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

Copyright Cl IFAC Nonlinear Control Systems Design,

Enschede, The Netherlands, 1998

GLOBAL STABILIZATION OF NONLINEAR


CASCADE SYSTEMS: LIMITATIONS IMPOSED BY
RIGHT HALF PLANE ZEROS

R. Sepulchre a , M. Arcak b

(a)Institut Montefiore, B28


University of Liege
B-4000 Liege Sart- Ti/man, BELGIUM
e-mail: sepulchre@montefiore.ulg.ac.be

(b) Center for Control Engineering and Computations


University of California
Santa Barbara, CA 93106-9560 USA
e-mail: murat@seidel.eee.uesb.edu

Abstract: This paper analyzes the stabilizability properties of nonlinear cascades in


which a nonminimum phase linear system is interconnected through its output to a
stable nonlinear system. It is shown that the instability of the zeros of the linear system
can be traded with the stability of the nonlinear system up to a limit fixed by the
growth properties of the interconnection term between the two parts of the cascade.
Below thi~ limit, global stabilization is achieved by smooth static state feedback.
Beyond this limit, various examples illustrate that controllability of the cascade may
be lost, making it impossible to achieve large regions of attractions.
Copyright© 1998lFAC

Keywords: Stabilization, nonlinear cascades, peaking.

1. INTRODUCTION escape time for the solution z(t) if the output


y(t) of the linear subsystem (A, B, C, D) is not
It has been widely acknowledged in the nonlinear properly controlled.
stabilization literature that the study of partially
Beyond invertibility conditions for the linear sys-
linear cascades
tem (for simplicity, we only consider square sys-
~ = f(z) + \It(z,~,y)y tems (m = p) and assume a uniform relative
~ = A~+Bu degree {r, .. . , r} ), successive contributions in the
y = C~ + Du, (z,~, y, u) E lR$ X lRn x lRP x lRm literature have revealed the prominent role played
(1) by the zeros of the linear system in the global
stabilizability of the cascade (1) .
illuminates the structural properties which make
With their celebrated analysis of the peaking phe-
possible or impossible to achieve large regions
nomenon, Sussmann and Kokotovic (1991) have
of attraction (Sussmann and Kokotovic,1991;
shown that the infinite zeros of the linear system
Saberi, et al., 1990; Lin and Saberi, 1993; Braslavsky
are the most harmful ones. Because of the large
and Middelton, 1996; Sepulchre, et al., 1997; Teel,
transients that they exhibit during the fast stabi-
1996). The general scenario in these references is
lization of the output, the states associated to the
that the nonlinear subsystem i = f(z) has a glob-
infinite zeros, that is, the output derivatives, must
ally asymptotically stable equilibrium z = 0 but
be excluded from the interconnection \It to render
that the perturbation \It(z,~,y)y may cause finite

597
the global stabilization of the cascade possible. In Section 2 describes our main result in the relative
a subsequent paper (Saberi, et al., 1990), Saberi degree zero case, that is when y = u. Section
and the same authors showed that if the out- 3 provides two examples of loss of global con-
put derivatives do not enter the interconnection trollability when the stabilizability boundary is
and all the finite zeros of the linear system are attained. The general cascade (3) is considered in
Lyapunov stable (the cascade is then said to be Section 4.
"weakly minimum phase"), then the stabilizabil-
ity of the linear system guarantees the global sta-
bilizability of the cascade. This result was further 2. MAIN RESULT
extended in (Sepulchre, et al., 1997) -see also
(Teel, 1996) for a different version and (Lin and If the cascade (1) has a uniform relative degree
Saberi,1993) for the semiglobal counterpart- to zero, that is, if D is invertible, then a preliminary
the unstable situation where repeated zeros lie on feedback puts the system in the simpler form
the imaginary axis. (with new matrices A, B, and new variable u)

