This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Select desired terminal condition changes dt/l., d.p2, ....d"';p to bring the next solution closer to the specified values t = 0 than were achieved by the nominal path. (g) Select a reasonable value of (dP)'/(T - to), which is a mean-square deviation of the control variable programs from the nominal to the next step. (h) Use the values of dq and dP to calculate (dP)t. d~'IH -ld,p; if this quantity is negative, automatically scale down dl/l to make this quantity vanish. If the quantity is positive, leave it as is. (i) Using the values of dP and dtk (modified by (h) if necessary) calculate &a(') from equation (50), (&x(lo) :IS 0). (j) If final timet T, is not specified or being extremalized, compute the predicted change, dT, for the next step: 0 on U when cp :;t: 0 and V(O) = 0'" and at the end " ... intersect the boundary of C; ...". This is clear from his proof and is necessary, since he wanted to generalize the usual statement of Cetaev's theorem to include the possibility that the equilibrium point be inside U as well as on its boundary. Corollary 6. Let p 0, the region G = {q;; V(qJ) < O} is nonempty, and no trajectory starting in G can have lp = 0 as a positive limit point nor can it leave G. Hence by Theorem 5, each trajectory starting in G must be unbounded. Since qJ = 0 is a boundary point of G, it is unstable. It is also easily seen [J] that if a < 0 and I b 1< I a I, then cp = 0 is asymptotically stable in the large. In [1] Hale has also extended this theory for systems with infinite lag (r = 00), and in that same paper gives a number of significant examples of the application of this theory. (M) consisting of .Notice also that observability is a global concept; it gl'· .. ,gn' and by § the smallest linear subspace of rmght be necessary to travel a considerable distance or for C~(M) containing go which is closed with respect to Lie a long time to distinguish between points of M. Therefore differentiation by elements of ~ 0. An element of § is a we introduce a local concept which is stronger than ob- finite linear combination of functions of the form servability. Let U be a subset of M and xo,x lEU. We say 1 O X is Uslndistinguishable from xl (xOlux ) if for every control (U(/),[tO,/I]), whose trajectories (xo(t),[tO,t 1]) and where fj(x) = f(x,u j ) for some constant u' EQ. (x1(t),[tO,t 1) from XO and Xl both lie in U, fails to If h1,h2EGX(M) and cpECCXl(M) then distinguish between XO and Xl, i.e., if xO(t)E U and x1(t)E l U for tE[tO,t ] , then L (L (cp)) - L (L (q») = L h ] ( cp). hl ' 0 *JT = ( t,x),p( x), v( I,A,e») restrictive than the usually given condition J)- Q(t,x2,q>~1 < X1(x,q>,p, v){lx l ables (not necessarily stationary or with zero means). X2) + I'PI - 9>21} for Xi E ~ (x,p) for some p= p(x) >0 where B.2: E le(t)I' exists and is bounded in I for each p > I. xEDR ; q>iE~(tp,v), v;>O. B.3.· The function Q(I,X,q» is Lipschitz continuous in C.2: The matrix functions A ( .) and B ( .) are Lipschitz x and tp: IQ(t,xl,cpl)- Q(/,X 2,q>2)1< :JC.(x,CP,p,v){lx l - continuous in DR' x 21+ 1'1- CJ>21} for x, E ~ (x,p) for some p== p(x) >0 where C.3.· Z(/,X) as defined by (18a) converges for all ie DR xEDR ; cp;E~('P,v), v>O. as 1--+00. Denote the limit by f(X). B.4.· l%l(x,CPI,p,vl)~ ~(X,CP2~P,V2)1 < ~(x,cp,p,v, w) C.4: ko(/,x,A,C) defined by (18b) converges to a finite ·{lfP.- 9>21 + Iv.- v21} for ~;E~ (tp, w) and v;E~(v.,w). limit as 1-+00 for all xEDR , ~< 1 and c j(x A+x B)+ 8 and QB analogously
~
{}
where ( · ) means
~
cos 'Y -V
9
(58)
V sin 'Y
(59)
= V cos 'Y
(60)
m = m(h, M)
(61)
:t
The computing procedures evolved over a period of time in solving many types of problems are summarized here. The computer used for all these problems was an IBM 704.
mV
h=
1 Compaling Procedures
dT = _
+ F(A, M) • - - - S I D at -
( ), and
F
=
F(h, M) is thrust, given &8 a tabular function, Fig. 4(4)
D
=
pV 28 C»(at, ill) - - is drag 2
L ". CL(a, M)
P~'S is lift
Cd,a., M}, CL(a, l{) are given as tabular functions, Fig.. 4(b) p
p(h.) is air density, given as a tabular function
g
acceleration due to gravity (taken as
CODSta.Dt
here)
V - is Mach number a
Jf
ALTITUDE
Lin
ZERO-UFTAX'S YELOCITY.V
DRAG WEIGHT
x~
h
r T In'GOadl
J~
)wr."""'''J''''P''f''''''''''''''''''''''''''''''''''''''~''''''''''''P7''I''......,.~,.....HORIZONTAL DISTANCE
If IdTI is greater than a preselected maximum allowable value, scale down &a(t) to achieve this maximum value.
Fig_ 3
225
Nomenclature used in analysis of sup......nlc interceptor
d+ - # - (( ~ ).
+ ~'F )
d.
+ ~'Gd« + ~.'dfo,
(98)
(83) where
where the nomenclature is explained in Section 4 of the peper, Now let us ehOOle ~' so that the eoefficientof dx vanishes; Le.,
(~). + ~'F" 0,
1.. ==
or (85)
dq
This reduces equation (83) to the expression
(86)
An exactly similar procedure yields
d'k
== ~'Gdcr
+ ~'dfo,
(87)
where
~'G ~.p'G !:2
i.e.,
-=
V~da
t-A.c..t M....... in Orcllftary Calculu. Us_... a WeJ.htecl 1... of the Control Varia"'.. to D......ln. Step Slz••
The problem, 88 stated in Section 4 of the text, is to choose du maximize dtP for given values of d~1 diG (usually zero), and tlP, where 80 88 to
dq, -=
~/GdQ
d~
~'Gda.
==
(dP)'
cz
+ ~'dfo + 'J..~/df.
-
~'Gdcr
+ ~'dfQ + Y'(d~ - l.,,'dfo - '-p'Gda.) + P«dP)1 - da'Wd.)
where v' is a 1 X fJ row matrix of eoDstants, and p is a constant, all to be determined for our convenience. Note that the quantities multiplying .' aad #j are both zero, by equations (90) and (91). Take the second differential of this quautity:
d't/J
:=
(~'G
- .''J..",'G - 2lUlcr/W)d t a
(93)
Thus the maximum of ~ occurs when the coefficient of ilia vanishes in equation (93). This will be the case if
da.
a&
1 2~ W-J(G/~ - G~.)
(94)
=:
IW-'hdal.
c14J)Z == (V~W_l/J)(W-'''V4J') ( dP 'P
= VcPW-1VlIIt. -
_ [(VepW-l/t)(W_I/tV~')]' (Vl/IW-1/t)(W_1/tVt/!')
(VtPW-IVI/I')f
(102)
VJ/lW-1V~'
."
:::: I.et» -
(92)
I~'I#-J'i4
which gives a geometric interpretation of equation (24). Clearly this quantity is positive unless Vt/J is parallel to in which case it iSlero.
V"'.
3 Ad;oi"' D ntia' Equations Given Control V.rIo "..,.",••
for Small Perturbation, About
For the optimum programming problem stated in Section 6 of the text, we wish to determine the changeJ!l in ., ~ and for small perturbatioDS, 8cr(t), in the control variable programs about given Dominal eontroJ variable programs ••( t), where n -= 0 is the stopping condition. To do this we consider the linear equations (39) describing small perturbations about the nominal path: d (103) dI. (elx) == F(t)6x + G(t)8cr
n
To determine the changes in tj), ~I and 11 we introduce the linear differential equations adjoint to (103) defined 8S
Substituting (94) into (90) we have
d:>-
1 d~ ~ (IH - I"'I/f~) 2p.
-:- = - F'(t)l at
(95)
wbere
(101)
Hence we must subtract the component of V¢W- 1 / , that is parallel to V.pW-I/s from Vq,W- l /a. Using the Pythagorean theorem we have simply then
(91)
We use the process of Seetion 1 of the Appendix again; Le., we ~
dP
(90)
.coDlider a linear combination of the three foregoing equa.tions:
(Vv.W-I/')W'ltdo.
V1k is the gradient of 'l!. For the moment we will consider q to be a lingle scalar quantity rather than a column matrix. If W is the metric in the a-hyperspace then dP is the infinitesimal distance from the present nominal point to a neighboring point a + d«. We wish to choose d« to maximize dq" keeping d1/J = 0, for a given value of
(89)
da.'Wda.
=a
ezVq, is the gradient of tP in the a-hyperspace, and
(88) 2 St Sctu. . '
(99)
Substituting (97) and (98) iDUJ (94) and (89) we obtain the results given in SectiODS 4 of the text in equations (21) and (23), respectively. These results have a simple geometric interpretation in the crhyperspace. Equations (89) and (90) with clio = 0 may be written 1 dtj) - V
(84)
dq, == :>-.. 'Gdcr + '-.,/ til.
~et»'GW-1G'~
( 104)
where ~. is an n X 1 matrix, 4J, is an n X p matrix, and A.o is an X 1 matrix. If (103) is premultiplied by).' and (104) is premultiplied by (8x)' and the transpose of the eeeond product is added to the first product, we have ft
d~ I~
=-
~/dI. J.p/GW-iG/~
IH ==
~/GW-l(;').#
:::I
dq -
(96)
~I
Solving equatioD (95) for Y we obtain " -= -2.uleJ",-ld~
+ IH-ll~.
SubltitutinR '-97) aud (94.1 into (jill and 101YiD2 for
(97)
d(8x) + rD.' ~x == l'F8x ell
tit
~'Fax + ~'G6u
or d - ().'6x) de
u. we obtain 230
= ).'GSu
(105)
11 we integrate equatioDl (105) 'rom It to T, the result is
Subetitutinl this into equations (110) and (111), we have the re-
latioDl
().'h)1-2' - f..1').'GBIfIlt + ().'Ix),_••
(106)
dt/I -
II we let
~'(T) - (~).
;
,- T
()X
~'(T)
...
(!t). , '-T
()X
'-o'(T) ..
(=r
dtl' -
8t-
(~'aX)"'T;
an ..
d4J -
(~'8.)...2' (108)
+ tkdT
II-A.cent M
(~/(Io) -
n. U
W••hW
D
SI•••
8ex
J"
dt -
1:
~/G8crdl + ~'(I0)&.(Ie)
'1'
t.
~'G8crdt
.
+
+ ~/(1o)8x(lo) + +IT
o ... dO ... /,.7' 19'G6..u + '-0'(10)6.(10) + OtlT (tIP)1 ...
f,.T 8«'WBad,
4)
...p. (110)
a
-
liT D "
-.
19'G8adt - - 1 10'(10)6.(10)
n
+ Y'd~ + lA(dP)1
(117)
v'~'G - 2~e'W)6lcrd~
(118)
1
&II
21£ W-I G'(~ -~,,)
=-
:!:
[1•• - IH'I~.,,-JI~ ] (dP)1 - d~/IH-Id~
(119) CaD
solve (120) (121)
where (111)
dll - d1i -
(112)
.## -
(114)
~D/(,.)&X(,.)
/..'1' ~'GW-JG~ (122)
(113)
II> ... f..1'~'GW-JG'~1
We use the analog of the process of Section 2 of the Appendix with only small differences arising from the additional term in tiT in the foregoiDl equatioDl. The first step is to elimiDate dT from equation (112): dT
y'l~'(Ie»8x(,.)
" - -2pIH- l dO+ I~~-JI~.
'-
dt; =
p&e'W)lcrdt
Subltitutinl (119) back into (116) and then (113) we for the CODltaDta " and 1':
The problem 88 stated iD Section 6 of the text is to choose 8u(t) as to maximise d4J for liven values of d.J (usually zero). and dP, with dO -= 0 where
(T
v'~o'G -
mum
ln Calcvlu... y 01 Centrol V• . . ,
10
-
where 1.(Ie), d~, and dP are coDlidered CODltaDta. The maxi~ will occur if the coef&eieDt of 'III in the integrand of (118) vauiaheaj i.e.,
(109)
where the nomenclature is given in equatioDS (48). SuhetitutiDl (107) and (109) into (106) we obtain equatioDl (42), (43), and (44) of the text. 4
(,.,'1' (~'G
J, -+
6(d4J) ... f..1' (~'G -
dn=80+DdT
M
(116)
Now we take the variatioD of this relation (117):
For unall perturbatioDS the value of T will be changed only a small amount d'l't 10 that ~:: 8.+~T d~ - 8q
(115)
where the nomenclature is pveD iD equatioDS (51). Next we con. eider & linear combiDatioD of (116) and (113) with (1115); i.e.,
(107)
where the nomenclature is given in equatioDB (47), it is clear that , . - (~'lx),-2';
f..T ~'G81f1lt + ).,.o'e4)8.(4) j',.' ~'G8" + ~0'(to)8.l")
I .. -
f..T ~o'GW-JG'~ndt
SubstitutinC (120) and (121) into (119) and (115) we obtain the relU1ta pven in Section 6 of the text in equatioDl (SO) and (51), respectively.
231
The Solution of Certain Matrix Inequalities in Automatic Control Theory V. A. YAKUBOVICH
IN
1961-62, Vladimir Yakubovich was well prepared to respond to the 1960-61 breakthrough of Vasile-Mihai Popov [14] with a breakthrough of his own. As a prominent young member of the Leningrad School, Yakubovich was investigating the solvability of Lurie's "resolving equations," that is, the problem of existence of Lyapunov functions consisting of a quadratic form plus the integral of a single sector nonlinearity. The nonlinearity was in feedback around the linear part of the closedloop system. Motivated by actuator nonlinearities, this problem was formulated in 1944 by Lurie and Postnikov [11] and remained an open problem until it was solved by Popov in [14]. Popov's solution was a frequency domain condition that restricted the Nyquist plot of the linear part of the feedback system. Yakubovich interpreted Popov's condition using linear matrix inequalities-LMIs, further refined by Kalman [8]. Taken together, these results made a tremendous impact on the development of feedback control theory. Yakubovich's papers [16] set in motion a process that culminated in the computationally efficient LMI's method. Today's proponents of this method, Boyd et al. [2], credit Yakubovich as its father. Yakubovich's four-page "doklad" (communication) is included in this collection because of its historical significance, although its laconic style is more challenging to the reader than the detailed presentation in Yakubovich's subsequent papers [16]. In the West, the most incisive reader of the papers by Popov and Yakubovich was Kalman, who employed his controllabilityobservability theory and the canonical structure theorem [8] to sharpen and clarify the LMI condition ofYakubovich. This led to a result which later became popular as the Kalman-YakubovichPopov (KYP) Lemma. KYP Lemma [8]: Given a number y 2: 0, two n-vectors b, c, and an n x n Hurwitz matrix A, if the pair (A, b) is completely controllable, then a vector q satisfying ATp
+ PA =
_qqT,
Pb-c=y'Yq,
(1)
for all real o). Moreover, the set {x : x T Px = O} is the unobservable subspace of the pair (A, cT ) . With a change of notation, and after some algebraic manipulations, the reader familiar with controllability-observability theory can deduce how this lemma evolved from Yakubovich's theorems in this paper. While (3) is a form of Popov's frequency domain condition, (1) and (2) represent the Lyapunov matrix inequality ATP + PA ::s 0, restricted by a relationship connecting b, c and P. When P > 0 and y = 0, this relationship is Pb = c, which implies c T b > 0, that is, the transfer function H(jOJ) = cT (jOJI - A)-lb has relative degree one, and, by (3), its real part is nonnegative. In other words, H (j OJ) belongs to the class of positive real functions extensively used in network theory for passive synthesis. This fact was recognized only a couple of years later independently by Popov [15] and Brockett [3]. With this positive real (passivity) property, the renamed PositiveReal Lemma soon led to the formulation of the circle criterion, as will be discussed in connection with the paper by Zames [17] included in this volume. Some of these developments are described in the 1966 survey by Brockett [4], while a comprehensive theory of multivariable positive real systems was presented by Anderson [1]. The impact of the PR lemma on model reference adaptive control was crucial as can be judged from the textbooks by Landau [10], Narendra and Annaswamy [13], and Ioannou and Sun [7]. It started with a proposal of Monopoli [12] (see also Fradkov [6]) that an unknown, minimum phase, relative degree one transfer function be made strictly positive real (SPR) by a feedback controller whose parameters are to be adaptively tuned to render the time derivative of a Lyapunov function nonpositive. The matrix P in the Lyapunov function was chosen to satisfy the SPR condition for a reference model. From adaptive control, the idea of achieving passivity by feedback spread to nonlinear control via Kokotovic and Sussmann [9] and Byrnes et al. [5].
(2) REFERENCES
exists if and only if y
+ 2Re[c T (jwI
- A)-lb] ~ 0
(3)
233
[1] B.D.a. ANDERSON, "A system theory criterion for positive real matrices," SIAM J. Contr. Optimiz., 5:171-182,1967.
[2] S. BOYD, L. EL GHAOUI, E. PERON, AND V. BALAKRISHNAN, Linear Matrix Inequalities in System and ControlTheory, SIAM Studies in Applied Mathematics,Vol. 15, SIAM (Philadelphia, PA), 1994. [3] R.W. BROCKETT, "On the stability of nonlinear feedback systems," IEEE Trans. Applications and Industry, 83:443-448, 1964. [4] R.W.BROCKETT, "The statusof stabilitytheoryfor deterministicsystems," IEEE Trans. Aut. Contr., AC-ll:596-606, 1966. [5] C.I. BYRNES, A. ISIDORI, AND lC. WILLEMS, "Passivity, feedback equivalence, and global stabilizationof minimumphase systems,"IEEE Trans. Aut. Contr., AC-36:1228-1240, 1991. [6] A.L. FRADKOV, "Quadratic Lyapunov functions in the adaptive stability problem of a linear dynamic target," SiberianMath Journal, pp. 341-348, 1976. [7] P.A. IOANNOU AND J. SUN, Robust Adaptive Control, Prentice Hall (EnglewoodCliffs, NJ), 1996. [8] R.E. KALMAN, "Canonical structure of linear dynamical systems," Proc. National Academy of Sciences of the United States of America, 48:596600,1962. [9] P.V.KOKOTOVIC AND H.J. SUSSMANN, "Apositivereal conditionfor global stabilizationof nonlinear systems," Syst. Contr. Lett., 19:177-185,1989. [10] 1.0. LANDAU, Adaptive Control: The Model ReferenceApproach, Marcel
Dekker,Inc. (New York), 1979. [11] A.I. LURIE AND V.N. POSTNIKOV, "On the theory of stability of control systems,"Prikl. Mat. Meh., 8:246-248, 1944. [12] R.V. MONOPOLI, "The Kalman-Yakubovichlemmain adaptivecontrolsystem design," IEEE IEEE Trans. Aut. Contr., AC-18:527-529, 1973. [13] K.S. Narendra and A.M. Annaswamy, Stable Adaptive Systems, Prentice Hall (Englewood Cliffs, NJ), 1989. [14] V.M. Porov, "Absolute stability of nonlinear systems of automatic control," AutomationandRemoteControl, 22:857-875, 1962. Translatedfrom Avtomatika i Telemekhanika, 22:961-979, 1961. [15] V.M. PoPov, "The solution of a new stability problem for controlled systems," Automation and Remote Control, 24:1-23, 1963. Translated from Avtomatika i Telemekhanika, 24:7-26, 1963. [16] V.A. YAKUBOVICH, "The matrix-inequality methodin the theoryof the stability of nonlinear control systems-Parts I-III," Automation and Remote Control, 25:905-923, 26:577-592, 26:753-763, 1965. Translated from Avtomatika i Telemekhanika, 25:1017-1029,1964; 26:577-590,26:753763,1965. [17] G. ZAMES, "On the input-output stability of time-varyingnonlinear feedback systems-Parts I and II," IEEE Trans. Aut. Contr., AC-ll:228-238 & 465-476, 1966.
P.V.K.
234
THE SOLUTION OF CERTAIN MATRIX INEQUALITIES 1M AUTOMATIC CONTROL TH EORY
" v. A. JAKUBOVIC 1. We shall denote square matrices by upper-ease italic letters, column-vectors by lower-case italic letters, and D1IIDbers by Greek letters. An asterisk will denote the Hermitian cODjusate, so that ab * is a matrix and b * a = (o, b) is the scalar product. The notatiee H > 0 will sipify that H i. a Hermitian positive-definite matrix. I is the unit matrix.
Let us consider the following problems: (I
J.
Given A,
G,
b; the eigenvalues of A lie in the left half-plane. For H = H * we define G=-(A*H+.HA),
g=-(HG+b).
(1)
We are required to find conditions whose satisfaction will indicate that the quadratic inequality
G-gg*>.o with respect to the matrix
(2)
H = H*, has a solution.
(fiJI). Given B, c ~ 0, d 1:-'0; the eigenvalues of B lie in the left balf-plane. We are required to
find conditions whose satisfaction will indicate that there exists
- Y == B * X + X B <0,
X == X* satisfying the relatioDs
X c + d = O.
(3)
The index v in both problems will denote the order of the corresponding vectors and matrices. Problem (IJ can be reduced to problem (nl/+ 1) and vice versa by simple transformations (see below, Reprinted with permission from Soviet Mathematics (by American Math. Society), V.A. Yakubovich, "The Solution of Certain Inequalities in Automatic Control Theory," 1962, pp. 620-623. 235
§ § 7,8). In applications the given problems (Iv) and (11 11) depend
00
parameter s which must be chosen such
that the corresponding problem has a solution. In connecrion with this, it is desirable to have rational solutions of problems (Iv) and (IIJ, Le., solutions which can be reduced to the verification of a finite
Dumber of conditions eJ.L> 0, '1p. = 0, where eJL and Tip, are polynomials associated with the given problems (i.e., associated with the real and imaginary parts of the elements of A, a, b, B, c, a). For tbe sake of generality and ease of solution, we shall assume that the elements of the vectors and matrices are complex, although in practice they may be real.
J
2. To problems (I and (II v ) can be reduced the problems of the determination of the global stability conditions, optimal ina defined sense, for nonlinear differential equations with one nonlinearity
of class (A) [1-5]; problems with fixed Donlinearities of the type studied in {6]; and, some others. The method of Lurie [1,2] gives rational sufficient conditions for solving problem (Iv). For II ~ 2 these conditions are also necessary [5], but it can be shown that for v
> 2 these conditions
do not coincide with
the necessary .oaes, * Let us note the effective sufficient conditions of Lefschetz [7], which can be CODsidered as sufficient conditions for the solvability of problem (Iv). In [8] V. M. Popov deduced a special global stability condition -for sysJemsnonlinearities of class (A) encompassing all th~ conditions which can be obtained by means of Ljapunov functions of the type: "a quadratic form plus an integral of the nonlinearity." Theorem 1 (see below) with minor additions gives a new proof of POPovJs condition, together with an answer to the converse problem posed in [8], p. 972. ·Namely, systems for which there exist Ljapunov functions- of the indicated type will satisfy the special condition of Popov [8]. 3~
Let us Introduce the following notations:
A co = A - iel , QUI == A;1a, bw == A(LJ*-1b,
ez».
(4)
The functions ¢I«(U)' ¢II(ti» will be called the characteristics of problems (Iv), (U), respectively. Theorem 1. In order thot mequQ,lity (2) have the solution H =-8* it is necessary and sufficient tAat
cI1 (Ct)
> 0 for -
DO
< (l) < +
:00.
Theorem 2. In order that there exist a matrix
sufficient that the following be satisfied: 1)
K
X = X!" satisfying relations
is real;
2)
f3
it is.necessary and ",0; 3) epII (Ct)) > 0 for - 00 < Ct) < + 00. (3)
Theorem 3. In order that there exist a masri» X =-X* satisfying relations (3) ana the inequality f). 0, it is necessary -th;at the following be satisfied: 1) K is real; 2) EO = 2 Inf CPu «(.c)/l c;' > 0; 3) £ ~ fO; and sufficient that the conditions 1), 2), and 3') £ < £0' be satisfied.
Y >f 1 for a gi ven
Obviously Theorems 1, 2, and 3 give rational solutions. **
4. The necessity of the conditions of Theorems 1-3. We have A:H +HAw = _·C, (CacJ aJ =
,=
J =·2 Re (b, aJ + 2 Re " where -(g, a J. From (2) and the first of relations (1) we successively deduce (GarJ aJ > I ,,2, . pI «(U) > , 1 + ,2 \.*** - 2 Re (Ha, a
*Thus, the problem posed in [5], p, 129, has a negative solution. ·.Otber, less convenient, solutions of problems (Iv) and (II,) were reported by the author in the spring of 1961 at the seminar of V. V. Nemycki1 at the Moscow State University. ···Let us note that the necessity of the condition 11(0) >0 was essentially proved in [3, 5]. The necessary condition (3.2) of [5], for complex A, a, b in the notation here takes the form 2 = 1 +:2 Re (b, .,4-1 a) > O. As long as problem (I ) does not change when A is substituted for A, we have from here that ""- (6) > o. v ~ ~
r
236
The necessity of the conditions of Theorems 2 and 3 follo.ws hom the relations (1 c, e), 2epII (lU) = (Y ccu J c > fie wI 2.
K
=·(Xc, e),
J
f3 =
5. The sufficiency of the conditions of Theorem 3 follows from the sufficiency of Theorem 2 (for matrices of fixed order v). Assuming X == Xo \£H O' where B*H 0 + HoB == - I, we get that problem (Ill) bas the desired solution if there exists Xo = Xo satisfying the conditions B* Xo + XoB == - Y0 < 0, Xoc + dO = 0, where dO = EH OC + d. The fulfillmeot of conditions 1), 2), 3) of Theorem 2 and, consequently, the existence of XO' follow from conditions 1), 2), 3') of Theorem 3.
6. We shall write fAI = p.
the matrix A has Il eigenvalues Inthe left half-plane and does have any purely imaginary or zero eigenvalues. jf
Dot
Lemma 1. Let us consider the matrix of order v + 1
B-
I"V _
(A P] •
(5)
q*a
Let Re a~ 0 and let e be a vecto~with v + 1 components 0, • - ~ ,0, 1. Let us assume XI «(c) = «A - iCd)-l p, q) - a, XII (Ct) = - «B - j(J)l)-l e, e). Then the conditions: al' Re XI (CIl) > 0 for - 00
< +,:, b I ) IAI b 2) IB I = It + 1. eu
C
Il' are equivalent to the conditions: 8 2) Re XII (Cd) > 0 for -
Proof. Solving the equation
(if -
icoI)
[~J
-
e, we find
00
~ =- XII == (XI + iCl))-l.
<
<(() < + 00, Therefore a 2 ) fol-
'"
~
lows from a 1). From a 1), b l ) and the relation XII = - det (A - iCcJn/det (B - ieu]) we conclude that B has 00 purely imaginary eigenvalues, As I CI) I -+ 00 we have.!e XII =- Re a/CI)2 - [1 +0 «(;.I -1)], Le.,
R~ a
<0
and 11 Arg XII (cu) I~ =,- 11. Therefore /1 Arg det (B - i(,)n The coeverse statement can be proved analogously.
I: = fT [(JI + 1) -
(p. + 1)], Le.,
IBI =' p. + 1.
7. The sufficiency of Theorem 1 follows lem (Ill') for B =
(t* _al/
from the sufficiency of 'Theorem 2. Let us consider prob-
]. c =e, d =- e. Applying Lemma 1 we fiod that the conditioos of Theorem 2
2 are satisfied. Therefore, there exists X = [:
~]
as the solution of problem (nv+l)' It is easy to
verify that the matrix H is the solution of problem (Ill)' 8. Proof of sufficiency ,0E Theorem 2. For
11
=1
the assertion of Theorem 2 is easily verified. Let
us assume that sufficiency has been 'proved for matrices of order v. Let us coosider problem (Jl v +1)
n
i
and assume that the conditions of Theorem 2 are satisfied. From the formula ,,1< =, lim
0-.+00 ..:0
cPII(6.)
a6J
it follows that I( > O. Therefore the matrix S whose last columa equals c and whose first v columns are some basis' in a subspace orthonormal to tI, is nonsinplar•.Multiplying the rust relatioD in (3) on the left· by S * and OD ·the right by S, and the second on the left by S*, we pass .to the equivalent problem ~
"'.'"
I"\.I'1"V
- Y == B X + X8 < 0, where
"'B"
Ii =.5- 1 BS.
Inasmuch as cPU «(;.I)
in 'the form (5), we find that Re a
Re XI () > 0, fA} = v. ~
'"
Xe
=P... [I + 0 «(;.I -1)] (JJ2
=·_fJ/K< 0,
Re XII (Cd)
d
I:
Cor
I(;.I I ,. . . oa, we have f3 >.0.
Representing
=thl (6.)/1< > O. By Lemma 1 we have
E The seem relation 'in '(6) signifies that ""X has th e rorm
237
(6)
'ICe,
~X -_
[08 0] I(
•.
From th e
first relation in (6) it follows that the desired matrix H is the solution of problem (Ix,,)' where a =.
p/JP, b = /(Q/,ff3.
The characteristic of this problem is
Let us consider the first case, when Re (A -1 a, b) that IDf
CPI (Cd) = ¢l (r).
Let us
ch (CLl) =~
< 0 for some
'" problem consider the auxiliary
According to the induction postulated in
Ci."
> O.
Then there exi.sts a r such
X1 a, + b r - 8HOar =o,
-Yl=A*X 1 + X1A<0, where H0 is the solution of the equation
Re Xl (CLl)
A. *H0 + HoA
= -
1, 8 > 0, and
§ 4 we "are justified in
(7)
a" b
T
are determined from (4).
applying Theorem 3. The value of
ap = (A.;1 a, b) + 8 (HOar' a r) is real since I/>i(r) = O. The characteristic of the auxiliary problem can be rewritten in "the form (tPI (Cd) -
11
Therefore EO
~ 8/2
,
,
I
and by Theorem 3 there exists a matrix Xl
= H such that Y 1 > ~ I.
We will show
that the matrix H is the solution of problem (Ill) if the Dumber 8 > 0 is chosen 5ufficiendy small. After some computation we obtain, using the second relation in (7), (C- 1 g, g) = - 2 Re (0" b) + 8 2 (C- 1 h, II) < 1 - tPI(,) + 38 , h t 2, where h = A;H Oar. For 0 < 8 < epI (')/3 , h 12 we get (G-l g, g) < 1, which for the condition G = Y 1 > 0 is equivalent to (2). The proof is complete for the first case.
In the second case, when Re (0tu' b) > 0, the auxiliary problem has the form - Y 1 i5 A*X1 + X1 A < 0, Xl a + b - 8H 0 a = 0, where 0 < 0 , H0 a 12 /3. All the above arguments are simplified. The Dumber leap == - (a, b) - 8 (8 a, a) is real since as t cu' --+ oa from the condition Re (aGJ' b) = 0 -Im(a, b)/fJJ + 0 (fJJ -2) > 0 it follows that 1m (a, b) = O. The characteristic of the auxiliary problem has the form I/>ap
= Re
(a""
b) +
~ IQ",1 2, Le., EO ~ 8/2.
