Robust Nonlinear Forwarding Design

Weizhou Su, Minyue Fu

Department of Electrical and Computer Engineering
The University of Newcastle, NSW 2308, Australia
Email: [email protected]; [email protected]

Abstract. This paper considers the problem of robust In [5] and [6], a robust control design approach is de-
stabilization for a class of nonlinear up-augmented sys- veloped for an upper-triangular nonlinear system with
tems with uncertain parameters in a very general form. large size uncertainties in both local linearizied model
This system involves a base system with a control input and global nonlinear model of the system. Apart from
and a forwarding structure. Several robust control design providing a robust control design method, the results in
methods are discussed for the cases in which these parts [5] and [6] also bridge a gap between linear robust control
of the system are in SISO and MIMO forms, respectively. theory and nonlinear robust control theory in the sense
The main assumptions required for these methods are that the design method coincides with the seminal work
quadratic stabilizability for the local linearized model of of Wei [9] on quadratic stabilization of linear systems.
the system, global asymptotic stabilizability for the base
system and some mild conditions on the up-augmentation In this paper, we will review the existing results for the ro-
and the nonlinearity of the system. bust stabilization problem of (1). Then the robust design
method in [5] and [6] is generalized to the system (1). The
Keywords: Nonlinear control; Forwarding; Robust con- main advantage of this method is that the key condition
trol; Quadratic stabilization. for robust stabilization of (1) boils down to quadratic sta-
bilizibility of the locally linearized model. Since there are
many well-established methods, such as linear quadratic
stabilization theory [2], H∞ control [1], Wei’s method [9],
1 Introduction etc. to solve quadratic stabilization problem for a wide
class of linear uncertain systems, the system satisified this
In this paper, we address the robust stabilization prob- assumption is much more general than the systems stud-
lem for a class of cascaded nonlinear systems with the ied in existing literature. The other advantage of this
so-called forwarding or feedforwarding structure. More method is that a simpler design method for robust con-
precisely, these systems have the following model: trollers and a simpler proof of the stabilizability of the
system are provided. Our design method considers two
ẋ1 = f1 (x1 , x2 , q) + g1 (x1 , x2 , u, q) notions of stability: global asymptotic stability and local
(1)
ẋ2 = f2 (x2 , q) + g2 (x2 , u, q) quadratic stability. The latter requires the existence of
where x1 ∈ Rn1 and x2 ∈ Rn2 are state variables, a locally quadratic Lyapunov function for the system (1)
u ∈ Rm is a control input, q is an uncertain parame- for all admissible uncertain parameters q ∈ Q. The sta-
ter vector belonging to a compact set Q ⊂ Rp , nonlinear bilizing controller to be proposed involves two steps: The
vector functions f1 (x1 , x2 , q), f2 (x2 , q), g1 (x1 , x2 , q) and first controller brings the state of the base system, x2 ,
g2 (x2 , q) are smooth in x1 , x2 and u and continuous in to a small neighbourhood of the origin, and the second
q, f1 (0, 0, q) ≡ 0, f2 (0, q) ≡ 0, g1 (x1 , x2 , 0, q) ≡ 0 and controller is used to bring both x1 and x2 to the origin.
g2 (x2 , 0, q) ≡ 0 for x1 ∈ Rn1 , x2 ∈ Rn2 and ∀q ∈ Q. A non-quadratic Lyapunov function is used to design this
robust controller.
Many similar cases of (1) have been studied; see, e.g.,
[3, 4, 5, 6, 8]. The work in these papers have led to sev-
eral design control methods. In [8], a saturation function
and input-output method are used to design a global as-
ymptotic stabilizer for an upper-triangular system. In
this case, the state variable x1 in (1) is a scalar vari- 2 Forwarding Control Design via
able. In [3, 4], cascaded systems similar to that in (1) are
studied and Lyapunov function based design methods are Lyapunov Functions
proposed. However, the aforementioned design methods
rely on very accurate knowledge of the system. The only
exception is [8] where uncertainties are allowed in high In this section, we review design methods given in [3, 4].
order (nonlinear) terms. Consider the following system (which is a special case of