The situation of unstable finite zeros was con- ~=Fz+W(z,~,u)u (3)


sidered for the first time in the recent paper ~ . = A~ + Bu .
(Braslavsky and Middelton, 1996). With simple For clarity, we tehiporarily assume that the z-
but illuminating examples, like the two dimen- subsystem is linear. We assume that W is smooth,
sional system that the pair (A, B) is stabilizable, and that F is
~ = -O'Z + zq+1y1' Hurwitz. Let 0' > 0 such that QF+FTQ ~ -20'Q
~ = v~ + u, (2) for some Q > 0 (that is, 0' := - maxRe>.(F) > 0
y=u z E JR, ~ E JR, u E JR if all the eigenvalues with real part -0' are simple) .
the authors showed that unstable zeros may con-
stitute an obstacle to semiglobal stabilization if Although local stabilization is achieved with a lin-
they are "too far" to the right . Indeed, their sta- ear feedback u = K~, a difficulty arises to achieve
bilization requires a finite output energy (Middel- large regions of attraction when the matrix A has
ton,1991) which is sufficient to cause finite escape some eigenvalues in the right half-plane and the
time for z(t). In the same paper, the authors growth of \11 is faster than linear in the variable z .
showed on an example (using discontinuous feed- Indeed, a finite energy of the controller (in the L2
back) that global stabilization might be possible sense) is then needed to stabilize ~(t) (see for in-
when the zeros are closer to the imaginary axis. stance (Seron, et ai, 1997). At the same time, any
finite perturbation in the z subsystem may cause
The present paper completes the analysis initiated a finite escape time of the solution z(t) so that the
in (Braslavsky and Middelton,1996) by quantify- controller energy must be minimized. To overcome
ing for the general cascade (1) a precise stabiliz- these contradictory requirements, we will use a
ability boundary in terms of structural parameters form of local input-to-state (ISS) stability (Son-
of the cascade. Below this boundary, we design tag,1995), for the nonlinear system and implement
a smooth static state feedback which achieves the strategy suggested in (Braslavsky and Mid-
global stabilization of the cascade. Beyond this delton, 1996): letting ~(t) temporarily diverge, a
boundary, various examples illustrate that struc- zero (or almost zero) control is first applied to let
tural limits are attained and may cause a loss z(t) converge to a small neighborhood of z = O.
of global controllability, making the semiglobal Then we apply the finite energy control needed to
stabilization of the cascade impossible. The sta- stabilize ~(t). This control will not destabilize z(t)
bilizability condition is expressed as an inequality thanks to the local ISS property of the nonlinear
between two ratios: a stability ratio between the system . Our main result quantifies when such a
local stability of the z-subsystem i =• (pa-
J(z) strategy is successful.
rameter 0') and the instability of the finite zeros
(parameter v), and a growth ratio between the Theorem 1. Suppose that there exist positive con-
growth of the interconnection term w(z,~, y) in stants p, q, C such that, for z sufficiently small,
the variable z (parameter q) and in the variables
(~, y) (parameter p). Our stabilizability boundary 1I\II(z,e, u)1I ~ Cllzllq+11l({, u)IIP - 1 (4)
of the cascade is attained when the stability ratio and there exist a constant 11 > 0 and a matrix
becomes equal to the growth ratio: P > 0 that satisfy
~='!.
0' P PA + AT P - 211 P ~0 (5)
Below this limit, the stability of the finite zeros A - BBT P Hurwitz. (6)
associated to the z-subsystem can be traded with
the instability of the finite zeros associated to the Then the equilibrium (z,e) = 0 of (3) is globally
~ -su bsystem. stabilizable by smooth state feedback provided

598
that one of the two following conditions is sat- for some constant Cl so that
isfied:
v q IIzTQ1lr(Z~~rBT P~)llr(z,~)IIBT P~II
-<- (7) z z
a p
or
~= C£ and 1lr(z,~, 0) = 0 (8) Rearranging the terms of (14) we obtain
0' P
. IIBT P~1I2 1.
W:::; -2W(", + Pr II~II~ - qC2 W') .
Proof: Let Q > 0 be solution of Q F + FT Q :::;
-20'Q and define IlzllQ := JzTQz . Then impos- This means that in the region nf , the control (13)
ing u = 0 in the region IIzllQ 2: 6 guarantees that ensures ItV < - ~ W provided that { is chosen
all solutions of the cascade satisfy after a finite small enough to satisfy
time T= T(z(0),~(0),6)
'" - qC2 {21 >
- -2 .
'"

