Professional Documents
Culture Documents
6
47th IEEE Conference on Decision and Control
Cancun, Mexico, Dec. 9-11, 2008
Abstract— This paper exploits the notion of S-regularity of Inequalities [5]) might be prefered. In that sense, one of
a matrix pencil to propose insights for the robustness analysis the first steps was made in [6]. The advantage of LMI is
of descriptor models of the form E ẋ = Ax (or Exk+1 = Axk also that they can easily enable ones to extend the results to
for the discrete case) subject to norm-bounded LFT (Linear
Fractional Transform)-based uncertainties on both matrices A various clustering regions D.
and E. The property to be studied is the robust D-admissibility
(robust D-stability togeteher with robust regularity and robust
impulse freeness). All the proposed conditions are expressed Many contributions deal with robust analysis or control
in terms of strict LMI (Linear Matrix Inequalities). Two of descriptor models, particularly through (unfortunately not
techniques are proposed and numerically compared. necessarily strict) LMI approach: see [22], the seminal work
of Masubuchi ([16], [17] and the references therein), and
Index Terms— Descriptor systems, S-regularity, Robust D- many others. But very few really consider uncertainty on E.
stability, Robust D-admissibility, strict LMI.
Let us quote [15] where A is however precisely known or
[14] for interval matrices. But the best insights can actually
I. I NTRODUCTION
be found in [12], [18].
It is now well admitted that systems of the form E ẋ = Ax
(or Exk+1 = Axk for the discrete case), that are called
singular systems, descriptor systems, generalized systems In this paper, we make an extensive use of the notion
or implicit systems, and so on, are of great interest for of S-regularity inspired from that of ∂D-regularity [2], [3]
the modelling of many practical devices (interconnected and of some versions of the so-called S-procedure (see [5],
systems, electrical networks, robotics). For conciseness, [11], [19], [10] and the references therein) to derive strict
rather than to quote many references, we urge the interested LMI (sufficient) conditions for a descriptor model subject
reader to examine [21], [7] and some references therein. to norm-bounded LFT (Linear Fractional Transform)-based
uncertainties on both matrices A and E (that are general
As for conventional models for which E = I (or at least enough to encompass many uncertainties encountered in
E is non singular), the D-stability, i.e. the clustering of the practice). Our purpose is to propose a tool, useful at once,
eigenvalues of (E, A) in some region D of the complex as simple as possible, that can be a basis for many other
plane, is of high importance to analyze the transient future works.
behaviour of the system, particularly to assess asymptotic
stability. But it does not suffice in the singular case. The paper is organized as follows. The next section is
Two other properties have to hold, namely the regularity dedicated to the mathematical problem statement, including
(existence of a unique solution to the generalized state-space basic definitions, the description of the uncertain matrix
equation) and impulse freeness (meaning that the infinite pencil, the formulation of the considered regions and the
eigenvalues of the pencil induce no impulsive terms in the actual condition to be checked by LMI. Section 3 proposes
response even when the control signals are not smooth [7]). a first reasoning to derive a sufficient strict LMI condition.
When these three properties hold, the model is said to be We insist on the importance of the strictness of LMI from a
D-admissible. computational point of view. This part of our contribution.
Another reasoning is followed in Section 4 and yields
It is important to derive some simple tools that enable another condition. In section 5, a discussion is led about
the designer to test whether an exactly known pencil is singular systems and their properties (regularity, impulse
D-admissible. For admissibility test, a big focus has been freenees) to analyse our conditions through the lens of those
put on the generalized Lyapunov equations [13], [20], [9]. fundamental properties. The two conditions are numerically
Though very interesting, some of those approaches require compared on an example in Section 6 before to conclude.
the systems to be transformed into equivalent forms, which
is not desirable in an uncertain context. In the presence
of uncertainties, the use of strict LMI (Linear Matrix Notations: M 0 is the transpose conjugate of M . σ̄(M ) is
the maximum singular value of M . I and 0 are identity and
The authors are with the University of Poitiers, Laboratoire zero matrices of appropriate dimensions respectively. In ma-
d’Automatique et d’Informatique Industrielle (LAII), de l’École trix inequalities, < 0, > 0, ≤ 0 and ≥ 0 must be understood
Supérieure d’Ingénieurs de Poitiers (ESIP), Bâtiment de Mécanique,
40, Av. du Recteur Pineau, 86022 POITIERS C EDEX, France, in the sense of Löwner (sign definition of matrices). i is the
Olivier.Bachelier@univ-poitiers.fr imaginary unit and ⊗ denotes the Kronecker product.
