ ) = ū(t + Tc ), t + Tc ≤ τ ≤ t + Tp , (D.35c)

where the objective functional is defined as


  t+Tp     
J x(t), Ut = ȳc (τ ) − r(τ )2 + ū(τ ) − ur (τ )2 dτ, (D.36)
Q R
t

and ȳc (·) and ȳb (·) are the predicted controlled and constrained output, respectively,
calculated through the following dynamic equation:
 
˙ ) = f x̄(τ ), ū(τ ) , t ≤ τ ≤ t + Tp , x̄(t) = x(t),
x̄(τ (D.37a)
 
ȳc (τ ) = gc x̄(τ ), ū(τ ) , (D.37b)
 
ȳb (τ ) = gb x̄(τ ), ū(τ ) . (D.37c)

In the above, Tc and Tp are the prediction and control horizons, satisfying
Tc ≤ Tp ; (r(·), ur (·)) are the references of the controlled output and the correspond-
ing control input; (Q, R) are weighting matrices, allowed to be time varying; ū(·)
is the predicted control input, and for τ ∈ [t, t + Tc ], it is defined as
 
τ −t
ū(τ ) = ūi , i = int , (D.38)
δ

where δ is sampling time and Tc = Nc δ. Then ū0 , . . . , ūNc −1 constitute independent

variables of the optimization problem, denoted as Ut
⎡ ⎤
ū0
⎢ ū1 ⎥
⎢ ⎥
Ut  ⎢ . ⎥ . (D.39)
⎣ .. ⎦
ūNc −1

Note that x(t), the system state, is also the initial condition of the prediction
model (D.37a).

Remark D.1 By Eqs. (D.38) and (D.39), the control input is treated as constant dur-
ing the sampling period, and thus the optimization problem is transferred into a
problem with limited independent variables.
D.2 Nonlinear MPC (NMPC) 233

Remark D.2 The constraint on the change rate of control input can be taken into
consideration through the following two methods. One is to add a penalty item into
the objective function:

  t+Tp     
J x̂(t), Ut = ȳc (τ ) − r(τ )2 + ū(τ ) − ur (τ )2 dτ
Q R
t
c −1
N
+ ūi − ūi−1 2
S. (D.40)
i=0

This is a somehow “soft” treatment. Another one is to use ūi −δūi−1 to approximate
the change rate of the control input, and add the following constraint into (D.35a)–
(D.35c):
ūi − ūi−1
dumin ≤ ≤ dumax . (D.41)
δ

Assume that the optimization problem D.3 is feasible at each sampling time, and
the solution is denoted as
⎡ ∗ ⎤
ū0
⎢ ū∗1 ⎥
⎢ ⎥
Ut∗  ⎢ .. ⎥ , (D.42)
⎣ . ⎦
ū∗Nc −1

then, according to the basics of MPC, the closed-loop control is defined as

u(τ ) = ū∗0 , t ≤ τ ≤ t + δ. (D.43)

Because Ut∗ depends on the values of x(t), (Q, R) and (Tc , Tp ), u(t) is an implicit
function of these variables, denoted as
 
u(τ ) = κ x(t) , t ≤ τ ≤ t + δ, (D.44)

where the dependence of (Q, R) and (Tc , Tp ) is ignored for simplicity. By substi-
tuting it into (D.32a)–(D.32c), we have the closed-loop system
  
ẋ(τ ) = f x(τ ), κ x(t) , t ≤ τ ≤ t + δ, t ≥ 0. (D.45)

It is clear that predictive control has the characteristics of sampled-date systems.

The solution of NMPC is always summarized as solving a nonlinear program-
ming (NLP) problem, and the procedure is shown in Fig. D.2. For a more detailed
discussion on various MPC formulations, and theoretic issues as stability and ro-
bustness, we refer to, for example, [8, 10–19].
234 D Model Predictive Control (MPC)