By Theorem 3 there exists Xl =
~
such that
Y 1 > ~ I. For H = Xl we have G = Y 1 > 0, - g = 8H oa, (G-1 g, g) < 1. Theorem 2 is proved. Received 11/DEC/61 BIBUOGRAPHY [1] A. I. Lur'e, Some nonlinear problems of the theory of automatic regulation, GITTL, Moscow, 1951. "(Russian) [2] A. M. Letov, Stability in nonlinear control systems, GITTL, Moscow, 1955. (Russian)
[3] [4] [5] [6] [7] [8]
v, A. Jakubovi~; Dokl, Akad. Nauk SSSR 117 (1957), 44. ---......, ibid. 121 (1958), 984. ............... , Vestnik Leningrad. Univ, (15) (1960), no. 7, 120. ~,Dokl. Akad, Nauk SSSR 135 (1960), 26 = Soviet Math. Dokl. 1 (1960), 1238. S. Lefschetz, RIAS Tech. Rep. 60-9, 1960. V. M. Popov, Avtomat. i Telemeh, 22 (1961), 961 = Automat. Remote Control 22 (1962), 857. Translated by: N. H. Choksy
238
Mathematical Description of Linear Dynamical Systems R.E.KALMAN
T
HE contrastand relationship between models that explain
phenomena by either (i) displaying the (internal) wiring diagram/event sequence in a system, or (ii) classifying the (external) input/output behavior, are central themes in many areas of science. Undoubtedly a good example is the modem computer, where "friendliness" is identified with the need to understand only the external behavior, but where design and efficiency require mastery of the internal structure. In the field of control, the input/output point of view dominated the development in the first half of the twentieth century, particularly in the western world. This is evident from the works of Nyquist and Bode (see their papers elsewhere in this volume) and that of other reserchers of the period. It is fair to say, in fact, that until the 1960s, systems and control was actually identified with input/output thinking; even nowadays researchers in other fields (such as biology and psychology) view the stimulus/response model as the essence of systems theory-a point of view strongly advocated, for example, by Wiener in his influential book on cybernetics [16]. The limitations of the input/output point of view became strikingly evident in the earlier work of Kalman ([8], [9]both included in this volume), and others. By using state space descriptions for systems, very effective new algorithms were put forward for filtering and prediction on the one hand, and feedback control design on the other. The filtering algorithms turned out to be dramatic improvements over those obtained earlier by Wiener and Kolmogorov (see the preambles to [15] and [8], elsewhere in this volume), and filtering and optimal feedback control became practical as a result of these developments. Soon after the appearance of these algorithms, the question was raised as to what the relationship was between state space and input/output models. In the paper that follows, Kalman examines this relationship, and especially the question of irreducibility, and the relevance of controllability and observability to the state space representation question. The results, while more general, appear to have been inspired by the work of Gilbert in a paper [5] that precedes Kalman's paper in the same issue of SIAM Journal on Control.
Kalman's paper discusses the relationship between the finitedimensional linear state space system d
dtX = F(t)x
+ G(t)u; y = H(t)x
(1)
and the convolution integral y(t) =
it
S(t, l:)u(-c) dt
(2)
to
Special attention is given to the situation in which the matrices F(t), G(t), and H(t) are constant (i.e., independent oft), and the input/output relationship (2) is shift-invariant, that is, Sit, r) 8(t - r, 0). (In mathematics, the term convolution is usually reserved for the shift-invariant case.) The paper contains a series of important results, among which are the following:
=
(i) The system (2) can be represented by a state space system (1) (assuming x(to) = 0) if and only if Stt , r) is factorizable as Sit, r) = P(t)Q(r), with P(t) and Q(t) finite matrices. (ii) The system (1) is irreducible (meaning that among all systems (1) that represent, in the above sense, the same system (2), it has minimal state space dimension) if, and only if, it is completely controllable and completely observable. (iii) Any two irreducible realizations are equivalent (in the sense that they differ only in the choice of the basis in the state space).
(iv) The system (1) can be decomposed into a controllable/ observable subsystem, a controllable/unobservable subsystem, an uncontrollable/observable subsystem, and an uncontrollable/unobservable subsystem (see Figure 4 of the paper). Important developments that were inspired by these results are generalizations to nonlinear systems (see, e.g., [14]) and to stochastic systems (see, e.g., [2]), and the associated algebraic theory using system descriptions in terms of matrices of
239
polynomial or rational functions [3], [18]. Further developments were in automata theory [10] and coding theory [11]. The issues raised in this paper and the results obtained form the basis of what has later been called realization theory, one of the central themes in system theory. The one important element that was added later is the Hankel matrix [7], [13], which turned out to be the most effective algorithmic tool for the computation of models from input/output data. This is already evident in the algorithms in [13], but was further accentuated by the work on model reduction, in which the paper [6] stands out. Model reduction has become a very effective tool for reducing the complexity (in the sense of reducing the number of state variables) of a linear system. Two of the most widely used methods for model reduction are (i) balancing (which may be traced back at least to [12]) and (ii) finding a Hankel matrix of lower rank that optimally approximates a given Hankel matrix (often called AAK-model reduction [1]). These methods are based on analyzing the singular values of the Hankel matrix, by exploiting effectively the fact that the basis of the state space may be chosen freely, and by considering the controllability and observability gramians of the system to be reduced. A germ of these ideas is of course present in the above mentioned results in Kalman's paper. The view that an external description of a system takes the form of an input/output map or a transfer function, and that an internal description leads to a state space model has recently been challenged by what is called the behavioral approach [17]. In this framework, a system is not viewed as an input/output object; instead all external ("manifest") variables are treated on an equal footing. Here irreducibility is equivalent to mere observability, since external behaviors can be uncontrollable (contrary to the system (2), where controllability is basically built in).
REFERENCES
[1] V.M. ADAMJAN, D.Z. AROV, AND M.G. KREIN, "Analytic properties of Schmidt pairs for a Hankel operator and the generalized Schur-Takagi problem," Math. USSRSbornik, 15:31-73, 1971. [2] H. AKAIKE, "Stochastic theory of minimal realization," IEEE Trans. Autom. Contr., AC-19(6):667-672, 1974. [3] P.A. fuHRMANN, Linear Systems and Operators in Hilbert Space, McGraw-Hill (New York), 1981. [4] ER. GANTMACHER, The Theoryof Matrices, Chelsea, 1959. [5] E.G. GILBERT, "Controllability and observability in multivariablecontrol systems," SIAM J. Control, 1:128-151,1963. [6] K. GLOVER, "All optimalHankel-normapproximationsof linear multivariable systems and their L oo bounds," Int. J. Contr., 39:1115-1193,1984. [7] B.L. Ho AND R.E. KALMAN, "Effective construction of linear statevariablemodelsfrominput/outputfunctions,"in Proc.ThirdAllertonConf., pp. 449-459, 1965. [8] R.E. KALMAN, "A new approach to linear filtering and prediction problems," Trans. ASME, J. Basic Engineering, 82D(I):35-45, 1960. [9] R.E. KALMAN, "Contributions to the theory of optimal control," Bol. Soc. Mat. Mexicana, 5:102-119, 1960. [10] R.E. KALMAN, P.L. FALB, AND M.A. ARBIB, Topics in Mathematical System Theory, McGraw-Hill(New York), 1969. [11] J.L. MASSEY AND M.K. SAIN, "Codes, automata, and continuous systems: Explicit interconnections," IEEE Trans. Autom. Contr., AC-12(6):644650,1967. [12] B.C. MOORE, "Principal component analysis in linear systems: controllability, observability, and model reduction," IEEE Trans. Autom. Contr., AC-26:17-32, 1981. [13] L.M. SILVERMAN, "Realization of linear dynamical systems, IEEE Trans. Autom.Contr., AC-26:554-567, 1971. [14] H. SUSSMANN,"Existence and uniqueness of minimal realizations of nonlinear systems," Math. Syst. Theory, 10:263-284, 1977. [15] N. WIENER, Extrapolation, Interpolation, and Smoothing of Stationary TimeSeries, MIT Press (Cambridge, MA), 1949. [16] N. WIENER, Cybernetics, MIT Press (Cambridge,MA), 1961. [17] J.C.WILLEMS, "Paradigmsandpuzzlesin thetheoryof dynamicalsystems," IEEE Trans. Automat.Contr., AC-36:259-294, 1991. [18] W.A. WOLOVICH, Linear Multivariable Systems, Springer-Verlag (New York), 1974.
J.C.W.
240
J.S.I.A.:M. CONTROL Ser. A, Vol. t, No.2 PrinUtl in U.S.A .• 1963
MATHEMATICAL DESCRIPTION OF LINEAR DYNAMICAL SYSTEMS·
R.E.KALMANt Abstract. There are two different ways of describing dyasmicel systems: (i) by meaDS of atate variables and (ii) by input/output relations. The first method may be regarded as an axiomatization of Newton's laws of mechanics and is taken to be the basic definition of a system. It is then shown (in the linear case) that the input/output relations determine only one part of a system, that which is completely observable and completely eontrollable. Using the theory of controllability and observabiJity, methods are given for calculating irreducible realizations of a given impulse-response matrix. In particular, an explicit procedure is given to determine the minimal number of state variables necessary to realize a given transfer-function matrix. Difficulties arising from the use or reducible realizations are dlaeussed briefly.
1. Introduction and summary. Recent developments in optimal control system theory are based on vector differential equations as models of physical systems. In the older literature on control theory, however, the same systems are modeled by transfer functions (i.e., by the Laplace transforms of the differential equations relating the inputs to the outputs). Two different languages have arisen, both of which purport to talk about the same problem. In the new approach, we talk about state variables, transition equations, etc., and make constant use of abstract linear algebra. In the old approach, the key words are frequency response, pole-zero patterns, ete., and the main mathematical tool is complex function theory. Is there really a difference between the new and the old? Precisely what are the relations between (linear) vector differential equations and transferfunctions? In the literature, this question is surrounded by confusion [1]. This is bad. Communication between research workers and engineers is impeded. Important results of the "old theory" are not yet fully integrated into the new theory. In the writer's view-which will be argued at length in this paper-the difficulty is due to insufficient appreciation of the concept of a dynamical 81/stem. Control theory is supposed to deal with physical systems, and not merely with mathematical objects such as a differential equation or a transfer function. We must therefore pay careful attention to the relationship betweenphysical systems and their representation via differential equations, transfer functions, etc.
* Received by the editurs July 7, 1962and in revised form December 9, 1962. Presented at the Symposium on Multivariable System Theory, SIAM, November I, 1962 at Cambridge, M88B&chuaetts. This research was supported in part under U. S. Air Force Contracts AF 49(638) . .382 and AF 33(616)-6952 as wellaa NASA Contract NASr-l03. t Researcb Institute for Advanced Studies (RIAS), Baltimore 12, Maryland. Reprinted with permission from SIAM Journal on Control, R. E. Kalman, "Mathematical Description of Linear Dynamical Systems" Vol. 1, 1963,pp.152-192.
241
To clear up these issues, we need first of all a precise, abstract definition of a (physical) dynamical system. (See sections 2-3.) The axioms which provide this definition are generalizations of the Newtonian world-view of causality. They have been used for many years in the mathematical literature of dynamical systems. Just as Newtonian mechanics evolved from differential equations, these axioms seek to abstract those properties of differential equationswhichagree with the "facts" of classical physics. It is
hardly surprising that under special assumptions (finite-dimensional state space, continuous time) the axioms turn out to beequivalent to a. system of ordinary differential equations. To avoid mathematical difficulties, we shall restrict our attention to linear differential equations. In section 4 we fonnulate the central problem of the paper: Given an (e%perimentally observed) impulse f'esponse matrix, how can we identify the linear dynamical system which generated it' We propose to call any such system a realiztJtion of the given impulse response. It is an irreducible realization if the dimension of its state space is minimal. Section 5 is a discussion of the "canonical structure theorem" [2, 14] which describes abstractly the coupling between the external variables (input and output) and the internal variables (state) of any linear dynami-
cal system. As an immediate consequence of this theorem, we find that a linear d1Jfl'Jmical system i8 an irreducible realizGlion oj an impu'lBe..re8pona8 matriz if and emly if the By8tem is completely controllable and completely ob· 8ervable. This important result provides & link between the present paper and earlier investigations in the theory of controllability and observability [3-5]. . Explicit criteria. for complete controllability and complete observability are reviewed in a convenient form in section 6. Section 7 provides a constructive computational technique for determining the canonical structure of a constant linear dynamical system. In section 8 we present, probably for the first time, a complete and rigorous theory of how to define the state variables of a multi-input/multi-output constant linear dynamical system described by its transfer-function matrix. Since we are interested only in irreducible realizations, there is a certain unique. well.. defined number 11, of state variables which must be used. We give a simple proof of a recent theorem of Gilbert [5] concerning the value of n. We give canonical forms for irreducible realizations in simple cases. We give a constructive procedure (with examples) for finding an irreducible realization in the general case. Many errors have been committed in the literature of system theory by
carelessly regarding transfer functions and systems as equivalent concepts. A list of these has been collected in section 9. The field of research outlined in this paper is still wide open, except 242
perhaps in the case of constant linear systems. Very little is known about irreducible realizations of nonconstant linear systems. It is not clear what additional properties-besides complete controllability and complete observability-are required to identify the stability type of a system from its impulse response. Nothing is known about nonlinear problems in this con-
text. Finally, the writer wouldlike to acknowledge his indebtedness to Professor E. G. Gilbert, University of Michigan, whosework [5] predates this and whose results were instrumental in establishing the canonical structure theorem.
2. Axiomaticdefinition of a dynamical system.. Macroscopic physical phenomena are commonly described in terms of cause-and-effect relationships. This is the "Principle of Causality". The idea involved here is at least as old as Newtonian mechanics.. According to the latter, the motion of a system of particles is fully detennined for all future time by the present positions and momenta of the particles and by the present 8J1d future forces acting on the system. How the particles actually attained their present positions and momenta is immaterial. Future forces can have no effect on what happens at present. In modem terminology, we say that the numbers which specify the instantaneous position and momentum of each particle represent the atate
of the system. The state is to be regarded always as an abstract quantity. Intuitively speaking, the state is the minimal amount of information about the past history of the system which suffices to predict the effect of the past upon the future. Further, we say that the forces acting on the particles are the inputs of the system. Any variable in the system which can be directly observed is an output. The preceding notions can be used to give a precise mathematical definition of a dynamical system [6]. For the present purposes it will be convenient to state this definition in somewhat more general fashion [14J. DEFINITION 1. A dynamical system is a mathematical structure defined by the following axioms: ( Dr) There is given a state space 1; and a set of values of time 9 at which the behavior of the system is defined; >; is a topological space and 9 is an ordered topological space which is 8 subset of the real numbers. (D2 ) There is given & topological space 0 of functions of time defined on 8, which are the admissible input., to the system. (Ds) For any initial time to in a, any initial state Xo in ~, and any input u in 11 defined for t ~ to. the future states of the system are determined by the transition function rp: n X e X 9 X ~ --. ~, which is written as ~u(t; to, %0) ~ z«, This function is defined 243
only for t ~ to. Moreover, any to ~ t 1 ~ tt in E), any Xo in 2;, and any fixed u in n defined over [ti), tIl n a, the following relations hold: (Da-i) ( D~-ii)
In addition, the system must be nonanticipatary, i.e., if u, n e we have
v f nand u == v on [/1), tIl
cp,,( tj to , xo)
==
\0,(t; to , xo).
Every output of the system is a function .p': e X X -+ reals. The functions fJ and 1/1 are continuous, with respect to the topologies defined for %, 8, and n and the induced product topologies. In this paper we will study only a very special subclass of dynamical systems: those which are real, jinite.. dimensional, continuous-time, and linear. "Real, finite-dimensional" means that l; = R" = n-dimensional real 1 linear space. "Continuous-time" means that 8 = R = set of real numbers. "Linear" means that rp is linear on n X }; and"" is linear on ~. By requiring qJ and'" to be sufficiently "smooth" functions, we can deduce from the axioms a set of equations which characterize every real! finitedimensional, continuous-timet and linear dynamic.al system. The proof of this fact is outside the scope of the present paper [14]. Here we shall simply assume that every such system is governed by the equations
dx
+ G(t)u(t),
(2.1)
dt = F(t)x
(2.2)
y(t) =- H~t)x(t),
defined on the whole real line - co < t < QO, where z, u, and yare n, m, and p-veotors" respectively, and the matrices F(t), G(t), and H(t) are continuous functions of the time t. We call (2.1-2) the dynamical equations of the system. It is instructive to check whether the axioms are satisfied. (Dr) is obviously true; we have 1; = R", = R 1• The state of the system is the vector z: To satisfy (D,), we must specify the class of all inputs, that is, a subclass of all vector functions u(t) = (Ut(t), ... , Um(t)). To define 0, we shall assume that these functions are piecewise continuous; this is sufficiently
e
* Vectors will he denoted by small Roman letters, matrices by Roman capitals. The components of & vector x are x, , components of a matrix-A are ai, . On the other hand, xl, %2, ••• , are veetors, and FAA, FAB are matrices . .4.' is the transpose of A. 244
general for most applications. We have exactly p observations on the system (the components of the vector y) and by (2.2) they are functions of t, z, Hence (D4) is satisfied. To check (Da) , we recall that the general solution of (2.1) is given by
(2.3)
CPu (t;
to, :Co)
==
+ 1 cI>(t, r )G( '0 1
X,
= tt(t, to)Xo
l'
)1£( 'T) dr,
where 4»(t, T) is the transition matrix of the free ·differential equation defined by F(t) [4, 7Jt. Since (2.3) is valid for any t ~ to (in fact, also for t < to), f/J is well defined. Property (Ds-i) is obvious. (Ds-ii) follows from the composition property [4, 7] of the transition matrix: c1J(t, a) = 4"(t,
(2.4)
T)~(r,
0"),
which holds for every set of real numbers t, T, IT. Indeed, (2.4) is simply the linear version of (Ds-ii). (Ds-iii) is obvious from formula (2.3). The continuity axiom (D&) is satisfied by hypothesis. Evidently tp given by (2.3) is linear on the cartesian product of 2; with the linear space of vector.. valued piecewise continuous functions. We call a linear dynamical system (2.1-2) constant, periodic, or analytic whenever F, GJ and H are constant, periodic, or analytic in t. It is often convenient to have a special name for the couple (t, %) lEe X ,;. Giving a fixed value of (t, x) is equivalent to specifying at some time (t) the state (x) of the system. We shall call (t, z) a phase and e X ~ the phase space. (Recall the popular phrase: "phases" of the
Moon.) To justify our claim-implicit in the above discussion-that equations (2.1-2) are 8. good model of physical reality, we wish to point out that these equations can be concretely simulated by a simple physical system: a general-purpose analog computer. Indeed, the numbers (or functions) constituting F, 0, and H may be regarded as specifying the "wiring diagram" of the analog computer which simulates the system (2.1-2) (see, for instance, [8D. 3. Equivalent dynamical systems. The state vector ~ must always be regarded as an abstract quantity. By definition, it cannot be directly measured. On the other hand, the inputs and outputs of the system (2.1-2) have concrete physical meaning. Bearing this in mind, equations (2.1-2) admit two interpretations: (a) They express relations involving the abstract linear transformations F(t), G(t), and II(t). (b) At any fixed time, we take an arbitrary but fixed coordinate system
t Le.,
if» is a solution of d~/dt
== 1 ~ unit matrix for all .,..
= F(l)tf>, subject 245
to the initial condition 4»(1', T)
2;. Then the symbol z == (Xl t •• • , :.eft) is the numerical n ..t uple consisting of the coordinates of the abstract state vector whichis also denoted by x. F, G, and H are interpreted
in the (abstract) vector space interpreted
88
as tile matrix representations of the abstract linear transformations denoted by the same letters under (a) .. To describe the behavior of a dynamical system in concrete terms, the second point of view must. be used. Then we must also ask ourselves the question: To what extent does the description of a dynamical system depend on the arbitrary choice' of the coordinate system ill the state space?
(No such arbitrariness occurs in the definition of the numerical vectors 'U, quantities.) This question givesrise to the next definition. y since the input and output variables 'Ui and YJ are concrete physical DEFINITION
2. Two linear dynamical systems (2.1-2), with state vectors
x, Z, are algebraically equivalent whenever their numerical phase vectors are related for all t as (t, x) . = (t, T(t)x),
(3.1)
where T(t) is a n X n matrix, nonsingular for all t and' continuously differ.. entiable in t. In other words, there is a 1-1 differentiable correspondence between the phase 8~ e X ~ and a x ~. Remark.· We could generalize this definition of equivalence to (1, i) = (r(t), T(t)x) where f' is an increasing function of t. But this involves distortion of the time scale which is not permitted in Newtonian physics. Algebraic equivalence implies the following relations between the defining matrices of the two systems: 1 ~(tJ"') = T)r ( r ) ,
tuw«,
F(t) = T(t)F
(3.2)
G(t)
1(t)
+ T(t)F(t)T- 1( t ) ,
= T(t)G(tJ,
H(t) = H(t)r1(t ) .
In general, algebraic equivalence does not preserve the stability properties of a dynamical system [7, 9, 10J. For this it. is necessary and sufficient to have topological equivalence: algebraic equivalence plus the condition (3.3) Cl and c! are fixed constants" and II 11 is the euclidean norm*. A nonconstant system may be algebraically and even topologically equivalent to a constant system. The latter case' is called by Markus (11 J
where
• Let 9, :t, and! have the usual topologies induced by the euclidean norm. Then the product topologies induced OD e X 2: and ex;; are equivalent if and only if (3.3) holds.
246
"kinematic similarity". Moreover, two constant systems may be algebraically and topologically equivalent without T(t) being a constant. To
bypass these complications, we propose DEFINITION 3. Two constant linear dynamical systems are 8trictly eqt,ivQ,lent whenever their numerical phase vectors are related for all t as (t, i) = (e, Tx), where T is a nonsingular constant matrix. Evidently strict equivalence implies topological equivalence. 4. The impulse-response matriz and its realization by a linear dynamical system. Sections 2-3 wereconcerned with mathematics, that is, abstract matters. If we now take the point of view of physics, then a dynamical system must be "defined" in terms of quantities which can be directly observed. For linear dynamical systems, this is usually done in the following way. We consider a system which is at rest at time to ; i.e., one whose input and outputs have been identically zero for all t ~ to. We apply at each input in turn a very sharp and narrow pulse. Ideally, we would take Ui'j)(t) = J 8(t - to), where I is the Dirac delta function, 8'i is the Kronecker symbol, and 1 ~ i, j :i nt. We then observe the effectof each vector input u(j)(t) on the outputs, which are denoted by u(t; j). The matrix Set, to) = (Sij(l, to)J = [Yi(ljj)] 80 obtained is called the impulBB..re8f.ltm86 maJ,riz of the system. Since the system was at rest prior to t = to, we must define 8(1, to) == 0 for t < to. We also assume, of course, that S is continuous in t and to for t > to. With these conventions, the output of a linear system originally at rest is related to its input by the well-known convolution integral:
'i
(4.1)
y(t) =
l''0 Set,
1')u(1') dr,
In much of the literature of system theory [12J (and also at times in physics) formula (4.1) is the basic definition of a system. The Fourier transform of S is often called "the system function" [13, p. 92]. Unfortunately, this definition does not explain how to treat systems which are not "initially at rest". Hence we may ask, UTo what extent, if any, are we justified in equating the physical definition (4.1) of a system with the mathematical one provided by (2.1-2)?" Suppose that the system in question is actually (2.1-2). Then (2.3) shows
that (4.2)
Set, r) == H(t)eJ!{t, r)G( r), == 0,
t ~ r, t
< r.t
t The right-hand side of the first equaeion (4.2) is defined also for I < "; then the left-hand aide may be regarded as the "backward impulse response", whose physical j n terpretation is left to the reader. 247
Thus it is trivial to calculate the Impulse-respouse matrix of a given linear dynamical system. The converse question, however, is non trivial and interesting. When and how does the impulse-resptm86 malrix determine the
dynamical eq'Uatiom of the system' This problem is commonly called the identification of the system from its impulse-response matrix. Having been given an impulse.. response matrix, suppose that we succeed in finding matrices' F. G, and H such that (4.2) holds. We have then identified a physical system that may have been the one which actually generated the observed impulse..response matrix. We shall therefore call (2.12) a,
realiz4tion of Set, r). This terminology is justified because the axioms given in section 2 are patterned after highly successful models of classical macroscopic physics; in fact, the system defined by (2.1-2) can be con .. cretely realized, actually built, using standard analog-computer techniques in existence today_ In short, proceeding from the. impulse-response matrix to the dynamical equations we get closer to "physical reality". But we are also left with & problem: Which one of the (possibly very many) realizations of S{t, .,.) is the actual system that we are dealing with? It is conceivable that certain aspects of a dynamical system cannot ever be identified from knowledge of its impulse response,as our knowledge of the physical world gained from experimental observation must always be regarded as incomplete. Still, it seems sensible to ask how much- of the physical world can be determined from a given amount of experimental
data. The first clear problem statement in this complex of ideas and the first results appear to be due to the writer (2, 14J. First of all we note THEOREM 1. An impulse..respome matrix Set, -r) i8 realizable b1J a finitedimensional dynamical 81J8tem (2.1-2) if and only if there ex·ist continuous matrices P (l) and Q(t) stl£h that S(t, .,.)
(4.3)
= P(i)Q(.,)
fur all
t,
T.
Proof. Necessity follows by writing the right-hand side of (4.2) as H(t)4'(t, O)cIJ(O, -r)G(.,.), with the aid of (2.4). Sufficiency is equally obvious. We set F(t) = 0, G.(t) = Q(t), and H.(t) = P(t). Then ~(t, r) == I and the desired result follows by (4.2). A realization (2.1-2) of S(t, .,.) is reducible if over some interval of time there is a proper (i.e., lower-dimensional) subsystem of (2.1-2) which also realizes S (t, 1'). As will be seen later, a realization of S (particularly the one given in the previous paragraph) i, often reducible. An impulse-response matrix I.'i is stationary 'whenever S ( t, r ) = Set (1, r + (1) for all real numbers t, T J and a, S is periodic whenever
+
248
the preceding relation holds for all t, T, and some e, An Impulse-response matrix is fJfUJl1lt~ whenever 8 is analytic in t and T; if (4.3) holds, then P and Q must be analytic in t. The main result, whose proof will be discussed later, is the following [14]: THEOREM 2. HvpotAesis: The impulse-response matriz S 8aJ,iBfiea (4.3) ani/, is eitAer periodic (and contiftUOUl) or a1UJlytic. ConclUBiom: (i) There ezist i1'Teducible realizati01f,8 oj 8, all oj wh.ich have the same conatGnt dimenaioft, n aM are algebraicaUyequivalent. (ii) If S i8 periodic [tJfl4lyticj 80 are us irreducUJle realimtionB. Topological equivalence cannot be claimed in general. It may happen
that S has one realization which is asymptotically stable and another which is asymptotically unstable [15]. Henceit may be impossibleto identify the stability of a dynamical system from its impulse response! This surprising conclusionraises many interesting problems whichare as yet unexplored [15J. If 8 is not periodic or analytic, it may happen that the dimension nCt) of an irreducible realization is constant only over finite time intervals. In the stationary case, Theorem 2 can be improved [14]. THBOltEll 3. Every Btationary impul8e-reapome matriz Set, T) == W(t - .,,) 8ati8fying (4.3) has ctm8lant irreducible realizations. AU auch realizations are 8trictl1l equiualm'. In view of this theorem, we may talk indifferently about a stationary impulse-response matrix or the dynamical system which generates it-as has long been the practice in system theory on intuitive grounds. But note that we must require the realization to be irreducible. For nonconstant systems, such a conclusion is at present not justified. The requirement of irreducibility in Theorem 3 is essential; disregarding it can lead-and has led-to serious errors in modeling dynamical systems. (See section 9.) In many practical cases, it is not the weighting-function matriz W (t - ,,) (see Theorem 3) which is given, but its Laplace transform, the tram/erfunction mattU Z(s) =- ..c(W(t)]. Then condition (4.3) has an interMting equivalent form, which is often used as a "working hypothesis" in engineering texts: TSEOBBM 4. A weighting-junc'ionmtIbU W(l - or) 8aJiBjie, (4.3) if and only if its Blementl arelinear combinati0'n8 oj terms of lAs type tie- i ' (i == 0, 1, · · · ,n - 1, j == 1, · .. J n). Hence etJery element oj the tro:ntJ/er..jundion matriz i8 a ratio of polynomials in 8 8UCh that the degree oj the dmominator polynomial alway, a:ceed8 the degree of the numerator polynomial. This result is provedd in [14J. It impliesthat the realizationof an impulseresponse matrix is equivalent to expressing the elements of F, G, and 11 as functions of the coeffic~ents of the numerator and denominator polynomials of elements of Z(8). (See section 8.) In the remainder of the paper, we wish to investigate two main problems 249
arising in the theory sketched above: (i) Explicit criteriafor reducibility. (ii) Constructionof irreducible realizations.
R6m4r1c. Elementary expositlons of systelJltheory often contain the statement that the operator d/dt (iiiiii 8) is a Usystem." Is a it system in the same sense as that word is used here? The answer is no. To define such a system rigorously in accordance with the axioms introduced in section 2, one must proceed as follows. The output of the system, which by definition is the derivative of the input, is given by (3.4)
y(t)
du(t)
== Cit == t(t, ~(t»,
so that at any fixed t, u(t) must be a poinJ function of (t, x(t». 'rherefore the state space ~ must include the space n of functions on which the operator d/dt is defined. It is simplest to let 1:. = D. Then Z is usually infinite dimensional because n is. Thus we define the state ~ - x(t) as the function u( T), definedfor all r ~ t. The mapping ",,(t; to,%1.) assignsto the function :to defined for 'T' ~ to the function x, , which is equal to %'0 on r ~ to and equal touonto < 'T' ~ t. In this paper, the finite dimensionality of ~ is used ill an essential way, which rules out considerationof the "system" d/dt in all but trivial cases. 6. Canonical structure of linear dynamical systems. The concept of irreducibility can be understood most readily with the help of the writer's "canonical structure theorem" for linear dynamical systems [2, 14]. Before presenting and illustrating this central result, it is necessary to recall some definitions and facts concerning the contf'ollabil£ty and obseroability of linear dynamical systems. DEFINITION 4. A linear dynamical system (2.1-2) is completely controllable at time to if it is not algebraically equivalent, for all t ~ to, to a system of the type (a)
(5.1)
(b) (0)
d~l/dt
= FU(t)x + F l
12
(t )X2
+ G1(t)u(t)
2
dz2/dt = F'lfJ(t)x y(t)
= H1(t)xt{t) + I/'(t)x 2(t ) .
(In (5.1), Xl and 3;2 are vectors of nl and ftt = n - ftl eomponents respeetively.) In other words, it is not possible to find a coordinatesystem in which the state variables x, are separated into two groups, ~1 == (Xl, • • • J %"1) and x' =- (X"I''''' .. · x,,), such that the second group is· not affected either by the first group or by the inputs to the system. If one could find such a t
250
1
xt - - - xn~
u
2
X."1+ 1- - -
xn -----
F,oUD 1.
coordinatesystem,wewould have the state ofaffairs depicted schematically in Fig. 1. Clearly, controllability is a system property which is completely independent of the way in which the outputs of the system are formed. It is a property of the couple (F(t), G(t»). The "dual" of controllability is observability, which depends only on the outputs but not on the inputs. DEFINITION 5. A linear dynamical system (2.1-2) is completely ob,ervable at time to if it is not algebraically equivalent, for all t ;ii to, to any system of the type (a)
(5.2)
(b) (c)
+ G1(t )u(t ) tb;1/dt == ~(t)Zl(t) + FtI(t),;' + G'(t)u(t)
dzl/dt = FlI(t ).zl{t)
y(t) .. H 1( t )Zl (t ) .
(Again, %1 is &11 nt-vector and Zl is an (n - "'I)-vector.) In other words, it is not possible to find & coordinate system in which the state variables %i are separated into two groups, such that the second group does not affect either the first group or the outputs of the system. If such a coordinate system could be found, we would have the state of affairs depicted in Fig. 2. The above definitions show that controllability and observability are preserved under algebraic equivalence. These properties are coOrdinate-
free, Le., independent of the particul~r choice of basisin the state space. The equivalence of the present definitions with other more abstract 251
1
..XI
J
Xn,
X.f\tl~
2
xn ---F'Gun 2.
definitions of controllability may be found in (4)~ As to observability, we note that the duality relatioo8
(5.3)
(a)
t - to = to - t',
(b)
F(t - to) ~ P'(to - t'),
(0)
G(t - to) ~ H'{lo - t'),
(d)
H(t - to) ~ a'(to - t'),
transform the system (5.2) into (5.1). Hence all theorems on controllability can be "dualized" to yield analogous results on observability. It can be shown that in applying definitions 4-5 to constant systems it is immaterial whether we require algebraic or strict equivalence (14). Henceas one would of course expect-for constant systems the notions of complete controllability and complete observability do not depend on the choice of to. EXAMPLE 1. A simple, well.. known, and interesting case of a physical system which is neither completely controllable nor completely observable is the so..called conatant-resistance network shown in Fig. 3. Let Xl be the magnetic flux in the inductor and %2 the electric charge on the capacitor in Fig. 3, while Ul(t) is 8. voltage source (zero short-circuit resistance) and Yl(t) is the current into the network. The inductor and capacitor in the network may be time-varying, but we assume-this is the constant-resistance condition-that L(t) and ci» are related by:
L ( t) / C(t) == RZ == 1 252
( L( l ), C( t)
> 0).
