Systems & Control Letters 57 (2008) 193 – 199

www.elsevier.com/locate/sysconle

Robust H∞ output-feedback control of systems with time-delay夡

V. Suplin, U. Shaked ∗
Department of Electrical Engineering-Systems, Tel Aviv University, Tel Aviv 69978, Israel

Received 12 September 2005; received in revised form 8 July 2007; accepted 9 August 2007
Available online 18 October 2007

Abstract
The problem of designing robust dynamic output-feedback controllers for linear, continuous, time-invariant systems with uncertain time-delay
in the measured output and/or the control input and with polytopic type parameter uncertainties is considered. Given a transfer function matrix
of a system with uncertain real parameters that reside in some known ranges, an appropriate, not necessarily minimal, state–space model of
the system is described which permits reconstruction of its states. The resulting retarded model incorporates the uncertain parameters of the
transfer function matrix in the state–space matrices and the uncertain time-delay that occurs in the control channel. To this model, the recent
theory of robust H∞ state-feedback control for retarded systems is applied. The theory is used to solve a benchmark problem of distillation
column robust control design.
© 2007 Elsevier B.V. All rights reserved.

Keywords: Robust control H∞ ; Time delay; Output feedback; LMI; Polytopic uncertainty; Distillation column

1. Introduction On the other hand, a comprehensive treatment of systems with

time-delay has been made based on the Lyapunov–Krasovskii
One of the main contributions of H∞ control theory is the approach [1]. While the latter provides a means for dealing
ability to construct controllers and estimators that are robust in with parameter uncertainties and constant unknown time-delay
the sense that they are built to cope with prescribed ranges of it entails an overdesign that stems from reducing a problem
parameter uncertainties. Applying the method of convex pro- which is of infinite dimensions in nature to one that is of finite
gramming, relatively simple solutions have been obtained to dimensions. In the presence of parameters uncertainties, these
the problems of state-feedback stabilizing and disturbance at- latter methods are suitable for designing state-feedback con-
tenuation in linear systems with polytopic or norm-bounded trollers; unfortunately, however, they cannot guarantee stability
uncertainties. Unfortunately, it has soon been realized that the and performance when applying output-feedback control.
corresponding output-feedback control problems are not con- Recently, a method has been proposed for the design of
vex [4], and various iterative methods have been suggested to robust output-feedback controllers for discrete-time systems
derive the required controllers [4]. These methods are known to without delay but with polytopic type uncertainties [10]. An
converge locally and they do not necessarily achieve a global augmented state–space model for the nonretarded system is
minimum for the disturbance attenuation level. first constructed which includes delayed versions of the sys-
This obstacle has also been encountered in the design of tem inputs and outputs. On this model robust state-feedback
output controllers in cases where a time-delay appears in the controllers are applied which satisfy the performance require-
control channel. In the case where the parameters of the sys- ments. These controllers are then translated into equivalent
tem are all known, a solution to the H∞ control problem was dynamic output-feedback controllers. A similar approach is
achieved by [8]. This solution, however, does not cope with pa- adopted in the present paper for the continuous-time case with
rameter uncertainties and it requires the knowledge of the delay. uncertain delay in the control channel. Also here, an augmented
state–space model is achieved to which a state-feedback con-
夡 This work was supported by C&M Maus Chair at Tel Aviv University.
trol corresponds to a dynamic output-feedback controller that
∗ Corresponding author. is applied to the original system. It is shown that the uncertain
E-mail address: [email protected] (U. Shaked). delay in the control channel can be translated to a state-delay
0167-6911/$ - see front matter © 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.sysconle.2007.08.007
194 V. Suplin, U. Shaked / Systems & Control Letters 57 (2008) 193 – 199

