Systems & Control Letters 116 (2018) 8–14

Contents lists available at ScienceDirect

Systems & Control Letters

journal homepage: www.elsevier.com/locate/sysconle

Robust pole placement under structural constraints

João Fábio S. dos Santos, Paulo C. Pellanda, Alberto M. Simões *
Department of Electrical Engineering, Instituto Militar de Engenharia, Rio de Janeiro, RJ, 22290-270, Brazil

article info a b s t r a c t
Article history: A new controller synthesis technique is presented which allows the design of output feedback control
Received 26 April 2017 systems achieving robust regional pole clustering in the presence of parametric uncertainties as well as
Received in revised form 14 March 2018 satisfying prescribed structural constraints. Such features are rarely jointly present in currently available
Accepted 18 March 2018
controller synthesis methods. The central idea in the proposed approach consists in reformulating the
Available online 17 May 2018
original robust pole placement problem into an equivalent robust stabilization problem involving highly
structured controller and uncertainty. A numerical application corroborating the applicability of the
Keywords:
Pole placement proposed synthesis technique is also presented.
Structured controller © 2018 Elsevier B.V. All rights reserved.
Robust control
Nonsmooth optimization

1. Introduction state feedback control and hence become inoperative in output

Closed-loop robustness in the presence of parametric uncer-
tainty represents a primary design objective in any modern control
system synthesis method, e.g. sliding mode [1], adaptive [2] or
robust [3] control. In the particular case of linear time invariant
systems, an appealing robust controller synthesis technique is
regional pole placement. As well-known, by clustering the closed-
loop poles in appropriate regions of the complex plane, the design
engineer can set bounds on the damping ratio, the decay rate or
the undamped natural frequency of the closed-loop modes, and
hence shape time-domain parameters of the system response,
e.g. rise time, settling time or overshoot. As noted in [4], from an
application viewpoint, regional pole clustering is more important
than exact pole assignment. The reader is referred, e.g. to [5,4] and
the references within for earlier research on the subject.
In the seminal work [6], sufficient conditions are obtained for
pole placement in a general class of convex regions of the complex
plane defined by linear matrix inequalities (LMI) constraints. Inter-
estingly, the resulting controller synthesis problem can be solved
very efficiently via semidefinite programming tools. Moreover, in
the LMI framework, pole placement constraints can be considered
simultaneously with other design criteria, e.g. H∞ constraints. That
technique has been expanded in [7] so as to address the so-called
robust D-stabilization problem, i.e. robust pole placement in LMI-
regions.
Different approaches have been proposed subsequently in the
literature dealing with robust D-stabilization. See, e.g., [8,9] and
references therein. Unfortunately, these techniques are limited to
* Corresponding author.
E-mail address: simoes@ime.eb.br (A.M. Simões).
feedback control problems.
In the output feedback case, a notorious limitation of LMI-based
controller synthesis techniques is the difficulty in handling struc-
tural constraints on the control law itself. The LMI-based robust
pole placement technique in [7], for instance, can only produce
full-order output feedback controllers. In [10], an LMI technique is
described to perform robust pole placement on second-order linear
systems, but the only controller structure that can be handled is
proportional-derivative control. The technique in [11] allows the
design of fixed-order H∞ controllers also achieving D-stabilization,
but the technique is limited to SISO systems.
The controller synthesis technique recently introduced in [3]
allows the design of reduced-order output feedback controllers
ensuring closed-loop robust D-stability, hence potentially over-
coming the difficulty of LMI-based techniques indicated above.
However, it is stated by the author that only systems with small
size can be handled, due to the associated computational burden.
The numerical applications discussed in [3], for instance, involve
only static controllers with a single tunable parameter.
The recurrent difficulty of LMI approaches in designing struc-
tured controllers is one of the main motivations behind the regain
of interest seen in the last decade in controller synthesis tech-
niques based on nonsmooth optimization, e.g. [12–19]. Another
motivation lies in the numerical difficulties that LMI and bilinear
matrix inequalities (BMI) [20] techniques tend to face for problems
of moderate size. Such a difficulty is mainly due to the presence
of Lyapunov variables, whose number grows quadratically with
the order of the closed-loop system [13]. Nonsmooth controller
synthesis techniques, on the other hand, can handle systems with
dozens of states, even hundreds, see e.g. [21,22].
Pole placement constraints appearing initially in the afore-
mentioned nonsmooth optimization-based synthesis techniques