1.(1)
+
c (t)
FIGURE
3.
The differential equations of the network are dzJ/dt = -[l/L(t)JXl ~/dt
+ Ul(t),
= -[1/O(t)1x2 + 'UI(t),
Yl(t) ~ (l/L(t)]xl - [l/C(t)]X2
+ Ul(t).
If we let
= %2 = il
(ZI
+ :i:z)/2,
(Xl -
%2)/2,
the dynamical equations become dZt/dt == -[l/L(t)]Zl (5.4)
dZt./dt y.(t)
+ 'Ul(t),
== -[l/L(t)]~,
== 2[1/L(t)]is + 1I1(t).*
Here the state variable Zl is controllable but not observable, while i2 is observable but not controllable. For obvious reasons, the subsystem (b) of (5.1) may be regarded as (completely) uncontrollable, while subsystem (b) of (5.2) is (completely) unobservable. In view of linearity, it is intuitiv~ly clear that it must be possible to arrange the components of the state vector-referred to a • Note that this equation does not correspond to (2.2) but to y(t) =- H(l)~(t) (0 4) may be generalized to: "(Dot): Every output is a function of t, z(t), sad u(l). II This entails only minor modifications as far &8 the result. and arlUments of the present paper are concerned.
+ J(t)u(t). This is & minor point. In fact, Axiom 253
suitable (possibly time-varying) coordinate system-into four mutally ex.. elusive parts, as follows: Part (A): Completely controllable but unobservable. Part (B): Completely controllable and completely observable. Part (0): Uncontrollable and unobservable.
Part (D): Uncontrollable but completely observable. The precise statement of this idea is [2, 14): THEOREM 5 (Canonical Structure Theorem). Comider aped linear d1fNJmical 8f/Btem (2.1-2). (i) At every fixed instant t of time, there is a coordinate B1J8Um in tAe state space relative to which the components 01 the 8late vector can be decomposed into Jour mutuaUy exluBive parts
x
:=
(x A , :eB ,
xC,
ZD),
which crm-espond to the 8cheme outlined above. (Ii) This decomposition can be achieved in many ways, but tJuJ number of state vtJriables n.t(t), .. · , nD(t) in each part is the same Jor any BUCk
decomposition. (iii) Relative to BUCk a choice of coordinate" the 'YBtem matrice, have the canonical 101m
and
B(t) = [0 HB(t)
0 HD(t»).
In view of this theorem) we shall talk, somewhat loosely, about ,cParts (A), · · · , (D) of the system." Thus the system (5.4) consists of Parts (A) and (D). The canonical form of F, G, and II can be easily remembered by reference to the causal diagram shown on Fig. 4. It is intuitively clear (and can be easily proved) that algebraically . equivalent systems have the same canonical structure. Unfortunately, the coordinate system necessary to display the canonical form of F, a, and H will not be continuous in time unless 1&A(l)t ••• ,nD(t) are constants. If these dimension numbers vary, we cannot call the various 254
4.
FlOUR.
parts of the canonical structure "subsystems.U For constant systems this difficulty does not arise. More generally, we have: TsEOUH 6. , . a ,,,iodie or analytic Ii"". dfIMmicGl 81/,tem (2.1-2) the dimemicm numbsr, J ftD are constanta, and tM canonical decom-
"A , ···
poaition i8 cmUinuoua with, rN'P'd to t. An illustration of the canonical structure theorem is provided by 2. Consider the constant system defined by
EXAMPLB
==[;: -; ~3 -~]
F
30 30
G
:=I
39 43
-2 -3
[ ~2o :1] 0
0
-27 ' -32
'
1
a.nd
H .. (-5
-8
1
We introduce new coordinates by letting where 3 0
T==[~
-2 -6
&nd
1
0
-3 0 -9 1
:f
5J.
== Tz,
-2] -1
3
'
6
[~ i ~ iJ 2
r: =
With respect to these new coordinates the system matrices assume the 255
canonical form:
P=
[~
rrr: =
G-TG-
4 -1 0
-3
0
0
[0 I] 1
~I]
1 0
-2 ' 1
1
~ ~'
and
8 == Hr 1
= [0
0
1
1].
On the other hand, if we define the new coordinates by
T==[~ i ~ =~] -5 -6
r
1 _
-
-7.5 9
0.5 1
6 6
J
[~3 -3 ~3 0~ -~'5] 1 ' 1
-1
1
-0.5
then the system matrices become
[~ i ~3 fJ l
F=
G=
[i il
and
R
= [0
1
o
IJ.
The numerical values of these two canonical forms are different, yet Theorem .5 is verified in both cases. In the second case the connections from Part (D) to Parts (A) and (C) are missing. This is not a contradiction since Theorem 5 does not require that all the indicated casual connections in Fig. 4 be actually present. The transfer-function matrix of the system is easily found from the canonical representation. The coordinate transformations affect only the 256
internal (state) variables, but not the external (input and output) variables; consequently the impulse response matrix is invariant under such transformations. We get by inspection: Z(s) =
[_1_ _1_] ,+1
8+1·
It would be rather laborious to determine these transfer functions directly from the signal-flow graph [16] corresponding to F, G, and H. EXAMPLE 3. A far less trivial illustration of the canonical decomposition theorem is provided by the following dynamical system, which occurs in the solution of a problem in the theory of statistical filtering [17). Let A be an arbitrary positive function of t and define
F=
G ==
- t«/ 4A
1 OJ
-t"/2A
0
-tL2A 0 1 , [ 0
[~~~~J, t / 2A l
and
H = [0
0).
1
We introduce new state variables i(t) == T(t)z(t),
where T(t)
=
[~-~ ~JJ o I-t 1/2J 1.
2
rl(t)
=
t /2 1/2 0
t
[1
o
0
Then
- t~/4A:- ~/4A:- t'l/4A
:------:. --
-_
o :
0
o
0
:
---
1
- - - - - .....1 __ .. • .. •• I . • .... - __ I I
!_'
t" /2A ] G(t) ... T(t)G(t) -
257
[
__
f
~
0
and H(t) == H(t)r 1( t )
=:
[t 10; 1].
Hence the system consists of Parts (B - D), with n. = ftC == nn = 1. It is interesting that the canonical decomposition is of constant dimension. even though the system may be neither periodicnor &D&1ytic. The preceding examples illustrate special cases of a noteworthy general relationship which exists between the canonical structure of a dynamical
system and irreducible realizations of an impulse-response matrix. The main facts here are the following: THEOREK 7. (i) TAl impuJae-rupome matriz oj a linear dynamical 8ystetn (2.1-2) depends solely on Part (B) the ayBtem andis given explicitly b:1J
0"
(5.5)
Set, T) == HB(t).BS(t, T)GB ( ,. ) ,
where flBB is the transition. maN corre8fJl1lllli1l{J to FBB• (ii) Any two comp1aelll controU4ble and completely oNervable realizations oj S are algebraictJllll equiVGlent. (iii) A reawation oj S i, in-educible if and only if at all times it con8ist8 of Part (B) alone; thus every irreducible realization oj S is completely controllable and completely OOBmJable. Proof. The first statement can be read off by inspection from Fig. 4. The second statement is proved in (14). The necessity of the third statement follows from Theorem 5, while the sufficiency is implied by (ii). It is clear that Theorem 2 is a consequence of Theorems 5-7. We can now answer the question posed in section 4 in a definite way: THEOREM 8 (Main RuuU)1l KMtDletlge oj 1M impulse.reaponse matrix S(t, -r) identifies the completely controllable and completely oo88TVable part, and this part alone, of the dynamical system which generated it. Thia part (UB" in Theorem 5) is it~lJ a dynamical B1/Btem and has the 8mGllut dimenBion G'In01&fI aU realizations oj S. Moreover, thi, part i8 identified by S uniquely 'Up to algebraic equivalence.
Using different words, we may say that an impulse-response matrix is a faithful representalirm of a dynamical system (2.1-2) if and only if the latter is completely controllable and completely observable. Remark. It is very interesting to compare this result with Theorem 4 of
E. F. Moore, in one of the early papers on finite automata (26): •'The cl4Bs oj aU machines which areindiBtingui.h4ble from a given strongly connected machine S by anysingle 6Zperiment has aunique (up toi8omorphism) member with 4 minimal number of stale,. This unique 'fTUJChine, called the reduced form of S, ia atrtmgly connected and has Ik property that any two oj its states arB di.tinguishable. " "Indistinguishable machines" in Moore's terminology correspond in ours to alternate realizations of the same input/output relation. "Strongly con258
nected" in his terminology means completely controllable in ours. "Indistinguishable states" in our terminology corresponds to states whose difference, not zero, is an unobservable state in the senseof [3]. Evidently the two theorems are concerned with the same abstract facts, each being stated in a different mathematical framework.
8. Exp6cit criteria for complete controllability and
observabillty~
The
canonical structure theorem is so far merely an abstnct result, since we have Dot yet given a constroctive procedure for obtaining the coordinate transformation which exhibits the ~tem matrices in canonical form. We shall do this in section 7. The method rests on the possibility of finding explicit criteria for complete controllability and complete observability. The following lemmas, proved in [4J, playa central role: LEIOIA 1. RA(lo) + n.(ia) III: rank Wet. , t1 ) for t 1 > to sufficiently large,
where
111 ~(to.1")G(1')G'(1')~'(to. 1') d1'
W{to, t1) =
(6.1)
'.
or
(6.2)
dlV/dte
LEMMA
smGll,
c:
2. ne(~)
F(t.)W
+
M(to,LI) =
r
'-l
= O.
- G(to)G'(to), W(t.)
no(to) = rank JI(t;., '-1) for l-l
tDMr6
(6.3)
+ WF'Cto)
<
to 8ujficiently
~'(1',")H'(1")H(1')~(1".to)d1"
or
(6.4) -dM/dto == F'(t.)M
+ MF(to)
- H'(lo)H(to), M(L1 )
:II
o.
For COnstallt systems, the preceding lemmas can be considerably improved (4]:
LmoIA 3. For a conBtant system,
n.. + n.
(6.5) LEMMA
· .. , F JI-
1G].
4. Fur a comtam syaletn,
(6.6) EXAMPLE
(6.7)
= rank [G, FG,
ftC
+ 1&D =
rank
[H', F'R',
.. , , (F')"-IH'J.
4. For F and G defined in Example 2, the matrix (6.5) is
[-~
3 -1 3 1
-3
-3
6
8 18 6
12 4
259
:J 2
12 4
~J
6 24 8
-3 14 36
12
-;~l
~J
=
The rank of this matrix is 2, which checks with the fact that nA 1 and nB = 1 in Example 2. The determination of the rank of (6.7), while elementary, is laborious.
For practical purposes it might be better to compute W; for instance, by solving the differential equation (6.2). In the constant case, there is another criterion of complete controllability which is particularly useful in theoretical.investigations. The most general form of this theorem (which may be found in [14]) is complicated; we state here a simplified version which is adequate for the present purposes: LEMMA 5. Hypothesi,: The matrix F is similar to a diagcmal matrix. In other word" there is a nonBingular coOrdinate tramjormation i = T» UJith the property thatin the new coOrdinate system F ha« theform
1
F =TFr =
~l III
0
lr
o
x, t.;
and the maJ,riz G haa the [orm
G = TG = a(r)
}
q,. rows.
CnnclU8ion: The system is completely controllable if and only if (6.8)
rank
0(1)
== ql , · .. , rank G(~)
==
qr.
We leave it to the reader to dualize this result to complete observability. 5. Consider the special case ql := ••• = qr = 1 of Lemma' 5. The eigenvalues of F arc then distinct. If condition (6.8) is satisfied, every element of the one-column matrix (} is nonzero; by a trivial transformation, all of these elements can be made equal to 1, without affecting F. Thus we can choose a coordinate system in which F, G have the representation: EXAMPLE
(6.9)
P=
:1
.~] ~i~ o, =
260
i = j)t 0 =
Dl
This is the canonical form of Lur'e (18). It is closely related to the partialfraction expansion of transfer functions. To illustrate this, consider the 1 X 1 transfer-function matrix:
() _
Zll 8
-
8+2
(8
_
%
J,~
%
+ 1)(8 + 3)(8 + 4) - 8 + 1 + 8+3 - 8+4·
This transfer function is realized by the (6.10)
F
=:
sys~m:
[-~ -~ ~J, o
0-4
(6.11) and
(6.12)
-f]
H == [t
which is in the canonical form of Lur'e. By Lemma5, (6.10-11) is completely eontmllable, by the dual of Lemma 5, (6.10-12) is completely observable. We can double-check these facts by means of Lemmas 3-4. For (6.9) the matrix (6.5) is
(6.13)
2
1
Al
).1
1
~2
).·l
1
~"
-:
Ai-Il ~i-l
oJ
>.:.-1
where the ~, are the diagonal elements (== eigenvalues) of F in (6.9). But the determinant of (6.13) is the well-known Vandermonde determinant. The latter is nonzero if and only if all the Ai are distinct, which is
what we have assumed.
1. Computation of the canonical structure. We show now how to deter.. mine explicitly the change of coordinates which reduces It', G, H to the canonical form. We consider only the ooDStant case of (2.1-2). The computations are elementary; it is Dot necessary to diagonalize the matrix F or even to determine its eigenvalues.. The procedure is as follows: (a) We compute the controllability matrix W = W(O, 1)* given by • It can be shown [4, Theorem 10] that in the constant case one may
'1 > to in Lemma 1.
261
0110088
any
(6.1); for instance, by solving the differential equation (6.2). Then we find
a nonsingular matrix T such that (7.1)
T'WT
=E
""
[101 ~J
where I-I is the 1&1 X nb 0 ~ nl ~ fI, unit matrix and the O's are zero matrices of appropriate size. Clearly nl = n.t + n. is the number of controllable state variables. The matrix T defines the change of coordinates (7.2)
in terms of the new coordinates, the system matrices are (7.3)
l'
(7.4)
Z =
G == riG,
=- riFT,
- [flJ i'
J
F =
R = BT,
1 ,12J til, (j == [00 ] ,and
[Pl10
H;
= E.
H __ [HI
This decomposition is trivial (and therefore omitted) if nl == n, i.e., when the system is completely controllable. (b) Next we consider the two subsystems defined by (7.5)
1'11, Gl ,
and HI;
Fa, 0,
and
trw
We compute the observability matrices Jll .. Jll( 0, 1) and 142 = if' (0, 1) given by (6.3) for both of these su6systems. Then we determine two nonsingular matrices 0 1, 0 1 such that
(O')'R'O' = 2 1 =
(7.6)
[~ I~J •
and (7.7)
These results define another change of coordinates f
[ZlJ u.!- == [010
== ~ ==
One or the other of these transformations is superfluous if n. fId
~
11. -
= nl or
1&1 •
Mter the coordinate changes (7.2) and (7.8), we obtain the following 262
matrices
FAA
o
(7.9)
FA"; FAt
FAd
F BB :
FBI,
,Ba
I ..... ··-··---1'--------·
!=
o o
olF
B
C
0 : 0
J
Fed.
ru
o
Clearly, n. is the number of state variables which are both controllable > nc . (c) It remains to transform the element into 0, if this is not already the case. (If p.e == 0, then ftc: = ftc , n4 = nD and (7.9) has the desired canonical structure.)
and observable.But, in general, n.. < nD and n c
,Bt:
We consider the subsystem (7.10)
r -
[-~~._~-~-] o ; pee
J
(j* =
[_Go-~]'
and 11* = [H B : 0].
The corresponding observability matrix given by (6.3) is .&1'*(0,1) ==
1•• : A ]
M* = ----: -....- , (Q = Q' nonnegative definite.) [
A' : Q
fl.
(The upper left element of M* is I in viewof (7.9) j all we know about the other elements is their symmetry properties.) Letting
'V*
:III
[-~·~-l.:'~-J o : In.
t
we find that
(1' ) M V == •
I -..
-. M ==
I •• : 0 ]
[
----:---- ,
o :
R
where R .. Q - A'A is & symmetric, nonnegative..definite matrix. 263
Now let
f** be a nonsingular matrix such that -1'1.* '-**17•• (y)M - [
where n. == rank R. Let
V == 17* V··
I na : 0
0
0: 0
0
o :0
In,
----:
Since f* and
relative to the partitioning in (7.10), so is upper triangular form FBB
I
V,
o
J'
V·· are upper triangular
which will take , . into the
BC
B
F
F
F· c
r:
• ]
~1-IF*V = --~-'ll;~---;;: [
l
·
where nc = n e - n.. But these transfonnations decompose 1* into & completely observable and an unobservable part. Hence = F'c = o. Moreover,
"C
R*f! =
[H
B
Orr =
!
[H
B
i 0
H·]
THEOREM 9. The explicit tramformation which takes the constant matrices F, G, and H into the canonical form required by Theorem (5-iii) is given by z --+ trlo-Irlx. We partition FAt
== [FA C FA.],
and partition
r: = [~4J Then we define nD = nil + 11, and find
r" ==
[Fo. o
FAd],
F B D == [F Be
F Bd ],
== [Fe.
F(.°ci],
IlcJD
F [r'o DD
HO
=
= [He
F~J p-tl' Hdl.
8. Construction of irreducible realizations. Now we give an explicit procedure for the construction of an irreducible realization of a weighting-function matrix Wet· - T). In view of Theorem 7, 264
part (iii), we can do this in two stages: (I) We construct a realizationof W, then (II-A) we prove, using Lemmas 1-5, that the resultant systenl is completely controllableand completely observable, hence irreducible; or (II-B) we carry out explicitly the canonical decomposition and remove all parts other than (B). Instead of the weighting-function matrix lV, it is usually more COilvenient to deal with its Laplace transform Z. Let us consider the problem with Method A in order of increasing difficulty. CaBe 1. m ::: p = 1. This is equivalent to the problem of simulating a single transfer function on an analog computer. There are several wellknown solutions. They may be found in textbooks on classical servomechanism theory or analog computation. Without loss of generali~y (sec Theorem 4) we nl&Y consider transfer functions of the form (8.1)
a. ,.-1 + ... + al N(B) = 8" + b. 811-1 + ... + b1 = D(8)
%11(8)
where the a,. J • • • oJ 01 ; b,., • •• , b1 are real numbers. Of course, at least one of the a, must bedifferentfrom sero, We assume also that the numerator N (8) and denominator D (B) of Zu(B) have no common roots. There are two basic realizations of (8.1). See Figs. 5-6, where the stand.. ard signal-flow-graph notation [16] is used. In either case, one verifies almost by inspection that the transfer functions relating Yt to 1'1 are indeed given by Z11 • In Fig. 5, the system matrices are
(8.2)
o
0 0
1
0
0
1
0 -hI
-bl
0 0
o
F= 0
0 -b3 0 0
(8.3)
G== 0
1
and (8.4)
H ==
[
as
0.1· 265
FIGURE
5.
In Fig. 6, the system matrices are.
(8.5)
F==
0 1 0
0 1
0 0 0
0 0 0
0
0
0
1
0
-b1l -bl
-:ba J' -b,.
al !J2
G= aa ,
(8.6)
aft
and (8.7)
H = [0
0
1J.
0
It is very easy to check by means of (6.5) and (6.6) that the system (8.2, 3) is completely controllable and (8.5, 7) is completely observable. However. if we attempt to check the controllability of (8.5, 6) by means of (6.5) we get a matrix whose elements are complicated products of the coefficients of N(s) and D(8). To prove that the detenninant of this matrix does not vanish, we have only one fact at our disposal: the assumption that N(B) and D(8) have no common roots. Guided by this observation, we find that the following is true: LEMMA 7. Suppose F has the/arm (8.5) and G has the/arm (8.6). Then (i) we have the relation
(8.8)
K(F, G) = [G
FG
... 266
F,,-lGJ
=
N(F),
and (ii) thepolvnomial, N(,) and D(a) MtM no roo'in comnum if tmd only if det K(F, G) sA O. The main fact to be proved is (ii), for then the complete controllability of (8.5, 6) follows by Lemma 3. A straightforward way of establishing (ii) is to transform the standard Euler.sylvester determinantal criterion [19, p. 84] for the nonexistence of common roota of N(a) and DCB) (the so-called raolwn' 01 N(s) and D(.» into the form (8.8). This can be easily done, but the details are not very transparent. Therefore we prefer to give another Proof. Let B, t i -= 1,···, n, be the set of n-vectors in which the j-th component of e, is Iii. Since F is given by (8.5), we see that 6'+1 -= Fe" 1 ~ n - 1, and K(F, '1) ::z (61, It, · · · , e.] =- I. Hence K(F, ei) == K(F, Fl-Jel) == F'-IK(F, '1) == r», when 1 ~ i :; n. Then (8.8) follows by linearity. Let ~,{A], i .. 1, ... ,n denote the eigenvalues (not necessarily distinct) of & square matrix A. Then
det K(F, G)
" ~i[N(F») =n ,.1
:II
nN(AdF]),
••1
where the second equality follows from' (8.8) by a well-known identity in matrix theory. Thus det K(F, G) == 0 if and only if N(~.{FJ) == 0 for some i; that is, when an eipnvalue of F is a root,of N()'). Since the eigenvalues of F are roots of D(A), this proves (ii).* It is interesting that (8.8) provides a new representation for the 18" solvent, which is preferable in some respects to the Euler-Sylvester determinant. The latter is a 2ft X 211, determinant, whereas det K(F, G) is n X n. The complete observabilityof (8.2, 4) is proved similarly. The systems given by (8.2-4) and (8.5-7) are duals of one another in • The present prool or Lemma 6 was suggested by Drs. John C. StuelpDqel and W. M. Wonham of RIAS.
267
the sense defined by (5.3). Fig. 6 is a reflection of Fig. 5 about the vertical axis, with all arrows reversed. A third type of realization in common use is obtained from the partial. fraction expansion of %11(') (see Example 5). Note, however, that this requires factorization of the denominator of ZI1('), whereas the preceding realizationa can be written down by inspection, using only the coefficients of %11(8). These coDSiderations may be summarized as the following result, which is a highly useful fact in control theory: THEOREM 10. CtmBidera linear constant d1fR/Jmical81/8temtoith m = p = 1, which i, completely controllable and completely observable. Then one may alway, choose a basis in the 8tate space 80 that F, G, H have the form (8.2-4) or (with respect to a different bais) (8.5-7). Proof. Let (8.1) be the transfer-function matrix of the given dynamical system. By Theorem 8, the given system is an irreducible realization of (8.1). So are the systems specified by (8.2-4) and (8Ji-7). By Theorem (7-ii)J all three systems are algebraically equivalent and by constancy (Theorem 3) they are even strictly equivalent. Extensions of this theorem may be found in [14). For an interesting application to the construction of Lyapunov functions, see (25). The procedure described here may be generalized to the non-constant case. Assuming the factorization (4.3) of 8 (I, If' ) is known ( with m == p == 1) J Batkov [20J shows how to detennine the coefficients of the differential equation dftYa/dtfl
+ b.(t)fI-lYt!dt"-l + ... + b1(t )YI
/~ ,,-1 == a.. ()d,,-I t 1'1 lit
(8.9)
+. .. + 01(t )Ul •
Laning and Battin [21, p. 191-2] show how one converts (8.9) into a system of first-order differential equations (2.1) with variable coefficients. We shall leave to the reader the proof of the irreducibility of the realization so obtained. CaseS-a. m == 1, P > 1. We have a single-input/multiooQutput system. We can realize Z(s), without factoring the denominators of its transfer functions, by the following generalization of the procedure given by Fig. 5 and (8.2-4). . First, we find the smallest commondenominator of the elements of Z (8) . (This can be done, of course, without factorization.) Z(s) assumes the form
[at"
1 Z(,) = ,,, + b" ,,,-I
+ ... + b1
8,,-1
apt 8
+ ... + an]
..-1 +'· · •••
+
ap1
·
Then the following dynamical system provides an irreducible realization 268
of Z(s): F and G are as in (8.2-3), while H given by (8.4) is generalized to H ==
at].
[a;1
OJ'"
aJ'l·
Complete controUabUity is trivial; complete observability is established by a straightforward generalization of Lemma 6. In this case we form p linear functions of the state) rather than merely one 88 in Fig. 5. CQ8sl..b. m > 1, p = 1. We can realize this multi-input/single-output system analogously to Case 2·a by generalizing the procedure given by Fig. 6 and (8.5-8.7). Let us write the elements of Z(8) in terms of their smallest common denominator: z(,)
=[
~ + ... au + bra 8,,-1 + .o ••. + b Sft-l
s-
1
•••
J.
+ ~ .. + + b; 8,,-1 + ... + b1
aft. 8.,,-1
s"
al M
Then the desired irreducible realization consists of Ii' and H as defined by (8.5-6), while
a~.] . .
au
G= [ : C%,al
a..
This case is the dual of Case 2-a. Even in Case 2, it is impractical to give a general formula which expresses the coefficients of F, 0, and H in terms of the coefficients of the transfer functions in Z (8) if the denominators are not all the same. When we pass to the general case, determination of F, G, and H often requires extensive numerical computation. CfJB6 S. m, p arbitrtJrJj. Here Me~hod (A) is very complicated if any tranafer function in Z(,) has multiple poles [14J.o In most practical appliea-
tiODS, however, such complications are of no interest. Ruling them out, E. G. Gilbert gave an elegant and relatively simple solution [5J. Let 81 , • .• J at be distinct complex Dumbers corresponding to the poles of all the elements of Z(B). Assumethat all polesare simple. Then
......
R(k) = lim (8 - S.)Z(B), k
=
I, · · · J q
8, =8i, then R(8t) = R(sl:). where the bar denotes the complex conjugate. In terms of the residue matrices, the weighting-function matrix Wet) corresponding to Z(s) bas the explicit form
is the k-th residue matrix of Z(s). If
q
Wet)
= .c- rZ (8)] = L R(k)e· k_l 1
k
'.
We have then: THEOREM 11. (Gilbert). H1fPOI,haeB: No element oj the tTa:nsfer-functioo 269
matrix Z (8) has multiple poles. Z (,) hCJ8 a total oj q distinct poles 81 , • • • , wifA corresponding residue matrice8 R ( 1») ... , R(q). Conclmitm8: (i) The dimension of irredU£ible realizations of Z(s) is
8'1 ,
fI
(8.11)
n
=L
~-l
rot , where TA: = rank R(k).
(ii) Write R(k) = H(Ic)G{k).
(8.12)
k =
whereH(k) iBOp X Tl:matrizandG(k) is anT. X Then Z (8) has the irreducible realization 81 Irk
(8.13)
1,··· ,q,
mmatriz,botAoJTankr~.
0
F=
(Ir
=r X
r unit 7natrix),
o (8.14)
G=
G
and (8.15)
II = [H(l)
•..
l[(q)].
Proof. This is one of the main results ill [5J. With the aid of machinery developed here, we can give & shorter (though more abstract) demonstration. The fa.ctorization (8.12) is well known in linear algebra. We give in the Appendix various explicit formulae (which are eaSily machine-computable) for G(k). and H(k). Applying Lemma 5 shows that the dynamieal systcnl defined by (8.13-15) is completely controllable and completely observable. Hence it is irreducible, which implies formula (8.11). By elementary changes of variables, (8.13-15) can be transformed into matrices which have only real clements. A serious disadvantage of Method (A), as expressed by Theorem 11, is that the denominators of the transfer functions in Z(s) must be factored in order to determine the poles. This is not easily done numerically. Moreover, the residue matrices R(k) corresponding to complex poles are COD1-
plex, which makes the factorizatioll (8.11) more complicated (see Appendix). Now we turn to Method (B). This method does not require computation of eigenvalues, and it is not bothered by multiple poles. This is a decided advantage in numerical calculations. On the other hand, the method is not convenient for simple illustrative examples. Nor is it possible to display the elements of F, G, and H as simple functions of the coefficients in %ij(S). 270
An easy way of realizing Z(a) (without guaranteeing irreducibility) is the following. Let a, be the number of distinct poles (counting each pole with its maximum multiplicity) in the ...th row of Z(s). and let fJi be the number of poles in the i-th column. Then the maximum number 'to of state variables required to realize Z(,) by repeatedly using the scheme given under Case 2-a or 2-b is
no ..
min {
t: a, , t: Ili} · ,_1
J-l
As before, we can determine the a, and (Jj without factoring the transfer functions oE Z (s). There is in pneral no simple way in this method to determine the dimension" ~ 110 01 irreduciblerealizations without performing the computatioDl outlined in Section 7. The two methods are best compared via an example. This example must be of fairly high order, since we wish to provide accurate numerical checks. ExAMPLE 6. Consider the transfer-function matrix
z(,)-
3(.+3)(.+5) (.+1)(.+2)(.+4)
8(.+1) (.+2)(.+4)
2 (.+3)(.+6)
1 (..+3)
2(,1+7.+18) (.+1)(.+3)(,+8)
2a (.+1)(8+3)
21+7 (_+3)(.+4)
21+5 (.+2)(.+3)
2(,-5) 8(.+2) (.+1)(.+2)(.+3) (,+1)(.+3)(.+6)
1
2(511+27.+34)
(.+3)
(,+1)(.+3)(.+5)
Applying Method (A) first, we find that the residue matrices are: R(l)
~ ~ ~J;
= [: 3
R(2)
III
R(3) =-
R(4) =
R(5)
[
rl = 3.
103
- 04.5 -30 o
0 0; IJ '1 == 2.
-6 0
0
01 01 IJ 1 2 2
[-3
[-~~ o
= [-~ 2
-3
1
0 j
ra
=
2.
1
~ ~J;
r. = 1.
~0 0~ -~J; 6
r, == 1.
:
000
ThusR = 9. 271
Employing the procedure given in the Appendix, we find the following factors for matrices R(k) (the products are accurate up to four places beyond the decimal point):
H(l)
==
[
8.ססOO 0.ססOO 0.ססOO] 0.ססOO 3.ססOO
0.ססOO,
4.1231
Oj276 an774 0(1) ==
H(2)
==
[
1 .ססOO
0.ססOO 0.ססOO
0.ססOO 0.9701
0.ססOO
0.3249
-0.2294
0.ססOO
505000 0.ססOO] 0.ססOO
6.ססOO
[ 0.ססOO
t
0.ססOO
G(2) == [-O.~182
1.3416
0.ססOO
G(3) = [ 0.2236 - 0.6708
9.0692] [ 0.ססOO
0.1818J. '
0.ססOO
-1.ססOO
0.ססOO
0.2236
0.6708 0.6708]. 0.2236 0.2236 '
0.4472] -0.4472, 4.4721
= [ 3.1305
H(4) =
0.ססOO
-0.5455
0.0000
H(3)
0.ססoo]
0.2425; 0.9175
G(4)
J
- 0.6708
== [-0.0551 0.9924 0.1103
0.ססoo];
0.ססOO
H(5)
0.ססOO ] == -3.1623 , [ 6.3246
0(5) == [0.3162 0.ססOO
0.ססOO
0.9487].