Fig. D.2 Schematic program of NMPC

References
1. Allgöwer F, Badgwell TA, Qin JS, Rawlings JB, Wright SJ (1999) Nonlinear predictive control
and moving horizon estimation—an introductory overview. In: Frank PM (ed) Advances in
control, highlights of ECC’99. Springer, Berlin, pp 391–449
2. Camacho EF, Bordons C (2004) Model predictive control. Springer, London
3. Maciejowski JM (2002) Predictive control: with constraints. Prentice Hall, New York
4. Mayne DQ, Rawlings JB, Rao CV, Scokaert POM (2000) Constrained model predictive con-
trol: stability and optimality. Automatica 36(6):789–814
5. Chen H (1997) Stability and robustness considerations in nonlinear model predictive control.
Fortschr.-Ber. VDI Reihe 8, vol 674. VDI Verlag, Düsseldorf
6. Kröner A, Holl P, Marquardt W, Gilles ED (1989) DIVA—an open architecture for dynamic
simulation. In: Eckermann R (ed) Computer application in the chemical industry. VCH, Wein-
heim, pp 485–492
7. Mayne DQ (1995) Optimization in model based control. In: Proc IFAC symposium dynamics
and control of chemical reactors, distillation columns and batch processes, Helsingor, pp 229–
242
8. Chen H, Scherer CW (2006) Moving horizon H∞ control with performance adaptation for
constrained linear systems. Automatica 42(6):1033–1040
9. Chen H (2013) Model predictive control. Science Press, Beijing. In Chinese
10. Bemporad A, Morari M, Dua V, Pistikopoulos EN (2002) The explicit linear quadratic regu-
lator for constrained systems. Automatica 38(1):3–20
11. Chen H, Allgöwer F (1998) A quasi-infinite horizon nonlinear model predictive control
scheme with guaranteed stability. Automatica 34(10):1205–1217
References 235

12. Chen H, Gao X-Q, Wang H (2006) An improved moving horizon H∞ control scheme through
Lagrange duality. Int J Control 79(3):239–248
13. Chisci L, Rossiter JA, Zappa G (2001) Systems with persistent disturbances: predictive control
with restricted constraints. Automatica 37(7):1019–1028
14. Grimm G, Messina MJ, Tuna SE, Teel AR (2007) Nominally robust model predictive control
with state constraints. IEEE Trans Autom Control 52(5):1856–1870
15. Grimm G, Messina MJ, Tuna SE, Teel AR (2004) Examples when nonlinear model predictive
control is nonrobust. Automatica 40:1729–1738
16. Lazar M, Muñoz de la Peña D, Heemels W, Alamo T (2008) On input-to-state stabilizing of
min–max nonlinear model predictive control. Syst Control Lett 57(1):39–48
17. Limón D, Álamo T, Salas F, Camacho EF (2006) Input to state stability of min–max MPC
controllers for nonlinear systems with bounded uncertainties. Automatica 42(5):797–803
18. Mayne DQ, Kerrigan EC, van Wyk EJ, Falugi P (2011) Tube-based robust nonlinear model
predictive control. Int J Robust Nonlinear Control 21(11):1341–1353
19. Mayne DQ, Seron MM, Rakovic SV (2005) Robust model predictive control of constrained
linear systems with bounded disturbances. Automatica 41(2):219–224
Appendix E
Linear Matrix Inequality (LMI)

Many problems arising from control, identification and signal processing can be
transformed into a few standard convex or quasi-convex (optimization or feasibility)
problems involving linear matrix inequalities (LMIs) [1, 2] which can be solved
efficiently in a numerical sense by the use of interior-point methods [3]. In this
book, for example, we formulate the calculation of the observer gain in Chap. 2
and the solution of the feedback gain in Chap. 4 as convex optimization problems
involving LMIs.

E.1 Convexity

Definition E.1 2 A set D in a linear vector space is said to be convex if

 
{x1 , x2 ∈ D} ⇒ x : αx1 + (1 − α)x2 ∈ D for all α ∈ (0, 1) . (E.1)

Geometrically, a set D is convex if the line segment between any two points in
D lies in D.

Definition E.2 (Convex hull) The convex hull of a set D, denoted as Co{D}, is the
intersection of all convex sets containing D. If D consists of a finite number of
elements, then these elements are referred to as the vertices of Co{D}.

The convex hull of a finite point set forms a polytope and any polytope is the
convex hull of a finite point set.

Definition E.3 A function f : D → R is called convex if

• D is convex and
• For all x1 , x2 ∈ D and α ∈ (0, 1),
 
f αx1 + (1 − α)x2 ≤ αf (x1 ) + (1 − α)f (x2 ). (E.2)

H. Chen, B. Gao, Nonlinear Estimation and Control of Automotive Drivetrains, 237

DOI 10.1007/978-3-642-41572-2,
© Science Press Beijing and Springer-Verlag Berlin Heidelberg 2014
238 E Linear Matrix Inequality (LMI)

Moreover, f is called strictly convex if the inequality (E.2) is strict for x1 , x2 ∈ D,

x1 = x2 and α ∈ (0, 1).

Geometrically, (E.2) implies that the line segment between (x1 , f (x1 )) and
(x2 , f (x2 )), i.e., the chord from x1 to x2 , lies above the curve of f .
Moreover, a function f : D → R is called affine if (E.2) holds with equality.