Vt > T: IIz(t)IIQ < 6 (9)


The constant 6 > 0 is a design parameter. We To have a smooth feedback, we choose for r a
will choose it small enough such that the growth smooth gain which satisfies r( z,~) = 0 for (z,~) t/.
condition (4) holds when IIzllQ :::; 6. =
nf and r(z,~) 1 for (z,O E nf / 2·
Next we show that ~(t) is bounded in nf . Suppose
Define 11~llp := Je P~ where P is as in (5) and
it is unbounded for some initial condition . Since
(6) . To analyze the convergence of the solutions
of (3), we will employ the nonnegative definite the product IIzll~q 1I~1I;r is decreasing, ~(t) ~ 00
implies z(t) ~ 0, therefore the solution (z(t) , ~(t))
function
enters the set nf / 2 in finite time. But the control
W(z,O = IIzll~ 1I~1I;r · (10) u = _BT P~ in nf / 2 stabilizes the ~-subsystem
and the result follows by contradiction.
Evaluating the time-derivative of W along the
solutions of (3) yields Having proved boundedness of solutions, we con-
. zTQ1lr ePB
clude the global attractivity of (z , O = (0,0)
W :::; -2(O'q - vp)W + 2W[q~Q + p~lu . by La Salle's invariance principle. Since the set
z z "P~ ItV == 0 is included in W == 0, we simply note
(11)
that the largest invariant set in W == 0 is the
We will treat conditions (7) and (8) separately. origin . The local exponential stability follows by
inspection of the J acobian linearization at the
Suppose first that ·the condition (7) is satisfied.
origin.
Then '" := (O'q - vp) > 0 and W decays ex-
ponentially along the solutions as long as u = Suppose now that we have condition (8) instead
O. Choosing the control u = 0 in the region of condition (7). Then (11) gives the weaker con-
where W(z,~) 2: { ensures that all solutions dition
of the cascade satisfy after a finite time T2 = . zTQ1lr ePB
T 2 ( {, z(O), ~(O)) W :::; 2W[q zTQz + p e P~ lu (15)
Vt 2: T2 : W(z(t),~(t)):::; { (12) so that the decay of W along the solutions is no
Combining (9) and (12), we will select the control longer guaranteed with the control u = O. We will
u = 0 outside the region now choose a control law in the form
2 T
n f := {(z,~)lllzIIQ :::; { and W(z,~):::; {} u = -rl(lIzIIQ)1 +r2(W)B P~, 0:::; r1:::; I , r2 2: O.
where the constant { > 0 is a parameter yet to (16)
be determined. By construction, the region n f is For the gain r1, we choose a smooth function
attractive and invariant. which satisfies r(8) = 0 for 8 2: 6 in order to
The control law in n f is selected in the form ensure (9), which is strictly positive for 8 < 6,
and which satisfies r(8) = 1 for 8 :::; ~ .
u = -r(z,~)BT P~, o :::; r(z,~) :::; 1 (13)
Because \)(z,~, 0) = 0 and 1lr is smooth, we can
With this choice, the time-derivative of W satisfies rewrite the_columns 1/Ii(Z,~,U) of1lr(z,~,u) in the
1 . IIBT p~1I2 ~ new form 1/Ii(Z,~, u)u, i = l..m. Then,
2W:::;-",W-PrW II~II~ +qC2W (14) T T - T-
I Z Q1lr ul < 11[ z Q1/IIU ... z Q1/Im u lllllull
where we have used the fact that, for { > 0 small zTQz - zTQz

< C 3 "m
enough, T - T - 2
L..ti-1 Iz Q1/Ii u l lIull < C 3 L~l IIz Q1/Iill lIull
IIzTQ1lr(z,~,r(W)BT P~)II:::; C11Izll~+211~1I~-1 - zTQz - zTQz

