Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

53rd IEEE Conference on Decision and Control

December 15-17, 2014. Los Angeles, California, USA

New LMI Conditions for Static Output Feedback


Synthesis with Multiple Performance Objectives

Hakan Koroglu and Paolo Falcone


Dept. Signals and Systems, Chalmers University of Technology, Gothenburg,Sweden
E-Mail : {hakanko,paolo.falcone}@chalmers.se
conditions for performance, which have proven useful in
multi-objective, structured and robust synthesis problems
(see e.g [1], [17] and the references therein). This approach
has been applied also to derive sufficient LMI conditions for
(single or multiple objective) static output feedback synthesis
problems for fixed or uncertain systems [2], [14], [7], [22],
[20], [6], [8].
The method developed in this paper also falls in the
category of suboptimal solutions to static output feedback
problems. Our approach is inspired by the solution of [4] to
the multi-objective static state-feedback synthesis problem.
By following the same way of dilation, we have been able
to modify the conditions in [5] in a way to reduce their
potential conservatism. One of the modified conditions is
actually quite similar to the recent work [8], in which the
considered system model is less general. The LMI conditions
depend (in the same way as in [4]) on scalar parameters that
need to be fixed beforehand (cf. [6]).
For the convenience of our presentation, we formulate
the problem as a multi-objective synthesis for linear timeinvariant systems in the following section. In Section III,
we derive two alternative sets of LMI conditions for H
performance objectives. We then adapt one of these to
generalized H2 performance in Section IV. These results
are combined in Section V to express our solution to the
multi-objective synthesis problem. Extensions to parameterdependent systems are also discussed briefly. The synthesis
is illustrated for the adaptive cruise control problem before
the concluding remarks.

Abstract The static output feedback synthesis problem is


considered with H and generalized H2 performance objectives. New sufficient LMI conditions are derived for guaranteeing the required performance objectives. These conditions also
depend on scalar parameters that need to be fixed beforehand.
The output feedback gain matrix is computed from the involved
matrix variables without the use of system matrices. Hence
the conditions can directly be used to solve the multi-objective
synthesis problem also for parameter-dependent systems with
constant or parameter-dependent gain matrices. The method is
illustrated in the adaptive cruise control problem.

I. I NTRODUCTION
Static output feedback synthesis is a conceptually simply
and yet theoretically challenging problem. It is well-known
that the dynamic, fixed-order output feedback controller
design problem can also be formulated as a static output
feedback synthesis by system augmentation. Referring the
reader to [21] for a survey through the relatively old literature, we cite below only some selected recent works that are
relevant for our discussion.
As is the case for various problems, the synthesis of static
output feedback controllers can also be based on matrix
inequality conditions that ensure stability as well as some
performance objectives (H , H2 , etc.). One is then faced
with optimization problems under bilinear matrix inequality
(BMI) constraints, which are non-convex and hence hard to
deal with. It is also possible to reformulate the constraints in
the form of linear matrix inequalities (LMI) accompanied by
a rank constraint or alternatively a coupling condition among
the matrix variables. Various approaches are developed in
the literature to deal with the non-convex constraints. One
typically identifies an iterative procedure in which a convex
optimization problem is solved repeatedly until a stopping
criterion is satisfied (see e.g. [12], [3], [15], [13], [10] and
the references therein). It would be fair to say that these
procedures can provide solutions only with limited efficiency.
It is hence natural to consider making a trade off between optimality and tractability in static output feedback
synthesis problems with specified performance objectives.
This is usually done by representing the systems in suitable
bases and then choosing the matrix variables with particular
structures that facilitate a reformulation of BMI constraints
in the form of LMIs [5], [16], [18]. Exact LMI formulations
can be made only when the plant satisfies some restrictive
assumptions [18], [9]. It is hence necessary to seek for
ways to reduce the potential conservatism. To this end, it
is possible to use dilated versions of the matrix inequality
978-1-4673-6090-6/14/$31.00 2014 IEEE