Definition E.4 (Local and global optimality) Let D be a subset of a normed

space X . An element x0 ∈ D is said to be a local optimal solution of f : D → R if
there exists  > 0 such that
f (x0 ) ≤ f (x) (E.3)
for all x ∈ D with x − x0 < . It is called a global optimal solution if (E.3) holds
for all x ∈ D.

Proposition E.1 Suppose that f : D → R is convex. If f has a local minimum at

x0 ∈ D, then f (x0 ) is also the global minimum of f . If f is strictly convex, then x0
is, moreover, unique.

Proposition E.2 (Jensen’s inequality) If f defined  on D is convex, then for all

x1 , x2 , . . . , xr ∈ D and λ1 , λ2 , . . . , λr ≥ 0 with ri=1 λi = 1 one has

f (λ1 x1 + · · · + λr xr ) ≤ λ1 f (x1 ) + · · · + λr f (xr ). (E.4)

E.2 Linear Matrix Inequalities

A linear matrix inequality (LMI) is an expression of the form

F (x) := F0 + x1 F1 + · · · + xm Fm > 0, (E.5)

where
• x = (x1 , . . . , xm ) is the decision variable;
• F0 , . . . , Fm are given real symmetric matrices, and
• The inequality F (x) > 0 means that uT F (x)u > 0 for all u ∈ Rn , u = 0.
While (E.5) is a strict LMI, we may also encounter non-strict LMIs which have the
form of
F (x) ≥ 0.
The linear matrix inequality (E.5) defines a convex constraint on x. That is, the
set F := {x|F (x) > 0} is convex. Hence, optimization problems involving the min-
imization (or maximization) of a performance function f : F → R belong to the
class of convex optimization problems, if the performance function f renders (E.2)
satisfied for all x1 , x2 ∈ F and α ∈ (0, 1). The full power of convex optimization
theory can then be employed [4].
E.3 Casting Problems in an LMIs Setting 239

E.3 Casting Problems in an LMIs Setting

There are three generic problems related to LMIs [1, 2]:

(a) Feasibility. The test whether or not there exist solutions x of F (x) > 0 is called
a feasibility problem. The LMI F (x) > 0 is said to be feasible if a solution
exists, otherwise it is said to be infeasible.
(b) Optimization. Let f : D → R be a convex objective function. The problem

min f (x)
x∈D

s.t. F (x) > 0

is called an optimization problem with an LMI constraint.

(c) Generalized eigenvalue problem. This problem amounts to minimizing the
maximum generalized eigenvalue of a pair of matrices that depend affinely on a
variable, subject to an LMI constraint. It admits the general form of

min λ
s.t. λF (x) − G(x) > 0,
F (x) > 0,
H (x) > 0.

Some control problems that can be easily casted in an LMI setting are given as
follows:

Stability An example of the feasibility problem is to test if the linear system

ẋ = Ax (E.6)

is asymptotically stable. This can be formulated as the following LMI feasibility

problem:
P > 0, AT P + P A < 0 (E.7)
with P as a variable. Indeed, with (E.7) feasible, we can easily show that the
quadratic function V (x) = x T P x decreases along every nonzero trajectory of (E.6),
and hence the stability property.
Moreover, if A is uncertain and varies in a polytope, i.e.,

A ∈ Co{A1 , A2 , . . . , Ar },

then the stability test (E.7) becomes

P > 0, ATi P + P Ai < 0, i = 1, 2, . . . , r. (E.8)

240 E Linear Matrix Inequality (LMI)

If the uncertain system is described as diagonal norm-bounded Linear Differen-

tial Inclusions (LDIs) as follows

ẋ = A0 x + Bp p, (E.9a)
q = Cq x + Dqp p, (E.9b)
 
pi = δi (t)qi , δi (t) ≤ 1, i = 1, 2, . . . , nq , (E.9c)

then the stability test (E.7) becomes

 T 
A0 P + P A0 + CqT ΛCq ∗
P > 0, diagonal Λ > 0, < 0,
BpT P + Dqp
T ΛC
q
T ΛD − Λ
Dqp qp
(E.10)
where P and Λ are variable. The S-procedure is used to get (E.10). A less conser-
vative test can be obtained by the use of the full-block S-procedure [2].

