2015 European Control Conference (ECC)

July 15-17, 2015. Linz, Austria

New Extended LMI Characterization for State Feedback Control of

Continuous-Time Uncertain Linear Systems
Luis A. Rodrigues1 , Ricardo C. L. F. Oliveira2 , Juan F. Camino1

Abstract— This paper provides synthesis conditions of robust parameter-dependent Lyapunov function, a set of parameter-
state feedback control law for continuous-time linear time- ized LMI conditions for state feedback control design that
invariant systems with polytopic uncertainty. By conveniently also provides an upper bound on the H∞ performance of the
expressing the Lyapunov stability condition of the closed-loop
system and using the elimination lemma, a new LMI synthesis closed-loop system has been proposed. A different approach
condition with scalar searches is proposed. As main novelty, the was proposed in [15], where parameter-dependent slack
scalar search is constrained to the interval (-1,1). Additionally, variables are introduced in the synthesis conditions. Recently,
it is shown that the proposed relaxation contains the less it has been shown in [18], by means of an exhaustive test,
conservative condition available in the literature as a particular that the state feedback stabilization condition from [19] is
case. Numerical examples illustrate the proposed approach.
the less conservative condition available in the literature for
continuous-time polytopic systems.
I. I NTRODUCTION The scope of this paper is the robust state feedback control
Robust control of polytopic linear systems has been widely design for continuous-time linear time-invariant systems with
investigated by the control community over the last decades. polytopic uncertainty. The elimination lemma [1, 23] is em-
It has been shown in [1] that many control design problems ployed to provide a new condition for robust state feedback
can be cast as convex optimization problems that involve controller synthesis. This condition generalizes the quadratic
linear matrix inequalities (LMIs), for which efficient com- stability condition [20] as well as some existing conditions
putational tools are available. However, LMI conditions for based on parameter-dependent Lyapunov functions [14, 19,
the design of robust controllers, based on Lyapunov theory, 21, 22]. The efficiency of the proposed relaxations is evalu-
are in general conservative [2–5]. This fact has motivated ated by means of numerical examples.
many researchers to look for less conservative LMI based
conditions [6–16]. II. P ROBLEM F ORMULATION
Particulary in the context of robust control synthesis for Consider the continuous-time linear time-invariant system
uncertain linear systems, a significant breakthrough was
the use of parameter-dependent Lyapunov functions. In [7], ẋ(t) = A(α)x(t) + B(α)u(t), (1)
an LMI representation for the discrete-time robust stability where x ∈ Rnx is the state vector and u ∈ Rnu is the control
condition that does not present the product between the input. It is assumed that the matrices A(·) ∈ Rnx ×nx and
Lyapunov matrix and the system dynamic matrix was pro- B(·) ∈ Rnx ×nu have the following polytopic uncertainty
posed by introducing a new matrix variable, called slack representation:
variable [17]. This decoupling between the Lyapunov matrix
N N
and the system dynamic matrix allowed to obtain a robust X X
A(α) = αi Ai , B(α) = αi B i ,
stabilization condition using parameter-dependent Lyapunov
i=1 i=1
functions, that encompasses the quadratic stability based
condition as a particular case. with α ∈ RN lying in the unit simplex Λ defined by
Differently from the stabilization conditions for discrete- n N
X o
time systems, the continuous-time stabilization conditions Λ := α ∈ RN : αi = 1, αi ≥ 0, i = 1, . . . , N .
based on parameter-dependent Lyapunov functions do not i=1
contain the quadratic stability condition, by means of a The aim is to provide a sufficient condition for the
given choice for the slack variable [11, 14]. To overcome synthesis of a robust gain associated to the state feedback
this weakness, researchers have resorted to parameterized control law
LMIs, that is, matrix inequalities that become LMIs by fixing
some additional scalar parameters. In [10, 12], using an affine u(t) = Kx(t), K ∈ Rnu ×nx ,
such that the closed-loop system
Supported by the Brazilian agencies CAPES, CNPq and FAPESP.
1 L. A. Rodrigues and J. F. Camino are with the School of Mechanical
ẋ(t) = Acℓ (α)x(t), Acℓ (α) := A(α) + B(α)K, (2)
Engineering, University of Campinas – UNICAMP, 13083-970, Campinas,
SP, Brazil. {larpig,camino}@fem.unicamp.br. is robustly stable for all α ∈ Λ.
2 R. C. L. F. Oliveira is with the School of Electrical and Computer
Engineering, University of Campinas – UNICAMP, 13083-852, Campinas, Based on the Lyapunov stability theory, the problem under
SP, Brazil. ricfow@dt.fee.unicamp.br. investigation can be characterized in terms of LMIs [1],