Using these numerical results, we find that the dynamical equations of the irreducible realization are given by 1 0 0 0
0
1 0 0
F= 0 0 0 0 0 0
0 0 0 0
0 0 1 0 0 0
0
0 0
0 0 0 2 0 0 0
0 0
0 0 0 0
2 0 0 0
0
0 0 0 0 0
0 0 0 0
3
0 3 0
0
0 0
272
0
0
0 0 0 0 0 0 0
4 0
0 0 0 0
o , 0
iJ
1.ססOO
G=
H -
[
I
0.ססOO
0.00001
0.0000
0.ססOO
0.ססOO
0.9701
0.2425
0.ססOO
-0.8182
0.3249 -0.5455
-0.2294 0.0000
0.9175 0.1818
0.ססOO
0.ססOO
-1.ססOO
0.0000 ,
0.2236
0.2236
0.6708
0.6708
0.6708
0.6708
0.2236
0.2236
-0.0551 0.3162
0.9924 0.ססOO
0.1103 0.0000
O.OOOOJ 0.9487
l
8.ססOO 0.ססOO 0.ססOO 5.5000 0.0000 1.3416 0.4472 0.0692 0.0000 ] 4.1231 0.ססOO 0.ססOO 6.ססOO 3.1305 -0.4472 0.ססOO -3.1623 • 0.7276 3.0774 0.ססOO 0.ססOO 0.ססOO 4.4721 0.ססOO 6.3246
0.ססOO 3.ססOO
Now we apply Method (B). First of all wenote that a, =: at == 4, Qa . . 3, while fJl == 5, ~2 == {JI == 4 (see p. 181). Hence it is best to choose for the preliminary realization three structures of the type discussed under Case 2-b. This will require no =: peal crt a,) == 11 dimensions.
+ +
Next, we find the least common denominator of the rows of Z(8). Sec Fig. 7. 2 3(.' ... 11. + 21, + ..,) 101' + ".2 + ~ +
.Ii:.
2"
6(,' + ,,2 .. 7, ... '1
.' ... a.2
z(.) •
2s'
+ 1....2 ... ;0,.
...
1~ .
?8' . .
15,2 + '" + 20
8(;2 + It. + If)
+ 17.... 10
2(.2 + 7.... 18L
2
. ' + 9. + 2,. .. l.5
FIGURE
7.
The desired realization of Z(a) can be read off by inspection from Fig. 7, using (8.5) and (8.6): I a 0 0 -24;J I 1 0 0 -50 1 I o 1 0 -35: 0 : 0
F=
r I~--~--~-=~~l------------l. .-------- I : I I
l
0
0 0 0 0 0 :0 1 0 '0 0 1
I1
-30: -61 i -41.:
-11
0
•
I
i
-- · --~--_· -I- .----.~ --_. ·l~- ~. =i~J ! 0 1
I
273
-9:
135 117 33 3-
18 42
30 6
14 25 13
20
2
2
50 20 2
32
33
15
--_ ... ------.-....----4
G=
6
10 17
8
2 0
-..
1
.......... --_ ...
~_
....
32
o.
~--
0 36 14 --10 2 -2
t
8 0
.....-.
5 68 6 54 1 10
and
0 0 0 1 0 0 0 0 0 0 0] ~~~--~-~-------~-~~-~--~
H= 0 0 0 0 0 0 0 1 0 0 0 .
[
~~--~~------~--~~~~~~-~~-
o
0 0 000 000 0 1
By virtue of its construction, this system is completely observable but we cannot tell by inspection whether or not it is completely controllable. (From the results obtained above with Method (A), we know that the system is not completely controllablesince 11 = no > n := 9.) Therefore the canonical decomposition may contain Parts (B) and (D). To see what the dimensions of these parts are, we compute numerically the decomposition of the system into completely controllable and uneontrollable parts according to the method described in Section 8. These calculations involve only the matrices F and 0, but the,resulting transformations must be applied also to the matrix H.
~.
-o.,,a.6
0.U8e
0.0119
O.OrJI1
-0.0001
O.2~"
-o.~
-0.0115
0.0268 -0.0101
O.DOOO
O.01!A 0.0120
1.Ci'"
..).oco,
-0.9'»8 . 0.02,.,
-0.029)
.().066'
.().on,
,
0.0000
O.S,3k
, o.tJCCX) -O.Ol2O • :u.546'
0.718'J
0.0000
I
0.0001 :
.,
-0.8""
.o.~
-0.102'
.0.29~'
U.20ja
0.0361
0.0022 ..().0610 -o.~
O.lT7'
..0.0569
..o.eaM
0.8Jl1
0.18;1
O.~
-e.~!t7
~.027'
1.1J99
0.0Ge"
0.008; : ...i.1M2
0.2-'77
1.~(
2.04» ..0.",., ..0.968, -o.~~
-o.~
o.~,.,
0.0199 1.a.62.24!21
a.cow -o.om
0.0016
..().~
-o.o~
..Q.oo~ :
O.tlOOU
o.L098
1.1616, -o.Ji2}} -0.0196
0.~7
-0.01180
-o.ooa.o : - 25.8717
- 2.1~1l
-0.&801 0.8992 0.896, ..o.~ .• -.•.• _.•2.a.60a. _. ______ . __ .a._ .. _. ___
.0. 26119 ~ .°(.1.0. m1
-0.11)8
0.01129
1.16'"
.0.0290
O.U14
2.~
-0.2787
-e.2'.!~
0.0'l5
. .1._
O."l~
I
_
1.6m,
I
~.96~
0.8726 2.,~
-1,.0068
••• __ .•-~.C17C6 - ..
._~_._.~
(1.0000
0.0000
,0.0000
o.~
c.ceoc
o.~'O
0.0000
0.0000
0.0000
~
:>., :'2
- 0.000'
0.0000
0.0000
0.0000
0.0000
0.0000
O.OCXJO
0.0000
0.0000
O.OOCO
~ - o.olaCU,
- 0.2-7)'
FIOUBE
274
8.
-
xlO
1.2012
~.
-0.7886 -0.,..,; -0.')20
1.6822
o.60:s
o.mll O.47k'
..1.79Q
O.86JIO
2.'381 -1.7927
.0.168, .o.2!f20
2.2981
-0.066,
O.,~
1.U"
2.7~
1.~12
1.273k
1.,8)0
-2.9,ea
-0.1117 -0.:»'6 -0.0'111
1.31,. -2.8589 -2.~
~.
~
-1.76yt
1.9252
O.a.O~
xlO
1.'»4 -o.498J
0.9281 -2.'137
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
..Q'2928
O.a.a16 0.0160 -0.011'
O.Ollal
0.0000
-0.0981 .0.0167 -0.0001
-0.0286
0.002' -0.0116 ..0.018]
0.0)7'
0.0022
-o.08~
.o.om
0.211'
0.0"" -0.060, -o.OOj]
o.61~
0.0l61
FIOUKI:
-0.00"
0.Dl78
0.0000
-0.0007 1.0000 O.OO2JJ
0.0000
l'OOOO~ 0.0000
0.0000
9.
The final results may be seen in Figs. 8-9, whichgive the matrices P, fl, and fl. Elements in the lower left-hand corner of I'should be exactly zero. In fact, they are zero to at least the number of digits indicated in Fig. 8. To check the accuracy of thesetwo irreducible realizations of the transfer function matrix on p. 181, we have computed the corresponding weighting.. function matrices W(1)(t) and W(2)(t). The equality WO)(t) == WCt>(t) was found to be correct to at least four significant digits. 9. Other applications to system theory. The literature of system theory contains many instances of errors, incomplete or misleading solutions of problems, ete., which can be traced to a lack of understandingof the issues discU88ed in this paper. This section presents somecasesof this known to the writer; other examples may be found in the paper of Gilbert [5]. AMlog computers. According to Theorem 8, a linear dynamical system (2.1-2) is a 4'faithful" realization of an impulse-response matrix if and only if it is irreducible. Suppose the dynamical equations (2.1-2) are programmed on an analog computer. (See [8].) Then it is clear from Theorem 8 that tM oompu,ter wiU simulate t1uJ impulse-reapons6 matriz ctm'6Ctly if and only if a minimal number of inUgrators are used. Otherwise the system programmed on the analog computer will have, besides Part (B), at least one of the Parts (A), (0), or (D). Since the impulse-response matrix determines Part (B), and that alone, the natureof the redundant parts will depend not on the impulse-response matrix but on the particular method used to ob275
tain the dynamical equations. It should be borne in mind that the canonical decomposition is an abstract thing; usuany it is not possible to identify the redundant integrators without a change of variables. The writer is not aware of any book or paper on analog computation where this is explicitly pointed out. But the facts of life seem to be well known (intuitively) to practitioners of the analog art. That redundancy in the number of integrators used can cause positive harm is quite clear from the canonical structure theorem. EXAMPLE 7. Let the simulated system consist of Parts (A) and (B) and suppose that Part (A) is unstable. Because of noise in the computer, Part (A) will be subject to perturbations; they will be magnified more and more, because of the instability. As longas assumptionsof linearity hold exactly, the unstable (A) component of the state vector will not be noticed, but soon the computer will cease to function because its linear range will be
exceeded. Lur'e canonical form. In his book on the Lur'e problem, Letov implies [18; equation (2.4) and (2.23)] that every vector system (9.1)
tk/dt
= Fz
+ g·a
(0'
= scalar)
can be reduced to the canonical form dxi/dt = ~.-z,
(9.2)
+ e,
i == 1, . · · J n
whenever the eigenvalues Ai of F are distinct. Since (9.2) is completely controllable, this assertion, if true, would imply that (9.1) is also com.. pletely controllable, which is false. In fact, the system defined by (9.3)
F
= [~:l
is obviously not equivalent to (9.4)F =
[~
:l
9
=
GJ
g=GJ
whenever X ¢ p.. In examining the derivation originally given by Lur'e for his canonical form [27; Chapter 1, §2-3], it is clear that the last step before equation (3.5) is valid if and only if det [H.(~p)] ~ 0 (in the notation of Lur'e [27].) It is easy to show that this condition is equivalent to complete controllability, whenever the eigenvalues of F are distinct. Unfortunately, the condition det [Hk(A p ) ] ~ 0 was not emphasized. explicitly by Lur'e [28] in the original publications, We may thus conclude that when F has distinct eigenvalues and there is a 276
Bing16 control variable, IAe Lur'e-Letov canonical form exists if and only if the pair {F, g} is completely controllable. It is interesting to note that (9.3) can be transformed into (9.4) when A = p.; in other words, when the eigenvalues are not distinct the Lur'e canonical form. may exist even if the system is not completelycontrollable. Cancellation. in the transfer-function. When a mathematical model is derived from physical principles. the equations of the system are in or near the form (2.1-2). Regrettably, it has become widespread practice in system engineering to dispense with differential equations and to replace them by transfer functions Z(a). Later, Z(s) must be converted back into the form (2.1-2) for purposes of analog computation. In the processof algebraic manipulations, some transfer functions may have (exactly or very nearly) common factors in the numerator and denominator, which are then canceled. This is an indication that a part of the dynamics of the system is not represented by the transfer function.
Such cancellationsare the basic idea of some elementary design methods in control theory. These methods do not bring the system under better control but merely "deeouple" some of the undesirable dynamics. But then the closed-loop transfer function is no longer a faithful representation of the (closed-loop) dynamics. Stability difficulties may arise. Similar criticisms may be leveled against the large, but superficial, literature on "noninteraeting" control system design. EXAMPLE 8. Consider the systemdefined by the matrices
(9.5)
F ==
[~-2 0~ -2~] ,
G =r
]
,
H == [- 2 1 0].
"1 is the sum of two terms:
= -2 2;1(') + XI(S} Ul(')
UI(')
(9.6)
~
0.5
The transfer function relating 111 to 1/1(8)
[
Ii'
1£1(8)
+ 2a1 ~ 68 (8 - 2)
== (8 + 1)(8 - 2)(8
6 + Ii'
+ 2B2 ~ 5. -
6
1
+ 3) = (8 +-1~)(s-+~3) •
Thus, by cancellation, the transfer function is reduced from the third to the second order. The system has an unstable "natural mode" (correspond. ing to 8a == 2) about which the transfer functions gives no information. Using (6.5) we see that the system (9.5) is compJetely controllable. By Theorem 5, the system cannot be completely observable: n. = 2 from (9.6) and Case 1, section 8. The canonical structure consists of Parts (A) 277
and (B). In canonical coordinates the system matrices can be taken as
F=
-1 0 OJ 0 -3 0 , [ 002
11 = (O.5
-0.5 0).
We can easily calculate the change of coordinates Z
:-=
Tz
by the method of partial fractions discussed in (8).. First we find r
1
,
then T.
The results are
r:
=
~[~5 10
:9 :J' 6
-1
T
=
[-;1 8
~14
:J. 2
Loss oj controllability and ooservalnlity d1MJ to BGmpZlng. Consider a single.. input/single-output constant linear I1Ystem. Suppose the output is observed only at the instants t = kT (k = integer, T > 0), and that the input is constant over the intervals kT ~ t < (k + 1 )T. This situation is commonly called "sampling"; it arises when a digital computer is used in control or data processing. T is the 8omplift(J period. We can regard such a setup as a discrete-time dynamical system. We define here 9 (Axiom (D 1» as the set of integers and replace (2.1) by a difference equation. AU theorems carry over to this situation with small modifications. The analysis of discrete-timesystems by eonventiona!techniques requires the computation of the so-called ,-transform of Z(8) [22J. The analysis using ~-tranRforms then proceeds in close analogy with analysis based on Laplace transforms.
A constant linear system whichis completely controllableand completely observable will retain these properties even after the introduction of sampling if and only if (4)
(9.7)
Re
8i -
Re s, implies
lin (8, -
8J) ~
rrr/T
where i, j == 1, · · · , .", and q == positive integer, If this condition is violated (the sampling process "resonates" with the system dynamic8) then cancellations will take place in the ~-transform. The z-transform will then no longer afford a faithful representation of the system, so that if (9.7) is violated, re8UUB based on formal manipulation of %-tTQ,ftIJ!orms may bs invaliil. This point is not at all clear in the literature. True, Barker (23) has drawn attention to a related phenomenon and called it "hidden oscillation." The textbooks, however, dismissthe problem without providing real insight [22. 15-3; 24, 12.13]. 278
A practical difficulty arises from the fact that near the "resonance" point given by (9.7) it is hard to identify the dynamical equations accurately from the z..transform. Small numerical errors in the computation of the ,-transform may have a large effect on the parametersof the dynamical equations. REFERENCES (I} R. E. KALMAN, DiBeuarion 01fltJper b1l1.Flv,gga..Lou, Proc. 1st International Conference on Automatic Control, Moscow, 1960; Butterworths, London, 1961, (2)
Vol. I, pp.398-7. . R. E. KALMAN, Canonical ,,""t'Mr6 o/liftBar dynamical '1I8eem., Proc. Nat. Acad. Sci. USA,48 (1962), pp. 596-600.
(3) It.. E . KALlUN, On tAt gmsrtJl tAsory of control '1j'ums, Proc. lIt International CODIfeII on Automatic Control, Moscow, 1960; ButterworthB, London,
1961, Vol. 1, pp. 481-492. (4] R. E. KALMAN, Y. c. Ho, AND K. S. N ARENDBA, Con,lrolltJbilit" ollinfKJrdyntJmi. col'r"''''', (to appear in Contributions to Differential Equations, Vol. 1, John Wiley, New York.) [5J E. G. GILBIJaT, Con.trollabilitr AM ob,,,Vtlbilily in multivario.ble control J. Soc. Induat. Appl. Math. Ser. A: On Control, Vol. 1, No.2 (1963) .. pp. 128-111. (6) V. V. NJJIlITSEII AND V. V. SUPANOV, (JUQlilatiue TMMlI Of Differential EflUJ'iou, PriDceton Univ. Press, Princeton, 1960. (7J R. E. KALMAN AND J. E. BBRTRAII, (,"cm'rol 81/,'em analt/na and dBaign vio the c,.ccmc! mIUaotl' 01 L7/GptmOtJ, J. Basic Engr. (Trani. A..8.M.E.), 82 D (1960), pp. 371-393. (8) R. E. KALMAN, Aft4lfJIi, tlnd delip principle, of leconcl tmd higher-order saturatin, .ervomecMnia"&a, Tr&DI. Amer. Inat. Elect. Engr8., 74, II (1906), pp.
'1/'""",
294.-310. [O} W. HAIIN, TlNone UM AfI,.Mung der direkttm MetMd" lion LjapufWu, Springer, Berlin, 1959. (10) J. P. LASALLB AND S. Lu8cJnITz, BltJbilily By LYGpunorJ'a Direct Met1&otl, Aoademio Prea., NewYork, 1961. [111 L. MARKVI, Ctmtinuou, matnee. arad the ,lGbili'lI of differential 'y8Um1, Math. Z., 62 (1956), pp. 310-319.. (12) L. A.. ZAD.B, A ,meral IA,orJ/ of linear signa.l 'rClumi"ion 1J1IBtema.. J. Franklin lut.,253 (1952), pp. 293-312. (13) D. MIDDLJlTON, Aft r,drodv.ction To 8tG'iltical Communication TMory, MaGrawHill, New York, 1960. [14) R. E. KALMAN, ()A COfttrol1c6ilit,l, ob,6nHlbiliI1l, and identifiability of linea.r clrMm,c:al (to appear). (15] R. E. KALMAN, On 'A, ,lability 01 time-fJtJrJl'ing linear 'yltem" Trans.. I.R.E. Prof. Gr. Circuit Theory, (CT-9 (1962), pp. 420-422.. ). [16) S. J. MASON, F,etlbac1c U&«w,:,ome prop,rll,. oj aipaljlotD gTA'P"", Proe. I.R.E., 41 (1983), pp. 1144-&6; Further prope,ti" 0/ M(Jft4l jlOtD grtJpM, ibid., 44 (19M), pp . 9~. (l7) R. E .. KALMAN, N,w ruulta in jUte""" and prediction .tMOTlI, RIAS Report 61-1, Researcb Institute for Advanced Studies (RIAS), Baltimore, 1961. (18) A. M. Lftov, Btabilit1l1ft Nonlinear Control 81/.tems, Princeton Univ. Press, Princeton, 1961.
."."mB,
279
(19) B. L.
VAN DEB WAEBDEN,
1I1odsrn Algebra, Vol. It 2nd Ed., Ungar, New York,
1949. BATltOV, On the problem oj ayntl&esu ulliMa, dyntunic syaletn, with two parameters, Avtomat. i Telelneh., 19 (1958), pp. 49-54. J. H. LANING, JR. AND R. H. BATTIN, RandO'" Proce',Bs In Alltuttlatic Control,
(20J A. M. [21]
McGraw..H ill, New York, 1956. G_ F. F~NJ[LIN, Sampled.Dala Control SYSUft18, McGrawBill, New York, 1958. R. H. BAlloR, 7'1&6 pula, trtJM/er !u,'ru;'ion and ill application to sampling servo 8y,tema, Proe, Inst. Elec. Engrs. 99 IV (191i2), pp. 302-317. E. I. JUBY, 8ampled-DataCtmlroI81/stem8, John Wiley, New York, 1957. R. E. KALMAN, L1IGpunov Junction. JOT tke problem 0/ Lur'« in fllIltltnatic control, Proc. Nat. Acad. Sci. USA, 49, (1963), pp. 201--205. E. F. MooD, Gedankm-a;peritMnt. Oft .eguentitJl mcchines, .1 utonuuo: ,.~tudie8. Princeton Univ. Prese, Princeton, 1956. A. 1. LUB'm, Certain N onlineGr P1'oblema in tA, Theory of Autmnatic Control. (in Russian), Gostekbizdat. M08cow, 1951;German translation Akademie· VerI'C. Berlin, 19m. Private communication, Academician A. I. Lur'e.
(22) J. R.
[23) [24]
[25J (26) {27]
(28]
RA.GAZZINI AND
APPENDIX Factorization of rectangular matrices. Given an arbitrary, real, p X 1n matrix R of rank q ~ min (m, p). We wish to find & p X q matrix Hand a q X m matrix G, both of rank q, such that R = RG. The existence of Hand G follows almost immediately from the definition of rank. We describe below a constructive procedure for determining Hand G numerically from numerical values of R. Let p ~ m. Form the p X p matrix S == ER'. As is well known, there exists a nonsingular matrix T such that
TRR'T'
(A-l)
=
T8T' = E,
where precisely q diagonal elements of E are 1, all other elements are O. T can be calculated by steps similar to the gaussian elimination procedure. Compute the generalized inverse R# (in the sense of Penrose [4]) of R. R1 is an m X p matrix. Using the properties of R# ([4]) we obtain (A-2)
R:= RR1R
== RR'R';' ==
SRII = r"lE'lrl'R1' = (7r- I E ) ( T -1E)'R".
Now FIE is & matrix which contains precisely p - q zero columns. De.. leting these columns, we obtain a p X q matrix ('T1E)O == H. Similarly, deleting p - q zero rows from (rIE)'R I ' = (R'riE)' we obtain a m X q matrix = (RlrIE)o. Evidently R HG. Since the ranks of II and G are obviously less than or equal to q, both ranks must beexactly q for other-
a
=
< q, contrary to hypothesis. Alternately, let T, U be Ilonsingular matrices such that
wise rank R
TRU == E; 280
then (A-3)
is the desired decomposition. However, the computation of (A-3) may require more steps than that of (A .. 2~. Suppose now that R is complex. Then S = RR' = RR* = . 4 + iB is complex hermitian: it corresponds to the 2n X 2n nonnegative matrix (A-4)
where A ;:;;: A' and B = -B'. III fact, if z·== x + iy, the hermitian form z*RR*z (which is real-valued) is equal to the quadratic form
As is well known, there exists a nonsingular complex matrix T such that TST* = E. If T = U + iV, it follows further that
[_Uv ~J [_AB ;] [~: ~;'] = [ :
;l
Hence the determination of the complex n X n matrix has been reduced to the determination of a real 2p X 2p matrix. Similar remarks apply to the calculation of R I . Thus the problem of factoring complex p X m matrices can be embedded in the problem of factoring real 2p X 21n matrices.
281
On the Input-Output Stability of Time-Varying Nonlinear Feedback Systems G.ZAMES
IN
the early sixties, a number of specific results regarding stability of feedback systems emerged. Especially important instances of such results are the Popov criterion (discussed elsewhere in this volume), and the circle criterion. The circle criterion has its roots in the work of Bongiorno [1] and Tsypkin [10], and was developed almost simultaneously by a number of authors. Popov's stability criterion and the circle criterion seem to be rather specific results with special purpose proofs, but they led researchers to discover interesting general underlying principles, and to develop a mathematical framework that allowed them to analyze stability of feedback systems. Three such general and intuitively very pleasing stability principles are formulated in the first part of the article by Zames: the small loop gain theorem, the positive operator theorem, and the conic sector theorem. Intuitively, the small loop theorem states that if the open loop of a feedback system attenuates signals, then the feedback system is stable. Intuitively, the positive operator theorem states that if a feedback system can be viewed as the interconnection of two passive systems, then the feedback system dissipates, and will hence be stable. The conic operator theorem provides a way of unifying and generalizing both the small gain and the positive operator theorem. These principles are stated in the summary of Part 1, and the paper itself contains a formalization and very clear proof of these results. These results were cast in terms of the properties of input/output relations on extended spaces. The concept of an extended space provided a rich and useful alternative to the Lyapunov methods which at the time were a more commonly used approach to stability. Extended spaces were introduced and formalized in these articles by Zames, with roots going back to his earlier work [14], [15]. They also occur in the work of Sandberg [8], [9]. Extended spaces allow one to analyze situations in which signals are not a priori assumed to belong to a normed space, but are a posteriori proven to belong to this normed space. The fact that all signals in a feedback loop belong to a normed space is what is called input/output stability. Often, the normed space may be L2[0, 00], and the associated extended space L 2e[0, 00] is the space of signals whose timetruncations are square integrable. This technique of approaching
stability is not unlike what is called Picard iteration in the theory of differential equations. In the second part of the paper, these general stability principles are used to obtain proofs of the circle criterion and the Popov criterion. The powerful general feedback stability principles discussed in Part I were further developed in a number of monographs [12], [7] and textbooks [5], [11] and led to a wide variety of stability criteria, often based on the use of multipliers, an instance of which is already used in Part II of the paper. Multipliers that are more flexible than the one used in the Popov criterion were used before in [2], and became a very effective technique for obtaining stability criteria. Many examples of such criteria are collected in the reprint volume [6]. This volume also contains an extensive bibliography related to the small gain theorem, the circle criterion, etc. The circle criterion as stated in Part II of the paper is a beautiful generalization of the Nyquist criterion (as a stability criterion) to systems with one time-varying nonlinear element. It is worth pointing out, however, that the circle criterion for time-varying systems can sometimes give quite conservative results. For example (see [3]), it can be shown that the system :(22 y + 2-fft y + k(t)y == is asymptotically stable whenever k is in the sector < k(t) < 11.6, whereas the circle criterion only predicts asymptotic stability for < k(t) < 4. The circle criterion has an instability converse, proven by Brockett and Lee [4] in a Lyapunov setting. The combination of the stability and the instability parts of the circle criterion makes for a full generalization of the Nyquist criterion. The question of how to generalize the instability part of the circle criterion in an input/output stability setting proved to be non-trivial, and was carried out in [13]. These input/output stability principles had a deep impact on the development of control theory in the decades that followed. This influence was perhaps most striking in the work on robust control. Indeed, the small loop gain theorem led to the insight that in order to obtain an adequate robustness margin, the controller should be designed so as to moderate the £2 -induced gain (the Hoo-norm) of the linear time invariant part of the system.
283
°
°
°
The small loop gain theorem was therefore instrumental for the development of H00 optimal control. The original paper [16] where the Hoo-control problem was first formulated will be discussed elsewhere in this volume. REFERENCES
[1] J. J. BONGIORNO, JR., "An extension of the Nyquist-Barkhausen stability criterion to linear lumped-parameter systems with time-varying elements," IEEE Trans. Aut. Contr., AC-8:167-170, 1963. [2] R. W. BROCKETT AND J. L. WILLEMS, "Frequency domain stability criteria, Parts I and II," IEEE Trans.Aut. Contr., AC-I0:255-261 & 407-413, 1965. [3] R. W. BROCKETT, "The status of stability theory for deterministic systems," IEEE Trans. Aut. Contr., AC-ll:596-606, 1966. [4] R. W. BROCKETT AND H. B. LEE, "Instability criteria for time-varying and nonlinear systems," IEEE Proceedings, 55:604-619,1967. [5] C. A. DESOER AND M. VIDYASAGAR, Feedback Systems: Input-Output Properties, Academic Press (New York), 1975. [6] A. G. J. MACFARLANE, Frequency-Response Methods in Control Systems, IEEE Press (Piscataway, NJ), 1979.
[7] M. G. SAFONOV, Stability and Robustness of Multivariable Feedback Systems, The MIT press (Cambridge, MA), 1980. [8] I. W. SANDBERG, "On the properties of some systems that distort signals," Bell Sys. Tech. J., 42:2033, 1963, and 43:91-112, 1964. [9] I. W. SANDBERG, "On the L2-boundedness of solutions of nonlinear functional equations," Bell Sys. Tech. J., 43:1581-1599,1964. [10] Y. Z. TSYPKIN, "On the stability in the large of nonlinear sampled-data systems," Dokl. Akad. Nauk., 145:52-55, 1962. [11] M. VIDYASAGAR, Nonlinear Systems Analysis, Prentice Hall (Englewood Cliffs, NJ), 1978. [12] J. C. WILLEMS, The Analysis of Feedback Systems, The MIT Press (Cambridge, MA), 1971. [13] 1. C. WILLEMS, "Stability, instability, invertibility and causality," SIAM J. Contr., 7:645-671, 1969. [14] G. ZAMES, Conservation of Bandwidth in Nonlinear Operations, Report MIT-RLE No. 55, 1959. [15] G. ZAMES, Nonlinear Operators for System Analysis, Report MIT-RLE No. 370, 1960. [16] G. ZAMES, "Feedback and optimal sensitivity: Model reference transformations, multiplicative semi-norms and approximate inverses," IEEE Trans. Aut. Contr., AC-26:301-320, 1981.
I.C.W
284
On the Input-Output Stability of Time-Varying Nonlinear Feedback Systems Part I: Conditions Derived Using Concepts of Loop Gain, Conicity, and Positivity G. ZAMES, Abstract-The object of this paper is to outUne a stabUity theory for input-output problems using functional methods. More particu.. lady. the aim is to derive open loop conditions for the boundedness and continuity of feedback systems, without, at the beginning, placing restrictions OD linearity or time invariance. It will be recalled that, in the special case of a linear time invariant feedback system. stability can be assessed using Nyquist's criterion; roughly speaking. stability depends on the amounts by which siguals are amplified and delayed in flowing around' the loop. An attempt is made here to shoW' that aimUu conlideratioDllovem the behavior of feedback systems in general-that stability of noalinear time-varying feedback systems can often be assessed from certain gro.s features of input-output behavior. which are related to amplification and delay. This paper is divided into two parts: Part I contains general theorems, free of restrictioDs on linearity or time invariante; Part D,I which will appear in a later iaue, contaiDa applications to a loop with one ncm1inear element. There are three maiD reault. in Part I, which foUow the introduction of concepts of gain, conlcitll. /JOsitiriit1l. and ,iro", jJositiuitg:
MEMBER, IEEE
terns we might ask: What are the kinds of feedback that are stabilizing? What kinds lead to a stable system? Can some of the effects of feedback on stability be described without assuming a very specific system representation? Part I of this paper is devoted to the system of Fig. 1, which consists of two elements in a feedback loop.' This simple configuration is a model for many controllers, amplifiers, and modulators; its range of application will be extended to include multi-element and distributed systems, by allowing the system variables to be multidimensional or infinite-dimensional.
THEOREM1: If the open loopgain is less than one, then the closed
Fig. 1. A feedback loop with two elements.
loop is bounded. THEOREM 2:
If the open loop CaD be factored into two, suitably proportioned, conic relatioDs. then the closed loop is bounded.
The traditional approach to stability involves Lyapunov's method; here it is proposed to take a different course, and to stress the relation between input-output THEOREM 3 : H the open loop can be factored mto two positive rebehavior and stability. An input-output system is one in lations, one of which is Itrongly positive and has' which a function of time, called the output, is required finite gain, then the closed loop is bounded. to track another function of time, called the input; more Results analogous to Theorems 1-3, but with boundednes8 regenerally the output might be required to track some placed by continuity, are also obtained. function of the input. In order to behave properly an input-output system must usually have two properties:
I.
INTRODUCTION
1) Bounded inputs must produce bounded outputs-
F
E E D BA CK , broadly speaking, affects a system in .one of two opposing ways:.depending on circumstances it is either degenerative or regenerative-either stabilizing or destabilizing. I n trying to gain some perspective on the qualitative behavior of feedback sysManuscript received December 29.1964; revised October 1, 1965~ February 2, 1966. This work was carried out at the M.I.T. Electronic Systems Laboratory in part under IUpport extended by NASA under Contract NsG-496 with the Center (or Space Research. Parts of this paper were presented at the 1964 National Electronics Conference, Chicago. Ill., and at the 1964 International Conference on Microwaves, Circuit Theory/ and Information Theory, Tokyo, Japan. The author is witn the Department of Electrical Engineering, MaS$l\t;husetts Institute of Teclinology, Cambridge, Mass.
i.e., the system must be nonexplosive. 2) Outputs must not be critically sensitive to small changes in inputs-changes such as those caused by noise. 1 The system of Fig. 1 has a singleinput x. multiplied by constants at and CSt and added in at 'wo points. This arranJement has been chosen because it is symmetrical and thus convenient for analysis; it also remains invariant under some of the transformations that will be needed. Of course, a single inpu t loop can be obtained by setting al or (J! to zero. The terms 'WI and w, are fixed bias functions. which will be used to account for the effects of initial conditions. The variables e.. es, "1, and,... are outputs.
Reprinted from IEEE Transactions on Automatic Control, Vol.AC-11,1966, pp. 228-238,465-476.
285
These two properties will form the basis of the definition of stability presented in this paper. I t is desired to find conditions on the elements HI and HI (in Fig. 1) which will ensure that the overall loop will remain stable after H1 and H 2 are interconnected. It is customary to refer to HI and H, prior to interconnection as the "open-loop" elements. and to the interconnected structure as the "closed loop." The problem to be considered here can therefore be described as seeking open-loop conditions for closed-laop slability. Although the problem at hand is posed as a feedback problem, it can equally well be interpreted as a problem in networks; it will be found, {or example, that the equations of the system of Fig. 1 have the same form as those of the circuit of Fig. 2, which consists of two elements in series with a voltage source, and in parallel with a current source."
t
t
~:I i - ~. i1:;~·
Fig. 2.