https://doi.org/10.1016/j.sysconle.2018.03.008
0167-6911/© 2018 Elsevier B.V. All rights reserved.
 J.F.S. dos Santos et al. / Systems & Control Letters 116 (2018) 8–14 9

involved essentially half-plane constraints via the spectral abscissa

function [12,23,13,21]. More recently, however, a more general
region of the complex plane has been considered in [17] specifi-
cally for pole placement. The synthesis technique in [17] allows the
design of structured controllers achieving regional pole placement,
but it unfortunately presents a serious inconvenience: similarly as
in [6], robustness of the pole clustering is not explicitly addressed,
but instead handled indirectly via an additional H∞ constraint. As
well known, unscaled H∞ constraints are not the most appropriate
way to deal with structured uncertainties, e.g. parametric uncer-
tainties [24].
In the present work, a new controller synthesis technique is
introduced that allows the design of output feedback controllers
satisfying prescribed structural constraints as well as achieving
robust pole placement in the presence of parametric uncertainties.
The central idea in the proposed approach is to reformulate the
original robust pole placement problem into an equivalent robust
stabilization problem, which can be interpreted as a particular Fig. 1. Region D(q, r , τ , θ, γ ) for robust pole placement.
µ-synthesis [25] problem involving highly structured uncertainty
and controller. The resulting synthesis problem can then be solved
efficiently by means of a recently introduced parametric robust
structured control design technique [18]. with γ , q ∈ R, τ , r ∈ R>0 , θ ∈ (0, π/2), and
The paper is organized as follows. In Section 2, it is discussed Ωc (q, r) ≜ s ∈ C : (Re(s) + q)2 + Im(s)2 < r 2 ,
{ }
(4)
how the pole clustering robustness can be assessed via an equiva-
lent robust stability condition. In Section 3, the considered robust
Ωhp (τ ) ≜ {s ∈ C : Re(s) < −τ } , (5)
pole placement synthesis problem is formulated, and its solution Ωw (θ, γ ) ≜ {s ∈ C : tan(θ ) (Re(s) − γ ) + |Im(s)| < 0} . (6)
is discussed. A numerical application illustrating the validity of
It is to be recalled that the uncertain system (1) is said to
the proposed approach is then presented in Section 4. Section 5
be robustly stable if it is stable for all allowable uncertainty, or
concludes the paper.
equivalently, if the eigenvalues of the state matrix (∆ ⋆ M) lie in
Notation: For a given real matrix A ∈ Rn×n , λi [A] stands for the set
the open-left half plane for all ∆ ∈ B∆. If, moreover, the poles
of n eigenvalues of A, i.e. λi [A] ≜ {λ ∈ C : det(λI − A) = 0}, whereas
α (A) denotes the spectral abscissa of A, i.e. α (A) ≜ max{Re(λ) : of the uncertain system (1) lie in D for all allowable uncertainty,
λ ∈ λi [A]}. If every eigenvalue of matrix A has strictly negative real i.e., λi [∆ ⋆ M ] ∈ D for all ∆ ∈ B∆, then the uncertain system is
part, then A is called a Hurwitz matrix. Symbol ⊗ stands for the said to be robustly D-stable.
Kronecker product. For a given matrix H ∈ Cn×m , σ (H) stands for The following theorem, which states the main result of this
its largest singular value. For a complex matrix P partitioned as section, provides a suitable necessary and sufficient condition for
[ ] the robust D-stability of the uncertain system (1).
P P12
P = 11 ∈ C(p1 +p2 )×(q1 +q2 ) ,
P21 P22 Theorem 1. Consider a non-empty region D(q, r , τ , θ, γ ), with q ̸ =
and a matrix ∆ ∈ C , notation ∆ ⋆ P stands for the classical
q1 ×p1 r, and let
⎡ ⎤
upper linear fractional transformation (LFT) [24] given by q+r 1
⎢q − r q−r⎥
∆ ⋆ P ≜ P22 + P21 ∆(I − P11 ∆)−1 P12 . Γ (q, r) ≜ ⎢ ⎥ ⊗ In . (7)
⎣ −2r −1 ⎦
For two transfers G and H, notation (G , H) stands for the closed-
q−r q−r
loop interconnection
Then, the uncertain system (1) is robustly D-stable if and only if the
y = G u,
{
system
u = Hy.
ẋ(t) = A∆ x(t), (8)
2. Robust D -stability analysis with
(
Consider the uncertain linear time-invariant system
A∆ ≜ diag Γ ⋆ (∆ ⋆ M), (∆ ⋆ M) + τ In ,
ẋ(t) = (∆ ⋆ M)x(t), (1) )
sin(θ ) − cos(θ )
[ ]
where M ∈ R(r +n)×(r +n) represents the nominal state matrix and ⊗ ((∆ ⋆ M) + γ In ) , (9)
∆ ∈ ∆ ⊂ Rr ×r represents a block diagonal uncertainty matrix cos(θ ) sin(θ )
whose structure is given by
is robustly stable.
∆ ≜ ∆ = diag(δ1r Ik1 , . . . , δm
r
I ) : δir ∈ R .
{ }
r kmr
(2)
Proof. For a given matrix A ∈ Rn×n , let
The unit ball in ∆ is denoted by B∆ ≜ {∆ ∈ ∆ : σ (∆) ≤ 1}.
Let D denote the non-empty region of the complex left-half Ac (q, r) ≜ Γ (q, r) ⋆ A, (10)
plane depicted in Fig. 1, constructed as the intersection of a disk, Ahp (τ ) ≜ A + τ In , (11)
a half plane and a wedge, as follows:
sin(θ ) − cos(θ )
[ ]
Aw (θ, γ ) ≜ ⊗ (A + γ In ). (12)
cos(θ ) sin(θ )
D(q, r , τ , θ , γ ) ≜ Ωc (q, r) ∩ Ωhp (τ ) ∩ Ωw (θ, γ ), (3)
10 J.F.S. dos Santos et al. / Systems & Control Letters 116 (2018) 8–14

