Professional Documents
Culture Documents
Sufficient Conditions for Dynamical Output Feedback Stabilization via the Circle Criterion
Abstract— This paper suggests sufficient conditions for alence principle can be relaxed. Instead, a fixed structure of
asymptotically stable dynamical output feedback controller an output feedback controller can be imposed, search for its
design based on the circle criterion. It is shown that a dynamic parameters and asymptotic stability can be approached by
output feedback stabilization problem with impending problems
of finite escape time, previously attacked by observer-based means of the circle criterion for the closed-loop system. It
design, can be successfully solved using circle criterion design. happens that for the large class of the systems considered
Stability of the closed-loop system is global and robust to in [1], [2], this argument works. Thus, time-varying systems
parameter uncertainty. and systems with structural uncertainties can be approached
whereas both extensions seem infeasible using the certainty
I. I NTRODUCTION
equivalence principle [1], [2].
Stabilizing a nonlinear system by output feedback is often The paper is organized as follows: Section II suggests
a difficult problem. Equipped with a range of available an illustrative example considered in details. The problem
asymptotic (robust) stability criteria, one could try to design a statement, assumptions and main result are given in Sec. III,
dynamical output feedback controller so that the closed loop a brief discussion of results being added in Sec. IV.
system satisfies one of these stability criteria. This simple II. M OTIVATING E XAMPLE
idea sometimes leads to a problem that could be solved,
but quite often it results in a problem that is intractable. Consider the following dynamical system [1]
This paper is devoted to the discussion of dynamic output d x1 0 1 x1 0
= + u − x52 (1)
feedback and the circle criterion, as a stability test for the dt x2 0 1 x2 1
closed loop system [6], [9]. One of standard initial ideas for y = x1 (2)
controlling dynamical system by output feedback is based
on the separation principle, that is, one needs to find a The relation between the relevant results of [1] and the
stabilizing full-state feedback controller and to determine an current development is discussed later. The problem is to
observer with asymptotically stable error dynamics; an output design a controller that renders the origin of the system (1)
controller is then chosen to coincide with the derived full- asymptotically stable. Let us consider a dynamical controller
state feedback controller where instead of the unmeasured of the form
true states of the dynamical system, the system states are d
z = λ3 x1 + λ4 z
substituted by observer states. This approach includes three dt
steps—design of a full-state controller, observer design, anal- u = λ1 x1 + λ2 z + (c1 x1 + c3 z)5 (3)
ysis of the closed loop system—where well-known stability
where λi , cj are real constants to be defined. With such a
criteria, like the circle criterion, could be applied to conclude
controller, the dynamics of the closed loop system are
stability. Indeed, one could use this test for checking stability
of the system with a full-state feedback controller; or use it x 0 1 0 x1 0
d 1
for checking that a particular structure of an observer results x2 = λ1 1 λ2 x2 + 1 w
dt
in stable error dynamics; or use this test at the final point z λ3 0 λ4 z 0
to verify that the closed-loop system derived via certainty w = (c1 x1 + c3 z)5 − x52 (4)
equivalence principle is stable. These arguments have been
One can easily check that the nonlinearity w of Eq. (4) and
used for checking stability via the circle criterion of error
the linear virtual output of the closed loop system (4)
dynamics for an observer [1], [2].
The main contribution of this paper comes from the v = c1 x1 − x2 + c3 z
observation that for the large class of systems treated in [1],
satisfies a passivity relationship for any x 1 , x2 , z
[2], there is no need to introduce an explicit observer and
assumptions and arguments relevant for the certainty equiv- v · w = (c1 x1 − x2 + c3 z)[(c1 x1 + c3 z)5 − x52 ] ≥ 0 (5)
A0 = 0 1 0 , A1 = 1 0 0 ,
0 0 0 0 0 0 −5.5
0 0 0 0 0 0
1
−6
λ
A2 = 0 0 1 , A3 = 0 0 0 ,
0 0 0 1 0 0
−6.5
0 0 0 0 x1
A4 = 0 0 0 , B= 1 ,X= x2 , −7
0 0 1 0 z
−7.5
one can rewrite the closed loop system (4) as follows 0.8 0.9 1 1.1
C3
1.2 1.3 1.4
d 4
X = (A0 + λi Ai )X + Bw Fig. 1. The red area corresponds to λ1 and c3 that satisfy the constraints
dt i=1 (6) (10)–(13) while the other parameters (9) have the nominal value (14).