978-3-9524269-4-4 © EUCA 1992

using, for instance, the quadratic stability approach [20], that (5) is always feasible with W (α) = W = X = X ′
reproduced below for convenience. and sufficiently small ξ when (3) is feasible. Furthermore,
Lemma 1 (Quadratic Stabilization): If there exist matri- Lemma 3 illustrates how the elimination lemma can be
ces W = W ′ > 0 and Z such that used to introduce new variables (X, ξ) in order to generate
less conservative robust stabilization conditions. Observe that
A(α)W + W A(α)′ + B(α)Z + Z ′ B(α)′ < 0, (3) condition (5) only becomes an LMI for constant ξ and α.
holds for all α ∈ Λ, then the state feedback gain K := Hence, the design of gain K through Lemma 3 relies on
ZW −1 ensures that the closed-loop system (2) is robustly solving (5) for several values of the scalar parameter ξ. One
stable. drawback of this search is that the domain where ξ must be
The above state feedback quadratic stabilization condition searched is unbounded.
uses a Lyapunov function that is quadratic in the states and In what follows, we propose an extended LMI characteri-
independent of the uncertainty. Hence, a fixed Lyapunov zation with a search on the scalar parameter ξ limited to the
matrix W is sought to guarantee the robust stability of interval (−1, 1), as in the case of discrete-time systems [16,
the whole domain, yielding in general conservative results. 25]. Moreover, a new degree of freedom, denoted by ǫ, is
As an alternative to reduce the conservatism, the use of a introduced, generalizing Lemma 3.
parameter-dependent Lyapunov matrix has been suggested in
the literature. However, such strategy generally requires that III. M AIN R ESULT
the product between the Lyapunov matrix and the dynamic The main result of this paper is based on the fact that the
matrix A(·) be decoupled. Among the techniques available robust stability condition of the closed-loop system, that is,
in the literature to perform this task, one can mention the W (α)Acℓ (α)′ + Acℓ (α)W (α) < 0, can be rewritten as the
Finsler’s lemma [22] and dilation lemma [19]. In this paper, following matrix inequality (see [26]):
the proposed methodology is based on the elimination lemma
[2, 3, 23]. Âcℓ (α)W (α)Âcℓ (α)′ − Ãcℓ (α)W (α)Ãcℓ (α)′ < 0, (6)
Lemma 2 (Elimination Lemma): Given matrices Q = where
Q′ ∈ Rn×n , A ∈ Rm×n and B ∈ Rr×n , there exists an
unstructured matrix X ∈ Rr×m that satisfies Ãcℓ (α) := ǫAcℓ (α) − I/(2ǫ), Âcℓ (α) := ǫAcℓ (α) + I/2ǫ,

Q + A′ X ′ B + B ′ X A < 0, (4) with ǫ 6= 0.

Applying the elimination lemma to (6), we obtain the
if and only if main contribution of the paper, which is a new robust
A′⊥ QA⊥ < 0, and ′
B⊥ QB⊥ < 0, state feedback stabilization condition for continuous-time
uncertain linear systems.
where A⊥ and B⊥ are arbitrary matrices whose columns Theorem 1: Let ǫ 6= 0 and ξ ∈ (−1, 1). If there exist
form a basis for the nullspaces of A and B, respectively. a parameter-dependent matrix W (α) = W (α)′ > 0 and
In [24], various extensions of the standard LMI conditions matrices X and Z such that
for stability and performance analysis of linear systems are  
Ψ(1,1) ⋆
reviewed and discussed in the light of elimination lemma. 0 > Ψ(α) := , (7)
Ψ(2,1) Ψ(2,2)
In [18], the main methods of robust state feedback control
design for polytopic systems are surveyed and compared. Ex- holds for all α ∈ Λ, where2
haustive numerical tests showed that the robust stabilization n o
condition from [19] is the less conservative one available Ψ(1,1) = He Ã(α)X + B̄(α)Z + W (α),
in the literature for the design of robust state feedback Ψ(2,1) = ξX ′ Ã(α)′ + ξZ ′ B̄(α)′ − Â(α)X − B̄(α)Z,
controllers. n o
Lemma 3 ([19]): Let ξ > 0. If there exist a parameter- Ψ(2,2) = He −ξ Â(α)X − ξ B̄(α)Z − W (α),
dependent matrix W (α) = W (α)′ > 0 and matrices X and
with
Z such that1
Ã(α) := ǫA(α) − I/(2ǫ),
A(α)X + X ′ A(α)′ +
  