Decay Rate The decay rate of a system is defined to be the largest α such that
 
lim eαt x(t) = 0 (E.11)
t→∞

holds for all trajectories. In order to estimate the decay rate of system (E.6), we
define V (x) = x T P x and require

dV (x)
≤ −2αV (x) (E.12)
dt
1
for all trajectories, which then leads to x(t) ≤ e−αt | λλmax (P ) 2
min (P )
| x(0) . By explor-
ing (E.12) for system (E.6), the decay rate problem can be casted in the following
LMI optimization problem

min α (E.13a)
α,P

s.t P > 0, α > 0, AT P + P A + 2αP ≤ 0. (E.13b)

Similarly, we can formulate the problem in an LMI setting for uncertain systems
with polytopic description as follows:

min α (E.14a)
α,P

s.t P > 0, α > 0, ATi P + P Ai + 2αP ≤ 0, i = 1, 2, . . . , r, (E.14b)

and for uncertain systems with diagonal norm-bounded description as follows:

min α (E.15a)
α,P ,Λ

s.t P > 0, α > 0, diagonal Λ > 0, (E.15b)

E.3 Casting Problems in an LMIs Setting 241
 
AT0 P + P A0 + CqT ΛCq + 2αP ∗
< 0. (E.15c)
BpT P + Dqp
T ΛC
q
T ΛD − Λ
Dqp qp

Designing a Feedback Controller The problem of designing a feedback con-

troller can be solved by the general procedure from analysis to synthesis [2]. For
example, design a stabilizing state feedback gain, one can easily perform the fol-
lowing steps:
• Replace A in (E.7) by AK = A + Bu K to get

(A + Bu K)T P + P (A + Bu K) < 0; (E.16)

• Define Q = P −1 and right- and left-multiply with Q and QT , to obtain an equiv-

alent inequality
QAT + QK T Bu + AQ + Bu KQ < 0; (E.17)
• And define Y = KQ to cast the problem in the LMI of

QAT + AQ + Y T BuT + Bu Y < 0 (E.18)

with Q > 0 and Y as variables.

A stabilizing feedback gain is then defined as K = Y Q−1 .
Similarly, by the use of the procedure from analysis to synthesis, one can design
a robust stabilizing feedback gain for the polytopic uncertainty, where (Q, Y ) is a
feasible solution of

Q > 0, QATi + Ai Q + Y T Bu,i

T
+ Bu,i Y < 0, i = 1, 2, . . . , r, (E.19)

or for diagonal norm-bounded uncertainty, where (Q, Y, M) is a feasible solution of

Q > 0, diagonal M > 0, (E.20a)

 
QAT0 + A0 Q + Bp MBpT + Y T BuT + Bu Y ∗
< 0. (E.20b)
Dqp MBpT + Cq Q + Dqu Y T −M
Dqp MDqp

The following lemmas are useful for casting control, identification and signal
processing problems in LMIs.

Lemma E.1 (Schur Complement) For Q(x), R(x), S(x) depending affinely on x
and Q(x), R(x) being symmetric, the LMI
 
Q(v) S(x)
>0
S(v)T R(x)

is equivalent to

Q(v) > 0, R(x) − S(v)T Q(v)−1 S(v) > 0,

242 E Linear Matrix Inequality (LMI)

or to
R(v) > 0, Q(x) − S(v)R(v)−1 S(v)T > 0.

Lemma E.2 (S-Procedure) Let F0 , F1 , . . . , Fp ∈ Rn×n be symmetric matrices. The

requirement of

ξ T F0 ξ > 0 for all ξ = 0 such that ξ T Fi ξ ≥ 0, i = 1, 2, . . . , p,

is satisfied if there exist λ1 , λ2 , . . . , λp ≥ 0 such that


p
F0 − λi Fi > 0.
i=1

For the case p = 1, the converse holds, provided that there is some ξ0 such that
ξ0T F1 ξ0 > 0.

References
1. Boyd S, El Ghaoui L, Feron E, Balakishnan V (1994) Linear matrix inequalities in system and
control theory. SIAM, Philadelphia
2. Scherer CW, Weiland S (2000) Linear matrix inequalities in control. In: Delft center for systems
and control. DISC lecture note, Dutch institute of systems and control
3. Nesterov Y, Nemirovsky A (1994) Interior point polynomial methods in convex programming.
SIAM, Philadelphia
4. Boyd SP, Vandenberghe L (2004) Convex optimization. Cambridge University Press, Cam-
bridge
Appendix F
Subspace Linear Predictor