The controller design problem considered here can be summa-

rized as follows: synthesize a possibly structured controller K (s)
so that the uncertain closed-loop system (G (s, ∆), K (s)) presents
robust D-stability. Besides the presence of parametric uncertainty
in the plant, another difficulty stems here from the fact the con-
troller K (s) to be designed is possibly subject to structural con-
straints, e.g. reduced-order, decentralized or PID. Such structural
Fig. 2. Uncertain closed-loop system.
constraints can be easily translated into constraints on the con-
troller realization (19), by letting the state–space matrices AK , BK ,
CK , DK depend smoothly on the actual design variable κ ∈ Rnc ,
Initially, it is shown that λi [A] ∈ D(q, r , τ , θ, γ ) if and only if the often referred to as vector of tunable parameters, see, e.g. [13].
matrix For notational simplicity, the dependency on κ will be omitted
whenever possible.
AΣ (q, r , τ , θ , γ ) ≜ diag(Ac , Ahp , Aw ) (13)
Let
is Hurwitz. [ ] [ ] [ ]
A 0 B1 0 B2
Clearly, AΣ is Hurwitz if and only if matrices Ac , Ahp and Aw  ≜ , B̂1 ≜ , B̂2 ≜ ,
0 0k 0 Ik 0
are all Hurwitz. Aw being Hurwitz is a well-known necessary and
sufficient condition for λi [A] ∈ Ωw , see, e.g., [26,27]. Moreover,
[ ]
0 Ik
0 ,
[ ]
since Ĉ1 ≜ C1 Ĉ2 ≜
C2 0
λi [Ahp ] = λi [A + τ In ] = λi [A] + τ , (14) 0
[ ]
0k
[
0
]
, ,
[ ]
D̂21 ≜ D̂22 ≜ D̂12 ≜ 0 D12
it follows that Re(λi [Ahp ]) < 0 if and only if Re(λi [A]) < −τ , and D21 0 D22
hence Ahp is Hurwitz if and only if λi [A] ∈ Ωhp . Thus, it only remains
and let
to show that λi [A] ∈ Ωc if and only if Ac is Hurwitz. ⎡ ⎤
Consider the bilinear transformation  B̂1 B̂2
(q + r) − z(q − r) N̂ ≜ ⎣Ĉ1 D̂11 D̂12 ⎦ . (20)
s= , s, z ∈ C , (15)
z−1 Ĉ2 D̂21 D̂22
which maps the interior of the disk with radius r and center (−q, 0) Consider, also, the hereafter called controller matrix
in the variable s onto the open left half-plane in z. It is easy to verify [ ]
that AK BK
K≜ (21)
CK DK
s −1
In = z −1
In ⋆ Γ (q, r). (16)
constructed from the matrices defining the state–space represen-
Consequently, with some abuse of notation, tation (19) for the controller. It can be shown that the uncertain
closed-system (G (s, ∆), K (s)) depicted in Fig. 2 admits the state–
(s−1 In , A) = (z −1 In ⋆ Γ (q, r), A) = (z −1 In , Γ (q, r) ⋆ A). (17) space realization
Thus, Ac = Γ (q, r) ⋆ A is Hurwitz if and only if λi [A] ∈ Ωc .
[ ] [ ]
ẋ(t) x(t)
Finally, the condition in Theorem 1 can be easily obtained by = (∆ ⋆ M(K)) , (22)
ẋK (t) xK (t)
applying the above result to the state matrix (∆ ⋆ M) of the
uncertain system (1). □ where
0 I 0 0 I 0
([ ] [ ])
In summary, Theorem 1 states that the robust D-stability of
the uncertain system (1) can be assessed via an equivalent robust M(K) ≜ I 0 0 N̂ I 0 0 ⋆ K. (23)
stability test on the auxiliary system (8)–(9). 0 0 I 0 0 I
Although the D-region considered here is less general that the Now, a straightforward application of Theorem 1 to the state equa-
LMI-region normally considered in LMI-based techniques e.g. [7,3],
tion (22)–(23) allows one to conclude that if a controller matrix
it is, however, general enough to address design specifications
K can be found such that the uncertain system (8) with matrix M
typically appearing in realistic engineering applications.
given by (23) is robustly stable, then it follows that the controller
3. Robust controller synthesis K (s) corresponding to K renders the original uncertain closed-loop
system (G (s, ∆), K (s)) robustly D-stable.
A new synthesis technique allowing the design of robust D- In summary, the original design problem involving the robust
stabilizing structured controllers is now obtained on the basis of D-stabilization of the uncertain closed-loop system (G (s, ∆), K (s))
the equivalent robust stability test introduced in Theorem 1. in (18)–(19) can be solved equivalently via the following robust
First, consider the uncertain closed-loop system depicted in stabilization problem:
Fig. 2. It is assumed that the uncertain plant G (s, ∆) admits the LFT Synthesis problem P1: find a static gain K with structure (21)
representation such that the uncertain system (8)–(9), with M(K) given by (23), is
robustly stable.
ẋ(t) = Ax(t) + B1 w(t) + B2 u(t), It is highly instructive to rearrange the uncertain closed-loop
z(t) = C1 x(t) + D11 w(t) + D12 u(t), system considered in problem P1 into the classical standard form
(18)
y(t) = C2 x(t) + D21 w(t) + D22 u(t), for µ-synthesis [25] depicted in Fig. 3. By following the procedure
w(t) = ∆z(t), often referred to as ‘‘pulling out the ∆s’’ in robust control the-
with x ∈ Rn , w, z ∈ Rr , u ∈ Rm , y ∈ Rp and ∆ ∈ B∆, whereas the ory [24, p. 100], it can be shown after some algebraic manipulations
controller K (s) to be designed admits the realization that the auxiliary uncertain system considered in P1, given by
(8)–(9) and (23), can be reorganized into the standard LFT inter-
ẋK (t) = AK xK (t) + BK y(t),
(19) connection (∆ ˆ , P(s) ⋆ C ) depicted in Fig. 3, with
u(t) = CK xK (t) + DK y(t),
with xK ∈ RnK . ˆ ≜ I6 ⊗ ∆ ,
∆ C ≜ I6 ⊗ K . (24)
 J.F.S. dos Santos et al. / Systems & Control Letters 116 (2018) 8–14 11