A circuit equivalent to the loop of Fig. 1.
1.1 Historical Note The problem of Lyapunov stability has a substantial history with which the names of Lur'e, Malkin, Yakubowitch, Kalman. and many others, are associated. On the other hand. functional methods for stability received less attention until relatively recently, although some versions of the well . . known Popov [1] theorem might be considered as fitting into this category. The present paper has its origin in studies [2a, b Jof nonlinear distortion in bandlimited feedback loops. in which contraction methods were used to prove the existence and stability of an inversion scheme. The author's application of contraction methods to 1110re general stability problems was inspired in part by conversations with Narendra during 1962-1963; using Lyapunov's method, Narendra and Goldwyn [3] later obtained a result similar to the circle condition of Part I I of this paper. The key results of this paper, in a somewhat different formulation, were first presented in 1964 [2d, e]. Many 2 I t is assumed that the source voltage v and the source current ·i are inputs, with v=aaX+wa and i=a:.~+'W~; the currents and voltages 'in the two elements are outputs.
of these results are paralleled in the work of Brockett and Willems [4], who use Lyapunov based methods. Several others have obtained similar or related results by functional methods: Sandberg [Sa] extended the nonlinear distortion theory mentioned above; later [Sb] he obtained a stability theorem similar to Theorem 1 of this paper. Kudrewicz [6] has obtained circle conditions by fixed point methods. Contraction methods for incrementally positive operators have been developed by Zarantonello [7], Kolodner [8], Minty [9], and Browder. [10]. A stability condition for linear timevarying systems has been described by Bongiorno [11
J.
2. FORMULATION OF THE PROBLEM
There are several preliminaries to settle, namely, to specify a system model, to define stability, and to write feedback equations. What is a suitable mathematical model of a feedback element? A "black box" point of view towards defining a model will be taken. That is to say, only input-output behavior, which is a purely external property, will be considered; details of internal structure which underlie this behavior will be omitted. Accordingly, throughout Part I, a feedback element will be represented by an abstract relation, which can be interpreted as a mapping from a space of input functions into a space of output functions. More concrete representations, involving convolution integrals, characteristic graphs, etc., will be considered in Part II. Some of the elementary notions of functional analysis will be used. though limitations of space prevent an introduction to this subject.' Among the concepts which will be needed and used freely are those of an abstract relation, a normed linear space, an inner product space, and the L; spaces. The practice of omitting the quantifier "for all" shall be utilized. For example, the statement Ie
-x! $
x:?(.~
E X)"
is to be read:
"for all .e E .Y, CONVENTION:
the type
_.\''J
< x~.~t
A ny expression containing a condit-ion of
t'xEX," free of quantifiers, holds for all xE.lY'.
2.1 The Extended Normed Linear Space .IY" In order to specify what is meant by a system, a suitable space of input and output functions will first be defined.' Since unstable systems will be involved. this space must contain functions which "explode, n i.e .• which grow without bound as time increases [for example, the exponential exp (t) J. Such functions are not contained in the spaces commonly used in analysis. for example, in the Ll1 spaces. Therefore it is necessary to I A good reference is Kolmogorov and Fomin (12). • The space of input functions will equal the space
functions.
286
or
output
The point of assumptions (2)-(3) on X can now be appreciated; these assumptions make it possible to deter.. mine whether or not an element xEX. has a finite norm, by observing whether or not linlt_co IIxtll exists. For ex..
construct a special space, which will be called X,. X" will
contain both elwell-behaved" and "exploding" functions, which will be distinguished from each other by assigning
finite norms to the former and infinite norms to the latter. X. will be an extension, or enlargement, of an associated normed linear space X in the following sense. Each finite-time truncation of every function in X. will lie in X; in other words, the restriction of xEX, to a finite time interval, say to [0, t], will have a finite norm-but this norm may grow without limit as t~ eo, First a time interval T and a range of input or output values V will be fixed. DEFINITION:
type
[to. co) or
T is a given subinterval of the reels, of the (- 00, (0)• V is a given linear space.
be any function mapping T into V, t be any point in T; then the symbol x, denotes the truncated function, x,: T-+ V, which assumes the values %,(".) =X(T) for 1"
where.
(A truncated function is shown in Fig. 3.) Next, the space X is defined.
v
>
t .....- - - - -....... T
Fig. 3. A truncated function. DEFINITION:
X is
CJ
space consisting of functions of the
cerning X:
(1) X is a normed linear space; the norm of xEX is denoted by 11%11. (2) If xEX then %,EX for all lET. (3) If x: T-+ V, and if %,EX [or all lET, then: (a) .IIXtll is a nondecreasingfunction of tET. (b) If lim, ... \\%t ll exists, then xEX and the limit
equals
1Ix ·
(For example" it can be verified that assumptions (1)(3) are satisfied by the L p spaces.) Next, X. is defined.
II xII
t-«
I
ee•
In other words, a relation is a set of pairs of functions in X •. It will be convenient to refer to the first element in any pair as an input, and to the second element as an output, even though the reverse interpretation is sometimes more appropriate. A relation can also be thought of as a mapping, which maps some (not necessarily all) inputs into outputs. In general, a relation is multivalued; i.e., a single input can be mapped into many outputs. The concept of state, which is essential to Lyapunov's method, will not be used here. This does not mean that initial conditions cannot be considered. One way of accounting for various initial conditions is to represent a system by a multi-valued relation, in which each input is paired with many outputs, one output per initial con .. dition, Another possibility is to introduce a separate relation for each initial condition. Note that the restrictions placed on X. tend to limit, a priori, the class of allowable systems. In particular, the requirement that truncated outputs have finite norms means, roughly speaking, that only systems having infinite "escape times. n i.e., systems which do not blow up in finite time, shall be considered. Some additional nomenclature follows: DEFINITION: If H is a relation on X., then the domain of H denoted Do(H), and the range of H denoted Ra(H) , are
the sets,
Do(H)=- {xl xEX" and there exists yEX, such that
The extension of X, denoted by X,. is the space consisting of those functions whose truncations lie in X, that is, X.= {xl x: T-+ V, and x,EX for all tET}. (NOTE: X. is a linear space.) An extended norm, denoted is assigned to each xEX. as follows: if
xEX, and Jlxll.= co if x(fX.
IlxlI,
DEFINITION: A relation H on X, is any subset of the product space X.XX•. If (x, y) is any pair belonging to H then y will be said to be H-related to x; y will also be said to be an image of % under H.5
DEFINITION:
(Ixll.,
r.r),
flx,rl or in other words, IlxU, =
The mathematical model of an input.. output system will be a relation on X e :
NOTATION: Let x that is, oX: T --+ V; kt
type x: T-+ V; the following assumptions are made con-
EXAMPLE 1: Let L 2[O, 00) be the norrned linear space consisting of those real..valued functions x on [0, oc) for which the integral Jo-o x 2 (t)dt exists, and let this integral equal I. Let X=L 2 [OI ee }, and let L 2-e = X . ; that is, L 2, is the extension of L 2[O, oc). Let x be the function finite, that iS on [0, 00) given by x(t) = exp (I). Is is x in X? No, because grows without limit as
2.2 Input-Output Relations
[For example, in the analysis of multielement (or V is multidimensional (or infinite-dimensional).] Second, the notion of a truncated function is introduced.
distributed) networks,
I - - - - -....-
ample:
(x, Y)EH}
Ra(H) = {yl yEX" and there exist$ xEX. s,l.tch that (x, Y)E'H}
IIxll.=lIxll
~
287
In general
%
can have many images.
NOTATION: If H is a relation on X., and if x is a git/en element of X., then the symbol Hx denotes an image of % under H.I
The idea here is to use a special symbol lor an element instead of indicating that the element belongs to a certain set. For example, the statement, "there exists Hx having property P" is shorthand for "there exists yERa(H), such that, is an image of x, and y has property P."· Observing that H» is, according to the definitions used here, a function on T, the following symbol for the value of Hx at time t is adopted:
DEFINITION: A relation H on X, is continuous if H has the following property: Given any xEX (that is, IIxfl.< 00), and any &>0, there exists 8>0 such that,for
all ,EX, ifllx-yll <8 then IIHx-Hyll
are:
The symbol Hx(t) denotes the value assumed by thefunction Hx at time t E T. NOTATION:
+ 41% + y2 e2 = W2 + 42% + 'Yl f:l
Occasionally a special type of relation. called an
=
WI
(la) (1b)
operator, will be used:
Y2
=
Hlez
(2a)
DEFINITION: An operator His a relation on X. which satisfies two conditions: 1) Do(H) =X•. 2) H is singlevalued; that is, if x, " and z are elements of X" and if , and I are images of x under H, then y == z.
Yl
= H1el
(2b)
in which it is assumed that:
HI and HI are relations in (R and a2 are real constants WI and W2 are fixed biases in X x in X. is an input 61 and e2 in X, are (error) outputs )'1 and )'2 in X. are outputs. 41
2.3 The Class
The assumption that H maps zero into zero simplifies many derivations; if this condition is not met at the outset, it can be obtained by adding a compensating bias to the feedback equations. If Hand K are relations in CR, and c is a real constant, then the sum (H+K), the product cH, and the composition product KH of K following H, are defined in the usual way,7 and are relations in CR. The inverse of H in at, denoted by B-1, ,always exists. The identity operator on X, is denoted by I.
2.4 Input-Output Stability
The term "stable" has been used in a variety of ways, to indicate that a system is somehow well behaved. A system shall be called stable if it is well behaved in two respects: (1) It is bounded, Le., not explosive. (2) it is continuous, i.e., not critically sensitive to noise. DEFINITION: A subset Y of X. is bounded if there exists A >0 such for all ,E Y, 11,11.
'1uJ'.
I In keeping with the usual convention used here, al!Y statement containing H% free of quantifiers holdsfor all » in Ra(H). For example, "Hx> 1 (xEX.)" means that "for all % in X., and for· all Hz in Ra(H), Hx>l." ., In particular, Do(H+K)=-Do(H)flDo(K). Note thata:lis nota linear space; for example, if Do(R) ,.eOo{K) then Do((H+K) -Xl
J1I.Do(H).
• This definition implies that inputs of finite norm produce outputs of finite norm. More than that, it implies that the sort of situation is avoided in which a bounded sequence of inputs, say IIx.1I< 1 where 11- I, 2, • · • • produces a sequence of outputs ha viog norms that are finite but increasing without limit, say URs." - n.
288
(The biases are used to compensate for nonzero zeroinput responses and, in particular, for the effects of initial conditions.) The closed-loop relations Et, E 2 , FI. and F 2, are now defined as follows. DEFINITION:
E 1 is the relation that relates el to x or,
more precisely. E 1 = {(x, et) I (x, et) EX,XX" and there exist e2, Y1, y" Hlel, and H 2e2, all in X" such that (1) and (2) are satisfied.} Similarly E 2 relates e" to X,· F 1 relates to x; F" relates )'2 to x.
'1
All the prerequisites are now assembled for defining the problem of interest which is: Find, conditions on HI and HI which ensure that E 1, E z, F., and F 2 are bounded or stable. In general it will be enough to be concerned with E 1 and E 2 only, and to neglect Fl and F 2 , since every F"x is' related to some E 1% by the equation F,.x=El% -alX-Wl, so that F2 is bounded (or stable) whenever E 1 is, and similarly for F 1 VS. E 2 • I t should be noted that by posing the feedback problem in terms of relations (rather than in terms of operators) all questions of existence and uniqueness of solutions are avoided. For the results to be practically significant, it must usually be known from some other source' that solutions exist and are unique (and have infinite "escape times"), • Existence and stability can frequently be deduced (rom entirely separate assumptions. For example, existence can often be deduced. by iteration methods, solely from the fact that (loosely speaking) the open loop delays signals: stability can not. (The connection between existence and generalized delay is discussed in G. Zames, "Realizability conditions for nonlinear feedback systems," IBEE Tran», 071 Ci,cuil TMcwy, vel. CF-l1, pp, 186-194, june 1964.)
3.
SMALL
Loop
jfH.f - ByH, ~
GAIN CONDITIONS
To secure a foothold on this problem a simple situation is sought in which it seems likely, on intuitive grounds, that the feedback system will be stable. Such a situation occurs when the open loop attenuates all signals. This intuitive idea will be formalized in Theorem 1: in later sections, a more comprehensive theory will be derived from Theorem 1. To express this idea, a measure of attenuation, i.e., a notion of gain, is needed.
-
yII. Ix, yEDo(H)J.
(9)
In the Feedback Equations (1)-(2), the product
g(H t ) 'g(H2) will be called the open-loop gain-product, and similarly, ,(HI)· g(H z) will be called the incremental open-loop gain-produ 3.2 A Stability Theorem
Consider the Feedback Equations (1)-(2). THEOREM 1 :10 a) If g(B!) ·g(Ht ) <1, then the closed loop relations E 1 and E 2 are bounded. b) If ;(HI ) · g(H2) < 1, then E 1 and' Et. are inpfi,t-output stable.
3.1 Gains Gain will be measured in terms of the ratio of the norm of a truncated, output to the norm of the related, truncated input.
Theorem 1 is inspired by the well known Contraction Principle." PROOF OF THEOREM
1: (a) Since eqs. (1)-(2) are syrn-
metrical in the subscripts 1 and I, it is enough to can.. sider E I • This proof will consist of showing that there are positive constants at b, and c, with the property that any pair (e, el) belonging to E 1 [and so being a solution of eqs. (1)-(2)], satisfies the inequality
The gain of a relation H in CR, denoted by
DEFINITION:
g(H) ·Ii.t
g(H), is
(3) where the supremum is taken over all x in Do (H) J all Hx in Ra(H). and all' in T for whieh x,7J60.
In other words, the supremum is taken over all possible input-output pairs, and overall possible truncations. The reason for using truncated (rather than whole) functions is that the norms of truncated functions are known to be finite Q, priori. It can be verified that gains have all the properties of norms. In addition, if Hand K belong to
s g(H) -11%,11
IIHxJl, s
g(H)
-flxll.
E Do(H); t E [x E Do(H)] [x
T]
It will follow that if x is confined to a bounded region, say IIxl( ~A, then el will also be confined to a bounded ~allwll1 +cA. Thus E 1 region, in this case will be bounded. PROOF OF INEQUALITY (10): If (x, el) belongs to E 1 then, after truncating eqs. (1a) and (Ib), and using the triangle inequality to bound their norms, the following inequalities are obtained:
Ilel!1
(5)
IIY2tll
1(H) = sup
II x, - yl "
(6)
where tlse supremum is taken overall" and, in Do(H) , all Hx and By in RtJ(H), and all t in T [or which x,;",. Incremental gains have all the properties of norms, and satisfy the inequalities
I(KH)
"
S
g(K)· g(H)
(7)
(HY),fI
SI(H)·Il%,-y,1I [x,yEDo(H);IET] (8)
(lOa)
(t E T)
(lOb)
s g(H ·lIe:ull
(lla)
2)
IIYuff s g(H1) -If eufl ·
(lib)
Letting g(Hl) ~ot and g(Ht ) ~fl. and applying (lla) to (lOa) and (lib) to (LOb), the following inequalities are obtained:
UeuJl ::; II wufl + I all ·lfxcJl + pIJe21lf fleuJl ~ IIwl,Jf + Ia'21 -lfXtJl + alleull
The incremental gain of any H in at, de-
(HY)tll
(t E T)
Furthermore, applying Inequality (4) to eqs. (2), the following is obtained, for each t in T:
noted by I(B). is, If (H%) , -
+bllWtll
lI euJf -s Ifwull + Iall ·J1XtU + IlY2,f1 lI e2,fI ~ I1 W2,J( + I Q21 ·llx,JI + 11"1,11
(4)
where (4) is implied by (3), and (5) is derived from (4) by taking the limit as t--+ ~ • If g(H) < co then (5) implies that H is bounded. In fact, conditions for boundedness will be derived using the notion of gain and inequalities such as (5). In a similar way, conditions for continuity will be derived using the notion of incremental gain, which is defined as follows: DEFINITION:
(10)
Applying (12b) to (1 - a~)lIellJf ~
E T) (I E T)_
(I
(12a) (12b)
ife2,1( in (12a), and rearranging,
HwuJI + .all w2tll + (I all + s] a21 )[fx,fI
(t E T).
(13)
la A variation of Theorem 1 was originally presented in {2dj. A related continuity theorem was used in [2c]. An independent, related result is Sandberg's {Sb). II If X is a comp~ete s2!lce, if all relations are in fact opcratC?rs. and if the hypothesIS of Theorem 1b holds, then the Contraction Principle implies existence and uniqueness of solutions-a rna tter that hal been disregarded here.
289
Since (l-~»O (as~
sup S
Fig. 5.
real
= Ie
,(m
N(x) /
/
/
4.1 Definitions of Conic and Positive Relations DEFINITION: A relation H in cR is interior conic if the,e are real constants r ~ 0 and, /OT which the inequality
II(Hx), - eXtli ~
rllx,1I
[~E Do(H); t
E r]
(14)
is satisfied. H is exterior conic if the inequality s'ign in (14) is reversed, H is conic if it is exterior conic or interior conic. The constant c will be called the center parameter of H, and r will be called the radius parameter.
The truncated output (Hx) , of a conic relation lies either inside or outside a sphere in X, with center proportional to the truncated input x, and radius proportional to The region thus determined in X. X X, will be called a "cone," a term suggested by the following special case: EXAMPLE 3: Let H be a relation on L 2• (see Example 1); let Hx(t) be a function of x(t), say Hx(t) = N [x(t)], where N has a graph in the plane; then, as shown in Fig. 6, the graph lies inside or outside a conic sector of the plane, with a center line of slope c and boundaries of slopes c-r and c+r. More generally, for H to be conic [without Hx(t) necessarily being a function of x(t), that is, if H has memory], it is enough for the point [x(l), Hx(t)] to be confined to a sector of the plane. I n this case, it will be said that H is instantaneously confined
IIx,lI.
/
41
-----..................----x /
transformation.
RESTRICTION: In the remainder of this paper, assume that X is an inner-product space, that (x, y) denotes the inner product on ..1{, and that (x, x) = 1fxIl 2 •
I :c: I N(x)
.~
This restriction is made with the intention of working mainly in the extended L 2[O, 00) norm;" with (x, y) =f:x(t)y(t)dt.
where the first sup is over [xEDo(H); HxERa(H); tET. x,~O]. That is, g(H) is k times the supremum of the absolute slopes of lines drawn from the origin to points on the graph of N. Here g(H) = k, so Theorem 1 implies boundedn~ss for k < 1. This example is trivial in at least one respect. namely, in that H has no memory; examples with memory will be given in Part II. (6) can be worked out to be k times the supremum of the absolute Lipschitzian slopes of N, that is, g(H) =k SUPa• ., r~d N(x) -N(y)/x-yl =2k. The closed leop is therefore stable for k < 1/2.
Fig. 4.
~x.w,
{~·N2[X(t)] dt / ~·XI(t) dt} 1/2
g(B1) = sup
=k
unaffected; however, HI is changed into a new relation Hi', as in effect -cl appears in feedback around Hi. Under what conditions does this transformation give a gain product less than one? I t will appear that a sufficient condition ·is that the input-ou tput relations of the open loop elements be confined to certain "conic" regions in the product space X.XX•.
/ Adenotes slope
Graph of the relation in Example 2.
to a sector of the plane.
4. CONDITIONS INVOLVING CONIC RELATIONS
Inequality (14)
The usefulness of Theorem 1 is limited by the condition that the open-loop gain-product be less than onea condition seldom met in practice. However. a reduced gain product can often be obtained by transforming the feedback equations. For example, if ,1 is added to and subtracted from H 2• as shown in Fig. S, then et remains
can be expressed in the Iorrn If norms are expressed in
II (Hx),-cx,IJ2- rl l x t!lz50.
12 However, in engineering applications it is often more interesting to prove stability in the L., norm. The present theory has been extended in that direction in the author's {2f]. The idea is (2f) is to transform L 1 functions into L. functions by means exponential weighting factors.
290
or
Hx(t)
CASE
la: If a>O then H-l is inside
{lib,
l/a} .
b-c.r
CASE
lb: If a
l/b} . CASE
2: If a=O then
(H-l-(l/b)I)
is
positifJe.
(v) Properties (ii) , (iii), and (if]) remain valid with the terms "inside { }" and " ou.t5ide { }" interchanged throughout. (vi) g(H)~max(lal, Ibl).HenceijHisin {-r,,} then g{H) ~r.
Interi or ot sector Is shaded.
The proofs are in Appendix A. One consequence of these properties is that it is relatively easy to estimate conic bounds for simple interconnections, where it might be more difficult, say, to find Lyapunov functions.
Fig. 6. A conic sector in the plane.
terms of inner products then, alter factoring, there is obtained the equivalent inequality «H~), - ax" (Hx), - bx,} ~ 0 [x
E
Do(H); t E T] (15)
where a==c-" and b==c+1'. It will often be desirable to manipulate inequalities such as (15). and a notation inspired by Fig. 6 is introduced: NOTATION: A conic relation H is said to be inside the sector {a, b}, if a5:b and if Ineq:ualit, (15) holds. His outside the sector {a, b} if a:S;b and if (15) holds with the inequality sign "etJers~d.
The following relationship will frequently be used: If H is interior (exterior) conic with center c and radius r then H is inside (outside) the sector {'-I', c+,}. Convenely, if H is inside (outside) the sector {a, s}, then H is interior (exterior) conic, with center (b+a)/2 and radius (b-a)/2. DEFINITION: A relation H in
[% E Do(H) i t E T].
4.3 A Theorem on Boundedness Consider the feedback system of Fig. 1, and suppose that Ht. is confined to a sector {at s}. It is desirable to find a condition on HI which will ensure the boundedness of the closed loop. A condition will be found, which places HI inside or outside a sector depending on a and b, and which requires either HI or H, to be bounded away from the edge of its sector by an arbitrarily small amount, 4 or ~. THEOREM: 2a: [In eqs. (1)-(2)] Let HI and Hz be conic relations. Let a and abe constants, of which one is strictly positive and one is sero. Suppose that
(1) -H2 is inside the sector {a+A, b-A} where si-o. and, (II) HI satisfies one of the following conditions. CASE
la: If 0,>0 then HI is outside
- s, - ~ + { _..:. 'a b
(16) CASE
A positive relation can be regarded as degenerately conic, with a sector from 0 to 00. [Compare (15) and (16).] For example, the relation H on Lt. is positive if it is instantaneously confined (see Example 3) to the first and third quadrants of the plane.
1b: If 0,<0 then HI is inside
{- ~ + Ii, CASE ~:
Hl+(+ - 8)1
Some simple properties will be listed. It will be assumed, in these properties. that H and HI are conic relations; that H is inside the sector {a, b}, with b>Oj that Hl is inside {aI, bt} with b1>O; and that 1e~O is a constant.
t
(iii) (iv)
SUM RULE: (H+H 1) INVERSB RULE
: - Ii} .
If a == 0 then
4.2 Some Properties of Conic Relations
(i) 1 is insitle 1, I}. (ii) kHisinside {ka,kb}; -His inside {-b.
Ii} .
-a}.
is inside {a+41' b+bl}.
U Short for "positive semidefinite," The terms Mpassive" and "nondissipative" bve allo been used,
is positive j in addition, if /);. = 0 then g(Ht )
< 00. Then E 1 and E, are bounded. The proof of Theorem 2a is in Appendix B. Note that the minus sign in front of HI reflects an interest in negative feedback. EXAMPLE 4: If HI and Hi are relations on Lt. instantaneously confined to sectors of the plane (as in Example 3), then the closed loop will be bounded if the sectors are related as in Fig. 7. (More realistic examples will be discussed in Part II.) 291
THEOREM 2b: Let H 1 and Hz be incrementally con-ic relations. Let d and be constants, of which one is strictly
a
positive and one is zero. Suppose that, (1) - H 2 is incrementallv inside the sector t a +A, b-6}, where b>O, and, (I I) H t satisfies one of the following conditions:
~'/b '-1/0
CASE 2: Q:aO
CASE
La: If a > 0 then HI is incrementally otits ide
{- : - a, ~~N~-)dt} ~1/b
+ o} . +
CASE
Ib: If a <0 then HI is inere'mentally inside
CASE
2: If a=O then
CASE 1b: 0<0 H,X(t) j-1/0
x(t)
Hi
~1/b
NOTE; IN ALL CASES A>OI6-0. AND b>O.
+ {; -
o} 1
is incrementally positive; in addition,
1
ADMISSIBLE REGIONS ARE SHADED.
then l(HI )
Fig. 7. Mutually admissible sectors for HI and H.
< :0 •
if d = 0
Then E 1 and E 2 are input-output stable. The proof is similar to that of Theorem Ia, and is omitted.
4.4 Incrementally Conic and Positive Relations
Next, it is desired to find a stability result similar to the preceding theorem on boundedness. To this end the recent steps are repeated with all definitions replaced by their "incremental" counterparts. DEFINITION: A relation H in CR is incrementally interior (exterior) conic if there are real constants r> 0 and c for whieh the inequality
y),f1 ~ ,11 (x - y),lf [x, Y E Do(H);t E T] (17)
rJ (Hx - Hy), - ,(x -
5. CONDITIONS
INVOLVING POSITIVE REI.ATIONS
A special case of Theorem 2, of interest in the theory of passive networks, is obtained by, in effect, letting a = 0 and b~ ee , Both relations then become positive; also, one of the two relations becomes strongly positive,
i.e. : DEFINITION: A relation H in CR is strongly (incrementally) positive if, for some a:» O. the relation (H - 0-1) is
(incrementally) positive"
is satisfied (with inequality sign reversed), An incremen-, tally conic relatio« H is incrementally inside (outside) the sector {a, b }, if a :Sb and if the inequality
«Hx - Hy), - a(x - y)" (Hx - Hy), - h(x - Y)t)
[x, Y E Do (H) ; t E
T]
:s; 0 (18)
is satisfied (with inequality sign reversed). A relation H
in
at
is incrementally positive" if
«x - Y)'t (Hx - Hy).) ;::: 0 [x, 'Y
E Do(H); t E T]. (19)
EXAMPLE 5: Consider the relation H on L 2e • with Hx(t) = N[x(t)], where N is a function having a graph in the plane. If N is incrementally inside {a, b J then lv satisfies the Lipschitz conditions, a(x - y) ~ N(x) - N(y) Sb(x-y). Thus N lies in a sector of the plane, as in the nonincremental case (see Fig. 6). and in addition has upper and lower bounds to its slope. Incrementally conic relations have properties similar to those of conic relations (see Section 4.2).
14
The terms "monotone" and "incrementally passive" have also
been used.
The theorem, whose proof is in Appendix C. is: THEOREM 3: 11 (a) [In eqs. (1)-(2)] I f HI and - H 2 are positive, and if - H 2 is strongly positive and has finite gain, then E 1 and E 2 are bounded. (b) I r Hi and - H 2 are incrementally positive. and if - H 2 is strongly incre-. mentally positive and has finite incremental gain, then E 1 and E 2 are input-output stable. For example, if H 2 on L 2• is instantaneously confined to a sector of the plane, then, under the provisions of Theorem 3, the sector of HI lies in the first and third quadrants, and is bounded away from both axes"
5.1 Positivity and Passivity in Networks A passive element is one that always absorbs energy.
Is a network consisting of passive elements necessarily stable? An attempt will be made to answer this question for the special case of the circuit of Fig. 2. First, an elaboration is given on what is meant by a 15 A variation of this result was originally presented in [Zd], Kolodner (8) has obtained related results, with n restriction of
linearity on one of the elements.
292
passive element. Consider an element having a current positive constant." In fact, the conic sectors defined i and a voltage v; the absorbed energy is the integral here resemble the disk-shaped regions, on a Nyquist J:i(t)v(t)dt, and the condition for passivity is that this chart. However, Theorem 2 differs from Nyquist's Cri .. integral be non-negative. Now, let Z be a relation map- terion in two important respects: (1) Unlike Nyquist's ping i into t/; by analogy with the linear theory, it is Criterion, Theorem 2 is not necessary, which is hardly, naturalto think of Z as an impedance relation; suppose surprising, since bounds on HI and H 2 are assumed in Z is defined on L 2et where the energy integral equals the place of a more detailed characterization. (2) Nyquist's inner product (i, v); then passivity of the element is criterion assesses stability from observation of only the equivalent to positivity of Z. Similarly, if Yon L z, is an eigenfunctions exp (jwt) , where Theorem 2 involves all admittance relation, which maps v into 1~, then passivity inputs in X,. There is also a resemblance between the use of the is equivalent to positivity of Y. Now consider the circuit of Fig. 2. This circuit con- notions of gain and inner product as discussed here, and sists of an element characterized by an impedance rela- the use of attenuation and phaseshift in the linear thetion Z2, an element characterized by an admittance rela- ory. A further discussion of this topic is postponed to tion Y 1, a voltage source v, and a current source i. The Part II, where linear systems will be examined in some detail. equations of this circuit are, One, of the broader implications of the theory de(20a) veloped here concerns the use of functional analysis for (20b) the study of poorly defined systems. I t seems possible, i2 = i - i1 (21a) from only coarse information about a system, and per. t'2 = Z 2i '!. haps even without knowing details of internal structure, (21b) i 1 = Y 1V l to make useful assessments of qualitative behavior. It is observed that these equations have the same form ApPENDIX as the Feedback Equations. provided that the sources i and v are constrained by the equations v = alX +Wl, A. Proofs of Properties (i-vi) and i = a2x+w2. (By letting at =0 the familiar "parallel Properties (i, ii). These two properties are immedi.. circuit" is obtained. Similarly, by letting a2= 0 the ately implied by the inequalities "series circuit" IS obtained.) Thus there is a correspondence between the feedback system and the network con«Ix), - I-Xt, (Ix), - 1,xt) = 0 sidered here. Corresponding to the closed loop relation «cHx), - cax" (CHX), - cbXt) E 1 there is a voltage transfer relation mapping II into = c2«Hx), - ax" (Hx), - bx,) ~ 0 VI. Similarly, corresponding to E 2 there is a current transfer relation mapping i into it- If Theorem 3 is now in which c is a (positive or negative) real constant. applied to eqs. (20)-(21) it may be concluded that: If Property (iii). It is enough to show that (H+H1) both elements are passive, and if, in addition, the relation has center !(b+b 1+ a + al) and radius !(b+b1-a-al); of one of the elements is strongly positive and has finite the following ·inequalities establish this: gain, then the network transfer relations are bounded. If [(H + Ht)x] , - !(b + hi + a + al)xtlf 6.
:s II (Hx) , - !(b + a)x,1/ + 11 (H1x) , - f(b. + a'I)1f
CONCLUSIONS
The main result here is Theorem 2. This theorem provides sufficient conditions for continuity and boundedness of the closed loop, without restricting the open loop to be linear or time invariant. Theorem 2 includes Theorems 1 and 3 as special cases. However, all three theorems are equivalent, in the sense that each can be derived from any of the others by a suitable transforrnation. There are resemblances between Theorem 2 and Nyquist's Criterion. For example, consider the following, easily derived, limiting form of Theorem 2 ~ "If HI = kl then a sufficient condition for boundedness of the closed loop is that HI be bounded away from the critical value - (11k)!, in the sense that
:S !(b - a)lf x tff + !(b 1 = !(b
+b
-
(Triangle Ineq.)
a- al)llx,ff
l -
(Ala)
at)lIx,1I (Alb)
where eq. (Alb) follows from eq. (Ala) since H has center !(b+a) and radius !(b-a), and since H 1 has center l(b 1+al) and radius t(b 1-al).
Property (iv). CASES
la
AND
l b: l·Iere a¢O and b>O, and
«H-IX)I = (YI -
+ +
X" (H-1x), (Hy)" YI -
:
x)
: (HY)I)
1
= ab «(By),
for all x in X. and t in T, where 6 is an arbitrarily small 293
- aYt, (Hy), - by,)
=,
where H-IX and x = H,. Since. by hypothesis, H is inside {a, b J and b> 0, the sign of the last expression is opposite to that of 0,. Thus the Inverse Rule is obtained. CASE 2: Here a = o. The property is implied by the
(III) (Using Fig. 5 as a guide,) define two new elements of X"
+ ce2 et' = el + CYI.