in the augmented state–space description. The recent result on We seek a dynamic output-feedback controller whose transfer
the robust state-feedback stabilization and control of systems function matrix is given by the following left matrix fractional
with state-delays of [9] are then applied to obtain the required description:
output-feedback controller.
H (s) = Ā−1 (s)B̄(s) (4)
The new theory is used to solve the benchmark problem of
robust control design of distillation columns [6]. This problem of an appropriate order.
has been studied by several authors, see [7] and references In the development below we consider an additional finite
therein. It is shown how an application of a simple PI controller energy disturbance vector w(t) ∈ Rq . We also consider an
and a simple notch prefilter satisfies the design specifications, objective vector signal z(t) ∈ Rp that will be defined below
in spite of large parameter uncertainty and the uncertain delay according to the performance requirements. We address the
in the control channel. following problem:
Notation: Throughout the paper, the superscript ‘T’ stands for The robust H∞ control problem: For a prescribed scalar
matrix transposition, Rn denotes the n dimensional Euclidean > 0, find a controller that asymptotically stabilizes the system
space with vector norm | · |, Rn×m is the set of all n × m and satisfies
real matrices, and the notation P > 0, for P ∈ Rn×n means ∞
that P is symmetric and positive definite. The space of square J∞ = (z(s)T z(s) − 2 w(s)T w(s)) ds < 0
0
integrable functions over [0 ∞) is denoted ∞by L2 , and · 2 ∀w(t) ≡
/ 0 ∈ L2 (5)
stands for the standard L2 -norm, u2 =( 0 u(s)T u(s) ds)1/2 .
· ∞ denotes the H∞ -norm. In is the unit matrix of order n, over the entire uncertainty polytope and for all allowed
0n,m is the n × m zero matrix and diag{A, B} denotes a block delays .
diagonal matrix with A and B on the diagonal. d (m) (t) denotes We note that the initial conditions of the system is assumed
the mth derivative of the vector function d(t) and by col{a, b} to be zero.
we denote the vector [a T bT ]T .
Remark 1. The theory below can readily be extended to the
case of different delays i , i = 1, . . . , q in the input channels,
2. Problem formulation where we have the following plant model:

We consider the following linear, time-invariant, multi-

q
G(s) = D̄(s) −1
N̄i (s)e−s i .
input–multi-output (MIMO) system:
i=1
−1 −s
G(s) = D̄(s) N̄ (s)e , (1) For the sake of simplicity of the notations and in order to avoid
1 1 −1 matrices of large size, we treat below the case of q = 1.
where D̄(s) = lk=0 D̄k s k and N̄ (s) = lk=0 N̄l1 −1−k s k are
m × m and m × r polynomial matrices, respectively, m is 3. The augmented state–space model
the dimension of output vector y and r is the dimension of
the control input u to the plant. Time-delay is an uncer- It follows from (1) and the stated model assumptions that
tain constant time-delay in either the control input channel l −1
1 −1
l 1
or in the measured output that, for a given scalar bound h, s Im +
l1
D̄k s Y (s) =
k
N̄l1 −1−k s e−s U (s). (6)
k
satisfies: k=0 k=0

0 h. (2) Note, that the last j rows in D̄i and N̄l1 −1−i , 0 i l1 −
lm−j +1 − 1, j = 1, . . . , m − 1, are identically zero.
We assume that D̄(s) is row reduced, namely that Choosing in (4):
deg det{D̄(s)} = m i=1 li = n̄, where li is the degree of ith row 1 −2
l 1 −1
l
of D̄(s) (it is assumed that li li+1 , i = 1, . . . , m) and n̄ is
Ā(s) = s l1 −1 Im + Āk s k and B̄(s) = B̄l1 −1−k s k (7)
the minimal order of G(s) [3]. We denote by D̄h the coeffi-
k=0 k=0
cient matrix of the highest order term in each row of D̄(s). To
simplify our derivation in the sequel we assume that D̄h does we obtain the following:
not incorporate uncertain parameters. Thus, we can assume, l −1
1 −2
l 1
without loss of generality, that D̄h = Im . l1 −1
s Im + Āk s U (s) =
k k
B̄l1 −1−k s Y (s). (8)
The system parameters are not completely known. We as- k=0 k=0
sume that they lie within a polytope
The configuration of the system and its controller is described
in Fig. 1.
= Co{1 , 2 , . . . , L̄ }, (3)
We define the following state vector (t)=col{y (l1 −1) (t), . . .
y(t), u(l1 −2) (t), . . . , u(t)}. Hence, we have the following
N̄0,j . . . N̄l1 −1,j
where j = , j =1, . . . , L̄, are the ver- state–space realization for the system of (1)–(4).
D̄0,i . . . D̄l1 −1,j
tices of the polytope. ˙ = A0 (t) + A1 (t − ) + B0 ū(t) + B1 ū(t − ),
(t) (9a)
V. Suplin, U. Shaked / Systems & Control Letters 57 (2008) 193 – 199 195

Fig. 1. The fractional description of the closed loop.

