Proceedings of the 19th World Congress
The International Federation of Automatic Control
Cape Town, South Africa. August 24-29, 2014
H∞ Filter Design for State Estimation and Unknown Inputs
Reconstruction of a Class of Nonlinear systems
Saleh S. Delshad, Andreas Johansson, Mohamed Darouach, and Thomas Gustafsson
Abstract— We consider a novel method to design a H∞ filter
for a class of nonlinear systems subject to unknown inputs.
First, we rewrite the system dynamics as a descriptor system.
Then, we design a robust H∞ reduced-order filter to estimate
both state variables and unknown inputs at the same time.
Based on a Lyapunov functional, we derive a sufficient condition
for existence of the designed filter which requires solving a
nonlinear matrix inequality. The achieved condition is further
formulated in terms of a linear matrix inequality (LMI) that
is straightforward to solve by popular methods. Finally, the
proposed filter is illustrated with an example.
I. INTRODUCTION
Observer design for nonlinear systems is a popular prob-
lem in control theory that has been studied in many angles.
Moreover, state estimation of nonlinear system in the pres-
ence of unknown inputs is one of the fascinating relevant
topics in the modern control theory. See for instance [1]–[6]
and references therein.
However, to the best of our knowledge, the idea of using
descriptor systems to estimate the state variables together
with unknown inputs is still an open question and it needs
more attention. Briefly, observer design for descriptor sys-
tems has been investigated in different aspects. In this field,
we can refer our readers to [7]–[10] where the authors
consider a variety of nonlinear methods on the descriptor
systems with Lipschitz nonlinearities.
On the other hand, the problem of the state estimation
for descriptor systems in presence of noise has also been
the subject of several studies in the past decades. There are
two popular methods in this field as Kalman filtering and the
H∞ approach. In Kalman filtering, briefly, the system and the
measurement noises are assumed to be Gaussian with known
statistics [11] while for arbitrary type of noise with bounded
energy, H∞ filtering can guarantee a noise attenuation level.
[12]–[14] and references therein are recent researches that
have been done in filtering for descriptor systems.
Inspired by [14], where the authors designed an H∞ filter
for a class of nonlinear singular systems, in this article
we try to rewrite our problem as a descriptor system and
use the features of the descriptor systems to estimate the
state variables of the system together with the unknown
inputs, simultaneously. To achieve this objective, a sufficient
Saleh S. Delshad, Andreas Johansson, and Thomas Gustafsson are
all with the Control Engineering Group at Luleå University of Tech-
nology, Sweden, {Saleh.S.Delshad,Andreas.Johansson,
Thomas.Gustafsson}@ltu.se
Mohamed Darouach is with CRAN-CNRS UMR7039, Université
de Lorraine, IUT de Longwy, Cosnes et Romain 54400, France,
{Mohamed.Darouach}@ univ-lorraine.fr
978-3-902823-62-5/2014 © IFAC
condition for existence of the designed filter is derived,
which requires solving a nonlinear matrix inequality. In
order to facilitate the filter design, the obtained condition
is formulated in terms of a linear matrix inequality (LMI)
that can be easily solved by well-known algorithms in this
area.
The rest of this paper is organized as follows. In Section 2,
we introduce the class of nonlinear systems with unknown
inputs. In Section 3, we propose a new method to design
a reduced-order filter for the systems under study. To vali-
date the proposed nonlinear filter, an example is shown in
Section 4. Conclusions are presented in Section 5.
II. P RELIMINARIES AND PROBLEM STATEMENT
Consider the following nonlinear system subject to un-
known inputs:
h
ẋ =Ax + Bu + D f (x) + D1 ω + ∑ Fi νi ,
i=1
h
y =Cx + D2 ω + ∑ Gi νi (1)
i=1
where x ∈ Rn , u ∈ Rm and y ∈ R p are the state vector,
known input and output vector, respectively. For i = 1, ..., h
(h ≤ p), νi (t) ∈ R are unknown inputs that can affect both
actuators and sensor. Matrices A, B, C, D, D1 , D2 , Fi
(for i = 1, ..., h) and Gi (for i = 1, ..., h) are real and with
appropriate dimensions. The function f is nonlinear. ω is
exogenous disturbance which belongs to L2 [0, ∞). The aim is
to design a nonlinear filter such that it can estimate the state
vector x together with the unknown inputs νi (for i = 1, ..., h)
asymptotically. To specify the class of nonlinear systems
under study, we use the following assumption.
Assumption 1. The function f (x) is nonlinear, and satis-
fies a Lipschitz constraint as follows:
 f (α) − f (β ) ≤ γα − β  (2)
√
where γ > 0 is the Lipschitz constant and x = xT x.
III. M AIN RESULTS
Assume that,
 T
ζ = xT ν1 ν2 ... νh (3)
then, the augmented system dynamics are,
η ζ̇ =Āζ + Bu + D f¯(ζ ) + D1 ω
y =C̄ζ + D2 ω (4)
61
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