1
the system (1)) rewritten as where the functions l(·) and k(·) are nonlinear weight
functions to be choosen.
ẋ1 = f1 (x1 , x2 ) + g1 (x1 , x2 )u
(2)
ẋ2 = f2 (x2 ) + g2 (x2 )u Under assumptions similar to Assumptions 2.1-2.3 and
Since f1 (0, 0) = 0 and f1 (x1 , x2 ) are smooth in x1 , x2 , we some local stabilizibility of the system, a global stabi-
can decompose f1 (x1 , x2 ) as follows: lizier is designed in [4] by applying the Lyapunov function
(9) and selecting approperiate functions l(·) and k(·). A
f1 (x1 , x2 ) = f¯1 (x1 ) + f12 (x1 , x2 )x2 . drawback in the methods in [3, 4] is that the Lyapunov
functions (7) and (9) are dependent on the precise knowl-
where f¯1 (x1 ) = f1 (x1 , 0) and f12 (x1 , x2 ) is a smooth func-
edge of the model of the system (2).
tion. Denote h(x1 , x2 ) = f12 (x1 , x2 )x2 .

Suppose the system (2) satisfies assumptions below:

3 Forwarding Design via Satura-
Assumption 2.1 Suppose W1 (x1 ) is a Lyapunov func-
tion of the system tion Functions
ẋ1 = f¯1 (x1 ) (3)
In this section, we will discuss another forwarding design
and Lf¯W1 (x1 ) ≤ 0. Moveover, there holds method in [8] which uses saturation functions. Compared
∂W1 with the results in the last section, the forwarding control
kx1 k ≤ kW1 (x1 ); ∀kxk ≥ c (4) via saturation function needs more structure information
∂x1
of the system (1). Suppose the system (1) without un-
for some constants k > 0 and c > 0. certain parameters q satisfies the following assumptions:

Assumption 2.2 The system Assumption 3.1 The state x1 is a scalar state variable
and f1 (x1 , x2 ) = x21 + f¯11 (x2 ). The functions f¯11 (x2 ) and
ẋ2 = f¯2 (x2 ) (5) g1 (x2 , u) involve only quadratic and higher order terms
is asymptotically stable and W2 (x2 ) is a Lyapunov func- of x21 , · · · , x2n2 and u, where x2 = (x21 , · · · , x2n2 )T .
tion of the system (5).
Assumption 3.2 Denote f2 (x2 ) = (f21 , · · · , f2n2 )T and
Assumption 2.3 There exist K functions γ1 (·) and γ2 (·) g2 (x2 , u) = (g21 , · · · , g2n2 )T . Then, for i = 1, · · · , n2 − 1,
such that
f2i = x2i + f¯2i (x2(i+1) , · · · , x2n2 ),
kh(x1 , x2 )k ≤ γ1 (x2 ) + kx1 kγ2 (x2 ) (6) g2i = x2i + ḡ2i (x2(i+1) , · · · , x2n2 , u)
where γ1 and γ2 are differentable at the origin.
where f¯2i and ḡ2i only have quadratic and higher order
terms of x2(i+1) , · · · , x2n2 and u only. Further,
In [3], under Assumptions 2.1-2.3 and some slightly
stronger assumption in the local structure of the system f2n2 = 0; g2n2 = u + ḡ2n2 (u).
(3) (see the assumption A3 in [3]), a Lyapunov function
V (x1 , x2 ) = W1 (x1 ) + W2 (x2 ) + Ψ(x1 , x2 ). (7) where ḡ2n2 (u) have quadratic and higher order terms of
u only.
is constructed by a design method, called cross term tech-
nique, for the system (2). The cross term Ψ(x1 , x2 ) is in Under Assumptions 3.1-3.2, there exists a linear non-
the form as below singular coordinate transformation (x, x21 , · · · , x2n2 ) →
Z ∞ (ζ1 , · · · , ζn2 +1 ) such that, in the ζ-coordinate the system
Ψ(x1 , x2 ) = Lh W1 (x̄1 (τ, x1 , x2 ), x̄2 (τ, x2 ))dτ (8) (1) can be written as
0

where x̄1 (τ, x1 , x2 ), x̄2 (τ, x2 ) are solution of (2) with ini- ζ̇ = Aζ + Bu + Φ (10)
tial state (x1 , x2 ) and u ≡ 0.
where
Applying the Lyapunov function (7), it has been proven 
0 1 ··· 1
  
1
in [3] that, under Assumptions 2.1-2.3 and some local .. .. .. .  .. 
. .. 

stabilizibility of the system (see the assumption A4 in [3]), A=
 . .  B = .
 