the introduction of structured robust controller synthesis tech-

niques based on nonsmooth optimization [16,18,28].
The structured µ-synthesis approach in [16,28] can tackle syn-
thesis problems like P1. However, since it relies on multipliers, it is
not the most efficient alternative for those real µ-synthesis prob-
lems as P1 where parametric uncertainties appear repeated many
times. The parametric robust structured control design technique
recently introduced in [18], on the other hand, dispenses with
multipliers, and hence is particularly tailored for real µ synthesis
problems such as P1. A numerical implementation of the technique
in [18] is currently available in routine systune in MATLAB.
Fig. 3. Standard form for µ-synthesis.
The key steps in the proposed controller synthesis technique
can be finally summarized as follows. First, given the uncertain
plant (18) and the regional specification D(q, r , τ , θ, γ ), the aux-
A realization for the corresponding augmented synthesis plant P(s) iliary uncertain system (∆ ˆ , P(s) ⋆ C ) considered in the synthesis
is given as follows. For a given quintuple (q, r , τ , θ, γ ), consider the problem P1 is constructed, according to (24) and (27). Next, prob-
notation lem P1 is solved via the robust structured controller design tech-
c θ ≜ cos(θ ), sθ ≜ sin(θ ), nique in [18]. Finally, from the resulting controller matrix K, the
q+r 1 −2r −1 sought robust structured controller K (s) can be straightforwardly
Γ11 ≜ , Γ12 ≜ , Γ21 ≜ , Γ22 ≜ . obtained.
q−r q−r q−r q−r
More precisely, problem P1 can be tackled by means of the
Also, let the order of the closed-loop system (G (s, ∆), K (s)) be nonsmooth optimization program
denoted by N = n + nK , and let
min h(κ ) ≜ max α (∆ ⋆ M(K(κ ))), (28)
κ∈Rnc ∆∈∆
R ≜ (IN − Γ22 Â), X ≜ (IN + ÂR−1 Γ22 ). (25)
with κ consisting in the vector of tunable parameters. The idea is
Finally, consider the static matrix that if a feasible solution κ ⋆ to program (28) can be found such
Ω11 Ω12 Ω13 that h(κ ⋆ ) < 0, then the controller given by K(κ ⋆ ) solves problem
[ ]
Ω ≜ Ω21 Ω22 Ω23 , (26) P1, and consequently the original robust D-stabilization problem.
Ω31 Ω32 Ω33 Optimization problem (28) is, however, recognizably difficult to
solve due to the nonsmooth, nonconvex objective function given
where
([ by the semi-infinite maximum. A particularly tailored nonsmooth
sθ −c θ
]
optimization technique allowing to tackle program (28) has been
Ω11 ≜ diag ⊗ (Â + γ IN ),
cθ sθ presented in [18].
) On a final note, it is worth mentioning that although only pole
placement is discussed here, the presented pole clustering crite-
