Professional Documents
Culture Documents
On The Robust Stabilizability of Linear Time-Invariant Plants With Unstructured Uncertainty
On The Robust Stabilizability of Linear Time-Invariant Plants With Unstructured Uncertainty
On The Robust Stabilizability of Linear Time-Invariant Plants With Unstructured Uncertainty
3, MARCH 1987 20 1
Abstract-We consider general families of linear time-invariant plants See Nikol'skii [16] for details concerning these definitions.
described by a nominal plant model with unstructured uncertuinty. We
show, under a rather weak assumption, that when a family of systems is 11. INTRODUCTION
not robustly stabilizable by any linear time-invariant feedback controller,
then no nonlinear time-varying controller can robustly stabilize the given Consider the family F A of finite-dimensional linear time-
family. invariant (FDLTI) plants, arising from an unstructured modeling
uncertainty:
I. NOTATION F A (respectively, F.4): = { P ( s ) : I P(j w ) -Po(j w ) l 5
R = Field of real numbers, and 8 : = field
1 W ( j w ) ( . where Po(s) is a givenstrictlyproper nominal
of complex numbers. plant model, thenumberofpolesof P ( s ) inthe right-half
planeisequalto (respectively, lessthan or equalto)the
Fl = Set of positive real numbers.
number of poles of P&) in the right-half plane, P ( s ) has no
RHP (open right-
poles on the imaginary axis: and W is a weightingfunction
half plane) = {s in C:Re (s) > 01.
( >' = "Transpose of."
'
such that Wand W- are both stable and proper (i.e., W is a
unit in RH=(j?l))]. (2.1)
Ld W
j = Hilbert space of square integrable hnctions
on j 2 . In this paper we prove, under a very weak assumption, that if
Hz(j3) = (Hardy) space of square integrable func- there does not exist an LTI controller that robustly stabilizes the
tions on J g i which admit an analytic exten- family F d (respectively, F4),then no nonlinear time-varying
sion to the RHP. (NLTV) controller can robustly stabilize F A (respectively, F A ) .
If K is a subspace of Lz( j R ) , We also prove a similar result for the families of plants F,w (Fb,)
KL = Orthogonal complement of K in L*( j s ) . characterized by multiplicative uncertainty, defined as follows.
( 1- = Projection from L2(j 2 ) onto H I (j;i). F,w (respectively. & ) = ( P ( s ) : I P ( j w ) - P , ( j w ) ( 5
L,( j R ) = Banach space of essentially bounded func- 1 W(jw)P,( j w ) l . where Po(s)is a strictly proper nominal
tions on j 2 , planr model, thenumber of poles of P ( s ) in theright-half
H,( j2) = Banach 'pace Of bounded equal
toplane
is (respectively.
than
less or equal to) the
O n j P 1 which admit an number of poles of P,(s) in the right-half
plane, P ( 5 ) has
no
sion to the RHP. poles
the
on
imaginary axis? and W is a unit in
RH,( j B ) = Space of rations1 functions which are RH,( j i j ) l . (2.2)
bounded o n j 2 and analytic in the RHP.
These results should be considered as a natural continuation of
If P is in L,(j?l), the earlier results obtained by Khargonekar et al. [l 11 and
I'(P) = Hankel 'perator mapping H2(ji2) into Khargonekar and Poolla [12]. In [ 111, the authors have shown that
H: ( j " ) . defined by r ( p ) x = ( p x ) - for for a class of robust control problems with structured uncertainty
all x in Hz(
j;i). (such as gain margin problems), time-varying controllers are
!![GI = nonzero singu1ar value Of an 'per- vastly superior over LTI ones. In contrast, Khargonekar and
ator G Poolla [I21 showed that for problems of uniformly (or H , -)
= (L~(G*G))"~ h m~
l n being the smallest
optimal control of LTI plants, NLTV controllers offer no
and *' the adjoint Of
nonzero
exists.)
'
.' (If is Of finite then d G ]
advantage over LTI ones. (See also [6] and [ll].)
The principal motivation for this research stems from the need
to understand the intrinsic limitations of feedback control.