where, A. Stability Analysis

  Since e = Jε for ω = 0, obviously the asymptotic stability
η = In×n 0n×h ,
  of ε is sufficient condition for limt→∞ e(t) = 0. The following
Ā = A F1 F2 ... Fh ,
  theorem gives the conditions for stability of the dynamics of
C̄ = C G1 G2 ... Gh (5) e(t).
and,
Theorem 1 Under Assumption 1, for ω = 0 and given a
f¯(ζ ) = f (ηζ ); f¯ : Rn+h → Rq (6)
scalar γ > 0, if there exists real matrices N, J, M, and P > 0
We can make the following assumption which will be used with appropriate dimensions such that the inequalities below
in the sequel of the paper.
  are satisfied,
η  T 
Assumption 2 rank = n + h. N P + PN + γ 2 J T J PMD
C̄ Γ − I < 0, <0 (15)
This assumption is necessary for the pair (η, C̄) to be DT M T P −Γ
observable [15]. then the state estimation error (14) produced by filter (7)
Now, the aim is to design a filter such that it can estimate tends to zero asymptotically.
ζ asymptotically.
Consider the following reduced-order filter for the aug- Proof. First, consider a Lyapunov functional as:
mented system (4), V = ε T Pε (16)
ż = Nz + Ly + Gu + MD f¯(ζ̂ ) where P is a positive definite matrix. From (14), Assump-
ζ̂ = Jz + Ey (7) tion 1, and by taking the derivative of V (t) along the
trajectory of (14) for ω = 0, we have
where vector z ∈ Rq1 and ζ̂ is the estimation of ζ . The
matrices N, L, G, M, J, and E must be determined such that V̇ =ε̇ T Pε + ε T Pε̇
the error dynamics e = ζ̂ −ζ converge to zero asymptotically. =[Nε + MDΔ f¯]T Pε + ε T P[Nε + MDΔ f¯]
By defining the error between z and Mηζ , =ε T (N T P + PN)ε
ε = z − Mηζ (8) + Δ f¯T DT M T Pε + ε T PMDΔ f¯
the error dynamics will be, + Δ f¯T ΓΔ f¯ − Δ f¯T ΓΔ f¯ (17)

ε̇ =Nε + (NMη + LC̄ − M Ā)ζ for Γ < I,

+ (G − MB)u + MD( f¯(ζ̂ ) − f¯(ζ )) V̇ ≤ε T (N T P + PN)ε
+ (LD2 − MD1 )ω (9) + Δ f¯T DT M T Pε + ε T PMDΔ f¯
on the other hand, + Δ f¯T Δ f¯ − Δ f¯T ΓΔ f¯ (18)

ζ̂ = Jz + Ey since,
= Jε + JMηζ + EC̄ζ + ED2 ω Δ f¯T Δ f¯ =Δ f¯(ζ )T Δ f¯(ζ )
= Jε + (JMη + EC̄)ζ + ED2 ω (10) =Δ f (η̄ζ )T Δ f (η̄ζ )
if there exists a matrix M such that, =Δ f (x)T Δ f (x)
≤γ 2 eT e (19)
G = MB, NMη + LC̄ − M Ā = 0, JMη + EC̄ = I (11)
so,
or,  T  T  
G = MB, (12) ε N P + PN + γ 2 J T J PMD ε
     V̇ ≤ (20)
Δ f¯ DT M T P −Γ Δ f¯
N L Mη M Ā
= (13)
J E C̄ I It should be noted that if (15) is satisfied then the state
estimation error (14) tends to zero asymptotically for any
then (9) and (10) become,
initial value ε(0). 
ε̇ = Nε + MD( f¯(ζ̂ ) − f¯(ζ )) + (LD2 − MD1 )ω
B. H∞ Design
e = Jε + ED2 ω (14)
The following theorem gives the sufficient conditions for
We will now formulate and solve an H∞ filter design problem (14) to be stable for ω = 0 and e2 < μω2 for ω = 0.
with the following conditions:
(1) The filter error (14) with ω = 0 is stable. Theorem 2 Consider the system (1) together with the non-
(2) Under zero initial condition, the induced L2 norm of linear filter (7). Under Assumption 1, given admissible
the operator from ω to e is less than μ , i.e. e2 < μω2 . Lipschitz constant γ and disturbance tuning parameter μ,

