43rd IEEE Conference on Decision and Control
3
December 14-17, 2004
Atlantis, Paradise Island, Bahamas
Generalized S -procedure for Inequality Conditions on
One-Vector-Lossless Sets and
Linear System Analysis
Yoshio Ebihara, Katsutoshi Maeda and Tomomichi Hagiwara
Department of Electrical Engineering, Kyoto University,
Kyotodaigaku-Katsura, Nishikyo-ku, Kyoto 615-8510, Japan.
Abstract— The generalized S-procedure, introduced by
Iwasaki et al., has proved to be very useful for robustness anal-
ysis and synthesis of control systems. This procedure provides
a nonconservative way to convert inequality conditions on loss-
less sets into numerically verifiable conditions represented by
linear matrix inequalities (LMIs). In this paper, we introduce a
new notion, one-vector-lossless sets, and propose a generalized
S-procedure to reduce inequality conditions on one-vector-
lossless sets into LMIs without any conservatism. By means of
the proposed generalized S-procedure, we can examine various
properties of matrix-valued functions over some regions on the
complex plane. To illustrate the usefulness, we show that full
rank property analysis problems of polynomial matrices over
some specific regions on the complex plane can be reduced
into LMI feasibility problems. It turns out that many existing
results such as Lyapunov’s inequalities for stability analysis
of linear systems and LMIs for state-feedback controller
synthesis can be viewed as particular cases of this result.
I. I NTRODUCTION
Recently, Iwasaki et al. [10], [11], [12] opened a new
horizon for robustness analysis and synthesis of control sys-
tems by introducing the generalized S-procedure. The gen-
eralized S-procedure concerns inequality conditions with
respect to a Hermitian matrix Θ and a subset S of Hermitian
Matrices described by
ζ ∗ Θζ > 0 ∀ζ ∈ G,
G := {ζ ∈ Cn : ζ = 0, ζ ∗ Sζ ≥ 0 ∀S ∈ S}
It can be easily seen that a sufficient condition to (1) is
given by
∃S ∈ S such that Θ > S
The procedure to replace the condition (1) by (2) is called
the generalized S-procedure [10], [11], [12]. Generally,
this replacement introduces conservatism; the condition (2)
is only sufficient for (1) and may not be necessary. The
significance of the studies in [10], [12] lies in the fact
that, the generalized S-procedure has been proved to be
nonconservative if the set S is lossless [9], [10]. If the
set S is lossless, then the set S is convex and hence the
linear matrix inequality (LMI) condition (2) can be verified
numerically via sophisticated interior-point methods [2], [4].
When we deal with linear system analysis and synthesis
problems by working with the generalized S-procedure,
the underlying idea is that inequality conditions on matrix-
valued functions G(λ) over lines λ ∈ Λ (Λ ⊂ C) can be
This work is supported in part by the Ministry of Education, Culture,
Sports, Science and Technology of Japan under Grant-in-Aid for Young
Scientists (B), 15760314.
0-7803-8682-5/04/$20.00 ©2004 IEEE
TuC10.
reformulated into a conformable form to the condition (1)
by considering an appropriate Hermitian matrix Θ and a
lossless set S [10], [12]. Various properties of linear sys-
tems can be characterized by inequality conditions on their
transfer functions in the frequency domain [2], [15], [16]. In
[10], [11], [12], it has been shown that frequency domain
inequalities for transfer functions can be reformulated in
the form of (1) so that the generalized S-procedure can
be applied. It follows that we can verify various properties
of linear systems without introducing any conservatism, by
solving LMIs resulting from the generalized S-procedure.
For linear system analysis and synthesis, however, we
also need to verify inequality conditions on matrix-valued
functions G(λ) over region λ ∈ D (D ⊂ C). For example,
full rank property analysis of polynomial matrices over
some specific regions on the complex plane forms an
important basis for the stability analysis of linear systems
[2], [6]. In view of these facts, it is natural to pose the
following question: Can we verify various properties of
matrix-valued functions G(λ) over region λ ∈ D (D ⊂ C)
by following similar lines to the generalized S-procedure?
To answer this question, in this paper, we first intro-
duce a new notion, one-vector-lossless sets, and provide
a nonconservative generalized S-procedure for inequality
(1) conditions on the one-vector-lossless sets. More precisely,
taking account of the facts that the properties of lossless sets
are fully used to represent lines on the complex plane in [9],
[10], [12], we first consider to relax the requirements for the
lossless sets given in [9], [10] and define one-vector-lossless
(2)
sets, which enables us to represent regions on the complex
plane. Then, we secondly clarify under what condition the
generalized S-procedure for inequality conditions on the
one-vector-lossless sets is nonconservative. It follows that
we can provide a counterpart result of [9], [10], [12] in
the case of the one-vector-lossless sets. We believe that,
similarly to the one in [9], [10], [12], the proposed S-
procedure could be a promising tool for linear systems
analysis and synthesis.
By working with the proposed generalized S-procedure,
we can verify some properties of matrix-valued functions
G(λ) over regions on the complex plane by solving LMIs
resulting from the proposed generalized S-procedure. To
illustrate the usefulness, we show that full rank property
analysis problems of polynomial matrices over some regions
D ⊂ C can be reduced into LMI feasibility problems. It
turns out that the well-known results such as Lyapunov’s
inequalities for stability analysis of linear systems [2] and
1272
LMIs for state-feedback controller synthesis [2], [13] follow
immediately from the full rank property analysis by means
of the proposed generalized S-procedure.
We use the following notations in this paper. For a matrix
A, its transpose, complex conjugate transpose and Moore-
Penrose inverse are denoted by AT , A∗ and A† , respectively.
For a matrix A ∈ Cn×m with rank r, A⊥ ∈ C(n−r)×n is a
matrix such that A⊥ A = 0 and A⊥ (A⊥ )∗ > 0. The symbols
Hn and Pn denote the sets of n × n Hermitian matrices
and positive-definite Hermitian matrices, respectively. For
matrices Ψ and P , we denote by Ψ ⊗ P their Kronecker
product. For λ ∈ C and Ψ ∈ H2 , we define a function
σ : C × H2 → R by