For a linear time invariant (LTI) system in question, assume that it can be described
in a state-space form as defined by the equations below:

xk+1 = Axk + Buk + Kek , (F.1)

yk = Cxk + Duk + ek , (F.2)

where uk ∈ Rm is the input variable, yk ∈ Rl is the output variable, xk ∈ Rn is

the state variable of the system and ek ∈ Rl is the white noise. The matrices A ∈
Rn×n , B ∈ Rn×m , C ∈ Rl×n , D ∈ Rl×m , and K ∈ Rl×l are the state, input, output,
feed-through, and Kalman gain matrices of the system, respectively.
From the state equation in (F.1), for t = k + 1, k + 2, . . . , k + M, we have

xk+2 = Axk+1 + Buk+1 + Kek+1

= A(Axk + Buk + Kek ) + Buk+1 + Kek+1
   
  uk   ek
= A xk + AB B
2
+ AK K , (F.3a)
uk+1 ek+1
xk+3 = Axk+2 + Buk+2 + Kek+2
 
= A A2 xk + ABuk + Buk+1 + AKek + Kek+1 + Buk+2 + Kek+2
⎡ ⎤ ⎡ ⎤
  uk   ek
= A3 xk + A2 B AB B ⎣ uk+1 ⎦ + A2 K AK K ⎣ ek+1 ⎦,
uk+2 ek+2
(F.3b)
..
.

H. Chen, B. Gao, Nonlinear Estimation and Control of Automotive Drivetrains, 243

DOI 10.1007/978-3-642-41572-2,
© Science Press Beijing and Springer-Verlag Berlin Heidelberg 2014
244 F Subspace Linear Predictor
⎡ ⎤
uk
 ⎢
⎢ uk+1 ⎥
⎥
xk+M = AM xk + AM−1 B AM−2 B ... B ⎢ .. ⎥
⎣ . ⎦
uk+M−1
⎡ ⎤
ek
 ⎢
⎢ e k+1 ⎥
⎥
+ AM−1 K AM−2 K ... K ⎢ .. ⎥, (F.3c)
⎣ . ⎦
ek+M−1

⎡ ⎤
uk+δ
 ⎢
⎢ uk+δ+1 ⎥
⎥
xk+M+δ = AM xk+δ + AM−1 B AM−2 B ... B ⎢ .. ⎥
⎣ . ⎦
uk+M+δ−1
⎡ ⎤
ek+δ
 ⎢ k+δ+1 ⎥
⎢ e ⎥
+ AM−1 K AM−2 K ... K ⎢ .. ⎥. (F.4)
⎣ . ⎦
ek+M+δ−1

Thus, for δ = 0, 1, . . . , N − M + 1, we can then collect the state variables in a

single block matrix equation as
 
xk+M xk+M+1 . . . xk+N +1
 
= AM xk xk+1 . . . xk+N −M+1
⎡ ⎤
uk uk+1 . . . uk+N −M+1
 ⎢
⎢ uk+1 uk+2 . . . uk+N −M+2 ⎥ ⎥
+ AM−1 B AM−2 B · · · B ⎢ .. .. . .. ⎥
⎣ . . . . . ⎦
uk+M−1 uk+M . . . uk+N
⎡ ⎤
ek ek+1 . . . ek+N −M+1
 ⎢
⎢ ek+1 ek+2 . . . ek+N −M+2 ⎥ ⎥
+ AM−1 K AM−2 K . . . K ⎢ .. .. .. .. ⎥.
⎣ . . . . ⎦
ek+M−1 ek+M ... ek+N
(F.5)

Next we will look at the output equation (F.2) and develop recursively an output
matrix equation. From (F.2), for t = k + 1, k + 2, . . . , k + M − 1, we have

yk+1 = Cxk+1 + Duk+1 + ek+1

= C(Axk + Buk + Kek ) + Duk+1 + ek+1
F Subspace Linear Predictor 245
   
  uk   ek
= CA1 xk + CB D + CK I , (F.6a)
uk+1 ek+1
yk+2 = Cxk+2 + Duk+2 + ek+2
 
= C A2 xk + ABuk + Buk+1 + AKek + Kek+1 + Duk+2 + ek+2
⎡ ⎤ ⎡ ⎤
  uk   ek
= CA2 xk + CAB CB D ⎣ uk+1 ⎦ + CAK CK I ⎣ ek+1 ⎦,
uk+2 ek+2
(F.6b)
..
.
⎡ ⎤
uk
 ⎢
⎢ uk+1 ⎥
⎥
yk+M−1 = CAM−1 xk + CAM−2 B CAM−3 B ... D ⎢ .. ⎥
⎣ . ⎦
uk+M−1
⎡ ⎤
ek
 ⎢
⎢ e k+1 ⎥
⎥
+ CAM−2 K CAM−3 K ... I ⎢ .. ⎥. (F.6c)
⎣ . ⎦
ek+M−1