Clearly, these results and their implications have a direct bearing
Manuscript received February 17, 1986: revised August 15, 1986. Paper on the subject of robust adaptive control. (see Remark (4.14) for
recommended by Associate Editor. R . Sivan. This work was supported in part Some comments in this direction.) It is hoped that further work
by the National Science Foundation under Grant ECS-8451519, and in part by along these lines will clarify the relevant issues. particular, the
grants from the Honeywzell Corporation. the 3M Company, and the MEIS
Center at the University of Minnesota. performance of NLTV controllers for the case of various types of
P. P. Khargonekar and A. M. Pascoal are with the Department of Electrical simukaneous structured and unstructured uncertainty needs to be
Engineering and the Center for Control Science and Dynamical Systems, addressed.
University of Minnesota. Minneapolis, MN 55455. The robust stabilization problem for LTI plants using LTI
T. T. Georgiou was with the Department of Electrical Engineering and the controllers has been considered by several researchers in recent
Center for Control Science and Dynamical Systems. University of Minnesota,
Minneapolis, MN 55155. He is now with Iowa State University. Ames, IA years. (See. for example, [20], [5]. [41? [14], [lo], and the
50010. references therein.) Kimura [14] gave necessary and sufficient
IEEE Log Number 8612620. conditions for the existence of robustly stabilizing LTI controllers
for the family FA in (2.1) of single-input single-output (SISO) projection operator ST,defined, for r > 0. by
plants. (For the multiinput multioutput case, see [24].) Adifferent
approach, that provides alternative conditions for robust stabiliza-
bility of the family F 4 , was recently developed by Glover [ 9 ] ,
Verma [ 2 I], and Verma et a/. [22]. It should be noted that robust
stabilization for LTI plants with LTI controllers can be formulated Let L2[Ol m; ?/) denote the subspace of norm square integrable
as an optimization problem in the Ha-norm. Thus. in this setting, functions in \Iif+.and define the extended space
robust stabilization is closely related to the Ha-optimal control
problem which was pioneered by Zames [27], Zames and Francis Lz,p[O.m; Y ) : = { f : f i s in Y i t ~ ,
[28], Francis et a/. [7], and Chang and Pearson [2]. However, if and srf belongs to Lz[O,m; Y ) for all r in ;l+ }.
one considers NLTV controllers for the robust stabilization of
families of plants such as in (2.1). (2.2), the equivalence between Any system G is thought of as an operator G:L2,,[0,03; ‘Y) -+
the H,-optimization problem and the robust stabilization problem L:.,[O, 03;?I).A causal system is simply an operator G:L2,JO,
does not hold. (See the recent survey paper by Francis and Doyle 03;V) Ll,,[O, 03; -\I)such that ~ , G s T=, T,G for all 7 in $IA. In
+
[SI for a fairly complete bibliography on robust and Ha-optimal 03;, Y) into Lz.,. The
the sequel we compress the notation L z , ~ [ O
control.) space .Y will usually be identified with $3 or it will be clear from
Recently, Poolla and Ting [ 171 have obtained some very the context. A causal system G is said to be (jinite gain) stable iff
interesting results for the problem of robust stabilization using
NLTV controllers. In their work, they considered families of n,f#O, in A,}
IIGII := sup {Ilr7Gf!lr/llr,f112: f i n Lz,,,
plants described by a nonlinear but stable perturbation of a
nominal plant model: is finite. In fact. if 11 GI1 is finite. then G is a bounded operator
from LJO, a)into itself, and 11 G 11 is the induced operator norm.
F , v ~= { P : P = Po + W A , where Po is an LTI nominal plant This definition of stability agrees withtheusual notion of
model, W is a stable LTI operatorwith a stable inverse,and A stability for linear time-invariant (LTI) systems. An LTI system P
is a possibly NLTV but stable perturbation with 11 All I1). is usually described by means of itstransfer function matrix P(s).
andprovedthat if F t r L can be robustly stabilized by a general In this case. P is stable if and only if P ( s ) belongs to H,( j??),
NLTV controller, then it can also be robustly stabilized using an and
LTI controller. Note that the family FA [in (2. l)] that we consider 11 PI1 =sup { 11 P(s)ll : s in the RHP}.