Y2'
inequality,
(X"
(H-I X). -
~
x,)
=
~
{(Hy)" by, - (Hy),)
~ o.
Property (v). Simply reverse all the inequality signs. Property (vi).
II (Hx) ,II :5 II (Hz), -
1(6
+ Hi(b + a)%,11 (Triangle Ineq.) s l(b - a)llxtlf + i l b + a 1·!fXt!l = max ( I a l, 'b I)·11 xtll
where eq (A2b) follows from eq (A2a) since, from the hypothesis, H has center !(b+a) and radius f(b -a). It follows that g(H) ~max bf). Q.E.D.
B. Proof of Theorem 2a The proof is divided into three parts: (1) The transformation of Fig. 5 is carried out, giving a new relation E,'; E,' is shown to contain EI. (2) The new gain product is shown to be less than one. (3) E I ' is shown to be bounded, by Theorem 1; the boundedness of E 2 and E1 follows. Let c=i(b+a) and r=l(b-a).
s.i
Transformation of Eqs. (1)-(2)
The proof will be worked backwards from the end; the equations of the transformed system of Fig. 5 are, ~l' e2
Y2'
"1
=
Wt'
= 'UJ2
+ al'x + Y'l' + + Y1
~
H 2'es
::z
H 1' el'
t%2X
where H 2'
H l'
= (H2 + cI) = (H1- 1 + c1)-1.
el'
(A2a) (A2b)
(A6a) (A6b)
It shall now be shown that there are elements Hl'el' and H 2'e2' in X. that satisfy eqs. (A3)-(A4) simultaneously with the elements defined in (1)-(111). Taking eqs, (A3)-(A4) one at a time: Equation (A3a). Substituting eq. (la) for el in eq. (A6b). and eq, (Ib) for
+ a)x,(I
= Y2
= (WI
- cw,.)
)'t,
+ (al
- ca2)X
+ (yz + ce2)-
(A7)
If Wl'=Wl-CWS and a!'=al-ca2. then. with the aid of eq. (A6a), eq. (A7) reduces to eq. (A3a). Equation (A3b): This is merely eq. (lb), repeated. Equation (A4a): Recalling that H 2' = H 2+cl, it follows, from eqs, (A6a) and (2a). that there is an H 2'ez in X. for which eq, (A4a) holds. Equation (A4b): If eq. (A6b) is substituted for el in eq. (2b), it is found that there exists H1(el' -CY1) in X. such that Yl=H1(el'-cYl). Therefore, (after inversion)
HI-IYl
(after rearrangement) (H1-l
= el'
-
CYI
+ ,1)Yl = et'
(after inversion)
Yl
=
(H1-
1
+ Cl)-l~l'.
That is, there exists HI/el' in X. for which eq. (A4b) holds. Since eqs, (A3)-{A4) are all satisfied, (x, e2) is in E 2' . Since (x t e2) is an arbitrary element of E 2, E s' contains E 2• B.2 Boundedness of E 2'
It will be shown that g(H1' ) · g(H.') < 1. The Case ~>O, a=o: g(H2 ' ) will be bounded first. (A3a) Since H 2 is in { -b+d, -a-~} by hypothesis, (Ht+cl) (A3b) . is in {-b+A+c. -a-~+c} by the Sum Rule of Sec.. (A4a) tion 4.2. Observing that (H,,+cI)=H./, that (-b+c) == -r, and that (-a+c) =r, it is concluded that H 2 ' is (A4b) in {-r+l1, r-t1}. Therefore g(HI')~r-.1.. Next, g(HI ' ) will be bounded. In Case la, where a>O and H 1 is outside (A5a) (ASh)
(I t may be observed that these equations have the same form as eqs. (1)-(2), .but HI is replaced by HI' and H 2 is replaced by HI'.) Let E t ' be the closed-loop relation that consists of all pairs (x, e2) satisfying eqs. (A3)(A4). It shall now be shown that E I ' contains E 2• that is that any solution of eqs, (1)-(2) is also a solution of e~s. (A3)-(A4); thus bounded ness of E 2' will imply bounded ness of E 2 • In greater detail
(I) let (x. e2) be any given element of E 2(II) Let elf Yh )'2, R 1e l , and H 2ez be fixed elements of X. that satisfy eqs, (1)-(2) simultaneously with x and e2.
the Inverse Rule of Section 4.2 implies that HI-I is out.. side t -b, -a'}; the same result is obtained in Cases lb and 2. In all cases, therefore, the Sum Rule implies that (H1-t+CI) is outside {-r, ,}. By the Inverse Rule again, (H1 -l+C!)-1 is in
Therefore g(H1' ) ::; l/r. Finally,
g(H1' ) • g(H2/ )
r -
T
294
.1
S -- <
1.
The Case A=O. <1>0: It shall be shown that this is a special case of the case 4> 0, a== O. In other words, it will be shown that there are real constants a*, b", and A· for which the conditions of the case .6>0,6=0 are fulfilled, but with a replaced by a·, b by b", and 4 bya*. Consider Case Ia, in which 4>0. (Cases Ib and 2 have similar proofs, which will be omitted.) It must be shown that: (1) - H 2 is in {a·+~, b* -~}. (2) HI is outside
-(Xc, (H2 x),) ~ O'IIXtll:!
II (H x) ,II s x 1lX t/l z 2
a*
a
0 and
- b·
-
2,0' + rZ)J1xtllz.
(All)
The author thanks Dr. P. Falb for carefully reading the draft of this paper, and for making a number of valuable suggestions concerning its arrangement and con .. cerning the mathematical formulation. REFERENCES
[1] V. M. Popov, "Absolute stability of nonlinear systems of automatic control," Automatic and Remot, Control, pp. 857-875.
March 1962. (Russian original in August 1961.)
by hypothesis, and since by construction
- - >- - -
~ (A 2
ACKNOWLEDGMENT
{- : - 6, - ~ + 6} 1
(AIO)
Equation (All) was obtained by expanding the square on its l.h.s., and applying eqs. (A9) and (AI0). Con.. stants X, r, and ~, are selected so that X>cr, ,=)...1/(/', and A=,[1-V1-(u/A)2]. Now it can be verified that, for this choice of constants, the term (~2_ 2rO"+,2) in eq, (All) equals (r-A)2; also, O<&
Since - HI is in {a, b} by hypothesis, and since a* b by construction, there must be a A*>O such that H s satisfies condition (1). Since HI is outside
1
2
for any :x in X. and for any t in T. Hence, for any r > 0,
H(H2x), + rXtll:z
Without loss of generality it can be assumed that ~ is smaller than either l/a or lib. Choose a* and b* in the ranges a b - - - < a* < a and b < b· < - - · 1 + a8 1 - bo
1
2
(A9)
[2] (a) G. Zames, "Conservation of bandwidth in nonlinear opera-
1
<- -
b
tiona," M.I.T. Res. Lab. or Electronics, Cambridge, Mass., Quarterly Progress Rept. 55, pp. 98-109, October 15, 1959. (b) - - , "Nonlinear operators for system analysis," M.l.T. Res. Lab. of Electronics, Tech. Rept. 370, September 1960. (c) - - , -Functional analysis applied to nonlinear feedback systems," IBEE Trans. on. Circuit Theory, vol. CT..IO, PP. 392404, September 1963. (d) - - , ·On the stability of nonlinear, time-varying feedback systems." Proc, NEe, vol. 20 pp, 725-730, October 1964. (e) - - , "Contracting transformations-A theory of stability and iteration for nonlinear. time-varying systems, Sumfll4,ies, 19641nU,nat'l Con!. Oft Mt€rt1flJtJ,fles, eireM" TMOf''Y, 4,", 1ft/or-
+ 6,
condition (2) is satisfied. Hence this is, indeed, a special case of the previous one. B.3 Conclusion of the Proof Since g(H1')-g(8.') < 1, El'is bounded by Theorem 1, and so is E I , which is contained in E I ' . Moreover, the boundedness of E 2 implies the boundedness of E 1 ; for, if (x, el) is in 8 1 and (x, e2) is in E I , then
mation TheM" pp, 121-122.
(f) - - , "Nonlinear time varying feedback systems-Condi-
tions for L II) -boundedness derived using conic operators on exponentially weighted spaces," Proe. 1965 Alkrlon Conf., pp. 460-471. [3] K. S. Narendra and R. M. Goldwyn, "A geometrical criterion for the stability of certain nonlinearnonautonomous s~tem8," IEEE Trans. on Cir,"u TJuOf'Y (CorrejpOfUlenu), vol. CT-!l, pp. 406408, September 1964. (4} R. W. Brockett and ]. W. Willems, "Frequency domain stability criteria, Jt pts, I and II, 1965 Proc. Joint Automatic Control Conf., pp. 735-747. (5J (a) I. W. Sandberg, "On the properties of some systems that distort signals, JI Bell Sys• .Tech. .L; vol, 42. p. 2033, September 1963, and vol, 43, pp. 91-112 January 1964.
tlelll S IIwIII + fall ·11%11 + g(H )lI e211. (A8) Thus, if.llxll ~const. and IleJII ~const.t then IIsll1 ~con5t. t
(Inequality (AS) was obtained by applying the Triangle Inequality and Inequality (4) to eq, (Ia), and taking the limit as ,-. co. It may be noted that g(R,.)
c. Proof of Theorem 2 This shall be reduced to a special case of Theorem 2 [Case 2 with 8==0]. In particular, it shall be shown that there are constants b>O and 4>0 for which (I) -Ht is inside {~, b-A}, and, (II) the relation [H1+Cl/b)I] is positive. [HI+(l/b)l] is clearly positive for any b>O, since by hypothesis H 2 is positive; the second condition is therefore satisfied. To prove the first condition it is enough to show that HI is conic with center - , and radius r-4, where ,==6/2. This is shown as follows: The hypothesis implies that, for some constant 0'>0 and for any constant). > g(H2) , the following inequalities are true
[6] {7] [8]
[9]
(101
(b) - - ,"On the L,..boundedness of .olutions of nonlinear functional equations," Bell. Sys. Tech. J., vol. 43, pt. 11, pp. 1581-1599, July 1964. J. Kudrewicz, "Stability of nonlinear feedback systems," Automa.ti~a i TelemecMnik4. vol. 25. no. 8, 1964 (and other papers). E. H. Zarantonello, "Solving functional equations by contractive averaging," U. S. Army Math. Res. Ctr., Madison, Wis. Tech. Summary Rept. 160, 1960. I. I. Kolodner, "Contractive methods for the Hammerstein equation in Hilbert spaces." University of New Mexico, Albuquerque, Tech. Rept. 35, July 1963. G. J. Minty, "On nonlinear integral equations of the Hammersteintype," survey appear!ng in N.tmlinea.' In!'frtd Eqll4twns, P. M. AnseJone, Ed. Madison, WIS.: UnIversity Press, 1964, pp.99-154. F. E. Browder, "The solvability o( nonlinear functional equations, Jt Duk, J{alii. J., vol. 30. pp. 557-566, 1963.
J. ]. Bongiorno, Jr., "An extension oE the Nyquist- Barkhausen stability criterion to linear lumped...parameter systems with timevarying elements," IEEE Trau. Oft Automtdit: Control (Cor,espontlcnce), vol. AC-8, pp. 166-170. Ap_ril 1963. [12] A. N. Kolmogorov and S. V. Fomin, Fu",tio,ud Analysis, vols. (11]
I and II. New York: Graylock Press, 1957.
295
On the Input-Output Stability of Time-Varying Nonlinear Feedback Systems-Part II: Conditions Involving Circles in the Frequency Plane and Sector Nonlinearities G. ZAMES,
MEMBER, IEEE
Abstract"':'-The object of this paper is to outline a stability theory based on functional methods. Part I of the paper was devoted to a
general feedback eonftguration. Part II is devoted to a feedback system consisting of two elements, ODe of which is linear time-in-
variant, and the other nonlinear. An attempt is made to unify several stability conditions, Including Popov's condition, into a single principle. This principle is based on the concepts of conicity and positivity, and provides a link with the notions of gain and phase shift of the linear theory. Part II draws on the (generalized) notion of a "sector Don-. linearity." A nonlinearity N is said to be INSIDB THE SECTOR {a, fJ} if it satisfies an inequality of the type «Nx-ax)" (Nx-fJx),)~O. If N is memoryless and is characterized by a graph in the plane. then this simply means that the graph lies inside a sector of the plane. However. the preceding definition extends the concept to include nonUnearities with memory. There are two main results. The first result, the CIRCLE THEOREM. asserts in part that: If the nonlinearity is inside CI sector {a, 13 }, and if the frequency response of the linear element avoids a 'lcritical region" in the complex plane~ then the closed loop is bounded; if a> 0 then the critical region is a disk whOle cente: is halfway between the points -I/a and -l/{J, and whose diameter is greater than the distance between these points. The second result is a method for taking into account the detailed properties of the nonlinearity to get improved stability conditions. This method involves the removal of a c'multiplier" from the linear element. The frequency response of the linear element is modified by the removal, and, in effect, the size of the critical region is reduced. Several conditions, including Popov·s condition, are derived by this method, under various restrictions on the nonlinearity N; the following cases are treated:
LINEAR TIME-INVARIANT a,x+~
NONUNEAR NO MEMORY
Fig. 1.
Nx(t)
A feedback system.
J
1m[H(juJ)]
---".::;;..",.----...x(t)
(a)
(b)
Fig. 2. If N:x(t) vs. x(l) and R(jw) lie in the shaded regions, and if the Nyquist diagram of H(j6J) does not encircle the cri tical disk, then the closed loop is bounded.
t
(i) N is instantaneously inside a sector a, P}• (ii) N satisfies (i) and ismemoryless and time-invariant. (iii) N satisfies (ii) and has a restricted slope.
1.
T
supposed, for the moment, that N has no memory.. These assumptions are among, the simplest which ensure that the system is both
INTRODUCTION
H E feedback system of Fig. 1 consists' of a linear time-invariant element H and a (not necessarily linear or time-invariant) element N. It will be
Manuscript received December 29,1964; revised October 1. 1965, and May 10, 1966. Parts of this paper were presented at the 1964 National Electronics Conference [Ia], Chicago. Ill. This work was carried out in part under support extended by the National Aeronautics and Space Administration under Contract N sG-496 with the l\1.I.T. Center for Space Research. The author is with the Department of Electrical Engineering, Massachusetts Institute of Technologr' Cambridge, Mass, 1 A single input x, multiplied by rea constants 41 and aI, is added in at two points. By setting at or a2 to zero, it is possible to obtain a single-input system. in which the element closest to the input is either the linear element or the nonlinearity. The terms Wl and W2 are fixed bias functions, which will be used to account for the effects of initial conditions. The variables tl and e2 are outputs.
296
(i) general enough to have many applications (ii) complicated enough to exhibit such characteristic nonlinear phenornena as jump resonances, subharmonics, etc.
The object here is to find stability conditions for the closed-loop system. For practical reasons, it is desirable to express these conditions in terms of q uan ti ties that can be measured experimentally, such as frequency responses, transfer characteristics. etc. In particular, the following question is of interest: Imagine that the graph of N lies inside a sector of the plane, as shown in Fig. 2(a), and that the frequency response of H is plotted in the complex plane; can the complex plane be divided
into regions that are "safe" or "unsafe" as far as stabil..
i ty is concerned? It will be shown that, with certain qualifications, such a division is possible. In fact it has already been shown in Part I that such regions, called 'Iconic sectors," exist in a quite general sense. Here these general results will be applied to some concrete situations, involving frequency responses, etc. (Fig. 2, which illustrates the simplest of the results to be obtained here, gives some idea of what is being sought.) 2.
STATEMENT OF THE PROBLEM
The purpose of this section is to define Hand N, and to write feedback equations. Hand N will be represented by input-output relations or by operators, in keeping with the theory outlined in Part I. DEFINITION: R [0. 00) is the space of real-valued functions on the interval [0, co). L" where p = 1, 2, · . . t is the space consisting of those x in R [0, «») for which the integral J: %(1) pdt is finite. In addition, for the case p == 2, it is assumed that L 2 is an inn"-product space, with inner-product
I I
cases, the ~ conditions imply L.-boundedness or continuity. For physical applications the most appropriate definitions of boundedness and continuity are, of course, obtained in the L norm. DEFINITION: Let
2.1. The Operator Classes
m. and. .c
DEFINITION: m is the class of operators on Lt. hafJing thefollowing property: If N is in ~ then there s a function. N: Reals-oReals, satisfying'
=
Nx(t)
(x E L 2 , ; t
~V(x(t)
~ 0)
(1)
and harling the following properties: (i) N(O) =: 0, (ii) f N(x) ~const. and (iii) for any real x, J:N(x')dx' is finite. (X, y) = ~ ..x(t)y(t)dt An operator in m. is memoryless, time ..i nvariant, not necessarily linear, and can be characterized by a graph and norm The symbol without subscript, will· in the plane. The letter N will indicate the graph of N. often be 'UsedJIIstead of l{xI12" DEFINITION. .c is the class of those operators H on L 2t L.. is tTJi space consisting of those functions x in R (0, co) that are measurable and essentia.lly bounded. L. satisfying an equation of the typel
lIxlit.
I
Ixl,
llxlL
is assum«l to be CJ ",ormed linear space, -with norm
Ilxll. = ess sup I :I:(t) I·
Hx(t) = h..x(t)
+
1:
I
h(t - 1')x(1')d1' (x
E £2.; t
~ 0)
(2)
r~O
No distinction wiU be made between functions diff~ing OfJeI' s6ls of .ero 1IIMS"", Those definitions which were introduced in Part I will only be summarized here. Following the convention of Part I, the subscripted symbol Xa denotes a function in R[O, 00) trunctJtetl after [0, t). The space L ••, where p = 1, 2J • • • , co t is the extension of L II, i.e., L p . == {zJsER[O, co) and %,ELpforallt~OJ.
p• IIxU•.
An utentktl norm lI~np. is defined on L t where JlxIlP. ==llxll , if xELp and 11%11".== ClCI if %EEL". The symbol Ie will usually be abbreviated to The concept of a relation H on L~•• with domain Do(H) and ,ange Ra(H) was introduced in Part I. A relation H on L __ is L,,-boundetl if H maps bounded subsets of L p • into bounded subsets of L p , . His Lp..continuous if, given any % in Do(R) and any ~>O, there is a 6>0 such that, for any 1 in Do(H)t if IIx-yllp.
llx
in which h. is a real constant, and lhe impulse response h isa function in L 1 with the property that, for some tTo < 0, h(t) exp (-aot) is also in L 1• Operators in £ are linear and time-invariant.
2.2. Feedback Equations Consider the feedback system of Fig. 1, but with two
modifications: (i) N is not necessarily rnernoryless: (ii) 41 and at are operators on Lt., multiplying s: (This amount of generality will be needed for some of the intermediate results; ultimately, the interesting case is that in which N has no memory, and at and al are real constants.) The equations of this system are
= 41 X + WI -
e2
::=
42 X
Ne2
(3a)
+ + He,
(3b)
W2
in which it is assumed that:
lIH%-H,IIJ1'<~'
Part II will be devoted entirely to finding L, conditions (for boundedness and continuity), since these are easier to derive than the other L. conditions. However, most of the results of this paper have been extended to the L. norm, in [Ib]. It has been found that, in most
et
H is
4ft,
N is
Q,
operator in ,£ relation in £Ro
tIt can be verified that every mapping ot the type N:LIJ-+R(O. GO ) satisfying (1) is in (act an operator on Lit. Similarly, every mapping H:L......R(O.IO) satisfying (2) is an operator on L,. {see (81) or
Appendix Bl. 297
X
111 Lt, is an input
(0) N IS INSTANTANEOUSLY INSIDE A. SECTOR {
'1 cnd 't t1l Lt. are outputs
and fDJ in L 1 are fixsd biases eithe, (11 and a2 are real constantsJ or, more generally,3 al is a relation on Lt. harJing the property that flalx fl . :Sconst . IIxll' and similarly for at. REMARK: If. to begin with, the linear element satisfies a state equation. then He« is set equal to the "zeroinitial-condition response" of the state equation. and WI is set equal to the "zero-input response." The closed-loop relations which map % into el and et will be denoted El and Bt.. The objective here is: Find conditions 011 Nand H which ensure that E1 and & are Lt-bountktl ana L t -continu,01J,s.
(b) N IS POSITIVE
Wi
Nx(t)
l
3.. CONDITIONS FOR CONICITY AND POSITIVITY
(x E
Do(B); t ~ 0)
Permissible regions (shaded) (or instantaneously confined nonlinearities.
(a) tot IS INTERIOR CONIC
This section has some preliminary results, which will be needed later in the analysis of stability. The following questions are fundamental to this analysis. Under what conditions is an operator conic or positive? Under what conditions is the composition product of two operators conic or positive? The definitions of conicity and positivity were introduced in Part I. They are repeated here, for the special case of relations on L 1... DEFINITION: Let H be Q, relation in (Ro. H is INTERIOR CONIC if c and r ~ 0 ar« rtal constants and the inequality
II (8%), - ex,1t Srllz,1I
Fig. 3.
(4)
holds. H is e%te,io, conic if (4) holds with the inequality sign '6fJ~sd. c will be called the center parameter oj H, and , will be CtJUetJ, the radius parameter of H. H is INSIDE (OUTSII>E) THE SECTOR {4, ~} if a ~{J a.nd if the inlquality
eb) H
1m
IS EXTERIOR CON'C
1m
IS POSITIVE
1m
Re
--1+It-It'~~Re
Fig. 4.
ee) H
Re
Permissible regions (shaded) (or the frequency response HljllJ).
inside {a, fJ} if. in addition. N satisfies the slope restrictions a~ [N(x)-N(y)]/(x-y)S/J.. N is positive if its graph lies in fhe first and third quadrants; N is incremen tally positive if, in a.ddition, N is nondecreasing. 3.2. Linear Time-Invariant Operators
Consider the operator class .c; it will be shown, roughly speaking, that a conic sector has a counterpart «H% - a2:)., (Hx - fJx),) S 0 (x E Do(H)j t ~ 0) (5) in the frequency plane, in the form of a circular disk (see Fig. 4). This disk degenerates into a half-plane in holds (with the inequality sign retJers~d). H is POSITIVE if it satisfies the ineguality (Xh (Hx).) " the case of a positive operator. DEFINITION: Let s c: tI +j~ denote a point in the complex ~ 0 for alZ :JC in Do(H) and all t ~ O. plane. The LAPLACE TRANSFORM H(s) of H in .e is In Part I, the concepts of conicity and positivity were subdivided into categories such as "instantaneous" (6) (u ~ 0) H(s) = h.. + ~ exp (-sl)dl conicity, "incremental" conicity, etc. The definitions of these terms are listed in Appendix A. REMARK.: The following conditions are equivalent: (The integral on the right-hand side of (6) exists and is (i) H is interior conic with parameters c and r, (ii) H analytic for (S 2: 0 [See (B 1) of Appendix B ]~) is inside {c-r, c+r}. DEFINITION: The NYQUIST DIAGRAM of H(s) is a curve in the complex plane consisting of: (i) the image of the 3.1. Memoryl,ss, Time-Invariant Nonlinearities jw-axis under the mapping H(s), and (ii) the point h•. Consider the operator class m; the conditions for N LEMMA 1. Let H bean operator in ~,andlet c and r~O in 9t to be conic. positive, etc.• are simply the "instan- be real constants. taneous" conditions of Appendix A. Some of these con... (a) If H(s) satisfies the inequality ditions are illustrated in Fig. 3. In particular, N is inside the sector {a. fJ} if its graph lies in a sector of the plane (7) I H(jw) - c :5' (w E (-00 ,00) bounded by lines of slopes a and fJ i N is incremen tally then H is incrementally inte1'io1 conic with center param-
\(t)
I
I
The more general assumption will be needed in Section S only.
eter c and radius parameter r, 298
(b) If H(s) satisfies the inequality
I H(jtJJ)
-
C
I ~,
(~ E (-00.00»
(0)
(8)
and if the Nyquist diagram of H(s) does not encircle the poin' (c, 0) t then H is incrementally exterior conic with center parameter c an~ radius param-
-- »-
~I0 ,
2
J,
2
0
I
-d
3
eterr, (c) If Re{H(j~)}~O for wE(-co,co) then H is incrementally positive.
-2
The Proof of Lemma 1 is in Appendix B. REMARK. The gains g(H) and ,(R) were defined, in Part I. I t follows from Lemma 1(a) that if HUw) ~ c, then i(H) =g(H) ~c.
I
Fig. 5.
-,
0
1
2
I
1
2
0
3
4
Products of intervals.
I
(0)
3.3. Composition Products and Sector Products
The composition product of two positive operators need not be positive. Those special cases in which the product is positive are of interest because they give a tighter bound on the composite behavior than would be obtained in general. (They form the basis of the factorization method of Section 5.2.) Similarly, those special cases in which the product of two sector operators lies inside the "product sector" are of in terest. DEFINITION: The PRODUCT SECTOR {al. ~l} X {at, {JI}
is the sector {a. (j} t where [a. ~] is the interval of the reels defined by [a, p] == xE [ai, #91] and yE [aI, ~I]}. In other words, product sectors behave like pointwise products of the corresponding real intervals (see
{%,I
Fig. S). It is easy to show that if both operators are memoryless, say if both operators are in 9l, then their product has the above mentioned properties. [This can be shown by expressing the ratio N 1(N1 (x» / x as the product (N1(y)/ , ) x (Nt(x)/x) , where ,~N,(x).] More difficult are cases in which one operator is in 9t and the other is in .e, as in Lemmas 2 and 3.
1m
(b)
Re
Fig. 6.
a:le operators.
(a) Typical pole-zero pattern. (b) Circle of
confinement of
KU,.,).
3.5. A Jlemoryless Nonlinearity and an CRe MuUiplier A situation resembling Lemma 2, but with N more restricted and K more general, is considered next. K is taken to be a sum of first-order terms, of a type that can be realized as the driving-point impedance of an RC network (see Guillemin [2], ch. 4). DEFINITION: Let
where ki~O, >.,>0, and K(ao)~O are real constants. An operator K in me has poles and zeros alternating 3.4. A Memoryless Nonlinearity and a First-Order on the negative-real axis, with a pole nearest but not at Multiplier the origin. The frequency response of K lies inside a The following lemma is the basis for Popov's condi- circle in the right-half plane, located as shown in Fig. 6(b); it follows that K is positive and inside the sector tion (Section 5.1). {K{ co), K(O) }. (Observe that LEMMA 2. Let N be an operator in m, K be an operator in .£, ani/, leI ehe Laolace transform of K be K(s) =11"/$+). who, k>O and ~>O. K(O) = K(co) + 2: ki ~ K(~).) (a) If N is posilifJe' then NK'is positifJe. (b) If N is insidel (I sector {a, ~} then NK is inside LE~IMA 3. Let N be an operator in m, and K be an ft
the product sector {a,
{j} X {O, s},
operator in me. (a) 1/ N is incrementally positive then NK is positi1Je. (b) If N is incrementally inside the sec/or {a, {j} then NK is inside the product sector {a, p} X {K( ee), '
The proof of Lemma 2 is in Appendix C. (Note that K itself is positive and inside {O, s}, since XU",) lies entirely in the right-half plane and since IK(jC&1)-ik! =;k.)
, Le., xN(x) ~O. , i.e., t.r ~ N(~)/%'5:~.
K(O)}. In other words, multiplication by K affects the composite sector as if K had no memory. The proof of Lemma 3 is in Appendix D. 299
4.
CIRCLE CONDITIONS FOR STABILITY
Consider now the main problem of this paper. namely, the problem of stability for the loop of Fig. 1. Suppose that ~ i~ a relation (which mayor may not be memoryless) Inside a sector {a, fj}. What conditions on the frequency response H(jCAJ) are sufficient to ensure boundedness of the closed loop? It will appear that the following "circle conditions" are sufficient: DEFINITION: H(j6J) will be sQ,id to SATISFY THE CIRCLE
{a, {3},
--.-o;,~--""
L'NDICATES SLOPE
1m
6, P>O, and a~o are real constants, if the fol-
CONDITIONS FOR THE SECTOR
WITH OFFSET
HOw) PLANE
where a5:~, lowing conditions hold: CASE lA. If a>O, then
I
H(jw)
I
+ ~ (~ + ~) ~ ~ (~
- ~) +
BOUNDEDNESS DISK FOR (0) AND (b),
&
(w E (- QO ,co))
(iO)
and the Nyquist diagram of HUw) does not encircle the point -1(l/a+l/~). CASE
I
1B. If a
H(jw)
< 0, then
I
+ ~ (~ + ~) s ~ (~ - ~) (w E (- cc,
&
00».
(11l
CASE 2. If a = 0, then Re {H(jCJJ) } ~ - (l/{j) +6 for CalE(-oo,oo). In other words, the complex plane is divided into two regions, shaped either like a circular disk and its complement, or like two half-planes. (The case a>O is illustrated in Fig. 2.) One of the regions will be called "permissible" and the other will be caned "critical." If H(je,)) does not enter or encircle the critical region, then the closed loop is bounded. If, in addition, N is incrementally inside {at fJ}, then the closed loop is con tinuous, These results are formalized in the following theorem:
A
CIRCLE THEOREM.
)(
(I) N is a relation in
The Circle Theorem is based on Theorem 2 of Part I. I t was assumed in Theorem 2 that al and at were real constants. However, with only minor changes in the proof, it can be shown that Theorem 2 holds more generally if al and a2 are relations on Lt., provided al and 42 S3;tisfy inequalities of the type lIalxll.~const. IIxll •. The Circle Theorem then follows immediately with the aid of Lemma 1 of Part II.
The Circle Theorem can be viewed as a generalization of Nyquist's criterion.! in which a critical region replaces the critical point. For a given N there are two critical regions, one for boundedness and one for continuity. It can be shown that the continuity region always contains the boundedness region (see Example 1 and Fig. 7). The Circle Theorem will serve as the generating thea.. rem for the rest of this paper; i.e., the remaining results will be obtained as corollaries to the Circle Theorem by variously constraining the form of N. In particular. the following corollary is obvious. 4.1. A Circle Condition for Instantaneous Nonlinearilies 7
Suppose that
which is greater than zero.
Fig. 7. Critical disks for Example 1. (Broken curve indicates edges of jump region in H(j",) plane.)
COROLLARY 1. If (I) N in CRo is instantaneously (inc,ementally)"inside the sector {a+A. ,g-a} where fJ>O'. and if conditions (II) and (III) of the Circle Theorem hold, then E1 and Et are L.,-bounded (L 2-conli nfl Ott s) . EXAMPLE 1. (a) Let No be the relation shown in Fig. 7(a), and Nbe the relation in
t
l\r~o~e accurately. o( the sufficient part of Nyquist's criterion. Similar or closely related circle conditions were found inde~ndently by the auttior [tal, Sandberg (51. Narendra. and Goldwy« [6), and Kudrewicz (71.
,
7
300
Observe that the nonlinearity N in Corollary 1 can be time-varying and can have memory. In fact, very little has been assumed about the detailed character of N. The price paid for this is that Corollary 1 is often conservative, i.e., the critical region is too large. This is especially true of the boundedness condition (see Example 2). The continuity condition probably gives a quite fair estimate of what to expect. In fact, an approximate analysis, based on the harmonic balance method (cf. Hatanaka [3]), suggests' that continuity breaks down in the following way: There is a zone, inside the critical continuity disk, in which jump-resonance phenomena occur. The zone is not much smaller than the continuity disk. Furthermore, the magnitudes of jump resonances depend on the Nyquist diagram behavior inside the continuity disk.
5. CONDITIONS WITH TRANSFERRED MULTIPLIERS The next two corollaries can be viewed as attempts to reduce the size of the critical region, at the cost of added restrictions on N. In certain cases, it will be possible to remove a "multiplier" K from the linear element, before applying the Circle Theorem. The removal of K will shift the frequency response of the remainder, H 1UCIJ), away from the critical region. Thus the effective size of the critical region will be reduced.
5.1. Popov's Condition Consider the feedback system of Fig. 1 t under the same conditions as in Corollary 1, but with the added constraint that N is a rnemoryless, time-invariant operator. The following condition for boundedness (not continuity) involves the removal of a first-order multiplier from H. COROLLARY 2.