62
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

there exist matrices N, L, M, J, E, and P > 0 with appro- Clearly, if Φ < 0, then
priate dimensions such that the following inequalities have
V̇ ≤ μ 2 ω T ω − eT e (28)
a feasible solution:
⎡ ⎤ Integrating of both sides of this inequality from zero to
a11 a12 a13
infinity yields
Γ − I < 0, Φ = ⎣aT12 −Γ 0 ⎦ < 0 (21)
aT13 0 a33 V (∞) −V (0) ≤ μ 2 ω(t)22 − e(t)22 (29)
with, for zero initial values, we obtain
a11 = N T P + PN + (1 + γ 2 )J T J, V (∞) ≤ μ 2 ω(t)22 − e(t)22 (30)
a12 = PMD, which leads to
a13 = P(LD2 − MD1 ) + (1 + γ 2 )J T ED2 , e(t)22 < μ 2 ω(t)22 (31)
a33 = (1 + γ 2
)DT2 E T ED2 − μ 2 I (22) and this completes the proof. 
According to [14], (13) has a solution if only if,
Then the state estimation error (14) produced by filter (7) ⎡ ⎤
tends to zero asymptotically for ω = 0 and e2 < μω2 Mη  
for ω = 0. ⎢ C̄ ⎥ Mη
⎢
rank ⎣ ⎥ = rank =n (32)
M Ā ⎦ C̄
Proof. From Theorem 1, we know that system (14) is I
asymptotically stable if (15) is valid. Now let ω = 0,
let R be any arbitrary full row rank matrix such that,
   
V̇ =ε̇ Pε + ε Pε̇
T T
R Mη
rank = rank =n (33)
=[Nε + MDΔ f¯ + (LD2 − MD1 )ω]T Pε C̄ C̄
+ ε T P[Nε + MDΔ f¯ + (LD2 − MD1 )ω] (23) then there always exist matrices M and K such that,
by adding and subtracting the right side of (23) by Δ f¯T ΓΔ f¯, Mη = R − KC̄ (34)

V̇ =ε T (N T P + PN)ε or,  
  η
+ ε T PMDΔ f¯ + Δ f¯T DT M T Pε M K =R (35)
C̄
+ ω (LD2 − MD1 ) Pε
T T
Now from Assumption 2, equation (35) always has a solution
+ ε T P(LD2 − MD1 )T ω given by,  +
+ Δ f¯T ΓΔ f¯ − Δ f¯T ΓΔ f¯   η
M K =R (36)
C̄
≤ε T (N T P + PN)ε
+ Δ f¯T DT M T Pε + ε T PMDΔ f¯ where [.]+ represents any generalized inverse operator satis-
fying [.][.]+ [.] = [.]. Thus,
+ ω T (LD2 − MD1 )T Pε  +    +  
+ ε T P(LD2 − MD1 )ω η I η 0
M=R and K = R (37)
C̄ 0 C̄ I
+ Δ f¯T Δ f¯ − Δ f¯T ΓΔ f¯ (24)
Also, the general solution for (13) is given by,
from (19),     +     +
N L M Ā Mη Y1 Mη Mη
= + (I − ) (38)
V̇ ≤ε T (N T P + PN)ε J E I C̄ Y2 C̄ C̄
+ Δ f¯T DT M T Pε + ε T PMDΔ f¯ where [Y1T Y2T ]T is an arbitrary matrix of appropriate dimen-
+ ω (LD2 − MD1 ) Pε
T T sion that will be identified later on. In order to facilitate the
+ ε T (LD2 − MD1 )T Pω filter design, we convert the provided nonlinear inequalities
in (21) to an LMI problem. For this end, let’s define,
+ γ 2 eT e − Δ f¯T ΓΔ f¯ (25)  +  
Mη I
Since, α1 = (39a)
C̄ 0
 +  
eT e =ε T J T Jε + ω T DT2 E T ED2 ω + ε T J T ED2 ω α2 =
Mη 0
(39b)
ω T DT2 E T Jε (26) C̄ I
  +  
  Mη Mη I
By letting λ = ε Δ f¯ ω , the following inequality is β1 =(I − ) (39c)
C̄ C̄ 0
obtained through (25) and (26)   +  
Mη Mη 0
β2 =(I − ) (39d)
V̇ + eT e − μ 2 ω T ω ≤ λ T Φλ (27) C̄ C̄ I