in this paper can be equivalently described in an analogous form:
In this setup, the set of all stable systems with the same input
+ TV(s)A(s), where P ( s )has the
FA : = { P ( s ) : P ( s )= Po(s)
and output spaces is a Banach algebra, where multiplication
same number of poles in the right-half plane as Po(s),
corresponds to composition of operators. This algebra is noncom-
W(s),and W(s)-’are both stable and proper. and A is mutative and it is not right distributive. (See. for example. [25].)
an
LTI
(possibly unstable)
transfer
function in In this paper. we consider feedback interconnections of systems
L = ( j $ i ) ,with llA!l,s I}. such as the one shown in Fig. 1 . The feedback interconnection of
Strictly speaking, the family F4 which we consider. and the theplant P and the controller K is defined by means of the
family F,vL considered by Poolla and Ting, are incomparable. following functional equations:
Hence, it is not easy to make a direct comparison of our results
with those in [ 171. It is clear that the family FA is more structured yI=Pel,y2=Ke2, el=ul-yz.
and e 2 = u z + y I .
than F:vL. Also, FA appears to be a more natural family, since all
plants in this family are linear and time-invariant. (See Remark The feedback system is we/[ defined when the above system of
(3.15) inPoolla and Ting [I71 onthe difficulties involved in equations has a unique solution in Lz,pX L:,, for all inputs (ul,
extending their results when A is linear time-invariant.) It should u2)’in L2,?X Lz.,. If the feedback system is well defined, denote
be noted that in FA we allow the locations of the unstable poles by G [ P , K ] the operator that maps the input pair (u,,N?) ’ to the
of theplunt to vary, whereas in FSr the unstable poles are fixed. (error) output pair ( e l , e 2 )’. The feedback system is called
(See also Section VI.) Finally, the methods of Poollaand Ting internal/y stable iff G [ P , K ] is stable. In this paper. we w i l l not
[I71 and our techniques bear no resemblance to each other. discuss well-posedness of the feedback systems i n detail. We note
In this paper, we only consider the continuous-time case. It is that if the plant P is strictly proper LTI and the controller is
very easy to check that all our results (with trivial modifications) locally Lipschitz continuous. thenthe feedback system iswell
extend to the discrete-time case. posed. (See [25, Corollary (4.1.2)].) In all our constructions. the
The organization of the paper is as follows. In Section I11 we plant turns out to be strictly proper LTI. Hence. the assumption
give a brief exposition of some background material. In Section that the controller is locally Lipschitz continuous guarantees well-
IV we present our main results for the case of plants with additive posedness of the feedback systems considered in this paper.
uncertainty, and in Section V these results are extended to the case Let P be an LTI plant. andlet K denote a nonlinear time-
of multiplicative uncertainty. Finally, we conclude with a discus- varying (NLTV) controller. In this case, G [ P , K] is given by a
sion of some directions for future research. fairly simple expression
+
and note that 11(1 KPJ'II = ~ ~ h l l ( P5,.)M,. ~ ~ By hypothesis,
h l l ( P )is bounded for all P in FA. This lmplles (for instance, see
[23]) that
6 : = )1(1+KP,)-1KWII < 1.
Hence, in this case 11(1 + (1 + KP,)-IK WA)-'ll 5 (1 - 6)-l,
and ~ ~ h l i (IP ()l -~6~) - i M , for all P i n FA. E
IV. MAINRF,SULTS
Our first main result is the following.