(11)
the system (2) is globally asymptotically stabilizable.  0 · · · 0 1 1
0 ··· 0 0 1
In [4], Mazenc and Praly obtained a Lyapunov function
in form and

ΦT = φ1 (ζ2 , · · · , ζn2 +1 , u) · · · φn2 +1 (u)

 
V (x1 , x2 ) = l(W1 (x1 )) + k(W2 (x2 )) (9) (12)

2
where φi , i = 1, · · · , n2 + 1 contains only quadratic and Lyapunov function, a robust controller can be designed
higher order terms of ζi+1 , · · · , ζn2 +1 and u (see [8] for for a much more general uncertain linear system (see, [9]).
details). So there is a big gap between linear and nonlinear systems
in terms of robust stabilization techniques. In this sec-
Denote σi (τ ), i = 1, · · · , n2 + 1 are continuous, nonde- tion we will merge this gap by using a local nonquadratic
creasing saturation functions such that Lyapunov function and a multi-step controller [5, 6].

Suppose the system (1) satisfies following assumptions:

1. σi (τ ) = τ ; |τ | < δi ;
2. σi (τ )| < εi for all τ ∈ R Assumption 4.1 The state x1 is a scalar state vari-
able. The function f1 (x1 , x2 , q) can be decomposed as
f1 (x1 , x2 , q) = θ(q)x21 + f¯11 (x2 , u, q) where θ(q) > 0,
where εi > δi . ∀q ∈ Q and f¯11 (x2 , u, q) has quadratic or higher order
terms of x21 only.
Select 1 >> εn2 +1 >> εn2 >> · · · >> ε1 and let con-
troller u such that Denote x2,n2 +1 = u is a new state variable and v is a new
control input. And also denote x̄2 = (x2 x2,n2 +1 )T . Then
u = −σn2 +1 (ζn2 +1 + σn2 (ζn2 + · · · + σ1 (ζ1 ) · · · )) (13)
the system (1) can be written as
Applying the controller (13) into (10), the state variable
ẋ1 = θ(q)x21 + f˜11 (x̄2 , q)
ζn2 +1 is driven into the region [−δn2 +1 , δn2 +1 ] after a fi- (16)
x̄˙ 2 = f¯2 (x̄2 , q) + bv
nite time tn2 +1 and the states ζn2 and ζn2 +1 satisfy fol-
lowing equations
where f˜11 (x̄2 , q) = f¯11 (x̄2 , q) + g1 (x̄2 , q) and
ζ̇n2 = −σn2 (ζn2 + · · · + σ1 (ζ1 ) · · · )    
(14) f (x̄ , q) g2 (x̄2 , q) 0
ζ̇n2 +1 = −ζn2 +1 − σn2 (ζn2 + · · · + σ1 (ζ1 ) · · · ) f¯2 (x̄2 , q) = 2 2 ; b=
1 0 1
In general, the controller (13) is a multi-step controller.
Assumption 4.2 The system
In every step, a new state of the system is driven into
a given samll region by the controller while the previous x̄˙ 2 = f¯2 (x̄2 , q) + bv (17)
states are kept in a given small region by the controller.
Since these states are very small so that this part of the is globally asymptotically stable and locally quadratically
system can be modelled as a linear system. Further it is stable. And
stabilizied by the linear part of the controller (13). In the
n2 + 1 step, there is a finite time t1 such that, after the V2 (x̄2 ) = x̄T2 P2 x̄2 (18)
finite time t, it holds
is a local quadratic Lyapunov function of the system (17)
ζ̇1 = −ζ1 in the region
..
. (15) Ω = { x̄2 | V2 (x̄2 ) < µ } (19)
ζ̇n2 = −ζ1 − · · · − ζn2
ζ̇n2 +1 = −ζ1 − · · · − ζn2 +1 where µ is a given positive constant.

Obviously, the system (15) is asymptotically stable. Assumption 4.3 Denote A2 (q)x̄2 is the local lineariza-
Thus, under Assumption 3.1-3.2, the system (1) is glob- tion of f¯2 (x̄2 , q). Then matrix A2 (q) has a structure as
ally asymptotically stabilizable if the system is without below:
uncertain parameters.  
0 A−
2 (q)
A2 (q) =
It should also be noted that the forwarding design via ? ?
saturation function allows some uncertainties in the sys-
where A−2 (q) is some uncertain matrix and ? can be some
tem but the precise locally linearized model of the system
uncertain scalar or vector elements.
is necessary.
In fact, the structure in Assumption 4.3 includes the
upper-triangular structure in Assumption 3.2. Thus,
4 Robust Forwarding Deisgn compared with Assumption 3.2, Assumption 4.3 is
weeker. Further, Assumption 4.1-4.3 allow uncertain pa-
rameter q to enter the system, so the controller to be
As we have seen in Section 3, using a very special struc- designed in this section will have more robustness.
ture assumption and the precise local linearizied model of
the system (1), this system can be stabilizied by a multi- Assumptions 4.2 leads to that the x̄2 converges into Ω in
step controller. On the other hand, applying a quadratic finite time while keeping x1 bounded. Hence, we assume