Compiling the above result for output equations into a single matrix equation gives
us
⎡ ⎤ ⎡ ⎤
yk C
⎢ yk+1 ⎥ ⎢ CA ⎥
⎢ ⎥ ⎢ ⎥
⎢ yk+2 ⎥ ⎢ CA2 ⎥
⎢ ⎥=⎢ ⎥xk
⎢ .. ⎥ ⎢ .. ⎥
⎣ . ⎦ ⎣ . ⎦
yk+M−1 CA M−1
⎡ ⎤⎡ ⎤
D 0 0 ... 0 uk
⎢ CB D 0 ... 0 ⎥ ⎢ ⎥
⎢ ⎥⎢ uk+1 ⎥
⎢ CAB CB D ⎢
. . . 0 ⎥⎢ uk+2 ⎥
⎥
+⎢ ⎥
⎢ .. .. .. .. . ⎥⎢ .. ⎥
⎣ . . . . .. ⎦⎣ . ⎦
CAM−2 B CAM−3 B CAM−4 B ... D uk+M−1
⎡ ⎤⎡ ⎤
I 0 0 ... 0 ek
⎢ CK I 0 ... 0⎥ ⎢ ⎥
⎢ ⎥⎢ ek+1 ⎥
⎢ CAK CK I ... 0⎥ ⎢ ek+2 ⎥
+⎢ ⎥⎢ ⎥.
⎢ .. .. .. .. .. ⎥⎢ .. ⎥
⎣ . . . . . ⎦ ⎣ . ⎦
CAM−2 K CAM−3 K CAM−4 K ... I ek+M−1
(F.7a)
246 F Subspace Linear Predictor

Adjusting the time for the variables in (F.7a) by an arbitrary discrete time δ will
result in a similar matrix equation

⎡ ⎤
yk+δ
⎢ yk+1+δ ⎥
⎢ ⎥
⎢ yk+2+δ ⎥
⎢ ⎥ (F.8)
⎢ .. ⎥
⎣ . ⎦
yk+M+δ−1
⎡ ⎤
C
⎢ CA ⎥
⎢ ⎥
⎢ 2 ⎥
= ⎢ CA ⎥xk+δ
⎢ .. ⎥
⎣ . ⎦
CAM−1
⎡ ⎤⎡ ⎤
D 0 0 ... 0 uk+δ
⎢ CB D 0 ... 0⎥ ⎢ ⎥
⎢ ⎥⎢ uk+1+δ ⎥
⎢ 0 ⎥⎢ uk+2+δ ⎥
⎥ ⎢
+ ⎢ CAB CB D ... ⎥
⎢ .. .. .. .. .. ⎥⎢ .. ⎥
⎣ . . . . . ⎦ ⎣ . ⎦
CAM−2 B CAM−3 B CAM−4 B ... D uk+M+δ−1
⎡ ⎤⎡ ⎤
I 0 0 ... 0 ek+δ
⎢ CK I 0 ... 0⎥ ⎢ ⎥
⎢ ⎥⎢ ek+δ+1 ⎥
⎢ CAK CK I ... 0⎥ ⎢ ek+δ+2 ⎥
+⎢ ⎥⎢ ⎥. (F.9)
⎢ .. .. .. .. .. ⎥⎢ .. ⎥
⎣ . . . . . ⎦⎣ . ⎦
CAM−2 K CAM−3 K CAM−4 K ... I ek+M+δ−1

Therefore, by collecting the variables for δ = 0, 1, . . . , N − M + 1, we can compose

the output equations in a single block matrix equations as shown below:

⎡ ⎤
yk yk+1 ... yk+N −M+1
⎢ yk+1 yk+2 ... yk+N −M+2 ⎥
⎢ ⎥
⎢ yk+2 yk+3 ... yk+N −M+3 ⎥
⎢ ⎥
⎢ .. .. .. .. ⎥
⎣ . . . . ⎦
yk+N −M+1 yk+M ... yk+N
⎡ ⎤
C
⎢ CA ⎥
⎢ ⎥ 
⎢ 2 ⎥
= ⎢ CA ⎥ xk xk+1 ... xk+N −M+1
⎢ .. ⎥
⎣ . ⎦
CAM−1
F Subspace Linear Predictor 247
⎡ ⎤
D 0 0 ... 0
⎢ CB D 0 ... 0⎥
⎢ ⎥
⎢ CAB CB D ... 0⎥
+⎢ ⎥
⎢ .. .. .. .. .. ⎥
⎣ . . . . . ⎦
CAM−2 B CAM−3 B CAM−4 B ... D
⎡ ⎤
uk uk+1 ... uk+N −M+1
⎢ uk+1 uk+2 ... uk+N −M+2 ⎥
⎢ ⎥
⎢ uk+2 uk+3 ... uk+N −M+3 ⎥
×⎢ ⎥
⎢ .. .. .. .. ⎥
⎣ . . . . ⎦
uk+N −M+1 uk+M ... uk+N
⎡ ⎤
I 0 0 ... 0
⎢ CK I 0 ... 0⎥
⎢ ⎥
⎢ CAK CK I ... 0⎥
+⎢ ⎥
⎢ .. .. .. .. .. ⎥
⎣ . . . . .⎦
CAM−2 K CAM−3 K CAM−4 K ... I
⎡ ⎤
ek ek+1 ... ek+N −M+1
⎢ ek+1 ek+2 ... ek+N −M+2 ⎥
⎢ ⎥
⎢ ek+2 ek+3 ... ek+N −M+3 ⎥
×⎢ ⎥. (F.10)
⎢ .. .. .. .. ⎥
⎣ . . . . ⎦
ek+N −M+1 ek+M ... ek+N

From the derivation of Eqs. (F.5) and (F.10), we can write the subspace I/O matrix
equations in the field of subspace system identification [1] as follows:

Yp = ΓM Xp + HM
d
Up + HNs Ep , (F.11)
Yf = ΓM Xf + HM
d
Uf + HNs Ef , (F.12)
Xf = AM Xp + dM Up + sM Ep , (F.13)

where the subscripts p and f denote the ‘past’ and ‘future’ matrices of the respective
variables, the superscripts d and s stand for the deterministic and stochastic part of
the system, respectively. Open-loop data uk and yk , k ∈ {0, 1, . . . , N} are available
for identification. Therefore, for the definition in (F.11)–(F.13), the past and future
data matrices are constructed as follows:
⎡y y2 ... yN −2M+1 ⎤ ⎡y yM+2 ... yN −M+1 ⎤
1 M+1
⎢ y2 y3 ... yN −2M+2 ⎥ ⎢ yM+2 yM+3 ... yN −M+2 ⎥
⎢
Yp = ⎣ . .. .. ⎥, Yf = ⎢ .. .. .. ⎥,
.. ..
. ⎦ ⎣ ..
. ⎦
. . . . .
yM yM+1 ... yN −M y2M y2M+1 ... yN
(F.14)
248 F Subspace Linear Predictor
⎡u u2 ... uN −2M+1 ⎤ ⎡u uM+2 ... uN −M+1 ⎤
1 M+1
⎢ u2 u3 ... uN −2M+2 ⎥ ⎢ uM+2 uM+3 ... uN −M+2 ⎥
⎢
Up = ⎣ . .. .. ⎥, Uf = ⎢ .. .. .. ⎥,
.. ..
. ⎦ ⎣ ..
. ⎦
. . . . .
uM uM+1 ... uN −M u2M u2M+1 ... uN
(F.15)
⎡e e2 ... uN −2M+1 ⎤ ⎡e eM+2 ... eN −M+1 ⎤
1 M+1
⎢ e2 e3 ... eN −2M+2 ⎥ ⎢ eM+2 eM+3 ... eN −M+2 ⎥
Ep = ⎢
⎣ .. .. .. .. ⎥,
⎦ Ef = ⎢
⎣ .. .. .. .. ⎥.
⎦
. . . . . . . .
eM eM+1 ... eN −M e2M e2M+1 ... eN
(F.16)

The matrices ΓM , dM and sM are defined as follows:

• Extended observability matrix
⎡ ⎤
C
⎢ CA ⎥
⎢ ⎥
ΓM =⎢ 2 ⎥
⎢ CA ⎥; (F.17)
⎣ ··· ⎦
CAM−1

• Reversed extended controllability matrix (deterministic)

 
dM = AM−1 B AM−2 B . . . B ; (F.18)

• Reversed extended controllability matrix (stochastic)

 
sM = AM−1 K AM−2 K . . . K . (F.19)
d and H s are given below:
The lower-triangular Toeplitz matrices HM M
⎡ ⎤
D 0 0 ... 0
⎢ CB D 0 ... 0⎥
⎢ ⎥
⎢ CAB CB D ... 0⎥
d
HM =⎢ ⎥, (F.20)
⎢ .. .. .. .. .. ⎥
⎣ . . . . . ⎦
CAM−2 B CAM−3 B CAM−4 B ... D
⎡ ⎤
I 0 0 ... 0
⎢ CK I 0 ... 0⎥
⎢ ⎥
⎢ CAK CK I ... 0⎥
s
HM =⎢ ⎥. (F.21)
⎢ .. .. .. .. .. ⎥
⎣ . . . . .⎦
CAM−2 K CAM−3 K CAM−4 K ... I