If
(I) N is an operator in 9t, inside the sector {a, (j} where 13>0. (II) H is an operator in .c that ai« be factored into a product H =KH1, where HI and K are in .c, and K(s) ==~/(s+") where >-->0. (III) H 1 satisfies the circle conditions for the product
sector {a, fJ} X {O, 1} with offset a, where 0> o. = 0 and W2 is in L 2, where t02 denotes the derivQ..
(IV) at
tifJe on
[0,
co).
Then E1 and &. are L 2-bountkd. (i) For a>O, Condition III simply means that Ref (j6J+'X)H(jllJ)} ~ -~/P+6. (ii) Condition IV limits the result to that configuration in which the directions of flow is from the input to H to N. PROOF OF COROLLARY 2. The feedback equations will be transformed, as illustrated in Fig. 8; i.e., H will be split into a product, H=KH1 , and the multiplier Kwill be transferred into a composition with N. It will then be shown, in Step 1, that the transformed equations are bounded, and, in Step 2, that they are equivalent to the original eq uations as far as stabili ty is concerned. REMARKS:
Fig. 8. A transformation.
Letting W2' =W2+X-l~, and recalling that at == 0, consider the equations of the transformed system of Fig. 8, el'
= WI + 01X -
NKe2'
e2' = wz' + Hlel'
(12a) (12b)
Let El' and E 2' be the closed-loop relations Cor (12a)(12b). STEP 1. E l ' and &' are Lt-bounded. This follows from the Circle Theorem whose hypotheses are satisfied because: wl is in L 2 by assumption IV; NK is in the prod.. uct sector {a, {j} X {O, 1} by Lemma 2; and NK satisfies the appropriate circle conditions. STEP 2. It will be shown (below) that (13a) (13b)
Since &' and E z' have been proved L 2-bounded, and since K is certainly L 2-bounded , it follows that E1 and 82 are Ls-bouilded. To prove (13b), recall that E s and ~' are subsets (of a product space), so that it is enough to establish that each contains the other. Suppose that (x, e2) is an element of E J ; by definition of E 2 , there is an el in L 2• satisfying (3a)-(3b); let el' = 81 and e2' = wt' +H1el. Direct substitution shows that (x, el', e2 /) satisfies (12a)-(12b), so that (x, et') belongs to Et ' . Substitution also shows that e2 = Ke2', so that (x, e2) is in KE,'. Since (x, e2) is an arbitrary element of Bt, it follows that KE,,' contains Et. It can similarly be shown that E,. contains KE'I.', so that (13b) holds. The proof of (13a) is similar. Q.E.D. EXAMPLE 2. Let N be the operator in et whose graph is shown in Fig. 9(a), and let H(s) = k/(s+'A) (s+it). For what values of k is the closed-loop L 2-bounded? Compare Corollaries 1 and 2. Here N is inside {O, 1 }, so the critical region is a half~ -1 +8, in both corollaries. In Corollary plane, 2 let K(s) =A/(S+A) and HI(s):=: kjA(s+Jl); the following estimates are obtained:
Re{.}
Corollary 1: -hlJ.
Nx(t)
5.3. A Slope-Restricted Nonlinearity and an
Im[·J
plier - - . -. . . .o J - - -... x(t)
-fH!~-~----+
(a)
Consider the feedback system of Fig. 1, under the conditions of Corollary 2, but with an added slope restriction on N, and a more general type of multiplier K. COROLLARY 3. 8 If:
Re[ .)
(I) N is an operator in
Cb)
{a,,B} where ~>O.
in (b) POPOV'S METHOD
incrementally
insid~9
(1I) H is an operator in .c, which can be factored into a product H == KH1, where K is in me and HI is
Fig. 9. An example of' Popov's Condition. Arrows indicate shift away from critical region. . (a)COROLLARY 1
m.,
.c.
(III) Hl satisfies the circle conditions fOT the product sector {a, ~} X {K( <Xl). K(O)} with offset a, where sc-o. (IV) There is 10 a W2' such that K(Wt') == 'U/2. (V) Either K( co) > 0 or at =0.
(c) CQROLLARY 3
Then E1 and E" are L'I,-bounded. Corollary 3 is obtained immediately by the factorization method, with the help of Lemma 3. Fig. 10. Nyquist diagrams for Corollaries 1-.1.
5.2. The Factorization Method The proof of Corollary 2 suggests a method for generating a class of Popov-like conditions. The method consists of a factorization of H into H=KH1, followed by the transformation of (3a)-(3b) into (12a)-(12b) followed by an application of the Circle Theorem. Various stability conditions are produced by variously choosing the multiplier K. The method has two preconditions:
(Ia) Either K-l exists or Q2=O. (Ib) There is a W2' in L 2 such that K(W2') =W2. These preconditions ensure that (3) are transformable into (12). Note that if K-l exists, then at need not be zero; however, in that case, (12b) must be modified by the addition of a term a./x where a,'x=K-1atX (that is, a·/ is a relation on Lt.). The method is worthwhile only if it gives a smaller effective critical region than Corollary 1. This happens if: (IIa) NK lies in a sector not greater than the product of the sectors of Nand K. (lIb) KH1 lies in a sector greater than the product of the sectors of K and HiIf Requirements (IIa)-(Ilb) are satisfied, then it is advantageous to transfer K from a composition with HI into a composition with N. Requirement (IIa) usually means that the multiplier K has a very special form, and the 'difficulty in finding suitable multipliers is the main problem in applying this method. Once K is fixed. Requirement (lIb) defines a (limited) class of operators H for which this method is useful. As an illustration of this method, a condition re-
sembling Popov's is derived next.
REMARK: For suitably restricted N, Corollary 3 has several advantages over Popov's method: (i) The shift away from the critical region, which depends on K(j",), can be controlled more flexibly as a function of C&J. This is likely to be useful where a negative magnitude slope d/tUaJf H(jw) I is followed by a positive slope at a larger w. (ii) 42 need not be zero if K ( ClO ) > o. (iii) If a>O, the critical region predicted by Corollary 3 (a disk) is sometimes smaller than by Popov's method (which always gives a half-plane for a>O). EXAMPLE 3. Let N be the operator in 9t whose graph is shown in Fig. 9(a) (the same as in Example 2), and let H have the Laplace transform
H(s)
= k,-l { S + 1 } s+ ,-1 { (s
+r
,-ts 2)(s
':$}
+ r + (s + r)(s + r 1 )
2)
where r»l. For what values of k is the closed loop bounded? Compare Corollaries 1, 2 and 3. Figure 10Ca) illustrates the significant features of the Nyquist diagram of H(jw) (not drawn to scale). Observe that H(jw) has two "pass bands," one for ,-2
• Corollary 3 and the factorization method. in a functional settinJt were introduced by the author in Reference [tal. A related method In a. Liapunov setting has been exploited by Brockett, Willems, and Forys (4aH4b). ••
N(x)-N(y)
I.e., a~ - - - %-Y~fJ.
10
X(
302
00
This condition is satisfied a.utomatically if Ki » »0. If )-0. then it is satisfied if till is in L".
Corollary 1 predicts boundedness for -1 < k < 8 approximately. Popov's method is useless here. A comparison of Figs. 10(a) and lOeb) shows the effect of removing the multiplier )./s+": BtU",) is moved away from the critical region in the lower left half-plane (in the decaying edge of the lower pass band, ,-1 <w < 1); however, this improvement is more than offset by the bulge introduced in the upper left half-plane (in the rising edge of the upper pass band, 1 <(I)
6.1. Circle Conditions The main result here is the Circle Theorem. The Circle Theorem is a sufficient condition for closed-loop stability, which requires the nonlinearity N to lie inside a sector, but which leaves N free otherwise. The other conditions are all corollaries of the Circle Theorem. Corollary 1 is probably the most useful result, since it roughs out the region of stability, with a minimum of restrictions on N. However, it is often conservative. Corollaries 2 and 3 provide a tradeoff between limitations on N and limitations on HU",,). Probably more significant than the actual conditions is the fact that there is a method of generating them, namely, the factorization method. The results derived in Part II hold for nonzero initial ccndltions in the linear element, provided the "zeroinput response" W2 satisfies the indicated restrictions. 6.2. Extensions of
th~
Theory
The theory has been extended in several directions (see [tb 1), notably, 1) to L., 2) to systems with a limited rate of time variation.
The extension to L. involves the use of exponential weighting factors, which transform L .. functions into k functions. The extension to time-varying systems involves the use of a shifted Nyquist diagram, H(tT+j6J), in which IT depends on the rate of time variation. 6.3. Gain and Phase Shift in Relation to Nonlinear TimeVarying Systems
of frequency response, have no strict meaning in nonlinear or time-varying systems. However, stability does seem to depend on certain measures of signal amplification and signal shift. Thus the norm ratio plays a role similar to the role of gain. Furthermore, the inner product (x, Hi), a measure of input-output crosscorrelation, is closely related to the notion of phase shift. For example, for linear time-invariant operators in £ the condition of positivity, (x, Hx)~O, is equivalent (by Lemma 1) to the phase condition,
llHxJ1/flxll
IArg{H(jw)}1 ~90°. I t may be worthwhile to see what the theorems of Part I mean in terms of gain and phase shift. This can be done with the help of Lemma 1. Theorem 1 of Part I can be viewed as a generalization to nonlinear timevarying systems of the rule that, "if the open-loop gain is less than one, then the closed loop is stable. " Theorem 3 can be viewed as the generalization of, "if the openloop absolute phase shift is Jess than 1800 then!' the closed loop is stable." Theorem. 2 places gain and phase shift in competition, permitting large gains at small phase shifts, etc. 6.4. Conclusions
Some of the salient features of the functional theory are: (i) It provides an alternative to the method of Liapunov, an alternative resembling the classical NyquistBode theory. (ii) I t is well sui ted to inpu t-outpu t problems. (iii) I t is free of state-space restrictions, and is therefore useful for distributed systems, hysteritic systems, etc. It also lends itself well to multivariable systems. (iv) I t unifies several results in stability theory. In particular, it is noteworthy that Popov's condition, the slope-restricted-N result, etc., can all be derived from the Circle Theorem. (v) It-has led to some new results, notably Corollary 3 and lib]. The theory outlined here is probably still far from its definitive form. Nevertheless, it provides enough insight to make possible a reasonably systematic design of stabilizers. ApPENDIX
A
DEFINITIONS OF CONICITY AND POSITIVITY
It wilt be assumed that H is a relation in CRo and that a5:~ are real constants.
c, ,~O. and GROUP
I. "Incremental" Conditions
H is incremen tally interior conic if
The stability of a linear time-invariant feedback system depends on the amounts of gain and phase shift introduced by the open loop. Are similar considerations involved in nonlinear, time-varying problems? Of course the classical defini tiODS of gain and phase shift, in terms 303
If (Hx
- By) t
-
c(x - y) ,lf ~
,H (x -
y) ,I[ ;
11 There are two positive elements in the 0cr.en loop; each contributes an absolute phase shift of less than 90 ; the open-loop absolute phase shift is therefore less than 180°.
{a, fJ}
H is incrementally inside the sector
if
The limits in the mean in (Bl)-(B2) exist, and x)(r) ==X(T).12
«(Hx - By), - a(x - y)" (Hx - By), - (j(x - y),} ~ 0;
Btl, Properties of Transforms
H is incrementally positive if
In the following properties, H is an operator in
«x - ,)" (Hx - Hy),) ~ 0; (where the inequalities of Group I hold for all x and y
(A) The integral defining H(s) [see (6) 1 converges and is bounded for <1">0'0' since
in Do(H) and all t2:0).
The definitions of an operator that is "exterior" conic or "outside" a sector, are identical to the preceding ones except for a reversal of the inequality sign. and will therefore be omitted. REMAlllt: If H is incrementally inside {a, 13} then H is inside {a, f1}. Similarly, each inequality in Group I implies a corresponding inequality in Section 3. GROUP
~ ..h(t) exp (-"I)(lt ~ (B) (C) (0) (E)
II. It Instantaneous" Conditions
H is instantaneously inside the sector
a ~ Hx{t}/x(t)
S fJ
(x
E
{a,
~l if
Lt,j t ~ OJ x(t) ;;C 0);
H is instantaneously positive if x(t) ·Hx(t)
~
0
H is instantaneously incrementally inside the sector if Hx(t) - Hy(t)
{a, s]
a<
{
00
1h(t)
I exp ( -fTot)dt == const,
H(s) is analytic for cr>cro. 13 For CT~ 0, lim,., ... H(s) = hfIQ.14 If x is in L 2 then H» is in ~.1fl If the Li.m, transforms of xEL 2 and HxEL 2 are X(s) and Y(s) respectively, then 16 Y(s)
III. "Instantaneous Incremental" Conditions
GROUP
~
having a Laplace transform H(s).
= H(s)X(s)
(B3)
One consequence of Property 0 is that every mapping of the type defined by the right-hand side of (2) is an operator" on L"" and belongs to
<~
A contour in the complex plane will be said to have Property N if it does not pass through or encircle the (x E L 2 . ; t ~ 0; x(t) - yet) ~ 0). origin. LEMMA 4. If the Nyquist diagram of H(s) has PropH is instantaneously incrementally positive if erty N then: (a) 1/H(s) is analytic for 0" 2:. o. (b) If the [Hx(e) - .,(1)]· [x(t) - y(t)] ~ 0 (x E £2.; t ~ 0). inequality 1/IH(C1+jw)1~M holds for (1=0 then it holds for all cr~O. REMARK: rf H is instantaneously inside {a, s}. then PROOF: (a) Since H(s) is analytic for (f ~ 0, it is H is inside {a, p}. Similarly. each inequality in Group enough to show that H(s) ~O for a~O to prove (a). For II implies a corresponding inequality in Section 3. Also this purpose several contours are defined ~ Let rill denote each inequality in Group III implies a corresponding inthe jCol)-axis (as shown in Fig. 11); for R ~ 0 a constan t, equality in Group II. let r R denote the clockwise contour bounding the semicircular region 10"+jw15R where (T~O; let rw-rR deApPENDIX B note the difference contour; and letH(r w) and H(r OI-r R) LEMMA 1 denote images of the respective con tours, each augThe proof of Lemma l(b) will be based on the Prin- mented with the point hac. ciple of the Argument, a theorem of Paley and Wiener, It will be shown that H(r s ) has Property :N' for' and Parseval's theorem. The proofs of Lemmas lea) R~Ro; since by hypothesis H(s) is analytic for O"~O. and l(c), being straightforward applications of Parse- and has no zeros on the j",..axis, Lemma 4(a) follows by val's theorem, will be omitted. the Principle of the Argument. Some preliminary lemmas and properties will now be introduced. DEFINITION: If x is a function in L 2 then its L.I.M. 1'See Widder (8), ch, II, Theorem 10. -
TRANSFORM
."J{{s)
=
x{t) - ,(t)
-
is
l.i.m. T--
la It
f
theorem {Titchmarsh (9a], Theorem 1). The general case follows
l' x(e)
exp (-st)dt.
«(1
~
0)
(BI)
0
The l.i.m. transform of X(s) is
Xl(T)
=~
2r
Li.m, W-..
Ibid.• Theorem Sa. The special case j "jw is implied by the Rieman-Lebesgue
f-wWX(jw) exp {jWT)dw.
(T real)
(B2)
from the special case and from Properties A and B by a theorem of Phragmen.Linde16f (Titchmarsh [9b), sec. 5.64). 11 This follows from Theorem 6S of Titchmarsh [9a], which implies
that the convolution of an L, function with an L 1 function is in L!, and has a transform of the type (83). 11 Suppose that x is in L,: x, is certainly in L 2, and H(x,) is in L s by Property D; since [H(x,»),-{Hx]" it follows that [Hxlcis in L 2 ; i.e., Hx is in L". Thus H maps L k into L~; since H also maps 0 into o, it follows that H is an operator on L2. and in
w
that is,
w
ex),H
II (Hx -
~
,1lx,ll·
(B4)
For this purpose, let y~(Hx-cx) and - [Hx],. Hence _ _ . - - . . . ..... 6
- - t - -........... 6
y, =
= H(x,,)
(Hx - ex),
- ex, -
lJ~H(x,)
o.
(B5)
Now x, is in L 2• Hence H{x,) is in L 2 by Property D of Appendix Bol. Since [H(x,) ],= (Hx) " it follows that (Hx), is in L 2 • Thus ~ is in L 2 , all terms in (BS) are in L t , and by Property E, Fig. 11. Contours.
y,(s) = H(s)Xc(s) - ,X,(s) - .6(s)
To prove that H(r R ) has Property N observe that H(rB)=H(r.,)+H(rR-r,.,); since H(r(#l) has Property N by hypothesis, it is enough now to show that H(rR-r.,) has Property N. This can be accomplished by showing that H(rR-f.,) lies in a circle centered at h. and not including the origin. The last assertion is a consequence of two facts. (i) There is an Ro>O for which, for R~Ro and s in (rR-r.,), jH(s)-hooJ ~ifhGOI· (ii) h.~O.
where Yt(s), Xt(s). and and ~. Hence
y,. x,
lfXt - q1l2 = -1
= -1
and since H(jCJJ) ;4 0 by Property N. (b) This is a special case of the Maximum Modulus Theorem of Phragrnen-Lindelof.P The theorem implies that a function analytic in a half-plane, and bounded on the boundary, is bounded throughout the half-plane.
(PW2)
f 00/ W(er + jw)
12dw
> 0,
s
and
const.
(er
>
The following lemma is a modification of Theorem 5 of Paley-Wiener [tal, and is stated without proof. LEMMA 5. (a) If w is in L z, w,=O, and W(s) is the l.i.m, transform of w(r) , then W(s) exp (st) satisfies the Paley-Wiener conditions. (b) Conversely, if t~O and W(s) exp (st) satisfies the Paley-Wiener conditions, then there is a function w in L" having the properties that w,=O and that W(s) is the l.i.m. transform of WO
B.4. Proof of Lemma. l(b) Let x in L 2t and t2:0 be given. Since H is linear, it is enough to show that H is conic with parameters c and r;
dw
C
2
Y,(jw) 1 d", (by (B7)] H(jw) - c
GO
-00
2
-00
[by inequality (8) of Section 3.2]
1 -IIY,1I ,
2
(Parseval's Theorem)
(B8)
It will be shown, in Assertion 1, that q(T) =0 for almost all T'
/lxc - q1l2
=
Jfx,fl2 + flqlf2 2: flXtJl2.
(B9)
(B8) and (B9) imply (B4). ASSERTION 1. The expression A(s)/[H(s)-c] is the l.i.rn, transform of a function q in L 2 : furthermore q(T) =0 for almost all T
It
2
1
(Parseval's Theorem)
f I
21("
0).
-110
A (iw) x,(jw) - . H(Jw) -
s ~f '" I y,(jw) 1 dw =
(f
I
-00
2"
B.3. A Pale,- Wieft~r Lemma
(PW1) W(s) is analytic for
f:lO
21r
lim H(jCAJ) ,
A complex-valued function W(s) will be said to satisfy the Paley-Wiener conditions if
(Bi)
cf
Now the braced terms in (B7) are l.i.m. transforms of functions in L s; for X,(s) this is true by definition; for the remaining terms, this can be proved by the reasoning given below in Assertion 1. Suppose that d(s)jH(s)-c is the l.i.m, transform of a function q(r); it follows that
I
=
are the I.i.m. transforms of
l = {YI(S)} . H(s) - c
.1(s) { H(s) -
{ .¥t(s) } -
(I) is obtained from Property C of Appendix B.l for sl ~Ro, and therefore certainly holds for s in (r R - r w). (ii) holds since '
h.
~(s)
(B6)
To prove (PW2). observe that
Titchmarsh [9bl. sec. 5.61.
305
f"_.. I .:\(s) exp oo I dw
2
J _
-IA(S) exp (st) 1 1 H(s) _ c dw ~ -;;
«(1 > 0)
~ canst.
(C6) is equivalent to the inequality
(by inequality (8)]
~ fJ ~
kfJ(Xh (Ny),} LEMMA
C Observing that, for
2
A preliminary assertion will be proved first. 2. If N in m. is positive, K is in y=Kx. then ASSERTION
j
t
o
dY -·N(y (.,.»d-r
aT
2.
PROOF OF ASSERTION
~
r:
Y d = f ' -·N(y(T»)dT dr
1co)
o
.c,
and
o.
(Cl)
N(y')dy'.
(C2)
C.I. Proof of Lemma 2 Part a) It is required to show that, for any given x in t~O,
the inequality
~' %(1'). [NKx(T) ]ar ~ 0
1 dy -
k"A dT
1
+-
k
kfJ(Xh (Ny),}
o
kX
dy
y(r).
k
Y(T)} . N(y(T)d1'
(C4)·
I
~0
LEMMA
D
3
Before proving Lemma 3, a few related assertions will be introduced. ASSERTION 3. Let K be an operator in me, x a fixed element of L 2. , and y~Kx. Then x has a "Foster expan.. sion" in )';18 that is, x can be expressed as a finite sum,
in which F, are operators mapping the image under K of Lt. into R[O, 00), as follows: CASE 1. F oy = K - l (O) · y . CASE 2. If i=1, 2, · .. , (m-1). then F, is in £ and has a Laplace Transform,
= his/(s + 8,),
CASE
(hi> 0,
8i
> 0).
3. F".y==h".y if K(oo)=O and F",y=O otherwise,
hm'~O. ASSER.TION
where
or (C5)
0
Now k and ~ are positive by hypothesis; the first integral in (C5) is non-negative by Assertion 2; the second integral is non-negative, since N is a positive operator; therefore (C5) is true. Q.E.D. Part b) It will be assumed, for simplicity, that (3)0. CASE A. Suppose a~O. It must be shown that NK is inside {a, p} X {O. k}. This is equivalent to saying that NK is inside {O~ kfJ} • or that «NK%),), (NKx - kfJ~)c) S O.
[N(Y(1')]2dr,
I
ApPENDIX
F,(s)
-2 f dy.ar N(y(r))dT + ~k f 'Y(1') · N(Y(1'»d1' ~ O. k"A 0
~ ~
which implies (C7). Q.E.D. CASE B. Suppose a <0. Decompose N(x) into two parts, N(x)=N+(x)+N_(x); let N+(x):=:N(x) for N(x) ~O and N+(x) =0 elsewhere, and let N_(x) be similarly defined. Since N+ is clearly inside {O, (j} t Case A implies that N+K is inside {O. k~}. Similarly N_K is inside {ka, o]. On summing the sectors of N+K and N_K (by the Sum Rule of Part I) it is found that NK is inside {kat k{i} ; that is, inside {a,~} X {O, k}. Q.E.D.
(C3) is therefore equivalent to
f ' {~ aT +.2..
N satisfies the inequality
(C3)
holds. For this purpose. make the following substitutions: Write ,~1U, and observe that, since K(s) ==k~/(s+~), y is differentiable and
x(r) == -
a~O,
Y(1')' N(Y(T»d1'.
I
~y(.,). N(y(r») ~ [N(y(1"») ]2, we get
Since K is in .c, Yet) is given by a convolution integral, whose kernel is fixed for a fixed x, and whose limits of integration are 0 and t; therefore yeO) = O. Furthermore, since N is positive, its graph lies in the first and third quadrants. It follows that the right-hand side of (C2) is non-negative. Q.E.D.
L 2• and any given
(C7)
where 'Y = Kx. Now recalling that (x" (Ny) c) equals the left-hand side of (CS), we get
(Since aEL s and a,lSO by construction, the last inequality is implied by Lemma S(a).) Q.E.D. ApPENDIX
2: It (Ny) ell 2
kf1(xc, (Ny) c}
4.. If N is incrementally positive, and (x" [Ny]&)~O then (Xc, [N(x+y)l,)~o. PROOF OF ASSERTION 4. I t is enough to show tha t (Xl,
[N(x
+ Y)r]) -
(Xl, [Ny],) ~ 0
(Dt)
But the left..hand side of (01) can be expressed as
«x
+ y), -
Y" [N(x
+ y)], -
(NY)t)
which has the form (XU-X2h Nxu-Nx2'), and is nonnegative, since N is an incrementally positive operator. Therefore (D1) holds.. Q.E.D.
(C6)
II
306
See Guillemin (2], p, 1t 5.
ASSERTION
5. If N is an operator in
m, incrementally
inside a sector {a. fJ} where a
{at O}, and N+ is inside {O, ~} . PROOF OF ASSERTION 5. Since N is incrementally inside a sector, its graph N is continuous and has bounded variation on every finite interval. Consequently N can be expressed as an integral. N(x) == !o*n(x')dx'. Let n+(x) =n(x) if n(x) ~O, and n+(x) =0 if n(x) <0; let N+(x) == !osn+(x')dx'. Clearly N+ has the desired property. N_ is constructed similarly. Q.E.D.
l-
Hence NK is inside {O, fjK (0) which equals {O, ~} X K(O)}. CASE B. If a>O then NK is decomposed into three parts
to,
NK= {[N-aI]K}
+ {alK-K(ce).]]}
+ {aK(oo).]J#
(D5)
Now the three parts lie in the sectors {O, (~~a)K(O)},
{OJ a[K(O)-K(oo)]}, and {aK(oo), aK(ao)}, respectively. (The first two of these sectors are determined by the rule formed in Case At after observing that [N -all is inside ~-a}, and that [K-K(oo)·l] is inside {O, [K(O) -K( co)]} ; the third sector is simply the sector of a constant times the identity.) On summing the three sectors (by the Sum Rule of Part I), it is found that NK is inside {aXe co), ~K(O)}: that is, inside ~} X {K( co), K(O)}. CASE C. If a
'to,
D.I. Proof of Lemma 3 Part a) Let x be any given element of L 2. , and t any given point in [0, 00); it is required to show that
(Xt, (NKx) ,)
~
o.
(D2)
Letting y ~ K», and recalling that x can be expressed by the Foster expansion
ta,
ACKNOWLEDGMENT
(see Assertion 3), (02) is equivalent to
.. L «Fiy)" (Ny),) ~ O.
(D3)
i-O
It will be shown that each component on the left-hand side of (03) is non-negative. CASE 1. Here F oy = K - l (O) .,. Hence «FoY)" (Ny),) =- [K-I(O)]· (Y" (Ny),); this is non-negative since N is a positive operator, and since K(O) is necessarily positive. CASE 2. Here F,(s) ==his/(s+e,). Let J(t)
= hrl ~
I
F.-y(T)d.,..
It follows that y;;;rj+8•.1 almost everywhere, and that F"y = hli almost everywhere. Hence
Now, observing that 8t>O, Assertion 2 implies that (i, IN(9i s) ],) is non-negative. Observing that hi is positive, the right-hand side of (D4) is non-negative by Assertion 4. Thus Case 2 is proved. CASE 3. Here FMy=n.j if K(co) =0. Hence «F",y)" (Ny),)==h.(;" (Ny),). Case 3 follows by Assertion 2. Since the inner product is non ..n egative in all three cases, (D3) holds. Q.E.D. Part b) Assume, for simplicity, that P>O. CASE A. If a=O then, by reasoning similar to that used in Lemma 2(b), (C5)-(C7), the foJlowing inequality is obtained:
The author thanks Dr. P. Falb lor correcting the manuscript, and for offering many valuable suggestions. He also thanks Dr. G. Kovatch and NASA's Electronic Research Center, Cambridge, Mass., for supporting the completion of the paper, and Mrs. Iris McDonald for typing it. REFERENCES
[tal G.. lames, ·On the stabili9' of nonlinear, time.varying feedback systems," Proe. 1964 NBC, vol. 20. pp. 725-730. [lb) - , ItNonlinear, time-varying (ee(fback systems-Conditions for L.-boundedneas derived using conic operators on exponentially weighted spaces," Proc. 1965 AUertDn Con!., pp.
460-471. [tel - . "On the input-output stability of time-varying nonlinear feedback systems-Part I. Conditions derived using concepts of loop pin, conicity. and positivity,· IEEE Tr6fts. Oft AfdomtUk CtmttDl, vel, AC-l1, pp, 223-239. April 1966. (2) E. A. Gulllemin, S"uhuis 0/ Passiv, NelVJor1es. New York: Wiley. 1957.
[3] H. Hatanaka, "The frequency responses and jump-resonance phenomena of nonlinear feedback control systems,» Trans. ASME, pp. 236-242, June 1963. [4a] R. W. Brockett and]. W. Willems. "Fr~uency domain stability criteria. Parts I and II, JJ Proc. JA ee, pp. 735-747, J965. {4b) R. W. Brockett and L. J. Forys, "On the stability of systems containing a time-varying gain," Proc. 1964 Allerton Con/., pp. 413-430.
(5] I. W. Sandberg, "A frequency domain condition (or the stability of ~ystems containing a single time-varying nonlinear element, " BeU Sys. Teel. J., vol. 43, p. 1601, 1964. [6] K. s. Narendra, and R. M. Goldwyn, MA ieometrical criterion for the stability or certain nonlinear, nonautonomous systems, n IEEE Tran«: Oft Ci,cuit TMO'Y, vol. CT-l1, pp. 406-408, September 1964. {7J J.. Kudrewicz, "Stability of nonlinear feedback systems," AfItomIIIiitJ i Tekmec1l4nilttJ, vel. 25, no. 8, 1964. [8] D. V. Widder, The LQ,pZa,e rrGns!orm. Princeton, N. J.: Prince. ton University Press. 1946. [9al E. C. Titchmarsh. Int,oductioftlo elu TMor, of Fourier I"teg'41s, 2nd ed. Oxford, England: University Press, 1962.. (9b] - . The Theory DI Functitms. 2nd ed. Oxford, England: University Press, 19M.. [10J R. Paley and N. Wiener. Fourier Tra"sforms in In. Complex DomtJ,in. New York: Am. Math. Soc. Colfoquium Publications, 19.14.
307
An Invariance Principle in the Theory of Stability JOSEPH P. LASALLE
T
HE stability theorems of Lyapunov have been among the oldest and strongest pillars of control theory. The centenary of Lyapunov's celebrated 1892 memoir was recently marked with its English translation [17], while the 1907 French translation was reprinted in 1949 by Princeton University Press [16]. Applications of Lyapunov stability concepts to control problems began to appear in Russia in the 1930s-1940s. Significant theoretical results from the 1950s were summarized in the books by Chetayev [3], Lurie [15], Malkin [18], Letov [14], Zubov [22], and Krasovskii [7]. In the post-1957 Sputnik era, English translations of these books, and of other Russian works, further increased the interest already stimulated by "Contributions to the Theory of Nonlinear Oscillations," a five volume series (1950, 1952, 1956, 1958, and 1960) edited by Lefschetz, and published by Princeton University Press. The fourth volume contained a detailed and rigorous survey of Lyapunov stability theory by Antosiewicz [1]. The 1960 survey by Kalman and Bertram [6] was more accessible and had a stronger impact on engineering audiences, as did the books by Lefschetz [12], Lefschetz and LaSalle [13], and a paper by LaSalle [10]. This activity continued in the early 1960s when several volumes of "Contributions to Differential Equations" were published, including the important results by Yoshizawa [20], whose 1966 book [21] presented a collection of advanced stability results. The status of stability theory in the pre-1965 period is summarized in the scholarly work by Hahn [4], which is the most comprehensive source covering that period. Among the innovations from that period are the results which settle the issue of existence of Lyapunov functions. Particularly important among the results of this type are the "converse theorems" of Massera [19], Krasovskii [14] and Kurzweil [8], which found applications in diverse areas of control theory. Radially unbounded Lyapunov functions were introduced in 1952 by Barbashin and Krasovskii [2] to guarantee stability in the large, that is, global stability. The same Barbashin-Krasovskii paper and the book by Krasovskii [14] initiated another line of research which culminated in this paper by LaSalle, and its extended 1968 version [11].