3
in the sequel that x̄2 (0) ∈ Ω, where x̄2 (0) is the initial Assumption 5.1 (Local Quadratic Stabilizibility):
value of x̄2 (t). Choose There exist a linear state feedback matrix
Z V2 (x̄2 ) K = [K1 K2 ]
V (x1 , x̄2 ) = (x1 − (γ 0)P x̄2 )2 + s(w)dw (20)
0 and a symmetric and positive-definite matrix P0 such that
as a local Lyapunov function for the system (1), where γ is
P0 [A(q) + B(q)K] + [A(q) + B(q)K]T P0 < 0, ∀q ∈ Q
a negative constant to be choosen and s(w) is a positive,
(24)
smooth, and monotonically non-decreasing function for
w ∈ [0, µ), with Without loss of generality, we let
Z V2  
P1 −P1 W
s(w)dw → ∞; as V2 → µ. (21) P0 = . (25)
0 −W T P1 P2 + W T P1 W

for some P1 = P1T > 0 and P2 = P2T > 0 and W . Equiv-

Remark 4.1 A particular choice of s(·) is given by alently, the quadratic Lyapunov function is given by
µ V0 (x1 , x2 ) = (x1 − W x2 )T P1 (x1 − W x2 ) + xT2 P2 x2
s(w) = .
µ−w (26)
In general, the Lyapunov function (20) is non-quadratic. Assumption 5.2 (Global Asymptotic Stabilizability of
However, as x → 0, V (x) becomes quadratic in x because the Base System): There exists a locally smooth con-
R V (x )
s(0) > 0. We also note that the function 0 2 2 s(w)dw troller u0 (x2 ) such that the system below
resembles a “potential barrier” and the Lyapunov func-
tion (20) is valid only for x2 ∈ Ω, i.e., x˙2 = f2 (x2 , q) + B2 (q)u0 (x2 ) (27)

V (x1 , x2 ) → ∞ as V2 (x2 ) → µ. (22) is globally asymptotically stable.

This implies that future x2 ∈ Ω as long as that V (x1 , x2 ) Assumption 5.3 (Local Boundedness of the Forwarding
remains bounded. 2 State): The matrix P1 in Assumption 2.1 is such that

P1 A1 (q) + AT1 (q)P1 ≤ 0; ∀q ∈ Q. (28)

Applying the Lyapunov function (20), a robust controller
is designed for the system (1) such that the system is ro- Assumption 5.4 (Smoothness Conditions): We require
bustly globally asymptotically stable and locally quadrat- f1 (x1 , x2 , q) = A1 (q)x1 + A12 (q)x2 + F12 (x, q)x2 (29)
ically stable (see [5, 6]).

f2 (x2 , q) = A2 (q)x2 + F2 (x2 , q)x2 (30)

5 Extension of Robust Forwaring for some F2 (x2 , q) which is continuous in q and smooth
in x2 with F2 (0, q) = 0, ∀q ∈ Q, and F12 (x, q) which is
continuous in q and smooth in x and satisfies
We new extend the robust forwarding method discussed
in last section to a more general form. max kF12 (x, q)k ≤ γ1 (x2 )kx1 k + γ2 (x2 ) (31)
q∈Q