Fig. 1 Theorem 4.1: Consider the family of (SISO) FDLTI continu-
ous-time plants
K is an internally stabilizing controller for P if an only if
FA := { P ( s ): I P ( j w ) - P , ( j o ) l I I W ( j w ) l , where Po(s)
K=(KO+MQ)(I-NQ)-' (3.1-a) is a given strictly proper nominal plant model, the number of
poles of P ( s )in the right-half planeis equal to the numbern of
where Q is a stable (possibly NLTV) system. Moreover, for each poles of Po(s)in the right-half plane,P ( s ) has no poles on the
stable Q , irnaginav axis: and W is a unitin RH,( j R ) ) } , (2.1)
[ P l ] R = [ P 1 11. and
that R' [A] [;]= . (4.6)
all E in (0,E , ] . In particular, the controller K2 stabilizes internally
the LTI plant Pi,Efor E = E , . It follows from Theorem 3.1 that K2
admits a (right) coprime representation
Internal stability of the feedback interconnection of P and K , is
K2 = ab-'
equivalent to stability of the mapping betweenthe inputs (ul, u2)' ,
and the (error) outputs ( e l ,e2)' , given by where a, b are stable operators satisfying a (right) Bezout identity
xa+yb= I (4.8)
and x , y are also stable operators. Using this representation we
Similarly, internal stability of the feedback interconnection of P I show that b l l , and consequently K2, are both stable.
and K2 is equivalent to the stability of the mapping between the Since G[P,, K2]is uniformly stable for E in (0,E , ] , it follows
inputs (u3, u4)' and the outputs (e3, e4)' (with uI = u2 = o), that the operators
given by
r
( l + ~ c K z ) - L = ( T-'n,ab-')-'
l+
=b[Tb+n,~]-'T (4.9)
and
Suppose G [ P , K is
] uniformly stable. We first note that
K2(l + P & - ' = . b - ' ( l + T-'niab-l)-'
Q := K ( l + P K ) - ' = K [ l + ( P , + P z ) K ] - '
=a(Tb+nla)-'T (4.10)
= K ( 1 +P2K)-'(1 +P,K(1 +P2K)-')-'
are uniformly stable. Multiplying the above expressions (4.9),
=Kz(l +P'K&'. (4.10) on the left, respectively, by y and x and using (4.8),we
deduce that
NOWconsider R # G [ P , K ] Rwhich is uniformly stable. since it is
a composition of uniformly stable operators. By straightforward
is stable. Define
C, : = Tb+nla,
and let T* denote the adjoint of T. Since T i s an isometry, T* is a
bounded but not causal operator, and T*T = I [where I is the
identity on Lz[O,w)]. It readily follows that
T*C,B, = T*T=I, and
number of poles ofP(s) in the right-half planeis equal W(s)-’ are both stable and proper, and A is a stable
to the number of poles ofP J s ) in the right-half plane, LTI transfer function in H - ( j R ) , with I(A / I rn 5 1 ) .
P(s) has no poles on the imaginary axis, and W is a
unit in R H , ( j l ) f . Clearly, FA contains FAS. Also, F& contains FAs. Now, it is a
fact that if FAs is robustly stabilizable by an LTI controller, then
Let n,(s) be an outer function in Ha(j 8 ) such that \Po(jw)( = the same controller stabilizes robustly both FA and FKL.We
lno(ju)l a.e. on j i l . (Then P,(s)n,(s)-’ isbounded on j;J.) conjecture that a result analogous to Theorem 4.1 can be obtained
Then, there exists an LTI controller that uniformly stabilizes Fw if for the family F 4 s . If this is obtained, then it would contain both
and only if rhe results of Poolla and Ting (17j and the results in this
paper.
a[r(P,(s)n,(s)-lW(s)-L)]>1. (5.2) iii) Further investigation is needed for the case of families of
plants with block diagonal structural uncertainty in the sense of
The above theorem follows readily from the work of Glover [9], Doyle [4]> as well as with parameter variations that arise in
Verma [21], and Verma et ai. [22]. Also, an analogous result adaptive control. The objective is to determine possible advan-
holds for F + f (defined in Theorem 5.5 below). Our main results in tages of NLTV controllers over LTIones. Finally, for the
this case are the following theorems. problem of robust performance of a set of plants (see [4]),similar
Theorem 5.3: Consider the family of (SISO) FDLTI continu- investigations should prove quite interesting.
ous-time plants FM.described in Theorem 5.1. If
REFERENCES
9 : = g[r(P,(s)n,(s)-l W ( S ) - I ) I < 1, (5.4) V. M. Adamjan, D. Z. Arov, and M . G. Krein, “Analytic properties of
Schmidt pairs for a Hankel operator and the generalized Schur-Takagi
then there exits no NLTV controller that uniformly stabilizes the problem,” Math. USSR Sbornik, vol. 15, no. 1. pp. 31-73, 1971.
family F M . B. C. Chang andJ. B. Pearson, “Optimal disturbance reduction in
Theorem 5.5: Consider the family of (SISO) FDLTI continu- linear multivariable systems,” IEEE Trans. Automat. Contr., vol.