Authorized licensed use limited to: Mehdi Driss. Downloaded on February 3, 2009 at 13:07 from IEEE Xplore. Restrictions apply.
47th IEEE CDC, Cancun, Mexico, Dec. 9-11, 2008 ThB11.6
II. P RELIMINARIES AND P ROBLEM STATEMENT Assumption (i) is the classical well posedness of LFT
A. Basic definitions forms and the term implicit refers to assumption (ii) which
is the only new concept introduced here.
In this section, we propose various definitions of properties Remark 1: The uncertainty on E can be useful in practice,
for matrix pencils. The reasonings in the paper are mainly for instance to take uncertain inertias into account.
followed on matrix pencils whereas the descriptor models
are only considered in Section 5. C. Formulation of S
Definition 1: Let (E, A) be a matrix pencil where In the paper, the set S is defined by
{A; E} ∈ Cl n×n . We denote by λ(E, A) the generalized »
s
–0 »
s
–
spectrum of the pencil (E, A) defined by S= s ∈ Cl : R = 0; &
1 1
» –0 » – ff
λ(E, A) = {λ ∈ Cl : det(A(λ)) = 0}, (1) s
Φh
s
≥ 0, ∀h ∈ {1, . . . , h̄} , (5)
1 1
with A(λ) = Eλ − A, and the elements of λ(E, A) are
where R and Φh , h = 1, . . . , h̄ are 2 × 2 Hermitian
referred to as the eigenvalues of (E, A).
matrices. This kind of description is borrowed from [3]
following insights proposed in [11], [10]. Special sets can
Definition 2: (inspired from [2]) Let (E, A) be a matrix be emphasized:
pencil where {A; E} ∈ Cl n×n . Also let S be any subset of » –
4814
Authorized licensed use limited to: Mehdi Driss. Downloaded on February 3, 2009 at 13:07 from IEEE Xplore. Restrictions apply.
47th IEEE CDC, Cancun, Mexico, Dec. 9-11, 2008 ThB11.6
» –
DE λ − D A − C A ∆¯ A BA C E ∆E
BE λ −(I − AE ∆E )
× At this stage, it is possible to apply the generalized Kalman-
» –« Popov-Yakubovich (KYP) lemma proposed in [10] (yet
I 0
−1
(I − AE ∆E ) BE λ I
6= 0 another application of a generalized version of the S-
„» –« procedure), with slight adaptations as in [3] (in order to
¯ A BA
DE λ − D A − C A ∆ C E ∆E
det
BE λ −(I − AE ∆E )
6= 0 encompass the case h̄ > 1), to claim that (15) holds if and
only if there exist an Hermitian matrix P and h̄ Hermitian
⇔ det(Ã(λ, ∆)) 6= 0, ∀{λ; ∆} ∈ S × ∆ (7) positive definite matrices Qh , h = 1, . . . , h̄, such that
with
Ẽ » –0 h̄ » –
z }| –{ CÃ DÃ X CÃ DÃ
» (R ⊗ P + Φh ⊗ Qh ) + Θ̃ < 0. (17)
DE 0 0 Ẽ 0 Ẽ
Ã(λ, ∆) = λ− h=1
BE 0
D C ˜
∆
The previous reasoning is summarized by the next theorem.
à Ã
z
„» }| –{ z
» }| –{»
z }| –{
DA 0 CA −CE ∆A 0
0 I
+
0 −AE 0 ∆E
× Theorem 1: Let an uncertain matrix pencil (E, A) comply
0 1−1 1 with the uncertainty described in §II-B and Assumption
A ˜ B
B z» }| –{z» }|
Ã
–{C
∆
z
» }|
Ã
–{C
1. Also let a set S be described as in §II-C. (E, A) is
robustly S-regular if there exist an Hemitian matrix P and
B C C
B AA 0 ∆A 0 C BA 0 C
BI− C C. (8)
0 0 0 ∆E C 0 I C
h̄ Hermitian positive definite matrices Qh , h = 1, . . . , h̄,
B
@ A A
such that the strict LMI (17) holds.
Let us notice that we recover an LFT-based uncertainty in
which ∆ ˜ is block-diagonal (the bad point) and Ẽ is not The conservativeness in the above theorem is due to the 2-
uncertain (the good point). It is clear that det(Ẽλ−Dà ) 6= 0 block diagonal structure of ∆˜ which makes the S-procedure
(the nominal augmented model has to be S-regular otherwise be pessimistic from (13) to (9) or in other words, it is due
why considering the uncertain case) and therefore (7) is ˜ ⊂∆ ˆ but ∆ ˜ 6= ∆.