II. P ROBLEM F ORMULATION


We consider a set of
with i = 1, . . . , as


xi (t)
A
i : zi (t) = Ci
yi (t)
C

linear time-invariant plants indexed

xi (t)
Bi B
Di Ei wi (t) , xi (0) = 0,
Hi 0
ui (t)

(1)

where xi (t) Rk denotes the state vector, while ui (t) Rn


and y(t) Rm represent the control input and the measured
output vectors respectively. These plants might refer to different systems or a single system expressed for the assessment
of different performance objectives. Indeed, the realizations
differ only in the matrices that relate to the generalized
disturbance input wi (t) Rqi and the performance output
zi (t) R pi . In fact, we can also allow for different (A, B,C)
866

matrices provided that ui s and yi s have identical dimensions


for all is. We avoid this only for notational reasons.
In this paper, we consider the synthesis of a controller
based on static output feedback. The control input to each
plant will hence be formed as
ui (t) = Kyi (t),

A. First Set of Conditions


The basic idea behind the dilation of the state-feedback
LMIs as in [4] is to assume a feedback gain matrix as
K = NW 1 ,

and construct a quadratic Lyapunov function with a matrix


that is different from W . Since we consider output feedback,
the dimensions of the matrices are identified in our case as
N Rnm and W Rmm . We represent the symmetric and
positive-definite matrix that will be used to construct the
Lyapunov function by Yi Sk+ . Being indexed with i, the
Lyapunov matrix Yi can be chosen different for different is.
On the other hand, since we need to have a common output
feedback design for all the plants (see (2)), we avoid any
index in N and W , which means that they will be identical
when we consider the performance objective for different is.
In terms of the matrices introduced so far, we now define
a transformed state vector i and a new signal i as

(2)

where K Rnm is the (common) gain matrix to be designed.


We emphasize at this point that static (partial) disturbance
feed-forward synthesis can also be included in this formulation thanks to the presence of Hi . Indeed, one can then extend
C and Hi matrices and thereby the measurement vector y in
a way to have the measurable disturbances as a part of the
extended measurement vector.
The closed-loop dynamics of the ith plant are obtained
with the control input of (2) as
[ ] [
][ ]
xi
A + BKC Bi + BKHi
xi
=
, xi (0) = 0, (3)
zi
Ci + Ei KC Di + Ei KHi
wi

i
i

where the time dependencies of the signals are suppressed


for notational simplicity.
In this paper, we will consider the static output feedback
synthesis problem with H and generalized H2 performance
objectives. Since we will work with time-domain formulations, we recall the definitions of L2 and L norms as
(
)1/2
w2 ,
,
(4)
w(t)2 dt

, sup z(t),

xi
zi

(5)

(10)

(11)

= (AYi + BNC) i BN i + Bi wi ,
= (CiYi + Ei NC) i Ei N i + Di wi .

(12)

We have thus obtained a description of the closed-loop that


will help us avoid a bilinear dependence on the design
variables when matrix inequality conditions are derived in
the usual way. This is achieved by the introduction of a new
input signal i , the relation of which to i and wi has to be
used to end up with useful conditions. By multiplying (10)
from the left with W , we derive this relation as

where a ,
We can now state the core problem
considered in this paper precisely as follows:
Problem 1: Given a set of linear time-invariant plants
i , i = 1, . . . , as in (1), find a matrix K Rnm such that
the closed-loop systems in (3) are all stable (i.e. A + BKC
is Hurwitz) and satisfy the following performance objectives
for all wi () with 0 < wi 2 < :
zi

yi .

This allows us to express the closed-loop plant dynamics in


(3) as

t0

< i wi 2 , i I
< i wi 2 , i I2

, Ci W

(9)
1

Kyi = NCi N i .

aT a.

zi 2

, Yi1 xi ,

Multiplication of (10) from the left by N leads to

(8)

oi , (WC CYi )i Hi wi W i = 0.