∗

λ λ
σ(λ, Ψ) := Ψ
1 1
II. G ENERALIZED S-P ROCEDURE FOR I NEQUALITY
C ONDITIONS ON O NE -V ECTOR -L OSSLESS S ETS
The notion of one-vector-lossless sets plays an important
role in this paper. In this section, we first describe its
precise definition and provide a nonconservative generalized
S-procedure for inequality conditions on the one-vector-
lossless sets.
Definition 1: (One-Vector-Lossless Sets) A subset S ⊂ Hn
is said to be one-vector-lossless if it has the following
properties:
(a) S is convex.
(b) S ∈ S ⇒ τ S ∈ S ∀τ > 0.
(c) For each nonzero matrix H ∈ Cn×n with rank r that
satisfies
H = H ∗ ≥ 0, tr(SH) ≥ 0 ∀S ∈ S, (3)
there exist vectors ζi ∈ C (i = 1, · · · , r) such that
n H=
 r
H= ζi ζi∗ and the condition
i=1
ζj∗ Sζj ≥ 0 ∀S ∈ S
holds for at least one index j.
It should be noted that the above definition has been
introduced by relaxing the requirements for the lossless sets
given in [9], [10]. Indeed, Definition 1 reduces to the one
for the lossless sets by replacing (c) by (c’) given in the
following.
(c’) For each nonzero matrix H ∈ Cn×n with rank r
that satisfies (3), there exist vectors ζi ∈ Cn (i =
1, · · · , r) such that
 r
H= ζi ζi∗ , ζi∗ Sζi ≥ 0 ∀i ∀S ∈ S
i=1
In contrast with the definition of the lossless sets, we see
that the condition ζj∗ Sζj ≥ 0 (∀S ∈ S) is required only for
one index j in the definition of the one-vector-lossless sets.
Hence, it is obvious that a lossless set is one-vector-lossless.
In the case where the set S is lossless, the condition
(1) can be converted into (2) without introducing any
conservatism [10], [12]. The following theorem gives a
counterpart of this result in the case where the set S is
one-vector-lossless.
1273
Theorem 1: (The Generalized S-Procedure for Inequality
Conditions on One-Vector-Lossless Sets) Let Θ ∈ Hn and
a one-vector-lossless set S ⊂ Hn be given. If Θ = Θ∗ ≥ 0,
then the following statements are equivalent.
(i) ζ ∗ Θζ > 0 ∀ζ ∈ G,
G := {ζ ∈ Cn : ζ = 0, ζ ∗ Sζ ≥ 0 ∀S ∈ S}
(ii) There exists S ∈ S such that Θ > S.
Proof: (ii) ⇒ (i) Suppose (ii) holds. Then, there
exists S0 ∈ S such that
ζ ∗ (Θ − S0 )ζ > 0 ∀ζ = 0
The above inequality implies
ζ ∗ Θζ > 0 ∀ζ ∈ G0 ,
G0 := {ζ ∈ Cn : ζ = 0, ζ ∗ S0 ζ ≥ 0}
Since G ⊂ G0 , we can conclude that the condition (ii)
implies (i).
(i) ⇒ (ii) Suppose (ii) does not hold, i.e., there is no S ∈
S such that Θ > S. Then, since S is convex, it follows from
the separating hyper-plane theorem [9] that there exists a
nonzero matrix H ∈ Cn×n such that
H = H ∗ ≥ 0, tr((Θ − S)H) ≤ 0 ∀S ∈ S (4)
In view of the property (b) of the one-vector-lossless set,
we see that the following conditions are necessary for the
second condition in (4) to hold.
tr(ΘH) ≤ 0, tr(SH) ≥ 0 ∀S ∈ S (5)
Since S is one-vector-lossless, it follows from the property
(c) of Definition 1 that the second condition above implies