′ ′ ⋆ Â(α) := ǫA(α) + I/(2ǫ),
 B(α)Z + Z B(α) 
 
 < 0, (5)
 W (α) − X+ ′
 B̄(α) := ǫB(α),
′ ′ ′ ′ −ξ (X + X )
ξX A(α) + ξZ B(α)
then the state feedback gain K := ZX −1 ensures that the
holds for all α ∈ Λ, then the state feedback gain K := closed-loop system (2) is robustly stable.
ZX −1 ensures that the closed-loop system (2) is robustly Proof: Inequality (7) can be rewritten as (4), with
stable.      
Unlike Lemma (1), the above condition allows the use W (α) 0 ′ Ãcℓ (α) ′ I
Q= , A = , B = ,
of parameter-dependent Lyapunov matrix. It can be shown 0 −W (α) −Âcℓ (α) ξI

1 The symbol ⋆ means a symmetric block. 2 The notation He {M } means M + M ′ .

1993
and X = X ′ , where Ãcℓ (α) := Ã(α) + B̄(α)K and holds. Noting that (1/ǫ)I = Âcℓ (α) − Ãcℓ (α), (10) can be
Âcℓ (α) := Â(α) + B̄(α)K. Hence, according to the elimi- rewritten as follows
nation lemma, Ψ(α) < 0 implies 
Ãcℓ (α)X + X ′ Ãcℓ (α)′ + W (α) < − ξX ′ Ãcℓ (α)′
0 > A′⊥ QA⊥ = W (α)Acℓ (α)′ + Acℓ (α)W (α),   
+ Âcℓ (α)X + − ξ Ãcℓ (α)X + X ′ Âcℓ (α)′ −
with A′⊥ = Âcℓ (α) Ãcℓ (α) , and
 
 
− ξ Âcℓ (α)X − ξX ′ Âcℓ (α)′ − W (α) .
′
0 > B⊥ QB⊥ = (ξ 2 − 1)W (α)
Finally, using (9), the following condition can be obtained
⇒ W (α) > 0, ∀ξ ∈ (−1, 1),

′ Ãcℓ (α)X + X ′ Ãcℓ (α)′ + W (α) < ξ Ãcℓ (α)X
 
with B⊥ = ξI −I .
Finally, the non-singularity of X is ensured by the

− X ′ Âcℓ (α)′ − ξ Âcℓ (α)X − ξX ′ Âcℓ (α)′ −
block (1,1) of (7) since Ãcℓ (α)X + X ′ Ãcℓ (α)′ < −W (α) <
−1 
0 requires that X has full rank.

W (α) ξX ′ Ãcℓ (α)′ − Âcℓ (α)X ,
Note that the stabilization condition of Lemma 3 is based
on a line search in the scalar ξ ∈ (0, ∞). On the other hand, which, by Schur complement, results in (7).
Theorem 1 provides a new controller synthesis condition A. Finite-dimensional set of sufficient LMI conditions
with ξ belonging to the bounded interval (−1, 1). Moreover,
it introduces an extra degree of freedom ǫ, which shows In order to solve the infinite-dimensional matrix inequali-
that our condition is less conservative than the conditions ties of Theorem 1 in terms of a finite set of LMIs, a particular
from Lemma 1 and Lemma 3. The following two theorems structure for the matrix W (α) must be provided. Considering
formalize this fact. that the Lyapunov matrix has a polytopic structure as the
system matrices, that is,
Theorem 2: If condition (3) in Lemma 1 holds, then
condition (7) in Theorem 1 also holds. N
X
Proof: If condition (3) is satisfied, then for all ǫ > 0 W (α) = α i Wi , (11)
i=1
 
0 > 2ǫ W Acℓ (α)′ + Acℓ (α)W = the following theorem can be derived.
 ′   Theorem 4: Let ǫ 6= 0 and ξ ∈ (−1, 1). If there exist
W Ãcℓ (α) + Âcℓ (α) + Ãcℓ (α) + Âcℓ (α) W, matrices 0 < Wi = Wi′ ∈ Rnx ×nx , i = 1, . . . , N , X ∈
Rnx ×nx and Z ∈ Rnu ×nx such that
or equivalently,  
Ψi(1,1) ⋆
0 > Ψi := , (12)
W Ãcℓ (α)′ + Ãcℓ (α)W + W < Ψi(2,1) Ψi(2,2)
− W Âcℓ (α)′ − Âcℓ (α)W + W, (8) holds for i = 1, . . . , N , where
holds. Using the matrix inequality Ψi(1,1) = Ãi X + X ′ Ã′i + B̄i Z + Z ′ B̄i′ + Wi ,