(Γ11 IN + Γ12 ÂR −1
Γ21 ), (Â + τ IN ) , rion can be easily incorporated into a multi-objective framework,
([ ) see e.g. [18, sec. V.D]. The multi-objective formulation typically
Ω12 ≜ diag sθ
−c θ ⊗ B̂1 , c θ sθ ⊗ B̂1 , (Γ12 X B̂1 ), B̂1 ,
] [ ]
considered in nonsmooth optimization approaches involves a fi-
([ ) nite set of objective functions belonging to a particular class of
Ω13 ≜ diag sθ −c θ ⊗ B̂2 , c θ sθ ⊗ B̂2 , (Γ12 X B̂2 ), B̂2 ,
] [ ]
semi-infinite maxima e.g. h(·) in (28). The application of a Cheby-
([ ]T ) shev norm-based scalarization of the original multi-objective de-
Ω21 ≜ diag I2 I2 ⊗ Ĉ1 , (Ĉ1 R−1 Γ21 ), Ĉ1 , sign problem then leads to a minimization program in which the
( ) objective function also belongs to the same class of semi-infinite
Ω22 ≜ diag I4 ⊗ D̂11 , (Ĉ1 R−1 Γ22 B̂1 + D̂11 ), D̂11 , maxima [15]. Consequently, the resulting minimization problem
( ) can be solved by the same nonsmooth optimization technique
Ω23 ≜ diag I4 ⊗ D̂12 , (Ĉ1 R−1 Γ22 B̂2 + D̂12 ), D̂12 , developed for the single objective case.
([ ]T )
Ω31 ≜ diag I2 I2 ⊗ Ĉ2 , (Ĉ2 R−1 Γ21 ), Ĉ2 ,
( ) 4. Numerical application
Ω32 ≜ diag I4 ⊗ D̂21 , (Ĉ2 R−1 Γ22 B̂1 + D̂21 ), D̂21 ,
( ) The structured controller synthesis technique introduced in
Ω33 ≜ diag I4 ⊗ D̂22 , (Ĉ2 R−1 Γ22 B̂2 + D̂22 ), D̂22 . Section 3 is now used to design a reduced-order controller for
the robust D-stabilization of the uncertain model of a missile.
Hence, a realization for the augmented synthesis plant P(s) is The present problem has been adapted from [7,29], where a more
detailed description of the missile model can be found.
P(s) ≜ s−1 I4N ⋆ Ω . (27) The uncertain dynamics of the roll axis of the missile is given by
From (24), it becomes clear that the robust controller synthesis ẋ(t) = (A + 0.9δ1 Aδ )x(t) + (B + δ2 Bδ )u(t),
problem P1 to be solved can be seen as a particular µ-synthesis y(t) = Cx(t),
(29)
problem involving a highly structured static output-feedback con-
troller and a parametric uncertainty block which is also highly where
structured. The ability to efficiently solve that class of robust con-
⎡ ⎤
−180 0 0 0 0
trol problems was out of reach of the control system engineers until ⎢ 0 −180 0 0 0 ⎥
recently, mainly due to fact that available synthesis techniques A = ⎢−21.23 0 −0.6888 −14.7 0 ⎥,
⎢ ⎥
were unable to satisfactory handle structural constraints on the ⎣ 256.7 0 122.6 −1.793 0 ⎦
controller. That scenario has substantially changed recently with −52.33 304.7 0 36.7 −9.661
12 J.F.S. dos Santos et al. / Systems & Control Letters 116 (2018) 8–14