599
12
< C W / ('Y11IBT P~II) 2
where (t) ~ i for some constant i > O. Using
Cauchy-Schwartz inequality, we have
- 4 II~II~ 1 + 1'2
where we have used the equivalence of norms and le P2 BBT pel ~ IIBT P2ellll BT pe 11
the growth condition (4) . Substituting this result and therefore
along with the control law (16) into (15) yields
d T T
dt 1I~1I~2 ~ -C611~1I~2 +(1-(t))IIB P2ellllB pell
W < -4P'Y1 ~ IIBT P~1I2
- 1 + 1'2 II~II~
(1 _12
Cs W / )
1 + 'Y2(W) Dividing both sides by lIe(t)II~, we obtain
for some positive constant Cs . By choosing for the
gain 1'2 a smooth function such that d
dt 10g(lIell~,) ~
S1/2 1 T
---<- < -C6 + (1 _ (t)) lIellp IIBT P2ellllB pe 11
1 + 'Y2(S) - 2Cs
- lie lip, lie 11 P, lIellp
and such that 'Y2(S) = 1 for S ~ (, with ( > 0, we ~ + Cd(t)
-C6
guarantee
for some constant C 7 > o. If /(t) ---> 0 as t ---> 00 ,
W< -2 W IIBT P~112 then -C6 + cdW ~ -~ for t large enough ,
- P'Y1 1 + 'Y2(W) II~II~ (17)
say for t ~ T . But 'then log(II~II~2) cannot grow
and we ensure that W(t) = IIz(t)lIbll~(t)lI~ expo- unbounded as t ---> 00, which contradicts the un-
nentially converges to zero. Then , the control law boundedness of e(t). This concludes the proof. 0
(16) is u = _BT P~ for W ~ ( and IIzllQ ~ ~ .
We now prove that the equilibrium (z,O = (0 , 0)
is globally attractive. Suppose first that ~(t) is It is clear from the proof that the exponential
bounded. Then by LaSalle Invariance Principle, stability of the z-subsystem is only used locally.
(z(t) , ~(t)) converges to the largest invariant set It is therefore a trivial adaptation of Theorem
where W == O. But W == 0 implies u == 0 so that 1 to replace the linear subsystem i = F z by
z(t) ---> O. Then 1'1 = 1 and u == 0 implies BT P~ == the nonlinear subsystem i = /(z) provided that
=
O. Because u _BT P~ stabilizes the pair (A, B), the equilibrium z = 0 is globally asymptotically
this implies ~(t) == O. We conclude that whenever stable and locally exponentially stable . In this
~(t) is bounded, the solution (~(t), z(t)) converges case, the norm IIzllb in the proof of Theorem 1
to the origin . must be replaced by a positive definite function
V(z), locally quadratic around the origin, which
Next suppose that ~(t) is unbounded for some locally satisfies V = L J V ~ -a V.
initial condition. Because W(t) = IIz(t)llbll~(t)lI~
is nonincreasing, this implies that z(t) ---> 0, so Assumption (6) can be relaxed to the stabilizabil-
that we can assume (for t large enough) that ity of the pair (A , B). This generalization requires
1'1 = 1. But then W(t) cannot converge to zero a different proof which will be presented in a
for, otherwise, we would have "12 == 1 for t large forthcoming publication (Sepulchre, et al., 1998) .
enough, which would imply that u = _BT P~ and
hence ~(t) ---> O. Hence W(t) must converge to
some positive constant W > 0, which implies, 3. OBSTACLES TO CONTROLLABILITY
from (17), that
The conditions of Theorem 1 are tight. In this
(>0 IIBT p~(t)1I2 section, we will show that they set a structural
10 lIe(t)lI~ dt < 00 .
(18)
limit to the possibility of achieving large regions of
· /() PHt)1I , ( 18) shows that /(t) E
attraction for the cascade (3). We will illustrate by
D efi mng t = IIBT
uw)llp two examples that relaxing any of the conditions
L 2 (0, 00), which, because j is bounded, implies in of Theorem 1 leads to situations in which the
particular that /(t) asymptotically converges to cascade (3) is no longer globally asymptotically
zero. controllable to the origin . (Global asymptotic con-
trollability to the origin is obviously a necessary
To end the proof, we show that /(t) ---> 0 leads
condition for semiglobal stabilization.) In each of
to a contradiction. Let P2 > 0 solution of the
Lyapunov equation the following examples, one condition of Theorem
1 is violated and we identify initial conditions
P2(A - BBT P) + (A - BBT pf P2 = -/ for which no open-loop control u(t) is able to
and define lIell~2 = er e.
P2 Then for t ~ 1, we
asymptotically drive the solution (z(t), e(t)) to the
origin.
have for some constant C 6 > 0
d Our first example is adapted from (Braslavsky
dt lIell~2 ~ -C6I1ell~, + 2(1 - ((t))e P2 BB P~
T
and Middelton, 1996) and illustrates a situation