AC-29. pp. 880-887. Oct. 1984.
ous-time plants R. F. Curtain andK.Glover, “Robust stabilizatjon of infinite
dimensional systems by finite dimensional controllers, Syst. Contr.
$,w:= { P ( s ) := ~ P ( j w ) - P , ~ w ) ( ~ ( W ( j w ) P , ( j wwhere
)~, Lerr., vol. 7, pp. 41-47. Feb. 1986.
Po@)is a strictly proper nominal plant model, the number J. C. Doyle, “Synthesis of robust controllers and filters.” in Proc.
of poles of P ( s ) in the right-half plane is less than or equal 22ndIEEE Conf. Decision Contr., San Antonio, TX:Dec. 1983, pp.
109-1 14.
to the number of poles of P,(s) in the right-half plane, J. C. Doyle and G. Stein, “Multivariable feedback design: Concepts
P(s)has no poles on the imaginary axis, andWis a unit in for a classicalimodern synthesis,” IEEE Trans. Automat. Contr.,
RHa(jFl)]. vol. AC-26, pp. 4-16, Feb. 1981.
A. Feintuch and B. A. Francis, “Uniformly optimal control of linear
Let n,(s) be an outer function in H , ( j 2 ) such that \Po(ju)( = time-varying systems,” Automutica, vol. 21, pp. 563-574, Sept.
1985.
In,( j w ) l a.e. on j R . If B. A. Francis, J . W . Helton, and G . Zames, “H”-optimal feedback
controllers for linear multivariable systems,” IEEE Trans. Automat.
:= _o[r(P,(s)n,(s)-lW(s)-l)]<l, Contr., vol. AC-29. pp. 888-900, Oct. 1984.
B. A. Francis and J. C. Doyle, “Linear control theory with an H ,
then there does not exist any NLTV controller K that pointwise optimality criterion,” preprint. Oct. 1985.
K. Glover, “Robust stabilization of linear multivariable systems:
stabilizes Fw. Relations to approximation,” Inr. J. Contr., vol. 43, no. 3, pp. 741-
The proofs of the above theorems are quite similar to the proofs 766, 1986.
of Theorems 4.1 and 4.13, and will therefore be omitted. P. P. Khargonekar and A. Tannenbaum, “Noneuclidean metrics and
Remark 5.6: Note that Theorem 5.1 can be generalized to the robust stabilization of systems with parameter uncertainty,” IEEE
Tram. Automat. Contr., vol. AC-30. pp. 1005-1013, Oct. 1985.
multiinput. multioutput systems. We also point out thatusing P. P. Khargonekar, K. Poolla, and A. Tannenbaum, “Robust control
Theorem 5.1, one can obtain alternative necessary and sufficient of linear time-invariant plants using periodic compensation,” IEEE
conditions for robust stabilization of families of distributed Trans. Automat. Contr., vol. AC-30, pp, 1088-1096. Nov. 1985.
systems described by multiplicarive uncertainty (see [13]. and P. P. Khargonekar and K. Poolla. “Uniformly optimal control of linear
time-invariant plants: Nonlinear time-varying controllers,” Syst.
PI). Contr. Lett., vol. 6 , pp. 303-308, Jan. 1986.
Remark 5.7: It is easy to obtain results analogous to Theorems .- . “Robust stabilization of distributed systems.” Automaticu, vol.
4.1, 4.13, 5.3, and 5.5 for additive and multiplicative families of 22. pp. 77-84, Jan.1986.
plants (similar to FA,F,w, F A , and FM),in which the uncertainty
H. Kimura, “Robust stabilization for a class of transfer functions,”
IEEE Trans. Automat.Contr., vol. AC-29, pp. 788-793. Sept.