63
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

so, 3) Compute α1 , α2 , β1 and β2 by (39)

4) Solve the LMIs defined by (41) for Ȳ1 , Y2 , and P
N = M Āα1 +Y1 β1 (40a)
5) Compute Y1 = P−1Ȳ1
L = M Āα2 +Y1 β2 (40b) 6) Using Y1 and Y2 , compute the filter gains as (40)
G = MB (40c)
IV. N UMERICAL EXAMPLE
J = α1 +Y2 β1 (40d)
In this section, we show that the proposed filter is effective
E = α2 +Y2 β2 (40e) on a simulation example. Consider system (1) with parame-
ters,
Theorem 3 Consider the system (1) together with the non- ⎡ ⎤ ⎡ ⎤
linear filter (7). Under Assumption 1 and Assumption 2 and −2 1 0 1
for a disturbance tuning parameter μ, assume that there A = ⎣ 0 −3 −1⎦ , B = ⎣0⎦ ,
exists real matrices Ȳ1 , Y2 , and P > 0 with appropriate 0 1 −2 0
⎡ ⎤ ⎡ ⎤
dimensions, such that the following LMIs have a feasible 0 1
solution: D = ⎣1⎦ , D1 = ⎣1⎦ ,
⎡ ⎤
a11 a12 a13 a14 0 1
⎢aT −Γ 0 0⎥ ⎡ ⎤T ⎡ ⎤T
Γ − I < 0, ⎢ ⎥
⎣aT13 0 a33 a34 ⎦ < 0
12 (41) 1 1
F1 = ⎣0⎦ , F2 = ⎣1⎦ ,
aT14 0 aT34 a44 1 0
   
where, 1 0 0 1
C= , D2 = ,
a11 = α1T ĀT M T P + β1T Ȳ1T + PM Āα1 + Ȳ1 β1 , 0 1 0 1
   