Initially, the design objective is to synthesize a static controller

achieving the closed-loop robust D-stabilization of the missile
dynamics with respect to the region D1 indicated in Table 1. The
concomitant presence of the structural and the pole clustering
constraints renders classical robust controller synthesis techniques
inoperative in this case. The popular D, G-K iteration [30], for
instance, can handle neither one of the constraints.
To begin with, the uncertain missile dynamics (29) must be
rearranged into the LFT form (18), e.g. following the lines in [24].
The resulting minimal LFT representation of the missile dynamics
is then given by (18), where

∆ = diag(δ1 I5 , δ2 I2 ), (30)

0.9Aδ
[ ]
Bδ , B2 = B, , C2 = C ,
[ ]
B1 = I5 C1 =
02×5
[ ] (31)
0
D12 = 5×2 , D11 = D21 = D22 = 0.
I2

The auxiliary synthesis problem P1 is then constructed and

solved via the nonsmooth technique in [18], which succeeds to find
Fig. 4. Plant pole diagram. a feasible solution. The resulting static controller

1.369 −0.363 −0.0986

[ ]
K1 = (32)
Table 1 −0.6826 −0.2373 −0.09961
Design specifications.
thus robustly D-stabilizes the closed-loop system with respect to
ζ r τ γ q
the region D1 . Fig. 5 depicts the pole diagram for the closed-loop
D1 0.5 250 20 0 0
D2 0.7 200 20 0 0
system with the static gain K1 , again for 500 random samples of
the uncertain parameters vector [δ1 δ2 ]T in the allowable region
δ1 , δ2 ∈ [−1, 1]. As expected, the imposed regional constraints are
⎡ ⎤ indeed respected.
27 0 0 0 0
Next, it is considered a more stringent design specification,
⎢ 0 27 0 0 0 ⎥
represented by the region D2 indicated in Table 1. The aim with the
Aδ = ⎢21.2 0 0.688 14.96 0 ⎥,
⎢ ⎥
⎣38.6 tighter specification is to reduce the closed-loop system bandwidth
0 122.6 0 0 ⎦
and the overshoot in the time response, in order to reduce the
52.4 304.8 0 36.8 9.66
risk of actuator amplitude and rate saturation. Unfortunately, no
⎤T feasible static control solution can be found with the new con-
40.5
⎡ ⎤ ⎡ ⎤ ⎡
180 0 0 0 0 0 straints set. A first-order controller is then selected, but once again
⎢ 0 180⎥ ⎢ 0 40.5⎥ ⎢0 0 0⎥ no feasible solution can be found.
B=⎢ 0 0 ⎥ , Bδ = ⎢ 0 0 ⎥ , C = ⎢1 0 0⎥ , When a second-order controller is considered, a feasible solu-
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣256.7 0 ⎦ ⎣57.9 0 ⎦ ⎣0 1 0⎦ tion satisfying the design constraints can be obtained. The resulting
0 0 0 0 0 0 1 dynamic controller is given by
⎤T ⎤
and with δ1 , δ2 ∈ [−1, 1] representing parametric
⎡
136.5 15.31
⎡
uncertainties.
40.24
[ ]
]T −2241 ⎣−44.36 −4.021⎦
The state vector in (29) is given by x = δr δp r ny p ,
[ ⎢ ⎥
13.62 −324.7
. .
⎢ ⎥
− 114 8 − 9 74
where δr and δp are, respectively, the yaw and roll control surface
⎢ ⎥
K2 (s) = ⎢ . (33)
⎢ ⎥
⎤T ⎥
deflections, r is the yaw rate, ny is the yaw acceleration and p is the . .
⎡
⎢ [ 2 31 − 0 6217 ⎥
3.112 −48.11
]
]T ⎢
⎣−0.3332
⎥
roll rate. The input vector is given by u = δrc δpc , with δrc and 0.0079 ⎦ ⎦
[
−0.1805 3.055
⎣
δpc representing the yaw and roll control commands, respectively. −0.5152 −0.01794
[ ]T
The output vector is y = r ny p . Fig. 6 depicts the location of the closed-loop system poles with
The main objective in this application is to design a robust reg- controller K2 (s), again for 500 random samples of the uncertain
ulator for the uncertain missile dynamics, i.e. a control system that parameters vector [δ1 δ2 ]T in the allowable region. Despite the
suitably takes the plant from any nonzero state to the zero state, severity of the new constraints, regional constraints are met by the
for any value of the uncertain parameters in the allowable region reduced-order controller.