Furthermore, the past and future state matrices are also defined by
 
Xp = x1 x2 . . . xN −2M+1 , (F.22)
F Subspace Linear Predictor 249
 
Xf = xM+1 xM+2 ... xN −M+1 . (F.23)

Taking Eq. (F.11) and solving for Xp will render us

 
Xp = ΓM† Yp − HM
d
Up − HM
s
Ep , (F.24)

where the subscript “†” denotes Moore–Penrose pseudoinverse of a matrix [2]. Sub-
stituting Eq. (F.24) into (F.13) will then give
  
Xf = AM ΓM† Yp − HM d
Up − HM
s
Ep + dM Up + sM Ep
   
= AM ΓM† Yp + dM − AM ΓM† HM
d
Up + sM − AM ΓM† HM
s
Ep . (F.25)

Therefore, substituting Eq. (F.25) into (F.12) will result in an equation for future
output as given below:
  
Yf = ΓM AM ΓM† Yp + dM − AM ΓM† HM d
Up
 s  
+ M − AM ΓM† HM s
Ep + HMd
Uf + HNs Ef
 
= ΓM AM ΓM† Yp + ΓM dM − AM ΓM† HM d
Up
 
+ HM d
Uf + ΓM sM − AM ΓM† HMs
Ep + HNs Ef . (F.26)

Due to the effect of Ef which is stationary white noise, and by the virtue of the
stability of a Kalman filter, for a set of measurements that is sufficiently large,
Eq. (F.26) above can then be written to give an optimal prediction of Yf as follows:
 
  Wp
Ŷf = Lw Lu = L w W p + L u Uf (F.27)
Uf

where “  ” denotes the estimate and

 
Up
Wp = . (F.28)
Yp

Equation (F.27) is thus known as the subspace linear predictor equation, with Lw
being the subspace matrix that corresponds to the past input and output data ma-
trix Wp , and Lu is the subspace matrix that corresponds to the future input data
matrix Uf .
In order to calculate the subspace linear predictor coefficients Lw and Lu from
the Hankel data matrices Up , Yp and Uf , we will solve the following least squares
problem, thus giving us the prediction equation for Yf :
  
   W p 2
min Yf − Lw Lu  , (F.29)
Lw ,Lu  Uf 
250 F Subspace Linear Predictor

where
 †
  Wp
Lw Lu = Yf
Uf
!  W  !−1
p
= Yf WpT UfT WpT UfT . (F.30)
Uf

In the control implementation, only the leftmost column of the matrix will be used
for the prediction of future output values. Therefore, after the subspace linear pre-
dictor coefficients Lw and Lu are found from the identification data, we can then
streamline equation (F.27) by taking only the leftmost column of matrices Ŷf , Yp ,
Up and Uf by defining
⎡ ⎤ ⎡ ⎤
yt+1 yt−M+1
⎢ yt+2 ⎥ ⎢ yt−M+2 ⎥
⎢ ⎥ ⎢ ⎥
ŷf = ⎢ . ⎥, yp = ⎢ .. ⎥,
⎣ . ⎦. ⎣ . ⎦
yt+M yt
⎡ ⎤ ⎡ ⎤ (F.31)
ut−M+1 ut+1
⎢ ut−M+2 ⎥ ⎢ ut+2 ⎥
⎢ ⎥ ⎢ ⎥
up = ⎢ .. ⎥, uf = ⎢ . ⎥,
⎣ . ⎦ ⎣ .. ⎦
ut ut+M

and
 
up
wp = , (F.32)
yp
then we will arrive at a streamlined subspace-base linear predictor equation, namely

ŷf = Lw wp + Lu uf . (F.33)

According to Eq. (F.33), we can predict the output of the system by the past input
and output data as well as the future input data that is applied. This result will be
utilized in the implementation of model predictive control algorithm, that is, data-
driven predictive control algorithm.

References
1. Overschee PV, Moor BD (1996) Subspace identification for linear systems: theory, implemen-
tation, applications. Kluwer Academic, Norwell
2. Van Overschee P, De Moor B (1995) A unifying theorem for three subspace system identifica-
tion algorithms. Automatica 31(12):1853–1864