the existence of the vectors ζi (i = 1, · · · , r) such that
r
ζi ζi∗ and ζj∗ Sζj ≥ 0 (∀S ∈ S) for some j, where
i=1
r is the rank of H. For those vectors ζi , the first condition
in (5) implies
ζi∗ Θζi = 0 (i = 1, · · · , r)
due to the assumption Θ = Θ∗ ≥ 0. These facts in
particular imply that ζj∗ Θζj = 0 and ζj ∈ G for at least
one index j. This clearly contradicts the condition (i).
In Theorem 1, we have shown a nonconservative gen-
eralized S-procedure for inequality conditions on the one-
vector-lossless sets. We see that, in comparison with the
case where the set S is lossless [10], an additional condition
Θ = Θ∗ ≥ 0 has been imposed in Theorem 1. This could
be regarded as a price to pay for relaxing the requirements
on the set S from a lossless one to a one-vector-lossless
one.
By means of the generalized S-procedure suggested by
Theorem 1, we can convert the multiple inequality condi-
tions (i) into the numerically verifiable LMI condition in (ii).
Hence, when we solve control system analysis and synthesis
problems at hand, a crucial step is to reduce those problems
into a form conformable to the condition (i). This step is
not obvious in general. When exploring such reduction, it
is indispensable to see concretely what sets are indeed one-
vector-lossless. In the next theorem, we will show a class
of one-vector-lossless sets that is relevant to control system
analysis and synthesis.
Theorem 2: Let Ψ ∈ H2 with det(Ψ) < 0 and Γ ∈ C2n×l
be given. Define a subset of Hermitian matrices by
S := {Γ∗ (Ψ ⊗ P )Γ : P ∈ Pn } (6)
Then the set S is one-vector-lossless.
Proof: The proof is omitted due to limited space. See
[3] for details.
It is meaningful to examine the properties of the one-
vector-lossless set S given by (6) in comparison with the
lossless set Sl described below [9].
Sl := {Γ∗ (Ψ ⊗ P )Γ : P ∈ Hn }
from its normal value, i.e., rank(M (z)) < m. In linear
system analysis and synthesis, it is of great importance to
determine whether the zeros of given polynomial matrix
M (s) belong to a specific region D ⊂ C. This can be
restated equivalently in the way that the polynomial matrix
M (s) is of full-column rank for all s ∈ Dc , where Dc
denotes the complement of the region D in C. In the
subsequent discussions, we restrict our attention to the
regions defined below.
Definition 2: For given Ψ ∈ H2 with det(Ψ) < 0, we
c
define a set DΨ and its complement DΨ by
DΨ := {λ ∈ C : σ(λ, Ψ) < 0}
DΨc
:= {λ ∈ C : σ(λ, Ψ) ≥ 0}
(10)

To see a significant difference between the above two sets, By selecting the Hermitian matrix Ψ in (10) appropri-
let us take Ψ = diag(−1, 1) and Γ = I2n for simplicity ately, we can obtain several regions that are relevant to
and consider the following set that concerns the condition linear system analysis and synthesis. In particular, by letting
(i) in Theorem 1.