ū = K(t), (9b)

where (t) ∈ Rn , nl1 (m + r) − r,

⎡ ⎤
−D̄l1 −1 . . . −D̄0 0 . . . 0
⎢ Im 0 . . 0 0 . . . 0⎥
⎢ ⎥
⎢ 0 I m . . 0 0 . . . 0⎥ Fig. 2. Disturbance entrance point.
⎢ ⎥
⎢ . . . . . . . . . .⎥
⎢ ⎥
⎢ 0 . . I m 0 0 . . . 0⎥ B̄l1 −1 , B̄l1 −2 , . . . , B̄0 and N̄0 , N̄1 , . . . , N̄l1 −1 as described in
A0 = ⎢ ⎥,
⎢ 0 . . . 0 0 . . . 0⎥ Fig. 2.
⎢ ⎥
⎢ 0 . . . 0 Ir 0 . . 0⎥ The obtained representation is nonminimal, but it has the
⎢ ⎥
⎢ 0 . . . 0 0 Ir . . 0⎥ merit that the original output-feedback problem for the uncer-
⎣ ⎦
. . . . . . . . . . tain plant has been transformed into a state-feedback problem
0 . . . 0 0 . . Ir 0 where the matrices A and B lie in the polytope defined by (3)
(10a) and any state-feedback control which is designed for this rep-
⎡ ⎤ resentation, say ū(t) = K̄(t) can be translated into a dynamic
0 . . . 0 N̄1 . . . N̄l1 −1
⎢0 0 . . 0 0 . . . 0 ⎥ output-feedback controller, following the definition of (10b) for
⎢ ⎥ the structure of K̄ and (4), using the definition of (7). Note that
⎢0 0 . . 0 0 . . . 0 ⎥
⎢ ⎥ the resulting controller H is proper but not necessarily strictly
⎢. . . . . . . . . . ⎥
⎢ ⎥ proper.
⎢0 . . 0 0 0 . . . 0 ⎥
A1 = ⎢ ⎥ (10b) If the state-feedback control law of (9b) stabilizes the aug-
⎢0 . . . 0 0 . . . 0 ⎥
⎢ ⎥ mented system (9a) the corresponding controller of (4) stabi-
⎢0 . . . 0 0 0 . . 0 ⎥
⎢ ⎥ lizes the original plant.
⎢0 . . . 0 0 0 . . 0 ⎥
⎣ ⎦ We consider next the problems of achieving a prescribed
. . . . . . . . . .
0 . . . 0 0 . . 0 0 bound on the H∞ performance level. The vector sequence
w(t) ∈ Rq is the disturbance input in L2 that acts through
B0 = [0̄ 0 0 . . . 0 Ir 0 0 . . . 0]T , (10c) the input matrix Bw , and the objective vector is z(t) which is
characterized by the output matrix C1 and the disturbance cou-
B1 = [N̄0T 0 0 . . . 0 0 0 0 . . . 0]T , (10d) pling matrix D11 . The corresponding state–space model of the
system is then given by the following:
K = [B̄0 B̄1 B̄2 . . . B̄l1 −1 − Āl1 −2
− Āl1 −3 − Āl1 −4 . . . − Ā0 ]. (10e) ˙ = A0 (t) + A1 (t − ) + B0 ū(t)
(t)
The representation above is appropriate for l1 > 1. In the case + B1 ū(t − ) + Bw w(t), (11a)
where l1 = 1, the input u does not appear in the above vector . z(t) = C1 (t) + D11 w(t), (11b)
In this case, the resulting controller will then be a static output-
feedback controller so that (8) is replaced by: U (s) = B0 Y (s). where ū(t) = K(t) and the entries of Bw , C1 , and D11 are
One way to describe the above state–space representation determined according to the type of the problem solved. Note
is to consider it as a simple backward shifting of the gains that the weight which is usually put on the control effort should
196 V. Suplin, U. Shaked / Systems & Control Letters 57 (2008) 193 – 199