(6)
(7)

(13)

Let us now form a quadratic Lyapunov function as


Vi (xi ) = xiT Yi1 xi ,

In these expressions, i s represent the desired levels of H


performance for i I and i s represent the desired levels
of generalized H2 performance for i I2 .

(14)

and recall that stability and H performance requirement


of (6) will be satisfied if the following condition is ensured
along the trajectories of (3):

III. S UFFICIENT C ONDITIONS FOR G UARANTEED H


P ERFORMANCE

dVi (xi (t)) zi (t)2


i wi (t)2 < 0, t 0.
+
dt
i

(15)

Indeed, stability will be ensured as the time-derivative of Vi


will necessarily be negative when wi = 0. On the other hand,
by integrating both sides of (15) from zero to infinity and
recalling that xi (0) = 0 as well as limt xi (t) = 0 (as follows
from stability and wi 2 < ), we easily infer (6).
In order to derive a linear matrix inequality condition that
ensures (15), we will use the description of the closed-loop
as in (12). Since i appears as an additional input in this
description, we will have to modify (15) in a way that will

In this subsection, we derive two alternative solvability


conditions for the generic H objective in (6). These conditions are both sufficient and are inspired by the work of [4]
on multi-objective controller synthesis based on static statefeedback. Moreover, they can be viewed as modified versions
of the conditions in [5], which are retrieved when particular
equality constraints are imposed on the design variables. We
describe the derivations of these conditions separately in the
following subsections.
867

Theorem 1 provides a potentially less conservative condition


than [5], which cannot be extended to the case in which Hi =
0. Note, however, that potential reduction of conservatism
comes with increased computational cost, especially because
a line search has to be performed over the positive scalar .
Remark 2: It is possible to modify the condition by defining i with C replaced by an arbitrary (yet fixed) matrix P
of compatible dimensions in (10). This leads to a condition
in which the terms WC and NC are replaced by W P and
NP respectively (see and cf. [8]). Theorem 1 is in fact an
extension of the result in [8], which is derived for plants with
Hi = 0 and Di = 0.

lead to a suitable condition. This is done by exploiting (13).


With oi = 0 by definition, we modify (15) as
zT zi
dVi (xi )
+ 2 iT oi + i i wTi wi < 0,
dt
i

(16)

where is an arbitrary positive scalar. We now introduce


]
[
iT , iT iT wTi i1 zTi ,
(17)
and express (16) along the trajectories of (12) as
)
(
zTi zi i T
T
T
wi wi
2 i xi + i oi +
2i
2
= iT (Mi + MiT ) i 0. (18)
{z
}
|

B. Second Set of Conditions


In this subsection, we derive a dual condition for guaranteeing the H performance objective. To this end, we assume
K to be of the form

Ni =He{Mi }

The H performance condition of (6) will be satisfied if Ni


is negative definite, which leads us to the following result:
Theorem 1: The closed-loop system in (3) is stable and
satisfies (6) with a gain matrix gain matrix obtained as in
(8), if there exist Yi Sk+ , W Rmm and N Rnm that
satisfy

He{W } (CYi WC) + N T BT Hi

He{AYi + BNC}
Bi

0,

i I

Ei N
CiYi + Ei NC
Di i I
(19)
where R+ is an arbitrary (yet fixed) positive scalar and
s represent entries that are identifiable from symmetry.
Proof: In completion of the proof, we first recall the
definitions of (9) and (10) together with (13) and (1) to derive
]
[
iT oi = iT W (CYi WC) Hi 0 i ,
[
]
iT xi = iT BN AYi + BNC Bi 0 i ,
]
[
i wTi wi
= wTi 0 0 0.5i I 0 i ,

2
]
[
zTi zi
= i 1zTi Ei N CiYi + Ei NC Di 0.5i I i ,
2i

K = V 1 M,

(21)
Rnm

where the dimensions are identified as M


and V
Rnn . By multiplying ui = Kyi from the left with V , we easily
obtain the relation
oi , MCxi + MHi wi Vui = 0.