0 1 1 0


 Ψc := , Ψd := , (11)
f1 2n f1 1 0 0 −1
G := ∈ C : f0 , f1 ∈ C , n
= 0,
f0 f0  (7) we see that DΨc and DΨd coincide with the open left half

∗

f1 f1 plane C− and the open unit disc D, respectively. These
S ≥ 0 ∀S ∈ S
f0 f0 regions are particularly important for stability analysis of
continuous- and discrete-time linear systems.
Then, we can show that the above set defined from the
A. Full Rank Property Analysis of Polynomial Matrices
one-vector-lossless set S coincides with

  We are now in a right position to show that the full
f1
L := ∈ C2n : f1 = sf0 for some s ∈ D (8) rank property analysis problems of polynomial matrices
f0 can be reduced into LMI feasibility problems by means
where D denotes the closure of the open unit disc D of the proposed generalized S-procedure. When exploring
on the complex plane. On the other hand, if we replace such reduction, as mentioned before, a crucial step is to
the one-vector-lossless set S in (7) by the lossless set reduce the analysis problem into a form conformable to
Sl , then the resulting set Gl coincides with the set Ll the condition (i) in Theorem 1 so that the generalized S-
obtained by replacing D in (8) by ∂D. These observations procedure can be applied. In the next theorem, we will show
clearly indicate that the lossless sets are related to lines explicitly how this reduction can be done via a sequence
on the complex plane, while the one-vector-lossless sets of equivalent restatement of the condition relevant to the
are related to regions on the complex plane. This is the analysis problem.
key observation to develop the generalized S-procedure Theorem 3: Let complex matrices
Mk ∈ Cn×m  (k =
ψ11 ψ12
for inequality conditions on the one-vector-lossless sets. 0, · · · , N ) with n ≥ m and Ψ = ∗ ∈ H2
We believe that, similarly to the one in [10], [12], the ψ12 ψ22
proposed generalized S-procedure could be a promising with det(Ψ) < 0 be given. Suppose either of the following
tool for analysis and synthesis of control systems, although conditions holds:
it is not yet clear to us how far it works effectively. To 1. MN is of full-column rank.
illustrate the effectiveness, we show in the next section that 2. ψ11 < 0.
full rank property analysis problems of polynomial matrices Define
over some regions on the complex plane can be dealt with N
M (s) := sk Mk , M := [ MN · · · M0 ] (12)
by the proposed generalized S-procedure.
k=0
Then the following conditions are equivalent.
III. L INEAR S YSTEM A NALYSIS U SING
(i) The polynomial matrix M (s) is of full-column rank
G ENERALIZED S-P ROCEDURE c
for all s ∈ DΨ .
For given complex matrices Mk ∈ Cn×m (k =
0, · · · , N ) with n ≥ m, let us consider the n × m complex (ii) M (s)∗ M (s) > 0 ∀s ∈ DΨ
c

polynomial matrix ⎡ N ⎤∗ ⎡ N ⎤
s Im s Im
⎢ ⎥ ⎢ ⎥

N
(iii) ⎣ ... ⎦ M∗ M ⎣ ... ⎦ > 0 ∀s ∈ DΨ
c
M (s) = sk Mk (9)
k=0
Im Im
∗ ∗
We assume that the normal rank of M (s) is m. Following (iv) f M Mf > 0 ∀f ∈ K,
∗
the discussions in [6], [14], we define a (finite) zero of M (s) K := f = [ fN · · · f0∗ ]∗ ∈ C(N +1)m : f0 = 0,
as a complex value z ∈ C for which the rank of M (s) drops c
∃s ∈ DΨ such that fk+1 = sfk (k = 0, · · · , N − 1)}
1274
 (v) f ∗ M∗ Mf > 0 ∀f ∈ L, 1. M1 is of full-column rank.
∗
L := f = [ fN · · · f0∗ ]∗ ∈ C(N +1)m : 2. ψ11 < 0.
fk ∈ C (k = 0, · · · , N ), f0 = 0,
m
Then, the matrix pencil M0 + sM1 is of full-column rank
c
f ∗ Sf ≥ 0 ∀S ∈ SW } ,
all∗ s∈ DΨ iff there exists P ∈ Pm such that
for
∗
SW := {W (Ψ ⊗ P )W : P ∈ PN m } , M1  