δ1 , δ2 ∈ [−1, 1]. More specifically, the design objective is to syn- The time response of the yaw acceleration for a one g lateral ac-
thesize a controller achieving closed-loop robust D-stabilization of celeration initial condition is depicted in Fig. 7 for both controllers
the missile dynamics. and for 50 random samples of the uncertain parameters in the
The controller to be designed admits the state–space realization allowable region. The more severe constraints on the damping rate
(19) and is to be placed in closed-loop with the uncertain missile and the natural frequency (bandwidth) considered in the design of
dynamics according to the interconnection depicted in Fig. 2, with K2 (s) result in smaller overshoots and large response times in gen-
the input vector u and output vector y as indicated above. Fig. 4 eral. The associated time response of the yaw control command is
depicts the location of the plant poles for 500 random samples of depicted in Fig. 8. As expected, controller K2 (s) leads to a significant
the uncertain parameters vector [δ1 δ2 ]T in the allowable region improvement in terms of smaller maximum amplitude and rate on
defined by δ1 , δ2 ∈ [−1, 1]. the yaw control command.
 J.F.S. dos Santos et al. / Systems & Control Letters 116 (2018) 8–14
Fig. 5. Closed-loop pole diagram with static controller K1 .
Fig. 6. Closed-loop poles with dynamic controller K2 (s).
5. Conclusion
References
A new robust pole placement technique has been presented,
which allows the synthesis of dynamic multivariable output feed-
back controllers meeting prescribed structural constraints and
achieving robust regional pole placement in the presence of para-
metric uncertainties. Since the presented technique dispenses with
Lyapunov variables, it can be applied to systems of more siz-
able order. Such features are rarely jointly present in currently
available controller synthesis methods. Moreover, the proposed
pole placement criterion can be easily incorporated into existing
multi-objective formulations based on nonsmooth optimization.
The applicability and validity of the proposed synthesis tech-
nique has been illustrated by a challenging numerical exam-
ple involving reduced-order constraint and repeated parametric
uncertainties.
13
Fig. 7. Yaw acceleration for controllers K1 (solid) and K2 (s) (dash–dot).
Fig. 8. Yaw control command for controllers K1 (solid) and K2 (s) (dash–dot).
[1] A. Tapia, M. Bernal, L. Fridman, Nonlinear sliding mode control design: An LMI
approach, Systems Control Lett. 104 (2017) 38–44.
[2] U. Montanaro, J.M. Olm, Integral MRAC with minimal controller synthesis and
bounded adaptive gains: The continuous-time case, J. Franklin Inst. B 353 (18)
(2016) 5040–5067.
[3] G. Chesi, Parameter and controller dependent Lyapunov functions for robust
D-stability and robust performance controller design, IEEE Trans. Automat.
Control 62 (9) (2017) 4798–4803.
[4] N. Sivashankar, I. Kaminer, P.P. Khargonekar, Optimal controller synthesis with
regional stability constraints, in: Proceedings of the 32nd IEEE Conference on
Decision and Control, 1993., Vol. 1, 1993, pp. 110–115.
[5] W.M. Haddad, D.S. Bernstein, Controller design with regional pole constraints,
IEEE Trans. Automat. Control 37 (1) (1992) 54–69.
[6] M. Chilali, P. Gahinet, H∞ design with pole placement constraints: an LMI
approach, IEEE Trans. Automat. Control 41 (3) (1996) 358–367.
[7] M. Chilali, P. Gahinet, P. Apkarian, Robust pole placement in LMI regions, IEEE
Trans. Automat. Control 44 (12) (1999) 2257–2270.
14 J.F.S. dos Santos et al. / Systems & Control Letters 116 (2018) 8–14
[8] D. Peaucelle, D. Arzelier, O. Bachelier, J. Bernussou, A new robust D-stability
condition for real convex polytopic uncertainty, Systems Control Lett. 40 (1)