600
of uncontrollability when the inequality (7) IS It is easily verified that z9 e=
0/ v is an invariant
reversed . manifold regardless of the choice of u because

Example 2. (Braslavsky and Middelton, 1996)


Consider the cascade
(24)
Z = -oz + z9+ 1 u 2
~ = v~ + u, z E JR, ~ E JR, u E JR Hence, initial conditions satisfying z9 e a/v
(19) cannot be controlled to the origin .

and assume that v - T


:= k > O. The change of
variables u = }. yields the equation
4. GLOBAL STABILIZATION OF PEAKING
u = qou - qu 2 CASCADES
which can be solved explicitely:
We now briefly return to the stabilization of the
u(t) = qe a9t (0) _ t e- a9 ·u 2(s)ds] . general cascade (1). Assuming a uniform relative
q lo degree r for the linear system, a preliminary
To avoid the finite escape time of z(t) , it is feedback puts the cascade in the normal form
necessary that u(t) > 0 for all t ~ 0, which (with new matrices A and B and new variable
Imposes u)
.z = J(z) + 1)(z,~o, y, iI, .. . , y(r-I))y
tOO e- a9'u 2(s)ds < lu(O)I. (20)
lo q ~o = A~o + By

On the other hand , the explicit solution


y(r) = u, z E IR! , ~o E JRn-mr, yE JRm , U E JRm
(25)
~(t) = e/lt[~(O) + lt e-/l'u(s)ds]
The peaking phenomenon (Sussmann and Koko-
tovic , 1991) imposes the restriction
imposes, if ~(t) -> 0, that
1)(z , ~o, y, iI, ... , y(r-l)) = q,(z,~o, y)
11 00

e-/l' u (s)dsI 2 = ~(0)2 in which case the cascade (25) becomes a relative
Using Holder 's inequality we obtain degree zero system of the form (3) controlled
through m chains of integrators oflength m . Using
~(0)2 = 110c> e-/l' u (s)dsI 2 the result of Theorem 1, a smooth control law
o( z , ~o) can be designed for the relative degree
00
zero subsystem, identifying the output y to a
= 11 e-be-T'u(s)dsI2 (21) virtual input. By backstepping this virtual control
law through m chains of integrators (Saberi, et
al., 1990) , a smooth stabilizing feedback is then
1 tOO a9 2 obtained for the overall system (25) .
~ 2k lo e- ·u (s)ds .
Standard backstepping requires, in addition to
Comparing (20) and (21) yields the constraint the stabilizing control law a( z, ~o) for the relati ve
degree zero subsystem, the knowledge of a pos-
(22) itive definite and radially unbounded Lyapunov
function V(~o, z) . Such a Lyapunov function is
We conclude that initial conditions which violate not provided by Theorem 1, which only ensures
(22) are uncontrollable to zero. the decay of the (semidefinite and not radially un-
bounded) function Wo(~o, z) = IIzllill~oll~ along
Our second example illustrates the necessity of the the solutions of the closed-loop system. Never-
condition 1)(z,~,O) = 0 in (8) of Theorem 1. theless, we show in (Sepulchre, et al., 1998) that
the knowledge of this function is sufficient for
Example 3. The cascade backstepping, yielding an explicit design for the
cascade (1). All the relative degree zero results of
~ = -oz - z9+1~P-IU Section 2 and 3 apply identically to relative degree
~ = v~ + u, z E JR, ~ E JR, u E R r cascades of the form (25), thus extending earlier
(23) stabilization results in the literature (Saberi, et
al., 1990); Teel, 1996; Sepulchre, et al., 1997).
with pv = oq , v -:/= 0, satisfies all the conditions of
Theorem 1 except that 1)(z,~,O) _zq+l~p-l # = The particular case of all zeros in the closed
O. left half-plane with possibly repeated zeros on