a12 = PMD, 1 0
G1 = , G2 = (44)
0 1
a13 = PM Āα2 D2 + Ȳ1 β2 D2 − PMD1 ,
and a nonlinear function,
a14 = α1T + β1T Y2T ,
a33 = −μ 2 I f (x) = 0.4sin(x1 ) + 0.45cos(x3 ) (45)
a34 = DT2 α2T + DT2 β T Y2T Since there are two different unknown inputs, by choosing
1 augmented states as,
a44 = − I (42)  T
1 + γ2
ζ = x1 x2 x3 ν1 ν2 (46)
and Y1 = P−1Ȳ1 . Then the state estimation error (14) pro-
duced by filter (7) tends to zero asymptotically for ω = 0 we follow our proposed algorithm to estimate both the real
and e2 < μω2 for ω = 0. states and unknown inputs at the same time. According to
Assumption 2, the augmented system is observable. For the
Proof. From Theorem 1, we know that system (14) is arbitrary matrix R (full row rank) as,
asymptotically stable if (15) is valid. Now let ω = 0. If we ⎡ ⎤
1 3 2 0 −1
substitute (40) into (15), we will obtain a new matrix new
R = ⎣−1 2 0 1 −2⎦ , (47)
submatrices M, Y1 , Y2 , and P. In other words, the problem
−3 2 3 0 4
of finding N, L, M, J, E, and P > 0 in Theorem 2 is now
equivalent to the problem of solving (37) for M, (21) for from (37), M is computed as follows:
Y1 , Y2 , and P. Since there is a multiplication between two ⎡ ⎤
1 4 2
unknown matrices Y1 and P, it is not yet an LMI in the M = ⎣−2 4 0⎦ (48)
variable Y1 and P. Define a new variable as, −3 −2 3
Ȳ1 = PY1 (43) Now, using the Matlab LMI toolbox, for γ0 = 0.45 and μ =
Now, using Schur complement, (21) is rewritten as (41). 2.5, we can solve the LMI defined in Theorem 3 with respect
Thus we can use the proof of Theorem 2 to show that state to P, Y1 and Y2 . One feasible solution is found as,
estimation error (14) tends to zero asymptotically and e2 < 0.114 −0.054 0.07
μω2 .  P = −0.054 0.194 0.065 ,
Based on Theorem 3, the reduced-order estimator algo- 0.070 0.065 0.361
rithm for state estimation and unknown inputs reconstruc- 2937.4 3 −1036.2 −380.2 −3327.2
tion in the class of nonlinear systems under study can be Y1 = 1133.8 −347.1 −416.4 −63.9 −1129.7 ,
−960.6 −106.6 411.8 −6.3 510.6
summarized in the following steps, ⎡ ⎤
97.9 9 220.4 −159.8 −236.5
1) Find a Lipschitz constant γ0 satisfying Assumption 1 ⎢ 9 −40.8 153.5 −17.9 65.8 ⎥
2) Fix the order q1 of the filter and choose a full row Y2 = ⎢
⎣ 220.4 153.5 −65.3 −35.1 −127.8⎥ ⎦,
rank matrix R such that (33) is satisfied, then compute −159.8 −17.9 −35.1 67.5 276.7
matrix M by (37) −236.5 65.7 −127.8 276.7 0.5

64
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

4
Therefore, according to (40), the H∞ observer dynamics are
as follows: 2
⎡ ⎤ ⎡ ⎤

x1
−4 0.5 0 3 5 0

ż(t) = ⎣−1.3 −2.9 −0.47⎦ z(t) + ⎣−2 2 ⎦ y(t) −2

0 2 4 6 8 10 12 14 16 18 20
1.1 0.63 −2.05 0 −5
⎡ ⎤ ⎡ ⎤ 2
1 4
+ ⎣−2⎦ u(t) + ⎣ 4 ⎦ (0.4sin(x̂1 (t)) + 0.45cos(x̂3 (t))) 1

x2
−3 −2 0
⎡ ⎤ ⎡ ⎤
0.176 −0.235 −0.118 0 0 −1
⎢ 0.088 0.132 −0.059⎥ ⎢0 0⎥ 0 2 4 6 8 10 12 14 16 18 20

⎢ ⎥ ⎢ ⎥
⎢ ⎥
ζ̂ (t) = ⎢ 0.235 −0.147 0.176 ⎥ z(t) + ⎢ ⎥
⎢0 0⎥ y(t)
4

⎣−0.176 0.235 0.118 ⎦ ⎣ 1 0⎦ 2

x3
−0.088 −0.132 0.059 0 1 0
(49)
−2
0 2 4 6 8 10 12 14 16 18 20
In order to simulate the designed filter, the following as- t (sec)
sumptions are made:
 T Fig. 2. Real (solid) and Estimated (dashed) State Variables
x(0) = 3 2 −2 ,
 T
z(0) = 0 0 0 , 8

u(t) = 0.1 (50)

7

Also we assume that the exogenous disturbance ω(t) is a

6
random signal with amplitude 0.5, frequency 0.01 Hertz, and
bias 0.1 as depicted in Fig. 1. 5

1 4
ν1

0.8 3

0.6
2

0.4
1

0.2
0
ω(t)