∗
∗ ∗ M1 M0 − Ψ ⊗ P > 0 (14)
W1 IN m 0 M0
W := , W1 := , W2 := m,N m If the matrices M0 , M1 and Ψ are all real, the equivalence
W2 0m,N m IN m still holds when we restrict P to be real.
(vi) f ∗ M∗ Mf > 0 ∀f ∈ L ,  B. Full Rank Property Analysis of Polynomial Matrices
L := f ∈ C(N +1)m : f = 0, f ∗ Sf ≥ 0 ∀S ∈ SW Using Linear Fractional Representation
(vii) There exists P ∈ PN m such that For given polynomial matrix M (s) of the form (9), let
M∗ M − W ∗ (Ψ ⊗ P )W > 0 (13) us define M̂ := [ M0∗ · · · MN ∗ ∗
] ∈ C(N +1)n×m . Then,
If the matrices Mk (k = 0, · · · , N ) and Ψ are all real, the we see that rank(M̂) = m is necessary for the condition
equivalence still holds when we restrict P in (v), (vi) and rank(M (s)) = m to hold at each fixed s ∈ C. In the
(vii) to be real. preceding subsection, Theorem 3 has been derived without
Proof: The equivalence of (i), (ii), (iii) and (iv) is paying much attention on this implicit condition. However,
apparent. The proof for the equivalence of (iv), (v) and (vi) by taking account of this condition explicitly, we can derive
is given in [3]. The main step of the proof, the equivalence other LMIs to verify the full-column rank property of the
of (vi) and (vii), follows immediately from Theorems 1 and polynomial matrix M (s).
2. Indeed, the set SW is one-vector-lossless by Theorem 2 To see this, let us recall that the polynomial matrix M (s)
while it is clear that M∗ M ≥ 0. Hence the generalized has a linear fractional representation (LFR) [15] of the form
S-procedure in Theorem 1 establishes the equivalence of
(vi) and (vii). M (s) = L0 + sL1 (I − sD)−1 L2
Noting that the real case results can be shown by fol- where
lowing similar arguments to [10], we complete the proof. ⎡ ⎤
M0 In 0 0 ··· 0
In Theorem 3, we have reduced the full-column rank ⎢ M1 0 In 0 ··· ⎥ 0
⎢ ⎥
property analysis problem of the polynomial matrix M (s)
 ⎢ .. .. .. .. .. ⎥ ..
⎢ . . . . . ⎥ .
over the region DΨ c
into the feasibility problem of the LMI L0 L1 ⎢ ⎥
:= ⎢ .. .. .. .. ⎥
(13). When exploring such reduction, we have shown that L2 D ⎢ . . . . 0 ⎥
⎢ ⎥
the generalized S-procedure given in Theorem 1 plays a ⎢ .. .. .. ⎥
⎣ . . . In ⎦
key role. The resulting LMI condition (13) is numerically
tractable and can be solved efficiently via interior-point MN 0 ··· ··· ··· 0
methods [2], [4]. We note here that the LMI condition (13) Then, from the well-known rank properties of LFR, we see
is not completely new and similar conditions have been that rank(M (s)) = m holds at each fixed s ∈ C iff
derived in [6] via rank-one LMI approach.  
Before proceeding, we will give brief comments on the rank [ M̂ sU − V ] = m + N n
implication of the assumptions 1 and 2 in Theorem 3. From
c
the definition of DΨ in (10), we see that the assumption where




c
ψ11 < 0 enforces the region DΨ to be bounded. To put it L1 IN n 0n,N n
another way, our assumptions require that the matrix MN to U= = , V = (15)
D 0n,N n IN n
c
be of full-column rank if the region DΨ is unbounded. When
studying the full rank property of polynomial matrices over Moreover, if rank(M̂) = m, the above rank condition is
unbounded regions, it is well-known that we have to take also equivalent to
a special care on zeros at infinity [6], [14]. If the matrix  
MN is of full-column rank, however, we can confirm that rank s(M̂⊥ U ) − M̂⊥ V = N n
the matrix polynomial M (s) given in (9) does not have
where M̂⊥ ∈ C((N +1)n−m)×(N +1)n . It follows that the
zeros at infinity [14]. Hence, by studying the full-column
following two statements are equivalent.
rank property of M (s) under the assumptions 1 and 2,
delicate problems stemming from zeros at infinity have been 1. M (s) is of full-column rank at each fixed s ∈ C.
excluded from our discussions. 2. rank(M̂) = m and s(M̂⊥ U ) − M̂⊥ V is of full-
In linear system analysis and synthesis, it is particularly column rank at each fixed s ∈ C.
important to verify full rank properties of matrix pencils. These observations lead us to the following theorem,
In the next corollary, we will show a special case result of which provides different LMIs from (13) for full-column
Theorem 3 when M (s) is a matrix pencil. rank property analysis of polynomial matrices.
Corollary 1: Let complex matrices M0 , M1 ∈ Cn×m with Theorem 4: Let complex matrices Mk ∈ Cn×m (k =
n ≥ m and Ψ ∈ H2 with det(Ψ) < 0 be given. Suppose 0, · · · , N ) with n ≥ m and Ψ ∈ H2 with det(Ψ) < 0
either of the following conditions holds: be given. Suppose either of the following conditions hold:
1275
 1. MN is of full-column rank. We readily obtain the following results from Theorem 4.
2. ψ11 < 0. Corollary 2: Let complex matrices M0 , M1 ∈ Cn×m with
Define n ≥ m and Ψ ∈ H2 with det(Ψ) < 0 be given. Suppose