G′ + G − P ≤ G′ P −1 G, P > 0, (9) Ψi(2,1) = ξX ′ Ã′i − Âi X + ξZ ′ B̄i′ − B̄i Z,

Ψi(2,2) = −ξ Âi X − ξX ′ Â′i − ξ B̄i Z − ξZ ′ B̄i′ − Wi ,
introduced by [7] (see also [15]), one has
with
− W Âcℓ (α)′ − Âcℓ (α)W + W ≤
 ′    Ãi := ǫAi − I/(2ǫ), Âi := ǫAi + I/(2ǫ), B̄i := ǫBi ,
− Âcℓ (α)W − W −1 − Âcℓ (α)W . −1
then the state feedback gain K := ZX ensures that the
Hence, condition (8) implies that closed-loop system (2) is robustly stable.
Proof: Let α ∈ Λ. Multiplying Ψi by αi and summing
Ãcℓ (α)W + W Ãcℓ (α)′ + W < for i = 1, . . . , N yields
 ′   
− Âcℓ (α)W − W −1 − Âcℓ (α)W , Ψ(α) = α1 Ψ1 + α2 Ψ2 + · · · + αN ΨN ,

holds. Finally, by Schur complement, the above inequality considering the structure (11) for the Lyapunov matrix.
is equivalent to condition (7) in Theorem 1 with W (α) = Therefore, the feasibility of the LMIs given in (12) ensures
W = X = X ′ and ξ = 0. that Ψ(α) < 0 for all α ∈ Λ.
Theorem 3: If condition (5) in Lemma 3 holds, then Theorem 4 provides a new sufficient condition for the
condition (7) in Theorem 1 also holds. synthesis of state feedback controllers that is numerically
tractable. Thus, a robust state feedback gain is obtained
Proof: If condition (5) is satisfied, then for all ǫ > 0
by solving a feasibility test based on a set of LMIs with
and ξ ∈ (−1, 1) the following
2N nx LMI rows and nx 2 (N + 2)/2 + nx (N + 2nu )/2 scalar
−(1/ǫ) (X + X ′ ) < (1/ǫ)ξ (X + X ′ ) , (10) variables. Note that the parameter ǫ lies in the unbounded

1994
 TABLE I
P ERCENTAGE OF SYSTEMS GIVEN IN E XAMPLE 1 STABILIZED BY A ROBUST STATE FEEDBACK GAIN SYNTHESIZED USING THE CONDITIONS OF
T HEOREM 4.

success
ξ\ǫ 10−6 10−5 10−4 10−3 10−2 10−1 100 101 102 103 104 105 106
(%)
-0.90 0 0 0.07 0.03 1.28 16.00 49.21 18.32 16.64 16.42 16.28 15.92 10.71 56.28
-0.54 0 0 0.03 0.21 3.82 33.67 26.78 17.32 16.50 16.28 16.32 16.14 14.00 51.17
-0.18 0.03 0 0.07 0.42 6.25 44.32 22.92 16.85 16.50 16.35 16.32 16.35 15.07 56.35
0.18 0 0.03 0.03 0.67 9.03 51.96 20.75 16.67 16.42 16.28 16.35 16.35 15.35 59.67
0.54 0 0.07 0.03 1.21 13.28 53.07 18.75 16.64 16.42 16.35 16.39 16.32 15.67 58.67
0.90 0 0.03 0.17 3.10 28.07 30.67 17.50 16.50 16.42 16.28 16.39 16.35 15.67 50.17
success(%) 0.03 0.07 0.21 3.14 28.07 63.35 52.60 18.42 16.64 16.42 16.39 16.35 15.85 65.14