(22)

In this section, we will derive a new condition by using


i = xi and i = 1 ui , where is an arbitrary positive scalar.
We hence construct the Lyapunov function as
Vi (xi ) = xiT Xi xi ,

(23)

where Xi Sk+ . The evolution of xi is then expressed as


xi = Axi + Bi wi + (B Xi1 BV )ui + Xi1 (BMCxi + BMHi wi ) .
{z
}
|
BVui

(24)

Based on this expression, we obtain


X xi = (Xi A + BMC)xi + (Xi Bi + BMHi )wi + (Xi B BV )ui .
(25)
By again expressing (16) as in (18), we are able to obtain
the following result:
Theorem 2: The closed-loop system in (3) is stable and
satisfies (6) with a gain matrix gain matrix obtained as in
(21), if there exist Xi Sk+ , V Rnn and M Rnm that
satisfy

He{V } (Xi B BV )T + MC MHi

He{Xi A + BMC}


0,

(XBi + BMHi )
i I

Ei
Ci
Di i I
(26)
where R+ is an arbitrary (yet fixed) positive scalar.
Proof: We first recall i = xi , i = 1 ui and use
equations (22), (25) and (1) to obtain
[
]
iT oi = iT V MC MHi 0 i ,
]
[
iT Xi xi = iT (Xi B BV) Xi A+BMC Xi Bi +BMHi 0 i ,
]
[
i wTi wi
= wTi 0 0 0.5i I 0 i ,

2
]
[
zTi zi
= i 1zTi Ei Ci Di 0.5i I i .
2i

where i is the vector defined in (17). By stacking the row


blocks in the right hand side of these equations on top
of each other and thereby forming a matrix Mi , we can
reexpress (16) as in (18). This inequality will be satisfied
for all i if Ni 0, which is the condition given in (19).
The construction of K as in (8) will then always be possible
since W is invertible as a result of He{W } 0, which is
necessarily implied by (19).
Remark 1: A comparable condition can be found in [5]
for the case in which Hi = 0 and C has full row rank. This
solvability condition is expressed by the constraints

He{AYi + BNC} Bi

i I
0 and WC CYi = 0,
CiYi + Ei NC
Di i I
(20)
where the LMI is expressed in a comparable form to the
condition in (19). If these are satisfied with some Yi ,W and
N, one can always find a sufficiently large such that (19)
is also satisfied with the same variables. This means that
868

exist Yi Sk+ , W Rmm and N Rnm that satisfy

He{W } (CYi WC) + N T BT Hi

He{AYi + BNC}
Bi 0, (32)

i I
[
]
Yi

0, (33)
CiYi i I

By using these expressions, we put (16) into the form (18),


from which a condition is obtained as in (26).
Remark 3: As derived in [5], a dual solvability condition
to (20) can be obtained for the case in which Ei = 0 and B
has full row rank. This condition is also formed by an LMI
accompanied by an equality constraint as

He{Xi A + BMC}

(Xi Bi + BMHi )T i I
0 and BV Xi B = 0.
Ci
Di i I
(27)
We again note that the feasibility of these conditions imply
the feasibility of (26).

where R+ is an arbitrary (yet fixed) positive scalar.