v = CX = c1 x1 − x2 + c3 z
Whereas the explicit form of the nonlinearity w—in this
case given as (4)—is not important, it is important that the then (7) is equivalent to the inequalities
passivity relation (5) be valid. By the circle criterion, the
γ < 0, α<0 (11)
system (6) is asymptotically stable provided that
1) the frequency condition b) Otherwise, (7) is equivalent to the inequalities
4
4αγ − β 2 > 0, α<0 (12)
Re {C(jωI3 − (A0 + λi Ai ))−1 B} < 0 (7)
i=1 4
2) The condition for the matrix (A 0 + λi Ai ) to be
holds for any ω ∈ IR + , the negative real notation i=1
convention being used in the inequality (7); strictly Hurwitz is equivalent to the inequalities
4
2) the matrix (A0 + λi Ai ) is strictly Hurwitz. λ4 < −1, λ1 < λ4 , λ1 λ4 > λ2 λ3 (13)
i=1
As known, these conditions are equivalent to the fact that The set of parameters (9) satisfying the constraints (10)–(13)
there exists the 3 × 3 matrix P = P T > 0, so that is not empty. It can be checked that the vector
4 4
(A0 + λi Ai )T P + P (A0 + λi Ai ) < 0 (λ1 , λ2 , λ3 , λ4 , c1 , c3 ) = (−5, −2, −7, −3, 2, 1) (14)
i=1 i=1 (8)
P B = −C T belongs to this set. This means that for this choice of
Thus, development of a dynamical stabilizing controller (3) the controller parameters (9), the corresponding BMI (8)
based on the circle criterion and the choice of quadratic becomes a Linear Matrix Inequality (LMI), and it has a
constraints (5) require determination of parameters solution. In fact, these values (14) have been found in [1]
via an appropriate observer design and checking the validity
λ1 , λ2 , λ3 , λ4 , c1 , c3 (9) of certainty equivalence principle. Below it will be shown,
so that all points 1)–2) are valid, or equivalently, the Bilinear what, in addition, to the asymptotic stability of the closed
Matrix Inequality (BMI) of (8) is solvable. The routine loop system found in [1], could be gained from the fact that
computations made for this example result in equivalent the BMI (8) is solvable.
statements written in terms of parameters λ i , cj : Firstly, the inequalities (10)–(13) have quite a rich set of
solutions. To show that the inequalities (10)–(13) suggest
1) Introduce the quantities
the controllers that cannot be obtained by the certainty
α = 1 − c1 equivalence principle elaborated in [1], let us check possible
β = c2 λ3 − λ2 λ3 + c2 λ3 λ4 − c1 λ24 − c1 λ1 + λ24 values for the parameter c 3 that is postulated in [1] to
be equal to 1. Figure 1 shows approximation for a set of
γ = (λ1 λ4 − λ2 λ3 )(c2 λ3 − c1 λ4 ) parameters λ1 and c3 that correspond to stabilizing controller
The validity of the frequency condition leads to the two provided that the rest of (9) have the nominal value (14). To
cases illustrate an advantage of the solution based on solvability of
a) if the parameters (9) are so that BMI (8) vs. the design via certainty equivalence principle,
let us include parametric uncertainty in the system. Consider
β < 0, (10) the system (1) with uncertainty factor ε in front of the
4683
20
is important to realize that this uncertainty cannot be treated
15
in the certainty equivalence design arguments elaborated in
10
[1]: for each new value of parameter ε, one need to change
5
the observer! Therefore, the value of ε should be known
0
precisely. Furthermore, the interval of an allowed uncertainty
x1, x2, z
4684
15
For asymptotic stabilization of the origin of the system (20)
10 by output feedback, try a dynamical controller of the form
5 u = Rz z + Ry y + R (Cy y + Cz z, t) (23)
d
0
z = Λz z + Λy y + Λu u + Λ (Cy y + Cz z, t) (24)
x1, x2, z
dt
where z ∈ IRk is the internal state of the controller; Λ z ,
−5
d x A + Ry N1 B1 Rz x
−20 = + (25)
0 5 10
time (sec)
15 20 25
dt z (Λy + Λu Ry ) N1 (Λz + Λu Rz ) z
Fig. 3. The solution of the closed loop system (15), (3) with the nominal B1 R B2
value ε = 1 for the initial conditions x1 = 1, x2 = −3, z = 10. + (Cy y + Cz z, t) + (d, t)
Λu R + Λ 0
Assumption 2: There exists a linear transformation
Necessary conditions for BMI solvability (8) suggest intro-
duction of a reduced order observer [1]. Tx Txz xn x
T = , =T (26)
Statement 2: Consider the nonlinear system (1), (2), (19). 0 Tz zn z
The frequency condition (7) takes the form
with det T = 0, and there exist matrices R , Λu , Λ of
q(λ) 2 such that Tx B1 R + Tzx (Λu R + Λ ) = −Tx B2 .