(2000) 21–30.
[9] D.H. Lee, J.B. Park, Y.H. Joo, K.C. Lin, Lifted versions of robust D-stability and
D-stabilisation conditions for uncertain polytopic linear systems, IET Control
Theory Appl. 6 (1) (2012) 24–36.
[10] D. Henrion, M. Sebek, V. Kucera, Robust pole placement for second-order
systems: An LMI Approach, IFAC Proc. Vol. 36 (11) (2003) 419–424 4th IFAC
Symposium on Robust Control Design 2003, Milan, Italy, 25-27 June 2003.
[11] F. Yang, M. Gani, D. Henrion, Fixed-order robust H∞ controller design
with regional pole assignment, IEEE Trans. Automat. Control 52 (10) (2007)
1959–1963.
[12] J.V. Burke, A.S. Lewis, M.L. Overton, A nonsmooth, nonconvex optimization
approach to robust stabilization by static output feedback and low-order
controllers, in: Proc. IFAC Symp. Robust Control Design, Milan, Italy, 2003.
[13] P. Apkarian, D. Noll, Nonsmooth H∞ synthesis, IEEE Trans. Automat. Control
51 (1) (2006) 71–86.
[14] J.V. Burke, D. Henrion, A.S. Lewis, M.L. Overton, HIFOO - A MATLAB package for
fixed-order controller design and H∞ optimization, in: 5th IFAC Symposium
on Robust Control Design, 2006.
[15] A.M. Simões, P. Apkarian, D. Noll, Nonsmooth multi-objective synthesis with
applications, Control Eng. Pract. 17 (11) (2009) 1338–1348.
[16] P. Apkarian, Nonsmooth µ synthesis, Internat. J. Robust Nonlinear Control
21 (13) (2011) 1493–1508.
[17] I. Yaesh, U. Shaked, H∞ optimization with pole constraints of static output-
feedback controllers - a non-smooth optimization approach, IEEE Trans. Con-
trol Syst. Technol. 20 (4) (2012) 1066–1072.
[18] P. Apkarian, M.N. Dao, D. Noll, Parametric robust structured control design,
IEEE Trans. Automat. Control 60 (7) (2015) 1857–1869.
[19] R.S. da Silva de Aguiar, P. Apkarian, D. Noll, Structured robust control against
mixed uncertainty, IEEE Trans. Control Syst. Technol. PP (99) (2017) 1–11.
[20] K.-C. Goh, M.G. Safonov, J.H. Ly, Robust synthesis via bilinear matrix inequali-
ties, Internat. J. Robust Nonlinear Control 6 (9) (1996) 1079–1095.
[21] V. Bompart, P. Apkarian, D. Noll, Non-smooth techniques for stabilizing linear
systems, in: 2007 American Control Conference, 2007, pp. 1245–1250.
[22] A.M. Simões, D.C. Savelli, P.C. Pellanda, N. Martins, P. Apkarian, Robust design
of a TCSC oscillation damping controller in a weak 500-kV interconnection
considering multiple power flow scenarios and external disturbances, IEEE
Trans. Power Syst. 24 (1) (2009) 226–236.
[23] M.A. Mammadov, R. Orsi, A nonsmooth optimization approach to H∞ synthe-
sis, in: Proceedings of the 44th IEEE Conference on Decision and Control, 2005,
pp. 6893–6898.
[24] K. Zhou, J.C. Doyle, K. Glover, Robust and Optimal Control, Prentice Hall,
Englewood Cliffs, New Jersey 07632, 1996.
[25] J. Doyle, Analysis of feedback systems with structured uncertainties, in:
IEE Proceedings D - Control Theory and Applications, Vol. 129 (6) 1982,
pp. 242–250.
[26] E. Davison, N. Ramesh, A note on the eigenvalues of a real matrix, IEEE Trans.
Automat. Control 15 (2) (1970) 252–253.
[27] S. Gutman, E. Jury, A general theory for matrix root-clustering in subregions of
the complex plane, IEEE Trans. Automat. Control 26 (4) (1981) 853–863.
[28] E.F.M. Menezes, R.S.S. Aguiar, A.M. Simões, P. Apkarian, Structured robust
controller design via non-smooth mixed µ synthesis, IET Control Theory Appl.
10 (17) (2016) 2186–2193.
[29] K.M. Sobel, W. Yu, J.E. Piou, J. Cloutier, R. Wilson, Robust eigenstructure assign-
ment with structured state space uncertainty and unmodelled dynamics, in:
1991 American Control Conference, 1991, pp. 3137–3141.
[30] P.M. Young, Controller design with real parametric uncertainty, Internat. J.
Control 65 (3) (1996) 469–509.