V. S OLUTION TO M ULTI - OBJECTIVE S TATIC O UTPUT
F EEDBACK S YNTHESIS P ROBLEM AND ITS E XTENSIONS
In this section, we first combine results of the previous
sections to state the main result of the paper. We then
briefly describe how this result can be extended to parameter
dependent systems.
Let us first combine Theorem 1 and Theorem 3 to state
the main result of the paper as follows:
Theorem 4: There is a solution to Problem 1 if there exist
Yi Sk+ , i = 1, . . . , ; W Rmm and N Rnm that satisfy
(19) for all i I with some fixed > 0 ; and (32)-(33) for
all i I2 with some fixed > 0. A gain matrix K Rnm
that solves the problem can then be computed as in (8).
In multi-objective synthesis, we typically need to consider
minimization over one particular i or i and fix all others to
feasible values. As is emphasized in the theorem statement,
we also need to fix ( , ) in order to obtain a genuine
LMI optimization problem. In order to reduce the potential
conservatism in our synthesis, we need to perform a plane
search over ( , ). In fact, we can also consider using
different scalars for different is. Nevertheless, this is not
quite desirable as it will lead to increased computational
complexity when we have to search over three or more
scalars.
Problem 1 can also be considered for parameter-dependent
systems. A parameter-dependent system is described by a
state-space model in which all the system matrices depend
on an uncertain parameter vector , which can take its values
from a known compact set R. The system matrices are then
described by continuous, matrix-valued maps, i.e. A() : R
Rkk , B() : R Rkn , etc. When is time-invariant or
arbitrarily time-varying, knowledge of R will be enough to
describe the uncertainty. On the other hand, when parameters
are assumed to vary slowly in time, it becomes convenient
to introduce an extended parameter vector as e , ( , ). In
this case, the uncertainty description can be captured by a
compact set U , which represents the set of admissible values
for e .
We will be faced with two different synthesis problems
for parameter-dependent systems, depending on whether (t)
is measurable during online operation or not. When (t) is
online-measurable, we can consider synthesizing an output
feedback gain that is also parameter-dependent, i.e. K() :
R Rnm . If this is not the case, we will have to consider
the synthesis of a constant matrix K Rnm with which the
performance objectives are achieved.
It is straightforward to extend Theorem 4 in a way to provide solutions to the multi-objective synthesis problems for

IV. S UFFICIENT C ONDITIONS FOR G UARANTEED


G ENERALIZED H2 P ERFORMANCE
In this section, we adapt the derivation in Section III-A
to obtain a set of sufficient LMI conditions for generalized
H2 performance. It is also possible to adapt the derivation
in Section III-B, which we avoid for reasons of space.
Throughout this section, we assume that the plant satisfies
Di = 0 and Ei = 0, i I2 .

(28)

This guarantees that the feed-through term of the closedloop system in (3) is zero, which is necessary for having a
bounded generalized H2 norm.
Let us now recall the Lyapunov function in (14) and
note that the conditions that guarantee the generalized H2
performance requirement of (7) are expressed as follows:
dVi (xi (t))
i wi (t)2 < 0, t 0,
dt
zi (t)2
> 0, t 0.
Vi (xi (t))
i

(29)
(30)

With x(0) = 0, we observe by integrating the first inequality


that Vi (xi (t)) < i wi 22 for any t 0. It then follows from
the second inequality that zi (t)2 < i2 wi 22 is ensured for
any t 0, which clearly implies (7).
In order to translate (29) into a matrix inequality condition
based on (12), we again need to modify it in a similar way
to (16). In this case, we use a different scalar > 0 to form
a condition that implies (29) as
dVi (xi (t))
+ 2iT (t)oi (t) i wi (t)2 < 0.
dt

(31)

We can again express this condition as in (18), by removing


zi from the expression of in (17) and forming Ni accordingly. This amounts to removing the fourth row and column
blocks from (19). Thanks to the assumption in (28), we dont
need to introduce any modification in (30) as it can directly
be translated into an LMI condition.
Along the steps we have thus outlined, we arrive at the
following result:
Theorem 3: The closed-loop system in (3) is stable and
satisfies (7) with a gain matrix obtained as in (8), if there
869

parameter-dependent systems. In this case, some or all of the


design variables will be chosen as parameter-dependent. As
a result of this, we will be faced with parameter-dependent
LMIs that need to be satisfied over R or U . The parameterdependent LMI conditions can then be relaxed into finitely
many LMIs by employing suitable relaxation schemes (see
[19] and the references therein). In this fashion, one arrives
at tractable optimization problems based on which synthesis
can be performed. If the parameter vector is not assumed to
vary arbitrarily in time, it will be possible to choose all Yi s
as parameter-dependent. In this case, we will have to add a
new term as j Yi(j ) j to the (2, 2) blocks of (19) and
(32). When is online-measurable, we can choose W and
N in a way to have parameter dependence as well. If this is
not the case, W and N need to be chosen as matrix variables
so that the resulting feedback gain K is a constant matrix.