Re {ṽ∗ w̃} = |w̃| < 0
p(λ) Assumption 3: The scalar nonlinearity (d, t) is such that
where ṽ = G(jω)w̃, λ = ω , 2 there exist a 2 × 2 matrix
Π11 Π12
q(λ) = (1 − c1 )λ(m+1) + qm λm + · · · + q1 λ + q0 Π= (27)
Π12 0
p(λ) = λ(m+2)
+ pm+1 λ (m+1)
+ · · · + p1 λ + p0
and there exists M ∈ IR1×k so that for any function d(t) ∈
and G(s) is the transfer function of linear part of the closed Lk2e : ∃ {tn }+∞
n=1 , tn → +∞ as n → +∞, such that ∀ n
loop system (1), (2), (19). Then, the frequency condition
tn T
holds at ω → +∞ only if the inequality 1 − c 1 < 0 holds. M d(τ ) M d(τ )
Π dτ ≥ 0 (28)
(d(τ ), τ ) (d(τ ), τ )
III. M AIN R ESULTS 0
A. Problem Formulation and Preliminary Comments Assumptions 1, 2 and 3 enable us to rewrite the closed loop
Consider a nonlinear control system of the form system (20), (23), (24) in the input-output form
d d xn xn
x = Ax + B1 u + B2 (d, t) (20) =A + B1 w1 + B2 w2 (29)
dt dt zn zn
y = N1 x, d = N2 x (21)
with
where x ∈ IR is the state vector; y ∈ IR is the measurable
n m
v1 = M d = M N2 Tx−1 xn − Txz Tz−1 zn (30)
output; d ∈ IRk is the vector of variables that serves as
input to the scalar nonlinear block ; A, B 1 , B2 , N1 , N2 v2 = Π (Cy y + Cz z − d) =
T
Π Cz Tz−1 zn
+ (31)T
4685
35
and v1 , v2 are scalar passive outputs (in integral sense) of
the closed loop system (29), i.e., ∃ {t ni } with tni → +∞ 30
as ni → +∞, i = 1, 2, so that
25
tn1 T tn2
v1 (τ ) v1 (τ )
dτ ≥ 0, v2 (τ )w2 (τ )dτ ≥ 0
20
Π
w1 (τ ) w1 (τ )
ε
0 0 15
Denote
10
C2 = ΠT {Cy N1 − N2 } Tx−1 , 0
4686
Theorem 1: Consider the system (39)–(41). Suppose As- only for this particular solution, while for different solutions
sumptions 1–6 hold. Take an output feedback controller as these Lyapunov functions are different. In turn, the solvability
of the BMI implies existence of only one positive definite
t),
u = K1 y + K2 d + K (d, (44) quadratic form that provides exponential convergence of any
where matrices K1 , K2 , K satisfy Assumption 4; the closed-loop solution to the origin.
variable d is defined as d = z + Φy with the matrix Φ from 5. The 2 × 2 quadratic form Π in Assumption 3 has a par-
Assumption 6 and z defined as a solution to ticularly simple structure for use in the example. In general,
one might consider any integral quadratic constraint, [7],
dz
= [(A32 − ΦA22 ) + (A33 − ΦA23 )Φ]y [8], that leads to a design of stabilizing controller when all
dt state variables are available. In turn, the nonlinearity (d, t)
+ (A33 − ΦA23 )z should not be scalar, but it seems that such generality would
+ (B31 − ΦB21 )u + B32 (z + Φy, t) (45) obscure the main contribution of the paper.
6. The computational algorithms for solving nonconvex prob-
Then, the closed-loop system (20)–(21), (44)–(45) is globally
lems (35) are still at a developing stage [5].
asymptotically stable.