ˆ
equivalent to to the fact that ∆
det(I − (Ẽλ − DÃ )
−1 ˜ − A ∆)
CÃ ∆(I Ã
˜ −1 B ) 6= 0
à IV. S OME “ NON - AUGMENTED LFT” SOLUTION
⇔ det(I − BÃ (Ẽλ − DÃ )
−1 ˜ − A ∆)
CÃ ∆(I Ã
˜ −1 ) 6= 0 The term “non-augmented LFT” solution refers to the
⇔ det(I − (AÃ + BÃ (Ẽλ − DÃ )
−1 ˜ 6= 0
CÃ )∆) fact that in this section, we preserve the two LFT-based
uncertainties of §II-B (the one on E and the other one on
˜ 6= 0), that one can write
(since det(I − AÃ ∆) A) without augmenting the size of the uncertainty matrix.
˜ 6= 0, ∀{λ; ∆} ∈ S × ∆,
det(I − G̃(λ)∆) A contribution is to emphasize the fact that this second
technique seems to be less conservative than the previously
with G̃(λ) = AÃ + BÃ (Ẽλ − DÃ )−1 CÃ . The previous presented one.
difference holds if and only if
» –0 » – It is clear that the uncertain pencil (E, A) can be S-regular
˜ ˜
∆
I
Q̃
∆
I
< 0, ∀{λ; ∆} ∈ S × ∆, (9) only if (E, DA ) is S-regular (if the property does not hold
for the nominal part of A, it is no use going further). So,
with » – » –0 necessarily, det(Eλ − DA ) 6= 0 for any {λ; ∆E } ∈ S × ∆E
G̃0 (λ) G̃0 (λ)
Q̃ = (−I) . (10)
I I and then (6) is equivalent to
ˆ as the set of all matrices ∆
Define ∆ ˆ such that det(I − (Eλ − DA )
−1 ¯ A BA ) 6= 0
CA ∆
» –0 » –
ˆ
∆ ˆ
∆ −1 ¯ A ) 6= 0
Ψ̃ ≥ 0, (11) ⇔ det(I − BA (Eλ − DA ) CA ∆
I I
−1 −1
⇔ det(I − BA (Eλ − DA ) CA ∆A (I − AA ∆A ) ) 6= 0. (18)
where
Ψ̃ = blockdiag(−γA I; −γE I, I) (12)
From Assumption 1, it comes det((I−AA ∆A )−1 ) 6= 0 which
˜ that is strictly contained in
˜ lies in a set ∆
It is clear that ∆ enables ones to write
ˆ
∆. Then, the S-procedure as proposed in [5] can be applied det(I − AA ∆A − BA (Eλ − DA )
−1
CA ∆A )det((I − AA ∆A )
−1
) 6= 0
to claim that (9) holds if there exists a scalar τ > 0 such
that
−1
⇔ det(I − AA ∆A − BA (Eλ − DA ) CA ∆A ) 6= 0
Q̃ + τ Ψ̃ < 0, ∀λ ∈ S, (13) −1
⇔ det(I − [AA + BA (Eλ − DA ) CA ]∆A ) 6= 0 (19)
which, by virtue of Finsler’s lemma (that can also be seen
as a special case of S-procedure), is equivalent to that one can write
» –0 » –
I I det(I − G(λ, ∆E )∆A ) 6= 0, ∀{λ; ∆} ∈ S × ∆. (20)
Ψ̃ < 0, ∀λ ∈ S. (14)
G̃(λ) G̃(λ)
Taking the definition of ∆A into account, the previous
The above inequality can also be written
» –0 » –
inequality is equivalent to
I I
(15)
q
Θ̃ < 0, ∀λ ∈ S,
(Ẽλ − DÃ )−1 CÃ (Ẽλ − DÃ )−1 CÃ inf { inf {σ̄(∆A ) : det(I − G(λ, ∆E )∆A ) = 0}} > −1
γA , ∀∆E ∈ ∆E
λ∈S ∆A
where » –0 » – q
I 0 I 0 ⇔ { sup (µCl (G(λ, ∆E ))}
−1 −1
(21)
Θ̃ = Ψ̃ . (16) > γA , ∀∆E ∈ ∆E
AÃ BÃ AÃ BÃ λ∈S
4815
Authorized licensed use limited to: Mehdi Driss. Downloaded on February 3, 2009 at 13:07 from IEEE Xplore. Restrictions apply.