In order to evaluate the method developed in this paper, we


considered the multi-objective synthesis problem with full
state-feedback, u = Ks f b x, and with partial state feedback,
[
]T
. Both synthesis are performed with the
u = Ko f b x1 x2
following numerical values:
h = 1, 0 = 0.5, min = 0 /1.1, max = 0 /0.9, = 0.05,
We used parameter dependent Yi = Yi0 + Yi1 ; i = 1, 2 with
= 0 / 1 and constant W, N in order to obtain constant
gain matrices. After an initial analysis of the system, we
set the value of to 1.1 and minimized over (19), (32)
and (33). Since these inequalities depend on the uncertain
parameter [0.1, 0.1], we needed to relax them into
finitely many LMIs. This is done by imposing them at
the extreme points = 0.1 and = 0.1. Since (19) and
(32) have quadratic dependence on , we also imposed
two convexity constraints (see [11]). In order to obtain the
smallest value for , we performed a plane search over
( , ). We could obtain various designs with different ( , )
choices that lead to the best value. We provide here the
designs for ( , ) = (9.3, 6.51) and ( , ) = (2.8, 2.52) used
in the cases of state and output feedback respectively:
[
]
Ks f b = 4.4291 4.4703 4.1786
s f b = 0.321,
[
]
Ko f b = 1.0736 0.7870
o f b = 0.384,

VI. A PPLICATION TO S AFE V EHICLE F OLLOWING


In this section, we apply the synthesis procedure to the
longitudinal vehicle following problem. This problem is
the basis of adaptive cruise control systems, which have
become widely available in current day automobiles. Let us
first describe the system model by referring to the leader
with index zero and to the follower with index one. The
absolute position (of the rear end), longitudinal velocity
and longitudinal acceleration signals are represented with p,
v = p and a = v respectively. We define the spacing error
between the two vehicles as x1 , e1 = p0 p1 s hv1 ,
where s is the desired spacing that also includes the length
of the following vehicle and h is headway time. By this
definition of the spacing error, we adopt a spacing policy
that depends on the velocity of the following vehicle. We
define the relative velocity of the leader with respect to
the follower as the second state x2 , v0 v1 . It is easy to
observe that x1 = p0 p1 ha1 = x2 ha1 and x2 = a0 a1
by definition. Choosing the acceleration of the follower as
the third state x3 , a1 and assuming a first-order, linear timeinvariant model for it with an uncertain time constant of
[min , max ], we can form a state-space model as