V. C ONCLUSIONS
Theorem 2: Consider the system (29)–(31). Suppose As-
In this paper, it was shown that a dynamic output feedback
sumptions 1–6 hold, then the BMIs (35) are solvable, one stabilization problem with impending problems of finite es-
solution being cape time, previously attacked by observer-based design, can
be successfully solved using circle criterion design. Stability
Λz = A33 − ΦA23 , Cz = Ik , of the closed-loop system is global and robust to parameter
Λu = B31 − ΦB21 , Λ = B32 , uncertainty.
Rz = K2 , Ry = K1 + K2 Φ, (46) VI. REFERENCES
R = K , Cy = Φ, [1] Arcak M. A Global Separation Theorem for a New Class
Λy = (A32 −ΦA22 )+(A33 −ΦA23 )Φ of Nonlinear Observers, Proc. 41st IEEE Conf. Decision &
Control, Las Vegas, NV, pp. 676–681, 2002
[2] Arcak M. and P. Kokotović. Nonlinear Observers: A Cir-
IV. D ISCUSSION cle Criterion Design and Robustness Analysis, Automatica,
37:1923–1930, 2001.
1. In Sec. III is shown that stabilization of nonlinear systems [3] Johansson R. and A. Robertsson. Observer-based Strict Posi-
by dynamical output feedback via the Circle Criterion could tive Real (SPR) Feedback Control System Design, Automatica,
be reformulated as a problem of solvability of Bilinear Ma- 38:1557-1564, 2002.
trix Inequalities (BMI) of particular type. In general, finding [4] Shiriaev A. S. Some Remarks on System Analysis via Inte-
gral Quadratic Constraints, IEEE Trans. Automatic Control,
even one solution for BMI could be a difficult problem, while 45(8):1527–1532, 2000.
for low dimensional system, all solutions could be found, [5] Tuan H. D. and P. Apkarian. Low Nonconvexity-rank Bilinear
and the result is written in the form of finite number of Matrix Inequalities: Algorithms and Applications in Robust
inequalities for the parameters of the controller. Controller and Structure Design, IEEE Trans. Automatic Con-
2. This approach allows not only to find a set of all solutions trol, 45(11): 2111–2117, 2000.
[6] Yakubovich V. A.. Frequency Conditions of Absolute Stability
(dynamical stabilizing controllers of particular structure for of Control Systems Having Hysteresis Nonlinearities, Dokl.
low dimensional systems), but also to analyze and quantify Acad. Nauk SSSR, 149(2), 1963.
a robustness of the closed-loop system in case of uncertainty [7] Yakubovich V. A. Frequency Conditions for Absolute Stability
and parameter variation. of Control Systems with Several Nonlinear or Linear Non-
3. To compare the results with related ones in [1], one stationary Blocks, Automatica i Telemekhanica, 6:5–29, 1967.
(English transl. in Automation and Remote Control, pp. 857–
should realize that methods in [1] are devoted to stabilization 880, 1968)
of more general time-invariant dynamical systems, than the [8] Yakubovich V. A. Necessity in Quadratic Criterion for Abso-
equations (20)–(21) represent. This generality reduces the lute Stability, Int. J. Robust & Nonlinear Control, 10:899–907,
ability to find a variety of stabilizing controllers, and only a 2000.
few controllers was found in [1]. [9] Zames G. On the Input-Output Stability of Time-varying
Nonlinear Systems—Part II: Conditions Involving Circles in
4. If Assumption 5 is omitted, then Theorem 1 is still the Frequency Plane and Sector Nonlinearities, IEEE Trans.
true, that is, the closed loop system (20)–(21), (44)–(45) Automatic Control, 11(3):465–476, 1966.
is globally asymptotically stable. In fact, this is the result
ACKNOWLEDGEMENTS
of [1, Theorem 1, p.677] whereas the main result—i.e.,
solvability of the particular BMI stated in Theorem 2— The work has been supported by the Danish Technical
is not. The reason is that convergence of any closed loop Research Council, Grant 26-01-0164. R. Johansson and A.
system solution to the origin can be proven in this case by Robertsson are grateful for financial support from the EC
a particular choice of a Lyapunov-type function, designed project NACO2 and the Vinnova project LUCAS.
4687