be incorporated in C1 , in the case where l1 > 1, since includes 4. The robust H∞ controller
the input u as a state. A term D12 u(t), where D12 is a weighting
matrix, will be added to the right-hand side of (11b) only in We consider system (11a),(11b) with the performance index
the case where l1 = 1. (5). A robust state-feedback controller is sought that achieves
We consider in the sequel the following types of problems. a minimum upper bound on the disturbance attenuation level
Complementary sensitivity: Many standard closed-loop re- over the entire uncertainty polytope. The basic tool in H∞
quirements refer to the norm of the closed-loop transference design is the bounded real lemma (BRL) [2].
that is called the ‘complementary sensitivity’ transfer function BRLs for stable linear time-invariant systems with state-
[5]. This matrix is given by T (s) = (Ir − L(s))−1 L(s) where delay have been derived by many authors (see [1] and the ref-
L(s) = H (s)G(s). The requirement is then to reduce the H∞ erences therein). Delay dependent and delay independent LMIs
norm of the transference T . This is achieved by choosing: have been obtained the solution of which guarantees that sta-
bility and the required disturbance attenuation of the closed-
Bw = [Im 0 . . . 0]T , (12a)
loop system. Of all the methods offered, those that are based
C1 = [0 . . . 0 N̄1 . . . N̄l1 −1 ] + N̄0 K (12b) on the descriptor approach seem to be the least conservative.
Substituting the control law (9b) in the BRL of [9] the follow-
and ing result is obtained for system (11a),(11b) for a single point
D11 = 0. (12c) (nominal value) in the polytope . This result also handles the
case of l1 = 1 which requires a nonzero weighting D21 .
In the above the entry point of the disturbance is described in
Fig. 2. On the other hand, measurement noise has been ne-
Lemma 1. Under the feedback law ū(t) = K(t), the sys-
glected. Instead of considering the real input point of the mea-
tem (11a),(11b) is asymptotically stable and, for a prescribed
surement noise signal one may consider the transference from
positive scalar , J∞ < 0 for all nonzero w(t) ∈ L2 [(0 ∞)
the measurement noise signal to the output {yk }. This trans-
and for any constant delay h, if for some scalar tun-
ference is the complementary sensitivity transference treated
ing parameters εi , i = 1, 2, 3 there exist n × n matrices
above.
0 < P1 , 0 < S, P2 , Y1 , Y2 , Z1 , Z2 , Z3 , R, and Ȳ ∈ Rp×n such
Sensitivity reduction: The sensitivity transfer function matrix
that
S(s) is defined by
S(s) = (I − L(s))−1 = I + T (s). + T < 0 (14a)

The minimization of the norm of the sensitivity transference and

is thus achieved by using Bw and C1 of (12a), (12b) and by
R Y1 Y2
taking D11 = Im . ∗ Z1 Z2 > 0, (14b)
Signal tracking: If one is interested in tracking a reference ∗ ∗ Z3
signal (say r), the integral of the tracking error can be added as
where
⎡ ⎤
A0 P2 +B0 Ȳ +Y1 + 21 S+ h2 Z1 P1 +ε1 (A0 P2 +B0 Ȳ )+Y2 +hZ 2 ε2 (A0 P2 + B0 Ȳ ) ε3 (A0 P2 + B0 Ȳ )C1T B
⎢ ⎥
⎢ −P2 −ε1 P2 + h2 (R + Z3 ) −ε2 P2 −ε3 P2 C1T 0 ⎥
⎢ ⎥
⎢ ⎥
= ⎢
⎢ −Y1 + A1 P2 −Y2 + ε1 (A1 P2 + B1 Ȳ ) − 21 S+ε2 (A1 P2 +B1 Ȳ ) ε3 (A1 P2 + B1 Ȳ )C1T 0 ⎥⎥ . (14c)
⎢ ⎥
⎢ C1 P2 + D12 Ȳ ε1 (C1 P2 + D12 Ȳ ) ε2 (C1 P2 + D12 Ȳ ) − 2 Im +ε3 (C1 P2 +D12 Ȳ )C1 D11 ⎥
2 T
⎣ ⎦
0 0 0 0 − 21 Iq
an additional state to the above augmented description. Namely, The state-feedback gain matrix is then given by
t
(t) = (r(¯) − y(¯)) d¯, (13a) K = Ȳ P2−1 (15)
0
and the transfer function matrix H (s) of the output-feedback
[ A ] 0
Ā = , (13b) controller is given by (4) and (7), applying (10e).
[ 0 . . . −Im 0 . . . 0] 0

[0] Remark 2. The result of Lemma 1 applies the scalar tuning
Bw = , (13c) parameters εi , i = 1, 2, 3. The question arises how to find the
Im
optimal combination of these parameters. One way to address
C1 = [0 . . . 0 Im ]. (13d) the tuning issue is to choose ε2 = ε3 = 0. Thus, the well known
Bw implies that the disturbance acts on the input of the inte- result of [1] is obtained and a linear search procedure can be
grator. applied. Another way is to apply a numerical optimization al-
In the next section we apply the recent theory of robust state- gorithm, such as the program fminsearch in the optimization
feedback stabilization and control of systems with uncertain toolbox of Matlab, to the cost function (5). A locally conver-
state-delay to our problem. gent solution to the tuning problem is thus obtained.
V. Suplin, U. Shaked / Systems & Control Letters 57 (2008) 193 – 199 197

The LMIs of Lemma 1 have been derived for a nominal value ◦ y2 (t) 0.5 for all t; and
of the system matrices in the uncertainty polytope of (3). ◦ −0.01 y2 (∞) 0.01.
Since these LMIs are affine in the system matrices the result of
Lemma 1 can be readily applied to the case where the system Corresponding requirements should hold also for a unit step
matrices reside in . demand in channel 2.