x1
0 1 h
x1
0
0
x2 = 0 0 1 x2 + 1 a0 + 0 u.
x3
0 0 1/
x3
0
1/

Since we have a linear time-invariant plant model, it is


meaningful to consider the transfer functions from w to z1
and w to z2 . We refer to these transfer functions respectively
as T1 (s) and T2 (s), adding sfb or ofb as a superscript,
depending on whether they are obtained with K = Ks f b or
K = Ko f b . The maximum singular values of these transfer
functions are plotted together in Figure 1 for a grid of in the
range [0.1, 0.1]. These plots illustrate that both synthesis
lead to robust designs for both transfer functions. This is
justified by the appearance of the curves simply as thick
lines (rather than wild variations) over certain frequency
ranges. The state feedback synthesis is better, to an extent
that wouldnt normally be expected based on the guaranteed
performance levels. This is due to the fact that the worst-case
value of the H2 -norm of T1 (computed over the considered
grid as 0.0602) is much smaller than s f b . For the case of
output feedback, this value (found as 0.3245) is comparable
to o f b . As is also visible from the singular value plots in
Figure 1 (bottom), the worst-case H norm of T2 is one in
both cases, although was set to 1.1.
The designs are also tested with simulations, in which
a particular deceleration and acceleration profile is applied
to the leading vehicle within a time range of ten seconds.
The resulting spacing errors and acceleration profiles of
the following vehicle are displayed in Figure 2, again in
the form of overlay plots for different time constants. It is
not surprising that the state feedback controller performs
significantly better. The static output feedback controller
shows acceptable performance in that it is able to keep the
spacing error within one meter in all the cases.
We conclude this section by some final remarks concerning our comparisons with the previous works on static output

In this model, u represents the desired acceleration signal


given to the follower as the control input.
The primary concern in vehicle following is to keep the
spacing error z1 = e1 = x1 as small as possible, in the
face of possible maneuvers by the leader, during which
w = a0 becomes nonzero. Of course, this is to be done
without requiring excessively large acceleration x3 = a1 .
For passenger comfort, it is also important to avoid large
x3 = a1 values.[ We hence ]form the second performance
T
output as z2 = x3 x3
, where is a positive scalar
that will determine the emphasis on the passenger comfort.
In this example, we prefer to consider a generalized H2
performance objective for z1 as z1 < w2 and an H
performance objective for z2 as z2 2 < w2 .
870

Singular Values

R EFERENCES

Singular Values (dB)