Denote the local linearized model of the system (1) as with some smooth functions γi (x2 ), i = 1, 2.
ẋ = A(q)x + B(q)u (23) Remark 5.1: It is obvious that Assumption 5.1 is neces-
sary for local quadratic stabilization. Since the quadratic
where
stabilization theory for linear uncertain system is well
established, we will not discuss methods of solving the
     
x1 A1 (q) A12 (q) B1 (q)
x= ; A(q) = ; B(q) = . linear quadratic control u = Kx and quadratic Lyapunov
x2 0 A2 (q) B2 (q)
function V0 (x1 , x2 ).
Without loss of generlity, we assume B1 (q) ≡ 0, ∀q ∈ Q.
Otherwise, we can extend x2 to a new state x̄2 by in- However, the condition in (28) is required for technical
troducing the control input u as a part of the new state. reasons (in the proof of Theorem 5.1). To justify this
As we showed in the last section, the control input ma- condition, we note several points: 1) When the forward-
trix corresponding to x1 will be equal to zero in the new ing state x1 is a scalar, this condition merely requires
system. A1 (q) to be non-positive. In fact, A1 (q) = 0 in the upper-
triangular structure. 2) When x1 is not a scalar, a con-
Suppose the system (1) satisfies assumptions below: dition similar to (28) is often used; see [4, 3]. 3) To show

4
that stabilizability may be impossible without (28), we with its inverse given by
consider the following example:
P + s−1 (V2 )W P2−1 W T
 −1 
s−1 (V2 )W P2−1
S= 1 . (40)
ẋ1 = x1 + 2x2 + x22 ,  > 0 s−1 (V2 )P2−1 W T s−1 (V2 )P2−1
(32)
ẋ2 = u

It is easy to verify that its local linearized model is stabi- Consider the following local controller:
lizable. However, 2x2 + x22 ≥ −1, implying that x1 (t) will
diverge if x1 (0) > 1/ regardless what control is used. ul (x1 , x2 ) = s−1 (V2 (x2 ))K1 x1
(41)
+[K2 + (1 − s−1 (V2 (x2 )))K1 W ]x2
Finally, we point out that Assumption 2.3, (29) and (31)
guarantee that x1 is bounded for bounded x2 (see [3]). 2 Then we have the following main result:

Now, we will pay attention to design a robust controller Theorem 5.1: Suppose the system (1) satisfies Assump-
for the system (1) under Assumptions 5.1-5.4. The closed- tions 5.1-5.4. Then the closed-loop system controlled by
loop system is required to have both global asymptotic the following controller
stability and local quadratic stability. 
u0 (x2 ), x2 ∈
/Ω
u= (42)
ul (x1 , x2 ), x2 ∈ Ω
From Assumption 5.1, we choose
is robustly globally asymptotically stable and locally
V2 (x2 ) = xT2 P2 x2 (33)
quadratically stable.
as a local quadratic Lyapunov function for the base sys-
tem and define a local region Ω as: Proof: As discussed before, we only need to consider the
case when x2 (0) ∈ Ω and ul (x1 , x2 ) is applied. Using the
Ω = {x2 |V2 (x2 ) < µ} (34) nonlinear coordinate transformation

where µ > 0 is to be specified. Denote z = [z1T z2T ]T = P x, (43)

 
A1 (q) A12 (q) + F12 (x, q) the equation (38) becomes
A(x, q) = . (35)
0 A2 (q) + F2 (x2 , q)
V̇ (x1 , x2 ) = z T [A(x, q)S + SAT (x, q)]z + 2z T B(q)u. (44)
From Assumptions 5.1 and 5.4, we know that the follow-
ing holds for sufficiently small µ > 0 and ε > 0: Denote s−1 (V2 (x2 )) by s−1 and let

P0 [A(x, q) + B(q)K] + [A(x, q) + B(q)K]T P0 < −εI; u = s−1 K1 x + K2 x2 + v. (45)

(36)
Then,
for all x2 ∈ Ω.
V̇ (x1 , x2 ) = 2z T (A(x, q) + B(q)K)Sz
Assumptions 5.2 and 5.4 guarantees that the state x2 −2(1 − s−1 )z T B(q)[K1 0]Sz + 2z T B(q)v(46)
can be driven into the region Ω in finite time by the con-
troller u0 (x) while keeping x1 bounded. Hence, we as- Rewrite (36) as
sume that x2 (0) ∈ Ω, where x2 (0) is the initial value of
x2 (t). Choose [A(x, q) + B(q)K]S0 + S0 [A(x, q) + B(q)K]T < −S02 ,
∀x ∈ Rn1 × Ω; ∀q ∈ Q (47)
Z V2 (x2 )
T
V (x1 , x2 ) = (x1 − W x2 ) P1 (x1 − W x2 ) + s(w)dw
0 where S0 = P0−1 . Rewriting S as
(37)  
P −1 0
S=s −1
S0 + (1 − s −1
) 1
as a local Lyapunov function for the system (1), where 0 0
s(w) is same as the function s(w) in (20).
we simplify (46) to
For x2 ∈ Ω, the derivative of V (x1 , x2 ) along the trajec-
tory of the system (1) is given by V̇ (x1 , x2 )
= 2s−1 z T (A(x, q) + B(q)K)S0 z
V̇ (x1 , x2 ) = xT [P A(x, q) + AT (x, q)P ]x + 2xT P Bu +2(1 − s−1 )z1T [A1 (q)P1−1 ]z1
(38)
−2(1 − s−1 )s−1 z1T B(q)K1 W x2 + 2z T B(q)v
where
  Applying Assumption 2.3 and choosing
P1 −P1 W
P = (39)
−W T P1 s(V2 )P2 + W T P1 W v = (1 − s−1 )K1 W x2 , (48)