interval (−∞, 0)∪(0, ∞), however, it can be proved that the {10−2 , 10−1 , 100 }, 65.03% of the systems are stabilized.
condition for some (ǫ,ξ) is the same for (−ǫ,ξ). Thus, without Interestingly, the number of stabilized systems does not de-
loss of generality, the parameter domain for ǫ can be limited crease considering only ǫ ∈ {10−1 , 100 }. In other words, all
to (0, ∞). Moreover, numerical experiments indicate that the systems stabilized through the conditions of Theorem 4 with
line search on ǫ can be limited to the interval ǫ ∈ [10−1 , 100 ]. ǫ = 10−2 are also stabilized using ǫ ∈ {10−1 , 100 }. More-
over, note that the total number of stabilized systems is an
IV. N UMERICAL E XAMPLES increasing function of ǫ into the interval {10−6 , . . . , 10−1 }
A. Example 1 and a decreasing function into the interval {100 , . . . , 106 },
indicating that ǫ has an optimal value between 10−1 and
This example uses the database of uncertain systems
100 . Therefore, we propose to limit the line search on ǫ to
from [18] which has 2800 unstable linear systems that
the interval [10−1 , 100 ].
are stabilized by a robust state feedback controller but are
2) Comparison with literature: Now, to evaluate the con-
not quadratically stabilized. The aim is to evaluate the
servatism of the proposed approach, the conditions given in
conservatism of the state feedback stabilization condition of
Theorem 4 are compared to the methods, for robust state
Theorem 4 by means of exhaustive tests. First, we analyze
feedback controller design, presented in [10, 11, 14, 15, 19].
the search space of the parameter ǫ. Next, we compare our
It has been suggested in [18] to limit ξ to the following 13
proposed approach with some existing methods.
logarithmically equally spaced points:
1) Analysis of the search domain of ǫ : The aim is to
determine a limited interval search for the parameter ǫ from ξ ∈ {10−6 , 10−5 , 10−4 , 10−3 , 10−2 ,
Theorem 4. To this end, instead of performing an exhaustive
10−1 , 100 , 101 , 102 , 103 , 104 , 105 , 106 },
line search on ǫ, we have selected 13 logarithmically equally
spaced points between 10−6 and 106 : instead of performing an exhaustive line search on the scalar
parameter ξ. These are the values of ξ used in the LMI con-
ǫ ∈ {10−6 , 10−5 , 10−4 , 10−3 , 10−2 , ditions from [10, 15, 19]. For the conditions in Theorem 4,
10−1 , 100 , 101 , 102 , 103 , 104 , 105 , 106 }. we have used the values of ǫ and ξ suggested in the previous
subsection, as follows (total of 12 searches):
Concerning the parameter ξ, we have chosen 6 linearly
equally spaced points between −0.9 and 0.9: ǫ ∈ {10−1 , 100 }, ξ ∈ {−0.90, −0.54, −0.18, 0.18, 0.54, 0.90}.
ξ ∈ {−0.90, −0.54, −0.18, 0.18, 0.54, 0.90}. The results are presented in Table II.
Table I shows the number of systems stabilized by
TABLE II
a robust state feedback gain synthesized through the
P ERCENTAGE OF SYSTEMS GIVEN IN E XAMPLE 1 STABILIZED BY A
conditions of Theorem 4. As can be seen, consider-
ROBUST STATE FEEDBACK GAIN SYNTHESIZED USING THE METHODS IN
ing all combinations of ǫ ∈ {10−6 , 10−5 , . . . , 106 } and
[10, 11, 14, 15, 19] AND T HEOREM 4.
ξ ∈ {−0.90, −0.54, . . . , 0.90}, the conditions of Theo-
rem 4 provided robust state feedback gains for 65.14%
Method success (%)
of the 2800 unstable system. However, its performance
Theorem 4ǫ={0.1,1},ξ={−0.9,0.54,...,0.9} 65.03
can vary abruptly depending on the value of ǫ. For ǫ ∈
[15, Theorem 4]ξ={10−6 ,10−5 ,...,106 } 07.46
{10−6 , 10−5 , 10−4 , 10−3 }, less than 3.57% of the systems
[19, Theorem 2]ξ={10−6 ,10−5 ,...,106 } 62.10
are stabilized. For ǫ ∈ {101 , 102 , 103 , 104 , 105 , 106 }, 18.42%
[14, Theorem 2] 23.85
of the systems can be stabilized. The most efficient values
[10, Lemma 3]ξ={10−6 ,10−5 ,...,106 } 18.78
of ǫ were 10−1 , 100 and 10−2 , that stabilized 63.35%,
[11, Condition 18] 15.64
52.60% and 28.07% of the systems, respectively. For ǫ ∈

1995
 TABLE III TABLE IV
M AXIMUM VALUE OF d SUCH THAT THE SYSTEM GIVEN IN E XAMPLE 2 N UMERICAL COMPLEXITY (V SCALAR VARIABLES ; R LMI ROWS ) OF
IS STABILIZABLE BY A ROBUST STATE FEEDBACK GAIN SYNTHESIZED THE METHODS IN [10, 11, 14, 15, 19, 20] AND T HEOREM 4, APPLIED TO
USING THE METHODS IN [10, 11, 14, 15, 19, 20] AND T HEOREM 4. CONTROL DESIGN PROBLEM GIVEN IN E XAMPLE 2.