A lies in the open unit ball in L,( jR). Since these extensions are 1984.
straightforward, we will not state them explicitly. B Mirtensson, “The order of any stabilizing regular is sufficient a
Driori information for adaDtive stabilization.”.~ Svst. Contr. Lett.. vol.
VI. COKCLUSIONS b. pp. 87-91, July 1985.’
N. K. Nikol’skii, Treatise on the Shift .Operator.
. New York:
In this paper we have shown, under a rather weak assumption, Springer-Verlag, 1986.
K. Poolla and T. Ting, “Nonlinear time-varying controllers for robust
thatwhen a family of plants as in (2.1), (2.2) is not robustly stabilization.” Dep.Elec.Eng.,Univ. Illinois. Urbana,Jan. 1986;
stabilizable by any linear time-invariant feedback controller, then also in this issue. pp. 195-200:
no nonlinear time-varying controller can robustly stabilize the 1. W . Sandberg,“On the L2-boundedness of solutions of nonlinear
family. Thus, for these robust stabilization problems, NLTV functional equations,” Bell Syst. Tech. J., vol. 43. pp. 1601-1608,
controllers offer no advantage over LTI ones. 1965.
S. Takeda and A. R. Bergen, ”Instability of feedback systems by
We conclude this paper with the following open problems. orthogonal decomposition of Lz,” IEEE Truns. Auromar. Confr.,
i) We have assumed in Theorems 4.1, 4.13, 5.3, and 5.5 that vol. AC-18, pp. 631-636. Dec. 1973.
< 1. Of course even if u = 1, there does not exist an LTI A. Tannenbaum. “Feedback stabilization of plants with uncertainty in
the gain factor.” Inr. J. Contr., vol. 32, pp. 1-16, July 1980.
controller that robustly stabilizes the family of plants. We M. S . Verma, ”Synthesis of infinity norm optimal linear feedback
conjecture that even in this case there does not exist a robustly systems,” Ph.D. dissertation, Dep. Elec. Eng.. Univ. Southern
stabilizing NLTV controller. However, new techniques are California, 1985.
needed to approach this problem. M . S . Verma, J. W . Helton, and E. A. Jonckheere. “Robust
ii) A very interesting family of plants is obtained by considering stabilization of a family of plants with vaqing number of right half
plane poles.” in Proc. Amer. Contr. Conf., Seattle. WA. June 1986.
M. Vidyasagar. Control Systems Synthesis: A Factorization Ap-
proach. Cambridge. MA: M.I.T. Press, 1984.
KHARGONEKAR etSTABILIZATION
ai.: ROBUST TIME-INVARIANT
PLANTS
OF LINEAR 207
[24] M. VidyasagarKimura,
H.
and“Robust
controllers
for uncertain Tryphon T. Georgiou (S’79-”83) was born in
linear multivariable systems,”
Aufomatica, vol. 85-94,
pp.
22,Jan.
Athens,
Greece,
October
on 18, 1956. He received
1986. the
Diploma
mechanical
in and
electrical
[25] J. C.Willems,TheAnalysis of FeedbackSystems.Cambridge, engineering from the National Technical University
MA: M.I.T. Press, 1971. of Athens, Athens, Greece, in 1979, and the Ph.D.
[26] G. Zames,“Functionalanalysis applied to nonlinear feedback sys-
terns,” ZEEE Trans.CircuitTheory, vol. 10, pp. 392-402, Sept. degree from the University of Florida, Gainesville,
1963. in 1983.
[27] G. Zames,“Feedback and optimal sensitivity: Model reference He was with Florida Atlantic University, Boca
transformations, multiplicative seminorms, and
approximate in- Raton, until 1985,and with the University of
verses,” IEEE Trans. Automat. Contr., vol. AC-26, pp. 301-320, Minnesota during the academicyear 1985-1986.
Apr. 1981. Since August 1986 he has been with Iowa State
[281 G. Zames and €3. A.Francis, ‘‘Feedback minimax sensitivity and University, Ames,as an Assistant Professor of Electrical and Computer
Optimal robustness,” IEEE Trans. Automat. Contr.* vOl. AC-28, Engineering. Hisresearchinterests lie in theareas of control and system
pp. 585-601, May 1983.
theory, signal processing, and applied mathematics.