N
either of the conditions 1 and 2 in Corollary 1 holds. Then,
∗ ∗
M (s) := sk Mk , M̂ := [ M0∗ · · · MN ] (16) the matrix pencil sM1 + M0 is of full-column rank for all
c
k=0 s
∈ DΨ iff there exists
Q ∈ Pn such that
Then the following conditions are equivalent. ∗
−M0 −M0
(i) The polynomial matrix M (s) is of full-column rank (Ψ ⊗ Q) <0 (21)
c
for all s ∈ DΨ . M1 M1
If the matrices M0 , M1 and Ψ are all real, the equivalence
(ii) rank(M̂) = m and the matrix pencil s(M̂⊥ U ) − still holds when we restrict Q to be real.
M̂⊥ V is of full-column rank for all s ∈ DΨ
c
, where We have shown two LMI conditions (13) and (18) for
U and V are given by (15). the full-column property analysis of polynomial matrices.
Since both conditions are necessary and sufficient, the LMI
(iii) rank(M̂) = m and there exists Q ∈ PN n such that

∗  ∗ condition (13) should be able to be rewritten in the form
U  
∗ M̂⊥ M̂⊥ U −V − Ψ ⊗ Q > 0 (17) of (18). However, it is not yet clear to us how we achieve
−V this reformulation straightforwardly in general cases. In the
(iv) There exists Q ∈ PN n such that subsequent discussions, it turns out that these two conditions

∗
 lead to well-known two types of LMIs when dealing with
M1 M1
(Ψ ⊗ Q) <0 (18) D-stabilizability analysis problems of matrix pairs.
M2 M2
where M1 , M2 ∈ C(N n)×((N −1)n+m) are C. D-Stability Analysis of Matrices
⎡ ⎤
0 ··· ··· 0 −M0 A matrix A ∈ Cn×n is said to be D-stable iff all
⎢ .. .. ⎥ the eigenvalues of A belong to the region D ⊂ C [1].
⎢ In . . −M1 ⎥
⎢ ⎥ This condition can be restated equivalently in the way that
⎢ ⎥
M1 := ⎢ 0 . . . . . . ... ..
. ⎥ (19) the matrix pencil sI − A is nonsingular for all s ∈ Dc .
⎢ ⎥
⎢ . . . ⎥ By applying Corollaries 1 and 2 to this nonsingularity
⎣ .. .. . . 0 −M
N −2
⎦
condition, we readily obtain the following results.
0 · · · 0 In −MN −1 Corollary 3: For given A ∈ Cn×n and Ψ ∈ H2 with
and
 det(Ψ) < 0, the following conditions are equivalent.
I(N −1)n 0 (i) The matrix A is DΨ -stable.
M2 := (20)
0 MN (ii) The matrix pencil sI −A is nonsingular for all s ∈ DΨ c
.
If the matrices Mk (k = 0, · · · , N ) and Ψ are all real, the (iii) There exists P ∈ Pn such that
equivalence still holds when we restrict Q in (iii) and (iv)

I  
to be real. ∗ I −A − Ψ ⊗ P > 0 (22)
−A
Proof: We have already shown that the condition (i)
(iv) There exists Q ∈ Pn such that
is equivalent to (ii). From Corollary 1, we see that the
∗