Corollary 1. Consider the system of (11a),(11b) where 5.3. The augmented model and the closed-loop controller
the system matrices reside within the polytope (3). For
a prescribed > 0, the cost function (5) achieves J∞ < 0 We solve this problem by constructing a two degree of free-
over for all nonzero ∈ Rq × L2 [0∞), and for dom controller. Namely, in phase one of the design we con-
all constant delays h, if there exist n × n matrices struct a controller that minimizes the integral of the tracking
j j j j j j
0 < P1 , 0 < S j , P2 , Y1 , Y2 , Z1 , Z2 , Z3 , R j , and Ȳ ∈ Rp×n error (see (13)). In phase two, we construct a prefilter to satisfy
that satisfy (14) for j = 1, . . . , L̄, where the matrices the input–output design specifications.
0 < S, Y1 , Y2 , Z1 , Z2 , Z3 , R, A0 , A1 and B1 , are taken with Obviously, the uncertainty polytope possesses L=4 vertices.
the superscript j . In order to incorporate the effect of a disturbance on the
output y we consider the following fourth order state vector
5. A solution to the distillation column control problem
(t) = col{y(t), (t)}, y(t), (t) ∈ R2 ,
The theory of Section 4 is applied in this section to the long where (t) was added in order to secure an integral action in
standing CDC distillation column benchmark control problem the controller, as in (13a)–(13d). It is readily found that
[6]. The plant model and the design specifications for this prob-
lem are as follows. D̄0,j = 1
75 I2 ,

0.878 −0.864 k1 0
5.1. The plant model N̄0,j = 1
.
75 1.082 −1.096 0 k2 .

The process to be controlled is a distillation column with re- By (10b) we obtain

flux and boilup flows as the manipulated inputs and the product
(j ) −D̄0,j 02×2 (j ) 0 02×2
compositions as the outputs. The resulting model is ill condi- A0 = , A1 = 2×2 ,
−I2 02×2 02×2 02×2
tioned; its transfer function matrix is given by
(j ) 0
B0 = 2×2 , B1 =
(j ) N̄0,j
, j = 1, . . . , 4.
1 0.878 −0.864 k1 e− 1 s 0 02×2 02×2
G(s) = ,
75s + 1 1.082 −1.096 0 k2 e− 2 s
ki ∈ [0.8 1.2], i ∈ [0.0 1.0], (16) Choosing

0 0 0.05I2
where time is measured in minutes and where i are the delays Bw = 2×2 , C1 = 2×2 ,
I2 02×2 02×2
in the input channels. In physical terms this is equivalent to a
gain uncertainty of 20% and a delay of up to 1 min in each input 0 0
D11 = 2×2 and D12 = 2×2 ,
channel. The set of possible plants defined by (16) is denoted 02×2 I2
by
in the sequel. where is a weighting put on the control effort. We apply
Corollary 1 for = 0.6, = 10−8 and for ε1 = 1.9981, ε2 =
Remark 3. In the following solution we consider, for sim- 0.4147, ε3 = −1.1695 (a close result was obtained for ε1 =
plicity, two equal delays in the control channel. The obtained 1.46, ε2 = ε3 = 0). We obtain the following controller, which
solution will then be tested to see whether the specifica- consists of the two integrators and constant gains (namely a PI
tions are still satisfied in the case where the two delays are controller). The latter simple controller scheme consists of a
independent.
constant gain matrix Kp = [ −1269.4 985.6
−1238.5 1020.2 ] from y to u and a

5.2. Design specifications feedforward gain matrix from the output of the integrators e to
−80.01
the input u that is given by Ki = [ 115.04
103.18 −94.50 ].
Specifications S1 and S2 below should be fulfilled for every
plant G ∈
: 5.4. Finding the prefilter