Method dmax Method V R

Theorem 4ǫ={0.1,1},ξ={−0.9,0.54,...,0.9} 14.2960 Theorem 4 40 16
[15, Theorem 4]ξ={10−6 ,10−5 ,...,106 } 11.6238 [15, Theorem 4] 120 84
[19, Theorem 2]ξ={10−6 ,10−5 ,...,106 } 13.3063 [19, Theorem 2] 40 24
[14, Theorem 2] 08.2749 [14, Theorem 2] 40 24
[10, Lemma 3]ξ={10−6 ,10−5 ,...,106 } 10.3738 [10, Lemma 3] 40 16
[11, Condition 18] 05.4469 [11, Condition 18] 40 24
[20, Theorem 1] 04.0000 [20, Theorem 1] 14 12

As can be seen, the LMI based methods from [11, 14] Table IV shows the numerical complexity associated with
provided robust stabilizing state feedback gains for 15.64% the methods from [10, 11, 14, 15, 19, 20] and Theorem 4. As
and 23.85% of the systems, respectively. On the other hand, discussed in [1, 27], the numerical complexity associated to
with 13 searches for ξ, the parameterized LMI based method an optimization problem based on LMIs can be estimated
from [19] improves the results from [14] in 160%, providing from the number V of scalar variables and the number R
feasible solutions for 62.10% of the systems. Regarding our of LMI rows. The methods based on parameter-dependent
proposed method, the conditions of Theorem 4 provided less Lyapunov functions, [10, Lemma 3] and Theorem 4, demand
conservative results than all the other methods. Thus, the the less numerical complexity followed by the methods from
number of stabilized systems provided by the less conser- [11, 14, 19] and [15], respectively. Note that, although the
vative method available in the literature (given in [19]) was methods from [11, 14] present a greater numerical complex-
increased by 4.6%. ity than [10] and Theorem 4, they do not have line search.
The greater number of scalar variables in [15, Theorem 4]
B. Example 2
is due to the additional slack variables, while the greater
Consider the system (1) for nx = 4, nu = 1 and N = 2 number of LMI rows is due to additional constraints on
with following matrices the slack variables. Finally, it is worth mentioning that
the additional number of LMI rows of the methods from
   
0 1 0 0 0
 0 0 1 0  0 [14, 19] with respect to [10] and Theorem 4 is due to an
A1 =   , B1 =   ,
additional condition to ensure that Lyapunov function is
 0 0 0 1 −6
−12 − 3d −12 − 3d −25 −1 6 positive definite, however, its condition can be removed
    without loss of generality, as shown in [24].
0 1 0 0 0
 0 0 1 0   0 C. Example 3
A2 =   , B2 =   ,
 0 0 0 1  0
Consider the system (1) with the following affine
−12 + 3d −12 + 3d −25 −1 6
parameter-dependent matrices
borrowed from [15]. The aim is to determine the maximum    
0.4 0.1 −0.7 0
value of d ≥ 0 such that the system is stabilizable by a robust
A(θ1 ) = −0.5 0.5 −2.1  , B(θ2 ) = 1 − θ2  ,
state feedback gain.
1.2 1.3 −0.4θ1 θ2
Table III shows the maximum value of d such that the
system can be stabilized by a robust state feedback gain where θ1 ∈ [0, θ̄] and θ2 ∈ [0, 1]. Note that it is possible to
synthesized through the methods from [10, 11, 14, 15, 19, 20] obtain a polytopic representation for these uncertain matrices
and Theorem 4. All the methods provide robust gains for d ≤ by combining the extreme values of θ1 and θ2 , yielding A(α)
4.0000. For d > 4.0000, the quadratic stabilization condition and B(α) with the following 4 vertices:
[20] becomes unfeasible. However, the methods based on    
0.4 0.1 −0.7 0.4 0.1 −0.7
parameter-dependent Lyapunov function still provide robust
A1 = −0.5 0.5 −2.1 , A3 = −0.5 0.5 −2.1  ,
gains. Using ξ = {10−6 , 10−5 , . . . , 106 }, the method in [19]
1.2 1.3 0 1.2 1.3 −0.4θ̄1
provides feasible solutions until d ≤ 13.3063. On the other    
hand, the conditions of Theorem 4 with ǫ = {0.1, 1} and 0.4 0.1 −0.7 0.4 0.1 −0.7
ξ = {−0.9, 0.54, . . . , 0.9} provide feasible solutions for all A2 = −0.5 0.5 −2.1 , A4 = −0.5 0.5 −2.1  ,
d ≤ 14.2960. Hence, the proposed approach increased by 1.2 1.3 0 1.2 1.3 −0.4θ̄1
′ ′ ′
B1 = 0 1 0 , B2 = 0 0 1 , B3 = 0 1 0 , B4′ = 0 0 1 .
       