0
−1
0 2 4 6 8 10 12 14 16 18 20
−0.2
t (sec)
−0.4

Fig. 3. Real (solid) and Estimated (dashed) Unknown Input ν1

−0.6

−0.8

−1
system with unknown inputs. The designed filter is able to
0 2 4 6 8 10 12 14 16 18 20
t (sec)
estimate both states and unknown inputs, simultaneously. In
the proposed method, first we rewrite the system dynamics
as a descriptor system and design an H∞ reduced-order filter
Fig. 1. Exogenous Disturbance ω(t)
for the new system dynamics. Then, we derive a sufficient
condition for existence of the designed filter which requires
After simulating the designed filter described as (49), solving a nonlinear matrix inequality. In order to facilitate the
results are presented in Fig. 2, Fig. 3 and Fig. 4. The Fig. 2 proposed filter design, the obtained condition is formulated in
shows the real and estimated state variables for initial condi- terms of LMIs that can be solved by well-known algorithms
tions given by (50) simultaneously. The estimated unknown easily. Finally, the proposed filter is validated by an example
inputs are shown in Fig. 3 and Fig. 4, respectively. Based on and simulation results are shown.
Fig. 2, Fig. 3 and Fig. 4, one can see that the filter performs
as expected. R EFERENCES
V. CONCLUSIONS [1] M. Darouach, M. Zasadzinski, O. A. Bassong, and S. Nowakowski,
”Kalman filtering with unknown inputs via optimal state estimation of
In this paper, we have presented a new and efficient singular systems,” International Journal of Systems Science, vol. 46,
approach to design a H∞ filter for a class of nonlinear no. 10, pp. 20152028, 1995.

65
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

2.5

1.5

0.5
ν2

−0.5

−1

−1.5

−2
0 2 4 6 8 10 12 14 16 18 20
t (sec)

Fig. 4. Real (solid) and Estimated (dashed) Unknown Input ν2

[2] Q. P. Ha, and H. Trinh, ”State and input simultaneous estimation for a
class of nonlinear systems,” Automatica, vol. 40, pp. 17791785, 2004.
[3] W. Chen and M. Seif, ”Unknown input observer design for a class
of nonlinear systems: an LMI approach,” in Proc. of IEEE American
Control Conference, Minnesota, 2006.
[4] M. Darouach, ”H∞ unbiased filtering for descriptor systems via LMI,”
IEEE Transactions on Automatic Control, vol. 54, no. 8, pp. 1966-
1972, 2009.
[5] S. S. Delshad, and T. Gustafsson, ”Nonlinear observer design for a
class of Lipschitz time-delay systems with unknown inputs: LMI ap-
proach,” XXIII IEEE Symposium on Information, Communication and
Automation Technologies (ICAT), Bosnia and Herzegovina, Sarajevo,
pp. 1-5, 2011.
[6] Fanglai Zhu, ”State estimation and unknown input reconstruction via
both reduced-order and high-order sliding mode observers,” Journal
of Process Control, vol. 22, pp. 296-302, 2012.
[7] D. Koenig, ”Observer Design for Unknown Input Nonlinear Descrip-
tor Systems via Convex Optimization,” IEEE Trans. on Automatic
Control, vol. 51, pp. 1047-1052, 2006.
[8] L. Zhou and G. Lu, ”Robust Stability of Singularly Perturbed Descrip-
tor Systems With Nonlinear Perturbation,” IEEE Trans. on Automatic
Control, vol. 56, pp. 858863, 2011.
[9] G. Lu, D. W. C. Hob, and L. Zhou, ”A note on the existence of a
solution and stability for Lipschitz discrete-time descriptor systems,”
Automatica, vol. 47, pp. 1525-1529, 2011.
[10] M. Ezzine, H. S. Alia, M. Darouach, and H. Messaoud, ”A controller
design based on a functional H∞ filter for descriptor systems: The time
and frequency domain cases,” Automatica, vol. 48, pp. 542-549, 2012.
[11] J. Ishiharaa, and M. H. Terrab, ”Robust state prediction for descriptor
systems,” Automatica, vol. 44, pp. 21852190, 2008.
[12] S. Xu, J. Lam, and Y. Zou, ”H∞ filtering for singular systems,” IEEE
Transactions on Automatic Control, vol. 48, no. 12, pp. 2217-2222,
2003.
[13] S. Xu, and J. Lam, ”Reduced-orde H∞ filtering for singular systems,”
Systems & Control Letters, vol. 56, pp. 48-57, 2007.
[14] M. Darouach, L. Boutat-Baddasand, and M. Zerrougui, ”H∞ observers
design for a class of nonlinear singular systems,” Automatica, vol. 47,
pp. 2517-2525, 2011.
[15] L. Dai, ”Singular Control Systems,” Springer-Verlag, 1989.

66