47th IEEE CDC, Cancun, Mexico, Dec. 9-11, 2008 ThB11.6
Notice that the columns of N(∆E ) span the kernel of of the E ẋ = Ax (39)
substitution associated with (or
FE Exk+1 = Axk (40)
z» }| –{
I 0
for the discrete case) where the matrices E and A comply
ˆ ˜ ˆ ˜
I −∆E = (I − ∆E AE ) −∆E BE (32)
AE BE
with (2). It is well known (see [7] and the references
Also notice that ∆E can be defined as therein) that the poles of such a model are the eigenvalues
» –0 » – ff
∆E = ∆E :
∆E
ΨE
∆E
≥0 , (33) of the pencil (E, A), including finite and infinite ones.
I I
The system reponse contains a term with modes related to
with finite poles (as for conventional models where E = I) and
another term with modes associated to infinite poles. As
» –
−γE I 0
ΨE = . (34)
0 I
for usual models, the transient behaviour of the first term
Taking these facts and the compactness of ∆E into account, is strongly related to the location of the finite poles in the
we apply the full block S-procedure in the version proposed complex plane. For this reason, D-stability (root-clustering
in [19] to claim that (30) holds if and only if there exists in some region D ⊂ Cl ) is of interest. But two other apsects
τE > 0 such that have to be considered. The descriptor model should be
4816
Authorized licensed use limited to: Mehdi Driss. Downloaded on February 3, 2009 at 13:07 from IEEE Xplore. Restrictions apply.
47th IEEE CDC, Cancun, Mexico, Dec. 9-11, 2008 ThB11.6
regular (meaning that there is only one solution to the state Regularity is actually not a real problem. The regularity
equation: some kind of well posedness of the model) and of a descriptor model is the regularity of the associated
it should be impulse free (meaning that the “infinite term” pencil as introduced in Defintion 3. So if S-regularity is
in the response does not convey impulses present in the assessed for some non empty S, regularity is proven.
control signals). We recall some classical definitions.
Impulse freeness is completely related to
Definition 4: The model (39) or (40) is said to be Assumption 1.(ii). Actually, this property holds when
D-stable if the finite eigenvalues of (E, A) lie inside some the number of finite poles equals r [7], [21]. Since the
region D. number of finite poles does not change, it is clear that
the sum of all geometric multiplicities of finite poles
Definition 5: The model (39) or (40) is said to be is preserved. This sum must equal r to ensure impulse
D-admissible if it is D-stable, regular and impulse free. freeness. Thus, Assumption 1.(ii) preserves the impulse
freeness (provided (DE , DA ) is impulse free of course).
For conventional models, D-stability is known to be re- This assumption is then fundamental. But, in practice, it
lated to S-regularity [2], [3]. Indeed, when S is the outside is not a drastic contraint because the rank deficiency of
of D then D-stability is the same as S-regularity. A more E is often due to structural properties of the model that
frequent case is when S is the boundary of D. Then D- are still valid in the presence of uncertainties. The only
stability is a special case of S-regularity for which all the exception might be when the descriptor model results from
poles are located on only one side of S. The same reasoning the idealization of a “singularly perturbed” system.
can nearly be followed with descriptor systems. However,
one has to be very careful with infinite poles. When S As a conclusion of the above discussion, when (DE , DA )
or D is unbounded, it is possible that the system be S- is D-admissible, Theorems 1 and 2 can be used to analyze
singular because of infinite poles although they should not the robust D-admissibility of (E, A).
be considered for D-stability. A solution is then to define D
or S not only as a subset of Cl but as a subset of C where VI. N UMERICAL ILLUSTRATION
C is disc centred around the origin and of radius ω possibly The uncertain model is as follows:
very large. In this case, S might be only part of the boundary
» –
AA BA
=
of D, the remaining part being bounded by the frontier of C. CA DA
4817
Authorized licensed use limited to: Mehdi Driss. Downloaded on February 3, 2009 at 13:07 from IEEE Xplore. Restrictions apply.
47th IEEE CDC, Cancun, Mexico, Dec. 9-11, 2008 ThB11.6
4818
Authorized licensed use limited to: Mehdi Driss. Downloaded on February 3, 2009 at 13:07 from IEEE Xplore. Restrictions apply.