[1] P. Apkarian, H. D. Tuan, and J. Bernussou. Continuous-time analysis,


eigenstructure assignment, and synthesis with enhanced linear matrix
inequalities (LMI) characterizations. IEEE Transactions on Automatic
Control, 46(12):19411946, December 2001.
[2] G. I. Bara and M. Boutayeb. Static output feedback stabilization with
H performance for linear discrete-time systems. IEEE Transactions
on Automatic Control, 50(2):250254, February 2005.
[3] Y. Y. Cao, J. Lam, and Y. X. Sun. Static output feedback stabilization:
An ILMI solution. Automatica, 34(12):16411645, December 1998.
[4] M. Coutinho, D. F. Coutinho, A. Trofino, and K. A. Barbosa. A new
strategy to the multi-objective control of linear systems. In Proc. 44th
IEEE Conference on Decision and Control, pages 37413746, Sevilla,
Spain, December 2005.
[5] C. A. R. Crusius and A. Trofino. Sufficient LMI conditions for output
feedback control problems. IEEE Transactions on Automatic Control,
44(5):10531057, May 1999.
[6] K. Dabboussi and J. Zrida. Sufficient dilated LMI conditions for H
static output feedback robust stabilization of linear continuous-time
systems. Journal of Applied Mathematics, pages 113, 2012. Article
ID: 812920.
[7] J. Dong and G. H. Yang. Static output feedback control synthesis
for linear systems with time-invariant parametric uncertainties. IEEE
Transactions on Automatic Control, 52(10):19301936, October 2007.
[8] J. Dong and G. H. Yang. Robust static output feedback control
synthesis for linear continous systems with polytopic uncertainties.
Automatica, 49(6):18211829, June 2013.
[9] Y. Ebihara, D. Peaucelle, and D. Arzelier. Some conditions for convexifying static H control problems. In Prep. 18th IFAC Triennial World
Congress, pages 92489253, Milano, Italy, August 28 - September 2
2011.
[10] J. Gadewadikar, F. L. Lewis, and M. Abu Khalaf. Necessary and
sufficient conditions for H static output feedback control. AIAA
Journal of Guidance, Control, and Dynamics, 29(4):915920, JulyAugust 2006.
[11] P. Gahinet, P. Apkarian, and M. Chilali. Affine parameter-dependent
Lyapunov functions and real parametric uncertainty. IEEE Transactions on Automatic Control, 41(3):436442, March 1996.
[12] L. El Ghaoui, F. Oustry, and M. AitRami. A cone complementarity
linearization algorithm for static output-feedback and related problems.
IEEE Transactions on Automatic Control, 42(8):11711176, August
1997.
[13] Y. He and Q. G. Wang. An improved ILMI method for static output
feedback control with application to multivariable PID control. IEEE
Transactions on Automatic Control, 51(10):16781683, October 2006.
[14] K. H. Lee, J. H. Lee, and W. H. Kwon. Sufficient LMI conditions
for H output feedback stabilization of linear discrete-time systems.
IEEE Transactions on Automatic Control, 51(4):675680, April 2006.
[15] F. Leibfritz. An LMI-based algorithm for designing suboptimal static
H2 /H output feedback controllers. SIAM Journal on Control and
Optimization, 39(6):17111735, 2001.
[16] M. Mattei. Sufficient conditions for the synthesis of H fixed-order
controllers. International Journal of Robust and Nonlinear Control,
10(15):12371248, December 2000.
[17] R. C. L. F. Oliveira, M. C. de Oliveira, and P. L. D. Peres. Robust
state feedback LMI methods for continuous-time linear systems:
Discussions, extensions and numerical comparisons. In Proc. IEEE
International Symposium on Computer-Aided Control System Design,
pages 10381043, September 2011.
[18] E. Prempain and I. Postlethwaite. Static output feedback stabilization
with H performance for a class of plants. Systems and Control
Letters, 43(3):159166, July 2001.
[19] C. W. Scherer. LMI relaxations in robust control. European Journal
of Control, 12(1):329, 2006.
[20] Z. Shu and J. Lam. An augmented system approach to static
output-feedback stabilization with H performance for continuoustime plants. International Journal of Robust and Nonlinear Control,
19(7):768785, May 2009.
[21] V. L. Syrmos, C. T. Abdallah, P. Dorato, and K. Grigoriadis. Static
output feedback - a survey. Automatica, 33(2):125137, February
1997.
[22] A. Trofino. Sufficient LMI conditions for the design of static and
reduced order controllers. In Proc. Joint 47th IEEE Conference on
Decision and Control and 28th Chinese Control Conference, pages
66686673, Shanghai, China, December 2009.

10
20
30
40
Tsfb

1
ofb

50

T1

60
1
10

10
Frequency (rad/s)

10

Singular Values

Singular Values (dB)

10

20

sfb

30

T2

ofb

T2
40
1
10

10
Frequency (rad/s)

Fig. 1.

10

Singular value plots of T1 ( j ) and T2 ( j ).

1
SFB
OFB

e [m]

0.5

0.5

10

5
Time [sec]

10

2
a0

a1 [m/s2]

Fig. 2.

Simulation results.

feedback synthesis. As we have noted in Remark 1, the H


synthesis condition in [5] is evidently more conservative than
(19). This is what we also observed in our numerical investigations. We have also made a comparison with Example 5.1
of [6], in which the optimum is found as 4.17 (and the
corresponding output feedback led to an H norm of 2.49).
Based on (19), we obtained the minimum value as 1.6716
and the associated output feedback led to an H norm of
1.4433. A theoretical investigation is needed to determine
whether (19) always provides less conservative results or not.
VII. C ONCLUDING R EMARKS
We have derived new sufficient LMI conditions for static
output feedback synthesis for guaranteed H and generalized H2 performance objectives. It is not necessary to
perform any state transformation to apply these conditions.
Hence they can be directly used to synthesize robust or
scheduled controllers for parameter-dependent systems. It is
also possible to use them for structured controller synthesis
as the feedback gain matrix is computed without the use of
system matrices. Seemingly the conditions need some improvement to be useful in reduced-order controller synthesis.
ACKNOWLEDGEMENT
This work is supported by SAFER Vehicle and Traffic Safety Center.
871

You might also like