5
we have [4] F. Mazenc and L. Praly, “Adding integrations, sat-
−1 T
urated controls, and stabilization for feedforward-
V̇ (x1 , x2 )= 2s z (A(x, q) + B(q)K)S0 z ing systems”. IEEE Trans. Aut. Contr, AC-41, pp.
−1 T
= 2s x P0 (A(x, q) + B(q)K)x (49) 1559-1578, 1996.
From Assumption 2.1, we can see that [5] W. Su and M. Fu, “Robust stabilization of a class
of nonlinear systems upper-triangular structure,”
V̇ (x1 , x2 ) < 0; ∀x ∈ Rn1 × Ω − {0, 0} . (50) Proceedings of IEEE Conference on Decision and
This implies that Control, Tampa, Florida, December, pp. 1983-1988,
1998.
V (x1 (t), x2 (t)) ≤ V (x1 (0), x2 (0)), ∀t ≥ 0.
[6] W. Su and M. Fu, “Robust Nonlinear Forwarding
Using (37) and monotonicity of s(·), we have with Smooth State Feedback Control,” Proceedings
of IFAC’99 World Congress, vol F, pp. 435-440, Bei-
V2 (x2 (t)) ≤ V (x1 (0), x2 (0))/s(0) =: ρ jing, July, 1999.
Hence, (49) leads to [7] W. Su and M. Fu, “Robust Nonlinear Control: Be-
V̇ (x1 , x2 ) ≤ −εs −1 T
(ρ)x x. (51) yond Backstepping and Nonlinear Forwarding,” Pro-
ceedings of IEEE Conference on Decision and Con-
Therefore, the system (1) is robustly globally asymptot- trol, Phonenix, Arizona, December, pp. 831-837,
ically stabilizible. Finally, the robust local quadratic 1999.
stability property follows from (51) and the fact that
V (x1 , x2 ) becomes quadratic as x2 → 0. ∇∇∇ [8] A. Teel, “Feedback stabilization: Nonlinear solutions
to inherently nonlinear problems,” Memorandum no.
UCB/ERL M92/65, June 12, 1992.

6 Conclusions [9] K. Wei, “Quadratic stabilizability of linear systems

with structural independent time-varying uncertain-
ties”. IEEE Trans. Aut. Contr, AC-35, pp. 268-277,
In this paper, we have studied the robust stabilization 1990.
problem for a class of uncertain nonlinear systems in a for-
warding structure. Several results on this problem have
been reviewed. Then a new design method generalized
from [5, 6] is introduced. This method guarantees robust
global asymptotic stability and local quadratic stability.
Compared with existing results, this method enables us
to simplify the design process and the required assump-
tions and provides better robustness. It should be noted
that this method can be combined with the backstep-
ping design method to give a recursive design for robust
controllers for a much larger class of uncertain nonlin-
ear systems involving both forwarding and backstepping
structures; see [7] for details.

References
[1] J.C. Doyle, K. Glover, P.P. Khargonekar and B.A.
Francis, “State-space solutions to standard H2 and
H∞ control problems’. IEEE Trans. Aut. Contr,
AC-34, pp. 831-847, 1989.
[2] P.P. Khargonekar, I. R. Peterson and K. Zhou,
“Robust stabilization of uncertain linear systems:
quadratic stabilizability and H∞ control tehory”.
IEEE Trans. Aut. Contr, AC-35, pp. 356-361, 1990.
[3] M. Jankovic, R. Sepulchre and P. V. Kokotovic,
“Constructive Lyapunov stabilization of nonlinear
cascade systems”. IEEE Trans. Aut. Contr, AC-41,
pp. 1723-1735, 1996.