601
the imaginary axis is of interest: previous re- Sepulchre, R. , M. Arcak , A.R. Teel (1998). Trad-
sults do not require local exponential stabil- ing the stability of finite zeros for global stabi-
it y of the z-subsystem but the unstable states lization of nonlinear cascade sytems. In prepa-
of the {-subsystem are excluded from the in- ration
terconnection because of the slow peaking phe- Seron , M., J . Braslavsky, P. Kokotovic , D . Mayne
nomenon(Sepulchre, et al. , 1997) . Theorem 1 (1997) . Feedback limitations in Nonlinear Sys-
yields different conditions: the unstable states are tems: From Bode Integrals to Cheap Control.
no longer excluded from the interconnection but Proc. of 36th Conf. on Decision and Control
the z-subsystem must be locally exponentially (San Diego) , pp . 2067-272.
stable and the interconnection must be at least Sontag, E.D . (1995) . State-space and I/O stabil-
quadratic in z near z = 0 ( the inequality (7) is ity for nonlinear systems . In : Feedback Con-
t hen satisfied for any values of 0: , p, and q, because trol, Nonlinear Systems, and Complexity, B.A.
1I can always be selected in the interval (0, ~)). Francis and A.R. Tannenbaum eds., Lectures
Thus, a form of local input-to-state stability prop- Notes in Control and Information Sciences 202 ,
erty for the nonlinear subsystem is sufficient to Springer Verlag.
overcome the slow peaking phenomenon in the {- Sussmann , H.J ., P.V. Kokotovic (1991) . The peak-
subsystem and make the global stabilization pos- ing phenomen6n \il-nd the global stabilization of
sible . nonlinear systems:' IEEE Trans . on Automatic
Control, vol. 36, pp . 424-439.
Teel , A.R.(1995) Semiglobal stabilizability of lin-
5. ACKNOWLEDGMENT ear null controllable systems with input non-
linearities. IEEE Trans . on Autom. Control,
The authors acknowledge several discussions with Vol. 40, No. 1, pp . 96-100 .
.J . Braslavsky, P. Kokotovic , M. Seron , and A. Teel Teel ,A.R.(1996) . A nonlinear small gain theorem
during the preparation of the manuscript . This for the analysis of control systems with satura-
work was completed while the first author was at tion . IEEE Trans. on Automatic Control, vol.
CESAME, University of Louvain , Belgium . The 41 , no. 9.
paper presents research results of the Belgian Pro-
gramme on Interuniversity Poles of Attraction ,
initiated by the Belgian State, Prime Minister's
Office for Science, Technology and Culture. The
scientific responsibility rests with the author. The
work of the second author was supported by the
Air Force Office of Scientific Research under grant
F49620-95-1-0409.

References
Braslavsky, J .H., R.H. Middleton (1996) . Global
and semi-global stabilizability in certain cas-
cade nonlinear systems . IEEE Trans . on Au-
tomatic Control, vol. 41 , pp. 876-880 .
Lin, Z., A. Saberi (1993). Semi-globalstabilization
of partially linear composite systems via feed-
back of the state of the linear part. Systems &
Control Letters, vol. 20, pp . 199-207.
Middelton, R. (1991). Trade-offs in linear control
system design. Automatica, 27(2), pp. 282-292.
Saberi, A., P.V. Kokotovic , H.J. Sussmann (1990).
Global stabilization of partially linear compos-
ite systems. SIAM J. Control and Optimization,
vol. 28, pp. 1491-1503.
Sepulchre, R., M. Jankovic, P.V. Kokotovic
(1997) . Constructive Nonlinear Control,
Springer-Verlag.
Sepulchre, R . (1997). Slow peaking and low-gain
designs for global stabilization of nonlinear sys-
tems. Proc. of 36th Conf. on Decision and Con-
trol (San Diego) , pp . 3491-3496.

602

You might also like