7.5% the maximum value of d. The resulting state feedback
gain for d = 14.2960, ǫ = 10−1 and ξ = 0.18 is given by
  The aim is to determine the maximum value of θ̄ ≥ 0 such
K = −12.7808 −27.1174 −0.5186 −2.0615 . (13) that the system is stabilizable by a robust state feedback gain.

1996
 TABLE V
[7] M. C. de Oliveira, J. Bernussou, and J. C. Geromel, “A new discrete-
M AXIMUM VALUE OF θ̄ SUCH THAT THE SYSTEM GIVEN IN E XAMPLE 3 time robust stability condition,” Syst. Control Lett., vol. 37, no. 4, pp.
IS STABILIZABLE BY A ROBUST STATE FEEDBACK GAIN SYNTHESIZED 261–265, July 1999.
[8] T. Shimomura, M. Takahashi, and T. Fujii, “Extended-space control
USING THE METHODS IN [10, 11, 14, 15, 19, 20] AND T HEOREM 4.
design with parameter-dependent Lyapunov functions,” in Proc. 40th
IEEE Conf. Decision Control, Orlando, FL, USA, December 2001, pp.
Method θ̄max 2157–2162.
[9] A. Trofino and C. E. de Souza, “Biquadratic stability of uncertain
Theorem 4ǫ={0.1,1},ξ={−0.9,0.54,...,0.9} 8.7681 linear systems,” IEEE T. Automat. Contr., vol. 46, no. 8, pp. 1303–
[15, Theorem 4]ξ={10−6 ,10−5 ,...,106 } 0.1819 1307, August 2001.
[19, Theorem 2]ξ={10−6 ,10−5 ,...,106 } 5.4860 [10] U. Shaked, “Improved LMI representations for the analysis and the
design of continuous-time systems with polytopic type uncertainty,”
[14, Theorem 2] 5.4860 IEEE T. Automat. Contr., vol. 46, no. 4, pp. 652–656, Apr. 2001.
[10, Lemma 3]ξ={10−6 ,10−5 ,...,106 } – [11] P. Apkarian, H. D. Tuan, and J. Bernussou, “Continuous-time analysis,
[11, Condition 18] 3.5276 eigenstructure assignment, and H2 synthesis with enhanced linear
matrix inequalities (LMI) characterizations,” IEEE T. Automat. Contr.,
[20, Theorem 1] – vol. 46, no. 12, pp. 1941–1946, Dec. 2001.
[12] Y. Jia, “Alternative proofs for improved LMI representations for the
analysis and the design of continuous-time systems with polytopic type
uncertainty: A predictive approach,” IEEE T. Automat. Contr., vol. 48,
Table V shows the maximum value of θ̄ such that the no. 8, pp. 1413–1416, August 2003.
system can be stabilized by a robust state feedback gain [13] C. E. de Souza, A. Trofino, and J. de Oliveira, “Parametric Lyapunov
function approach to H2 analysis and control of linear parameter-
synthesized through the methods from [10, 11, 14, 15, 19, 20] dependent systems,” IEE Proc. — Control Theory and Appl., vol. 150,
and Theorem 4. As can be seen, the system is not quadratic no. 5, pp. 501–508, September 2003.
stabilizable [20] for any value of θ̄. Based on a parameter- [14] S. S. Chughtai and N. Munro, “Robust stability condition for
continuous-time systems,” Electron. Lett., vol. 40, no. 16, pp. 978–
dependent Lyapunov function, although the method from 979, Aug. 2004.
[10] is not able to provide robust gains for any value of θ̄, [15] J. C. Geromel and R. H. Korogui, “Analysis and synthesis of robust
the methods from [11, 14, 15, 19] provide feasible solutions control systems using linear parameter dependent Lyapunov func-
tions,” IEEE T. Automat. Contr., vol. 51, no. 12, pp. 1984–1989, Dec.
for limited values of θ̄. The best result, θ̄max = 5.4860, 2006.
was provided simultaneously by the methods from [14, [16] S. Salhi and G. Garcia, “H2 and H∞ robust control for continuous-
19]. Regarding the proposed approach, the conditions of time linear systems,” Technical Report 08146, LAAS-CNRS, Mar. 2008.
[17] Y. Ebihara, D. Peaucelle, and D. Arzelier, S-Variable Approach to
Theorem 4 increased this value by 56.4%, providing robust LMI-Based Robust Control, London, PA: Springer-Verlag, 2015.
gains for all θ̄ ≤ 8.7681. [18] R. C. L. F. Oliveira, M. C. de Oliveira, and P. L. D. Peres, “Robust