equivalence of (ii) and (iii) readily follows by showing that A A
(Ψ ⊗ Q) <0 (23)
M̂⊥ U is of full-column rank if MN is. If MN is of full- I I
column rank, we have If the matrices A and Ψ are real, the equivalence still holds
⎡ †
⎤ when we restrict P in (iii) and Q in (iv) to be real.
In 0 · · · 0 −M0 MN
⎢ ⎥ The inequality (23) is known as Lyapunov’s inequality
⎢ 0 . . . . . . ... .. ⎥ [5], [9]. Corollary 3 indicates that this well-known result
⎢ . ⎥
⊥ ⎢ . . ⎥ can be derived via nonsingularity analysis of matrix pencils
M̂ = ⎢ . . .. .. ⎥
⎢ . . . 0 . ⎥ by means of the proposed generalized S-procedure. Indeed,
⎢ † ⎥
⎣ 0 · · · 0 In −MN −1 MN ⎦ the equivalence of (ii) and (iii) is a direct consequence of
⊥
0 ··· ··· 0 MN Corollary 1, while the equivalence of (ii) and (iv) follows
immediately from Corollary 2. By noting that ([ I −
Hence we have M̂⊥ U = [ IN n 0N n,n−m ]∗ , which clearly A ]∗ )⊥ = [ A∗ I ], the equivalence of (iii) and (iv) can also
shows that M̂⊥ U is of full-column rank. be verified by Finsler’s Lemma [2]. Hence, Corollaries 1
To prove the equivalence of (iii) and (iv), we first note and 2 lead to the same LMI when dealing with D-stability
that analysis of complex matrices.

 ∗ ⊥
M ∗
U∗ ⊥ 1
M̂ = D. D-stabilizability Analysis of Matrix Pairs
−V ∗ M2
We first define the stabilizability of matrix pairs (A, B)
when rank(M̂) = m. If we note that the condition with respect to a region D ⊂ C.
rank(M̂) = m is ensured by (18) in view of the form Definition 3: For given A ∈ Cn×n and B ∈ Cn×m , the
of M1 and M2 given by (19) and (20), respectively, the pair (A, B) is said to be D-stabilizable if there exists K ∈
equivalence can be verified by Finsler’s Lemma [2]. Cm×n such that A + BK is D-stable.
1276
 From the Popov-Belevitch-Hautus (PBH) tests [15], [16], In view of the results in Corollaries 3 and 4, we see that
we see that the matrix pair (A, B) is D-stabilizable iff the the proposed generalized S-procedure forms an important
matrix pencil [ sI − A B ]∗ is of full-column rank for all basis for linear system analysis and synthesis using LMIs.
s ∈ Dc . By applying Corollaries 1 and 2 to this full-column
IV. C ONCLUSION
rank condition, we obtain the following results.
Corollary 4: For given A ∈ Cn×n , B ∈ Cn×m and Motivated by the studies in [10], [11], [12], in this
Ψ ∈ H2 with det(Ψ) < 0, the following conditions are paper, we first introduced a new notion, one-vector-lossless
equivalent. sets, and provided a generalized S-procedure for inequality
(i) The matrix pair (A, B) conditions on the one-vector-lossless sets. By means of the

is DΨ-stabilizable.

 proposed generalized S-procedure, we next showed that
In −A∗
(ii) The matrix pencil s + is of full- full rank property analysis problems of polynomial matrices
0m,n B∗
c
column rank for all s ∈ DΨT . over some regions on the complex plane can be converted
(iii) There exists P ∈ Pn such that into LMI feasibility problems without introducing any con-

∗
 servatism. It turned out that many existing results such
In −A∗ In −A∗
− ΨT ⊗ P > 0 as Lyapunov’s inequalities for stability analysis of linear
0m,n B ∗ 0m,n B ∗ systems and LMIs for state-feedback controller synthesis
(iv) There exists P ∈ Pn such
that  can be viewed as particular cases of this result.
  A∗ The generalized S-procedure established in this paper
B ⊥ A I (ΨT ⊗ P ) (B ⊥ )∗ < 0 (24) is expected to be used as a new tool for control system
I
analysis and synthesis in the future work. A promising
(v) There exists Q ∈ Pn+m such that direction is polynomial methods in control in view of the
⎡ ⎤∗ ⎡ ⎤
A∗ A∗ recent intensive studies on this topic [7], [8]. In addition
⎢ −B ∗ ⎥ ⎢ ∗ ⎥ to providing new insights to existing results, it is expected
⎢ ⎥ (ΨT ⊗ Q) ⎢ −B ⎥ < 0 (25) that we can obtain novel results for control systems analysis
⎣ In ⎦ ⎣ In ⎦
and synthesis by working with the proposed generalized S-
0m,n 0m,n
procedure.
If the matrices A, B and Ψ are all real, the equivalence still
holds when we restrict P in (iii), (iv) and Q in (v) to be R EFERENCES
real. [1] B. R. Barmish, New Tools for Robustness of Linear Systems, Macmil-
Proof: The equivalence of the conditions (ii) and (iii) lan Publishing Company (1994).
[2] S. P. Boyd, L. E. Ghaoui, E. Feron, and V. Balakrishnan, Linear
is apparent from Corollary 1. In view of the fact that Matrix Inequalities in System and Control Theory, SIAM (1998).