S1 Closed-loop stability. The controller H that was found above may achieve a good
S2 For a unit step demand in channel 1 at t = 0 the plant attenuation of disturbances and small sensitivity to the param-
outputs y1 (tracking) and y2 (interaction) should satisfy: eter uncertainty. It may not be able, however, to satisfy the
◦ y1 (t)0.9 for all t 30 min; input–output transference requirements and it is the task of the
◦ y1 (t)1.1 for all t; second degree of freedom controller to achieve the required
◦ 0.99y1 (∞) 1.01; shape of the response.
198 V. Suplin, U. Shaked / Systems & Control Letters 57 (2008) 193 – 199

Step Responses

From input no. 1 From input no. 2

1

To output no. 1
0.5

0
Amplitude

-0.5
1.5
To output no. 2

0.5

-0.5
0 10 20 30 40 50 0 10 20 30 40 50
Time (minutes)

Fig. 3. Simulation results.

We choose a simple cascaded notch prefilter with the transfer uncertainty in the parameters of its transfer function matrices
function matrix: and uncertain, but bounded, time-delay in the control channel
is introduced. This method is based on combining two recent
s 2 + 0.161s + 1.152 s 2 + 0.11s + 0.552 results. The first is a continuous-time counterpart of the design
I2 .
s 2 + 2.3s + 1.152 s 2 + 1.1s + 0.552 method introduced in [10] for nonretarded uncertain systems,
This simple prefilter together with the PI controller of Section and the second is the recently developed theory of robust state-
5.1 is shown in the simulations below to satisfy the benchmark feedback design for systems with state-delay. The combination
problem requirements. of these two results yields a powerful design procedure that
applies solutions of linear matrix inequalities. The strength of
the results of the present paper is demonstrated in an example
5.5. Simulations of distillation columns control where a solution is obtained
by a simple low order controller that satisfies the benchmark
The worst case performance simulation results for all the requirements.
four vertices are given in Fig. 3. There, the responses of the two This paper deals with robust H∞ design of the retarded sys-
outputs to a unit step in one of the inputs are depicted for each tem. A similar approach can be applied to the corresponding H2
vertex. The responses in this figure are derived by applying the problem where a prescribed H2 -norm bound on the closed-loop
maximum delay of 1 min to both channels. It is verified that system should be satisfied. This is readily obtained by choosing
all the S1 and S2 requirements are indeed satisfied for all the an a priori very large in the design procedure of Section 4.
vertex points of the uncertainty polytope.
The case where the two delays are independent one of the
other has also been simulated using the obtained controller. References
This has been done by taking various independent values for 1
[1] E. Fridman, U. Shaked, Delay dependent stability and H∞ control:
and 2 in the given interval. Also in this case the worst results constant and time-varying delays, Internat. J. Control 76 (2003) 48–60.
have been obtained for 1 = 1 and 2 = 1. [2] P. Gahinet, P. Apkarian, A linear matrix inequality approach to H∞
control, Internat. J. Robust Nonlinear Control 4 (1994) 421–448.
[3] T. Kailath, Linear Systems and Control Theory, Prentice-Hall, Englewood
6. Conclusions
Cliffs, NJ, 1980.
[4] S. Kanev, C. Scherer, M. Verhaegen, B. De Schutter, Robust output-
A systematic method for designing robust output-feedback feedback controller design via local BMI optimization, Automatica 40
controllers for linear, continuous, time-invariant systems with (2004) 1115–1127.
V. Suplin, U. Shaked / Systems & Control Letters 57 (2008) 193 – 199 199

[5] H. Kwakernaak, A polynomial approach to H∞ optimization of [8] L. Mirkin, On the extraction of dead-time controllers and estimators from
control systems, in: Proceedings of NATO Workshop on Modelling, delay-free parametrizations, IEEE Trans. Automat. Control 48 (2003)
Robustness and Sensitivity Reduction in Control Systems, Gröningen, 543–553.
The Netherlands, December 1986. [9] V. Suplin, E. Fridman, U. Shaked, H∞ Control of linear uncertain time-
[6] D.J.N. Limebeer, The specification and purpose of a controller design delay systems—a projection approach, IEEE Trans. Automat. Control
case study, in: Proceedings of CDC 1991, Brighton, UK, 1579–1580. 51 (2005) 680–686.
[7] P. Lundstrom, S. Skogestad, J. Doyle, Two-degree-of-freedom controller [10] V. Suplin, U. Shaked, Robust H∞ output-feedback control of linear
design for an ill-conditioned distillation process using -synthesis, IEEE discrete-time systems, Systems and Control Lett. 54 (2005) 799–808.
Trans. Automat. Control 44 (1999) 12–21.