state feedback LMI methods for continuous-time linear systems:
V. C ONCLUSION Discussions, extensions and numerical comparisons,” in Proc. 2011
IEEE Int. Symp. on Comput. Aided Control Syst. Des., Denver, CO,
The paper provides a new robust state feedback synthesis USA, Sept. 2011, pp. 1038–1043.
condition for time-invariant continuous-time uncertain linear [19] Y. Ebihara and T. Hagiwara., “New dilated LMI characterizations
systems. It was shown, from a theoretical point of view and for continuous-time multiobjective controller synthesis,” Automatica,
vol. 40, no. 11, pp. 2003–2009, Nov. 2004.
corroborated by numerical examples, that the proposed LMI [20] J. Bernussou, P. L. D. Peres, and J. C. Geromel, “A linear programming
condition, with a scalar search, is less conservative than the oriented procedure for quadratic stabilization of uncertain systems,”
others conditions available in the literature. The proposed Syst. Control Lett., vol. 13, no. 1, pp. 65–72, 1989.
[21] J. C. Geromel, M. C. de Oliveira, and L. Hsu, “LMI characterization
approach can also be extended to cope with output feedback of structural and robust stability,” Linear Algebra Appl., vol. 285, no.
control and filter designs, as well as to include performance 1–3, pp. 69–80, Dec. 1998.
requirements like the H2 and H∞ norms. [22] M. C. de Oliveira and R. E. Skelton, “Stability tests for constrained
linear systems,” in Perspectives in robust control, ser. Lecture Notes
R EFERENCES in Control and Information Sciences, S. R. Moheimani, Ed. New
York, NY: Springer-Verlag, 2001, vol. 268, pp. 241–257.
[1] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix [23] R. E. Skelton, T. Iwasaki, and K. Grigoriadis, A Unified Algebraic
Inequalities in Systems and Control Theory, ser. Stud. Appl. Math. Approach to Linear Control Design, Bristol, PA: Taylor & Francis,
Philadelphia, PA: SIAM, 1994, vol. 15. 1998.
[2] T. Iwasaki and R. E. Skelton, “All controllers for the general H∞ [24] G. Pipeleers, B. Demeulenaere, J. Swevers, and L. Vandenberghe,
control problem: LMI existence conditions and state-space formulas,” “Extended LMI characterizations for stability and performance of
Automatica, vol. 30, no. 8, pp. 1307–1317, August 1994. linear systems,” Syst. Control Lett., vol. 58, no. 7, pp. 510 – 518,
[3] P. Gahinet and P. Apkarian, “A linear matrix inequality approach to 2009.
H∞ control,” Int. J. Robust Nonlin., vol. 4, no. 4, pp. 421–448, July- [25] C. F. Morais, M. F. Braga, R. C. L. F. Oliveira, and P. L. D. Peres,
August 1994. “Robust state feedback control for discrete-time linear systems via
[4] I. Masubuchi, A. Ohara, and N. Suda, “LMI-based controller synthe- LMIs with a scalar parameter,” in Proc. 2013 Amer. Control Conf.,
sis: a unified formulation and solution,” Int. J. Robust Nonlin., vol. 8, Washington, DC, USA, June 2013, pp. 3876–3881.
pp. 669–686, 1998. [26] A. Trofino, D. F. Coutinho, and K. A. Barbosa, “Improved H2 and
[5] C. Scherer, P. Gahinet, and M. Chilali, “Multiobjective output- H∞ conditions for robust analysis and control synthesis of linear
feedback control via LMI optimization,” IEEE T. Automat. Contr., systems,” SBA: Cont. Autom., vol. 16, no. 4, pp. 427–434, October,
vol. 42, no. 7, pp. 896–911, July 1997. November and December 2005.
[6] P. Gahinet, P. Apkarian, and M. Chilali, “Affine parameter-dependent [27] P. Gahinet, A. Nemirovski, A. J. Laub, and M. Chilali, LMI Control
Lyapunov functions and real parametric uncertainty,” IEEE T. Automat. Toolbox User’s Guide. Natick, MA,USA: The MathWorks Inc., May
Contr., vol. 41, no. 3, pp. 436–442, March 1996. 1995.

1997