∗ ⊥ [3] Y. Ebihara, K. Maeda and T. Hagiwara, “Generalized S-procedure
In −A∗   for Inequality Conditions on One-Vector-Lossless Sets and Linear
= B⊥ A I System Analysis,” Technical Report No. 04-03, Department of Elec-
0m,n B ∗ trical Engineering, Kyoto University (2004).
[4] P. Gahinet, A. Nemirovskii, A. J. Laub and M. Chilali, LMI Control
the equivalence of (iii) and (iv) follows from Finsler’s Toolbox, The Mathworks Inc. (1994).
Lemma [2]. On the other hand, the equivalence of (ii) and [5] D. Henrion and G. Meinsma, “Rank-One LMIs and Lyapunov’s
(v) is a direct consequence from Corollary 2. Inequality,” IEEE Trans. Automat. Contr., Vol. 46, No. 8, pp. 1285–
1288 (2001).
It is meaningful to examine the conditions (iv) and (v) [6] D. Henrion, O. Bachelier and M. S̆ebeek, “D-stability of Polynomial
in the light of existing LMI conditions for state-feedback Matrices,” International Journal of Control, Vol. 74, No. 8, pp. 845–
stabilizing controller synthesis [2], [13]. Let us take Ψ to 856 (2001).
[7] D. Henrion, M. S̆ebeek and V. Kuc̆era, “Robust Pole Placement
Ψc given in (11) for simplicity and consider the continuous- for Second-Order Systems: An LMI Approach,” Proc. IFAC Robust
time stabilizing controller synthesis cases. Then, the LMI Control Design (2003).
condition (24) reduces to [8] D. Henrion, M. S̆ebeek and V. Kuc̆era, “Positive Polynomials and
Robust Stabilization with Fixed-order Controllers,” IEEE Trans.
P > 0, B ⊥ (AP + P A∗ )(B ⊥ )∗ < 0 (26) Automat. Contr., Vol. 48, No. 7, pp. 1178–1186 (2003).
[9] T. Iwasaki, LMI and Control (in Japanese), Shokodo, (1997).
On the other hand, the LMI condition (25) comes to be [10] T. Iwasaki, G. Meinsma and M. Fu, “Generalized S-procedure
and Finite Frequency KYP Lemma,” Mathematical Problems in
Q11 > 0, AQ11 − BQ∗12 + (AQ11 − BQ∗12 )∗ < 0 (27) Engineering, Vol. 6, pp. 305–320 (2000).
[11] T. Iwasaki and S. Hara, “From Generalized KYP Lemma to Engi-
where neering Applications,” Proc. Conference on Decision and Control,

 pp. 792–797 (2003).
Q11 Q12 [12] T. Iwasaki and S. Hara, “Generalized KYP Lemma: Unified Char-
:= Q, Q11 ∈ Cn×n , Q22 ∈ Cm×m
Q∗12 Q22 acterization of Frequency Domain Inequalities wit Application to
System Design,” Technical Report of The Univ. Tokyo, METR2003-
The condition (26) is known as the elimination-of-variables 27 (2003). http://www.keisu.t.u-tokyo.ac.jp/Research/techrep.0.html.
[13] R. E. Skelton, T. Iwasaki and K. Grigoriadis, A Unified Algebraic
type LMI condition [13] for state-feedback stabilizing con- Approach to Linear Control Design, Taylor & Francis (1997).
troller synthesis, while the condition (27) is known as the [14] A. I. G. Valdulakis, Linear Multivariable Control. Algebraic Analysis
change-of-variables type LMI condition [2]. Corollary 4 and Synthesis Methods, John Wiley & Sons (1991).
[15] K. Zhou, K. Glover and J. C. Doyle, Robust and Optimal Control,
suggests that these well-known LMIs can be readily ob- Prentice Hall (1996).
tained from full rank property analysis of polynomial ma- [16] K. Zhou and J. C. Doyle, Essentials of Robust Control, Prentice Hall
trices by means of the proposed generalized S-procedure. (1998).

1277