This line of research was aimed at extracting from Lyapunov 's stability theorem more information about the asymptotic behavior of the solutions x(t) E IRn of
x=
f(x, t),
f(O, t) = 0 for all t
(1)
With a positive definite function Vex, t), a theorem of Lyapunov establishes asymptotic stability of the equilibrium x = 0, if V(x, t) is negative definite along the solutions x(t) of (1). If V is only nonpositive, the same theorem guarantees stability, but does not reveal whether x(t) converges to zero or to a set in IRn . For autonomous systems, x = f(x), a theorem of Barbashin and Krasovskii [2] (see also Theorem 14.1 onp. 67 of[14]) examines the set E in which V = O. If this set does not contain any positive semi-trajectory (solution x(t) for all t ::: 0) other than x(t) == 0, then x (t) ---+ 0 as t ---+ 00. This result gave a practical test for asymptotic stability when V is only nonpositive, instead of negative definite. In his 1960 paper [10], LaSalle extracted further information on the asymptotic behavior of x(t). Using the limit sets and the largest invariant set M of x = f (x) contained in the set E where V is zero, he showed that x(t) converges to the set M. This set need not be an equilibrium, but can be a compact limit set like a limit cycle. This result of LaSalle, which he later termed I nvariance Principle, had a significant connection with the then new concept of observability. For the system x = Ax, a positive definite Lyapunov function V = x' P x has the derivative V = A' P + PA = -x'Qx. Suppose that Q is positive semi-definite, so that it can be expressed as Q = C'C for some matrix C. It then follows that V = - y' y, where y = Cx can be treated as an output of x = Ax. Clearly, the set E where V = 0 is y(t) == O. If the pair (A, C) is completely observable, then y(t) == 0 implies x(t) = O. Hence, no nonzero positive semi-trajectory x(t) is contained in E, which proves, via Barbashin-Krasovskii [2], that x = Ax is asymptotically stable. However, if the system is only stable with a pair of purely imaginary eigenvalues unobservable from y = Cx, then LaSalle's Principle shows that x(t) converges to the periodic solution in the unobservable subspace. The Invariance Principle was subsequently extended to periodic and almost periodic systems, but it does not hold for more
309
general nonautonomous systems (1). To obtain similar information on the asymptotic behavior of x(t), Yoshizawa [20] derived a set of conditions under which V(x, t) .:s W (x), where W (x) is "positive definite with respect to a set M," implies that x(t) converges to M. The main theorem in this paper by LaSalle modifies and improves this result of Yoshizawa, and provides the strongest convergence result for nonautonomous systems. To go beyond this result, further restrictions on f (x, t) are needed. One of them is the so-called "persistency of excitation" condition in adaptive identification and control. Typically, an adaptive algorithm guarantees that V .:s -e 2 , where e is the scalar tracking error. The Yoshizawa-LaSalle theorem provides the conditions under which e(t) converges to zero. It has thus become an indispensable tool in adaptive control design. It has also become instrumental in deducing stability from passivity properties as in feedback passivation and backstepping designs of nonlinear systems. Other settings where LaSalle's Invariance Principle has been studied are infinite dimensional systems, as in the work of Hale [5], and stochastic systems governed by continuous-time Markov processes, as discussed by Kushner in [9]. REFERENCES [1] H. ANTOSIEWICZ, "A survey of Lyapunov's second method," in Contr. to Nonlinear Oscillations, S. Lefschetz, Ed., 4:141-166 (Princeton Univ. Press, Princeton NJ), 1958. [2] E.A. BARBASHIN AND N.N. KRASOVSKII, "On the stability of motion in the large," Dokl. Akad. Nauk USSR, 86:453-456,1952. [3] N.G. CHETAYEV, The Stability ofMotion, Pergamon Press (Oxford), 1961. (Russian original, 1946.) [4] W. HAHN, Stability ofMotion (Springer-Verlag, New York), 1967. [5] J.K. HALE, "Dynamical systems and stability," J. Math. Anal. Appl., 26:3959,1969. [6] R.E. KALMAN AND J.E. BERTRAM, "Control system analysis and design
via the 'second method' of Lyapunov, I Continuous-time systems," 1. Basic Engineering (Trans. ASME), 82D:371-393, 1960. [7] N.N. KRASOVSKII, Stability ofMotion, Stanford Univ. Press (Stanford, CA), 1963. (Russian original, 1959.) [8] J. KURZWEIL, "The converse second Liapunov's theorem concerning the stability of motion," Czechoslovak Math. 1.,6(81):217-259 & 455-473, 1956. [9] H,J. KUSHNER, "Stochastic stability," in Stability of Stochastic Dynamical Systems, R. Curtain, Ed., Lect. Notes in Math., Springer-Verlag (New York), 294:97-124,1972. [10] J.P. LASALLE, "Some extensions of Liapunov 's second method," IRE Trans. Circuit Theory, CT·7:52G-527, 1960. [11] J.P. LASALLE, "Stability theory for ordinary differential equations," 1. Differential Equations, 4:57-65,1968. [12] S. LEFSCHETZ, Differential Equations: Geometric Theory, Interscience, Wiley (New York), 1957. [13] S. LEFSCHETZ AND J.P. LASALLE, Stability by Liapunov's Direct Method, with Applications, Academic Press (New York), 1961. [14] A.M. LETOV, Stability in Nonlinear Control Systems (English translation), Princeton Univ. Press (Princeton, NJ), 1961. (Russian original, 1955.) [15] A.E. LURIE, Some Non-linear Problems in the Theory ofAutomatic Control (English translation), H.M.S.O., London, 1957. (Russian original, 1951.) [16] A.M. LYAPUNOV, "Probleme general de la stabilite du mouvement" (in French), Ann. Fac.Sci. Toulouse, 9:203-474, 1907. Reprinted inAnn. Math. Study, No. 17,1949, Princeton Univ. Press (Princeton, NJ). [17] A.M. LYAPUNOV, "The general problem of the stability of motion" (translated into English by A.T. Fuller), Int. J. Control, 55:531-773,1992. [18] I.G. MALKIN, Theory ofStability ofMotion, AEC (Atomic Energy Commission) Translation 3352, Dept. of Commerce, United States, 1958. (Russian original, 1952.) [19] J.L. MASSERA, "Contributions to stability theory," Ann. Math., 64:182-206, 1956. [20] T. YOSHIZAWA, "Asymptotic behavior of solutions of a system of differential equations," in Contributions to Differential Equations, 1:371-387, 1963. [21] T. YOSHIZAWA, Stability Theory by Liapunov's Second Method, The Mathematical Society of Japan, Publication No.9 (Tokyo), 1966. [22] V.1. ZUBOV, Mathematical Methods for the Study of Automatic Control Systems, Pergamon Press (Oxford), 1962. (Russian original, 1957.)
P.V.K. & J.B.
310
An Invariance Principle in the Theory of Stability JOSEPH P. LASALLEl Center for Dynamical Systems Brown University, Providence, Rhode Island
1. Introduction The purpose of this paper is to give a unified presentation of Liapunov's theory of stability that includes the classical Liapunov theorems on stability and instability as well as their more recent extensions. The idea being exploited here had its beginnings some time ago. It was, however, the use made of this idea by Yoshizawa in [7J in his study of nonautonomous differential equations and by Hale in [1] in his study of autonomous functional-differential equations that caused the author to return to this subject and to adopt the general approach and point of view of this paper. This produces some new results for dynamical systems defined by ordinary differential eq uations which demonstrate the essential nature of a Liapunov function and which may be useful in applications. Of greater importance, however, is the possibility, as already indicated by Hale's results for functional-differential equations, that these ideas can be extended to more general classes of dynamical systems. It is hoped, for instance, that it may be possible to do this for some special types of dynamical systems defined by partial differential equations.
In Section 2 we present some basic results for ordinary differential equations. Theorem I is a fundamental stability theorem for nonautonomous systems and is a modified version of Yoshizawa's Theorem 6 in [7]. A simple example shows that the conclusion of this theorem is the best possible. However, whenever the limit sets of solutions are known to have an invar.. iance property, then sharper results can be obtained. This "invariance principle" explains the title of this paper. It had its origin for autonomous and periodic systems in [2] and [4],. although we present here improved versions of those results. Miller in [5] has established an invariance property
1 This research was supported in part by the National Aeronautics and Space Adrnini . . stration under Grant No. NGR-40-002-015 and under Contract No. NAS8-11264, in part by the United States Air Force through the Air Force Office of Scientific Research under Grant No. AF-AFOSR-693-65, and in part by the United States Army Research Office, Durham, under Contract No. DA-31-124-ARO-D-270.
Reprinted with permission from Differential Equations and Dynamical Systems (New York: Academic Press, 1967),1. Hale and 1. P. LaSalle, eds., Joseph ~ LaSalle, "An Invariance Principle in the Theory of Stability," pp. 277-286.
311
for almost periodic systems and obtains thereby a similar stability theorem for almost periodic systems. Since little attention has been paid to theorems which make possible estimates of regions of attraction (regions of asymptotic stability) for nonautonomous systems results of this type are included. Section 3 is devoted to a brief discussion of some of Hale's recent results [I] for autonomous functional-differential equations.
2. Ordinary Differential Equations Consider the system i =f(t, x)
(1)
where x is an n-vector,fis a continuous function on Rn,+l to R" and satisfies anyone of the conditions guaranteeing uniqueness of solutions. For each x in Rn, we define I x I == (Xl! + ... + X n2)1I2, and for E a closed set in Rn we define d(x, E) =: Min { I x - y I; y in E}. Since we do not wish to confine ourselves to bounded solutions, we introduce the point at CX) and define d(x, 00) == I X 1-1. Thus, when we write E* == E U {oo}, we shall mean d(x, E*) = Min {d(x, E), d(x, oo)}. If x(t) is a solution of (1), we say that x(t) approaches E as t ~ 00, if d(x(t), E) --.0 as t 00. If we can find such a set E, we have obtained information about the asymptotic behavior of x(t) as t -+ CXJ. The best that we could hope to do is to find the smallest closed set !J that x(t) approaches as t ~ 00. This set Q is called the positive limit set of x(t) and the points p in Q are called the positive limit points of x(t). In exactly the same way, one defines x(t) -+ E as t --+- - 00, negative limit sets, and negative limit points. This is exactly Birkhoff's concept of limit sets. A point p is a positive limit point of x(t), if and only if there is a sequence of times t n approaching CX) as n ~ 00 and such that x(t n ) -+ p as n ----. CXJ. In the above, it may be that the maximal interval of definition of x(t) is [0, r ), This causes no difficulty since, in the results to be presented here, we need only, with respect to time t, replace 00 by T. We usually ignore this possibility and speak as though our solutions are defined on [0, (X)) or -)00
( - 00,
(0).
Let V(t, x) be a Cl function on [0, oo] x R" to R, and let G be any set in R". We shall say that V is a Liapunov function on G for Eq. (1), if V(t, x) ~ 0 and V(t, x) ~ -
W(x) ~ 0 for all t
>
0 and all x in G, where W is
continuous on Rn to R, and
. av "r i: + ~ n
U
i-I
312
av ;lfi· ox;
(2)
We define (G is the closure of G)
E
=
{x; W(x)
== 0, x in G}.
The following result is then a modified but closely related version of Yoshizawa's Theorem 6 in [7].
Theorem 1. If V is a Liapunov function on G for Eq. (I), then each solution x(t) of (1) that remains in G for all t > to ~ 0 approaches E* = E u {oo} as t ~ oo, provided one of the following conditions is satisfied: (i) For each pinG there is a neighborhood N of p such that If(t, x) is bounded for all t > 0 and all x in N.
I
(ii) W is Cl and W is bounded from above or below along each solution which remains in G for all t > to ~ o.
If E is bounded, then each solution of (1) that remains in G for t > to ~ 0 either approaches E or 00 as t -+ 00. Thus this theorem explains precisely the nature of the information given by a Liapunov function. A Liapunov function relative to a set G defines a set E, which under the conditions of the theorem contains (locates) all the positive limit sets of solutions which for positive time remain in G. The problem in applying the result is to find "good" Liapunov functions. For instance, the zero function V = 0 is a Liapunov function for the whole space Rn and condition (ii) is satisfied but gives no information since E = R". It is trivial but useful for applications to note that if VI and V2 are Liapunov functions on G, then V = VI + V2 is also a Liapunov function and E = £1 () £2 . If E is smaller than either £1 or £2' then V is a "better" Liapunov function than either E. or £2 and is always at least as "good" as either of the two. Condition (i) of Theorem 1 is essentially the one used by Yoshizawa. We now look at a simple example, where condition (ii) is satisfied and condition (i) is not. The example also shows that the conclusion of the theorem is the best possible. Consider x + p(t)i' + x =--..:: 0, where p(t) ~ (j > o. Define 2V = x 2 + y2, where Y == ..i'. Then V === - p(t)y2 ~ - by 2 and V is a Liapunov function on R2. Now W === l5y2 and W == 2t5YJi === - 2c5(xy + p(t)y 2 ) ~ - 2f5xy. Since all solutions are evidently bounded for all t > 0, condition (ii) is satisfied. Here E is the x-axis (y:=: 0) and for each solution x(t), yet) = X(/) --+- 0 as t ~ 00. Noting that the equation .t -~ (2 + exp[t)) + x == 0 has a solution x(t) = 1 .t- exp( - t), we see that this is the best possible result without further restrictions on p.
x
313
In order to use Theorem 1, there must be some means of determining which solutions remain in G. The following corollary, which is an obvious consequence of Theorem 1, gives one way of doing this and also provides, for nonautonomous systems, a method for estimating regions of attraction.
Corollary 1. Assume that there exist continuous functions u(x) and II(X) on Rn to R such that u(x) ~ V(t, x) ~ v(x) for all t ~ O. Define Q,/+ = {x; u(x) < 1J} and let G+ be a component of Q'1 +. Let G denote the component of Q'7 = {x; v(x) < 'YJ} containing G+. If V is a Liapunov function on G for (1) and the conditions of Theorem 1 are satisfied, then each solution of (1) starting in G+ at any time to ~ 0 remains in G for all t > to and approaches E* as t -+ 00. If G is bounded and EO = E () G C G+, then EO is an attractor and G+ is in its region of attraction. In general we know that if x(t) is a solution of (I)-in fact, if x(t) is any continuous function on R to Rtl-then its positive limit set is closed and connected. If x(t) is bounded, then its positive limit set is compact. There are, however, special classes of differential equations where the limit sets of solutions have an additional invariance property which makes possible a refinement of Theorem 1. The first of these are the autonomous systems
x =f(x).
(3)
The limit sets of solutions of (3) are invariant sets. If x(t) is defined on [0, 00) and if p is a positive limit point of x{t), then points on the solution through p on its maximal interval of definition are positive limit points of x(t). If x(t) is bounded for t > 0, then it is defined on [0, 00), its positive limit set Q is compact, nonempty and solutions through points p of Q are defined on (- 00, 00) (i.e., (J is invariant). If the maximal domain of definition of x(t) for t > 0 is finite, then x(t) has no finite positive limit points: That is, if the maximal interval of definition of x(t) for t > 0 is [0, fJ), then x(t) --+ 00 as t -+ fJ. As we have said before, we will always speak as though our solutions are defined on (- 00, 00) and it should be remembered that finite escape time is always a possibility unless there is, as for example in Corollary 2 below, some condition that rules it out. In Corollary 3 below, the solutions might welt go to infinity in finite time. The invariance property of the limit sets of solutions of autonomous systems (3) now enables us to refine Theorem I. Let V be a Cl function on R!" to R. If G is any arbitrary set in R", we say that V is a Liapunou function on G for Eq. (3) if V =: (grad V) • f does not change sign on G. Define E = {x; V(x) = 0, x in G}, where G is the closure of G. Let M be 314
the largest invariant set in E. M will be a closed set. The fundamental stability theorem for autonomous systems is then the following:
Theorem 2. If V is a Liapunov function on G for (3), then each solution x(t) of (3) that remains in G for all I > 0 (I < 0) approaches M* = M u {oo} as t --. 00 (I --+ - 00). If M is bounded, then either x(t) -+ M or x(t) ~ 00 as 1---. 00 (I --+ - 00). This one theorem contains all of the usual Liapunov like theorems on stability and instability of autonomous systems. Here, however, there are no conditions of definiteness for V or V, and it is often possible to obtain stability information about a system with these more general types of LiapUDOV functions. The first corollary below is a stability result which for applications has been quite useful, and the second illustrates how one obtains information on instability. Cetaev's instability theorem is similarly an immediate consequence of Theorem 2 (see Section 3). Corollary 2. Let G be a component of Q" = {x; V(x) < 1]}. Assume that G is bounded, V;;£ 0 on G, and MO = M n G c: G. Then MO is an attractor and G is in its region of attraction. If, in addition, V is constant on the boundary of MO, then MO is a stable attractor. Note that if MO consists of a single point p, then p is asymptotically stable and G provides an estimate of its region of asymptotic stability. Corollary 3. Assume that relative to (3) that V V > 0 on G and on the boundary of G that V = O. Then each solution of (3) starting in G a pproaches 00 as t -+ 00 (or possibly in finite time). There are also some special classes of nonautonomous systems where the limit sets of solutions have an invariance property. The simplest of these are periodic systems (see [2])
x = .((t,
x),
f(1
+ T, x) = .f(t)
for all t and .r.
(4)
Here, in order to avoid introducing the concept of a periodic approach of a solution of (4) to a set and the concept of a periodic limit point, let us confine ourselves to solutions x(t) of (4) which are bounded for t > O. Let!J be the positive limit set of such a solution x(t), and let p be a point in Q. Then there is a solution of (4) starting at p which remains in Q for all t in (- 00, 00); that is, if one starts at p at the proper time, the solution remains in Q for all time. This is the sense now in which Q is an invariant set. Let V(t, x) be Cl on R x R" and periodic in t of period T. For an arbitrary set G of RIJ. we say that V is a Liapunou function on G for the periodic system (4)
315
if V does not change sign for all t and all x in G. Define E· = {(t, x); V(t, x) = 0, x in G} and let M be the union of all solutions x(t) of (4) with the property that (t, x(t» is in E for all t. M could be called "the largest invariant set relative to E." One then obtain the following version of Theorem 2 for periodic systems: Theorem 3. If V is a Liapunov function on G for the periodic system (4), then each solution of (4) that is bounded and remains in G for all t > 0 (t < 0) approaches M as t --+ 00 (t -+ - 00). In [5] Miller showed that the limit sets of solutions of almost periodic systems have a similar invariance property and from this he obtains a result quite like Theorem 3 for almost periodic systems. This then yields, for periodic and almost periodic systems, a whole chain of theorems on stability and instability quite similar to that for autonomous systems. For example, one has
Corollary 4. Let QFI + = {x; V(t, x) < 'Y}, all t in [0, T]}, and let G+ be a component of Q,,+. Let G be the component of QFI = {x; V{t, x) < TJ for some r in [0, T]} containing G+. If G is bounded, V ~ 0 for all t and all x in G, and if MO = M n G c G+, then MO is an attractor and G+ is an its region of attraction. If V(t, x) = ,(t) for all t and all x on the boundary of MO, then MO is a stable attractor. OUf last example of an invariance principle for ordinary differential equations is that due to Yoshizawa in [7] for "asymptotically autonomous" systems. It is a consequence of Theorem 1 and results by Markus and Opial (see [7] for references) on the limit sets of such systems. A system of the form
x=
F(x)
+ g(t, x) + h(t,
x)
(5)
is said to be asymptotically autonomousif (i) g(t, x) -+ 0 as t --. 00 uniformly I h(t, 9'(t» I dt < 00 for all for x in an arbitrary compact set of Rn, (ii) q; bounded and continuous on [0, 00) to R",. The combined results of Markus and Opial then state that the positive limit sets of solutions of (5) are invariant sets of x = F(x). Using this, Yoshizawa then improved Theorem I for asymptotically autonomous systems. It turns out to be useful, as we shall jIlustrate in a moment on the simplest possible example, in studying systems (I) which are not necessarily asymptotically autonomous to state the theorem in the following manner:
f:
Theorem 4. If, in addition to the conditions of Theorem 1, it is known that a solution x(t) of (1) remains in G for t > 0 and is also a solution of an 316
asymptotically autonomous system (5), then x(t) approaches M* == M u {oo} as t ~ 00, where M is the largest invariant set of == F(x) in E. It can happen that the system (1) is itself asymptotically autonomous, in which case the above theorem can be applied. However, as the following example illustrates, the original system may not itself be asymptotically autonomous, but it still may be possible to construct for each solution of (I) an asymptotically autonomous system (5) which it also satisfies. Consider again the example
x
x==y y ==
-
x - p(t)y,
o < D ~ pet) ~ m for all t > o.
(6)
Now we have the additional assumption that pet) is bounded from above. Let (x(t), yet»~ be any solution of (6). As was argued previously below Theorem 1, all solutions are bounded and yet) ~ 0 as t ~ 00. Now (X(/), yet»~ satisfies x == Y, rV = - X - pet) yet), and this system is asymptotically autonomous to (*) .¥ ~ y, y :=: - x. With the same Liapunov function as before, E is the x . . axis and the largest invariant set of (*) in E is the origin. Thus for (6) the origin is asymptotically stable in the large.
3. Autonomous Functional-Differential Equations In this section we adopt completely the notations and assumptions introduced by Hale in his paper in these proceedings and present a few of the stability results that he has obtained for autonomous differential equations
(7) A more complete account with numerous examples is given in (1). For the extension to periodic and almost periodic functional-differential equations by Miller see [6]. We continue where Hale left off in Section 2 of his paper, except that we shall assume that the open set Q is the whole state space C of continuous functions. We also confine ourselves to solutions x of (7) that are bounded and hence defined on [- r, 00). Except that we are in the state space C, the definition of the positive limit set of a trajectory x f of (7) is essentially the same as for ordinary differential equations, and the notion of an invariant set is modified to take into account the fact that there is no longer uniqueness to the left. A set M c C is invariant in the sense that if qJ E M, 317
then x(tp) is defined on [-',00), there is an extension on (- 00, - r], and x,(fP) remains in M for all t in (- 00, 00). With these extensions of these geometric notions to the state space C, Hale then showed that the positive limit set of a trajectory of (7) bounded in the fu-ture is a nonempty, compact, connected, and invariant set in C. He was then able to obtain a theory of stability quite similar to that for autonomous ordinary differential equations. Let V be a continuous function on C to R and define relative to (7)
· 1 V(tp) == lim - [V(xr(fP» - V(
(8)
1:
With G an arbitrary set in C, we say that V is a Liapunovfunction on G for (7) if V(qJ) ~ 0 for all f/J in G. Define E = {
of (7) which remains in G and is bounded for t > 0, then x, -+ M as t --+ 00. Hale has also given the following more useful version of this result.
Corollary 5. Define Q" = {cp; V(
E C be an equilibrium point of (7) contained in the
closure of an open set U and let N be a neighborhood of p. Assume that: (i) V is a Liapunov function on G = U n N, (ii) M n G is either the empty set or p, (iii) V(qJ) < 11 on G when cp =1= p, and (iv) V(P) = 'YJ and V(<
its boundary or approach p. Conditions (i) and (iv) imply that it cannot reach or approach that part of the boundary of Go inside No nor can it approach p as 1-+ 00. Now (iii) states that there are no points of M on that part of the boundary of No inside G. Hence each such trajectory must leave No in finite time. Since p is either in the interior or on the boundary of G, each neighborhood of p contains such trajectories, and p is therefore un-
stable. In .[1] it was shown that the equilibrium point q; = 0 of
X(/) = ax 3( / )
+
bx3(1 - r)
was unstable if a > 0 and I b I < I a I. Using the same Liapunov function and Theorem 6 we can show a bit more. With
V(q» = - q>4(O) 4a V(x,) = --
+..!- JO
2-r
x4~) + ~
q;6(6) dO,
£-r x6(8)dO,
and
which is nonpositive when I b J < I a I (negative definite with respect to ,(0) and tp(- r»; that is, V is a Liapunov function on C and E = {
REFERENCES (I) Hale, I., Sufficient conditions for stability and instability of autonomous functional differential eauations, J. Diff. Eqs. 1, 452-482 (1965). [2] LaSalle, J., Some extensions of Liapunov's second method, IRE Trans. Circuit Theory CT-7, 520-527 (1960).
319
[3] LaSalle, J., Asymptotic stability criteria, Proc. Symposia Applied Mathematics, vol. 13, "Hydrodynamic Instability." 299-307. Amer. Math. Soc., Providence, Rhode Island, 1962. [4] LaSalle, J., and Lefschetz, S~, "Stability by Liapunov's Direct Method with Applications." Academic Press, New York, 1961. (5) Miller, R., On almost periodic differential equations. Bull. Amer. Math. Soc. 70, 792-79S (1964).
[6] Miller, R., Asymptotic behavior of nonlinear delay-differential equations, J. Diff. Eqs. 3, 293-305 (1965). [7] Yoshizawa, T., Asymptotic behavior of solutions of a system of differential equations. Contrib. Diff. Eq. 1, 371-387 (1963).
320
Decoupling and Pole Assignment in Linear Multivariable Systems: A Geometric Approach W. M. WONHAM AND A. S. MORSE
IN
the theory of isolated dynamical systems (flows on manifolds, etc.) the notion of an invariantset serves a crucial role for classifying the nature of a system. A subset of the state space is said to be invariantif startingin it, the trajectorymust remainin it for all time. For control systems, this property is not particularly useful. For example, if a system is controllable, in the sense that any initial state can be drivento any terminal state, then there are obviouslyno non-trivial invariantsets in this sense. However, if one modifies"must remain" into "can remain" in this definition, then a very useful notion is obtained. This is the key observation that underlies the notion of invariance introduced in the paper by Wonhamand Morse (see also [1]). Wonham and Morse consider finite dimensional linear sys= tems. The linear subspace V of the state space X of Ax + Bu is said to be controlledinvariant[or (A, B)- invariant, as they call it] if for any Xo E X there exists an input u(.) such that the resulting solution x(·) with x(O) = Xo satisfies x(t) E V for all t E IR. It turns out that this condition is equivalent to (A, B)-invariance (AV S; V + im(B)) and to the possibility of making the subspace invariantby state feedback (3F such that
1tx
(A
+ BF)V
S; V).
This concept, together with the related notion of controllability subspace and the dual notions of conditional invariance, opened the road to the elegant "geometric" theory of linear systems and a wide range of applications,especially to problems of disturbance decoupling,non-interactingcontrol, and regulation. Indeed, the decoupling problem was the main motivationfor the developmentof the geometric approach.The early contributions to the study of this problem are reviewed in Wonham's seminal book [20, Sec. 9.10]. The problem of disturbance decoupling with output feedback was completely solved in the linear case in [16] and, incorporating stability,in [18]. Soon after the appearance of the notion of controlled invariance for finite dimensional linear systems, it was generalized in many directions, notably to infinitedimensional systems [4], to "almost" versions of these notions [19] (that made it possible to treat approximate decoupling problems and high gain feedback), and to nonlinear systems. It also sparked a totally new approach to discrete event systems; see, e.g., [15].
In the late 70s and in the early 80s the notion of controlled invariance was extended to nonlinear systems, and this extension opened the way to the systematic development of methods for the design of nonlinear feedback laws. In the context of nonlinear systems, the notion of invariantsubspace generalizes in two distinctways:invariantsubmanifoldsandinvariantdistributions. Givena vector field f, an f -invariantsubmanifoldis a "surface" with the property that any integral curve of f which intersects this surfaceis entirelycontainedin it, whilean f -invariantdistribution is, essentially, a partition of the state space into a "family of surfaces" with the property that the flowof f carries surfaces into surfaces (if, in particular, one of the surfaces of this family contains an equilibrium of f, then this surface is an invariant submanifold,otherwiseit is not). In light of this, the extensionof the concept of controlled invarianceto nonlinear systems leads to the notions of controlledinvariantsubmanifolds/distributions .as objects that can be rendered invariant by means of suitable feedback laws. Thus, for instance, in a control system d dt x
= f(x) + g(x)u
with state x E lRn and u E R, a submanifold/distribution is controlled invariant if there exists a feedback law u = a(x) which renders it invariantfor the "closed-loop" system d
dt x
= f(x) + g(x)a(x)
In the nonlinearcase it is the notionof controlledinvariantdistribution (in contrast to invariantlinear manifold, around which the linear theory is centered)that plays a fundamentalrole in the analysis and solution of the problems of disturbancedecoupling and noninteracting control. Decoupling a fixed output from a fixed disturbance input is achieved by imposing, by feedback, the existence of an invariant distribution that renders the influence of the disturbance unobservable by that particular output. This fruitful domain of research was opened by the works [9] and [6], and then continued by various authors (see, e.g., [13], [10], [5]). The problem of noninteracting control with stability required some extra effort in view of the existence of certain
321
obstructions that do not have a counterpart in the case of linear systems, but eventually was solvedin [8], [17],and [2]. The notion of invariant manifold led to the introduction of the conceptof zero dynamics, the nonlinearanalogue of the dynamicsassociated withthe numerator of a transferfunction. The zero dynamics of a nonlinearsystemis a dynamical systemdescribingall (forced) statetrajectories that are consistentwiththe constraint that the output is identically zero. As in the case of linear systems, it can be shown (under appropriate "regularity" hypotheses) that those trajectories are actuallyfree trajectories of a feedback-modified system,whosestate-space is a submanifold (a controlled invariant submanifold) of the original state space. If all such trajectories converge as time tends to 00 to an equilibrium, the system has properties analogous to those of a linear minimum-phase system. This notion, developed in a few preliminary conference paperswhoseresultsare summarized in [3],hada significant impactin theanalysis of theproblemof output trackingand, aboveall, in the systematic designof adaptive feedback laws for systemsaffected by parameteruncertainties, as shownin the monographs [12], [11]. REFERENCES
[1] G. BASILE AND G. MARRo, "Controlled and conditioned invariant subspaces in linear system theory," J. Optimization Th. & Appl. 3:306-315, 1969. [2] S. BA'ITILOTII, "A sufficient conditionfor noninteracting controlwith stability via dynamicstate feedback," IEEETrans. Aut.Contr., AC-36:10331045, 1991. [3] C.I. BYRNES AND A. ISIDORI, "Asymptotic stabilization of minimum-phase nonlinearsystems," IEEETrans. Aut. Contr., AC-36:1122-1137, 1991. [4] R.F. CURTAIN AND H. J. ZWART, An Introduction to Infinite-Dimensional LinearSystemsTheory, Springer-Verlag (Berlin), 1995. [5] W.P. DAYAWANSA, D. CHENG, T. J. TARN, AND W.M. BOOTHBY, "Global
([, g)-invariance of nonlinearsystems," SIAM J. Contr. Optimiz.,26:11191132, 1988. [6] R.M. HIRSCHORN, "(A, B)-invariant distributions and disturbancedecoupiing of nonlinearsystems," SIAM J. Contr. Optimiz., 19:1-19. 1981. [7] A. ISIDORI, Nonlinear Control Systems: An Introduction, 3rd ed., SpringerVerlag (Berlin), 1995. [8] A. ISIDORI AND J. W.GRIZZLE, "Fixed modesand nonlinearnoninteracting control with stability," IEEETrans. Aut. Contr., AC-33:907-914, 1988. [9] A. ISIDORI, A. J. KRENER, C. GORI GIORGI, AND S. MONACO, "Nonlinear decouplingvia feedback: A differential geometricapproach," IEEETrans. Aut. Contr., AC-26:331-345, 1981. [10] AJ. KRENER, "(Adf,g), (adf,g) and locally (adf,g)-invariant and controllabilitydistributions," SIAM J. Contr. Optimiz., 23:523-549, 1985. [11] M.KRSTIC, I. KANELLAKOPOULOS AND P.KOKOTOvIc,NonlinearAdaptive Control Design, Wiley(NewYork), 1995. [12] R.MARINO AND P.TOMEI, Nonlinear Control Design:Geometric, Adaptive, Robust,PrenticeHall (New York), 1995. [13] H. NUMEUER AND AJ. VAN DER SCHAFf, "Controlledinvariance for nonlinear systems," IEEETrans Aut. Contr., AC-27:904-914, 1982. [14] H. NUMEIJER AND AJ. VAN DER SCHAFf, Nonlinear Dynamical Control Systems, Springer-Verlag (New York), 1990. [15] PJ. RAMADGE AND W.M. WONHAM, "On the supremal controllable sublanguageofa givenlanguage," SIAMJ. Contr. Optimiz., 25:637--659,1987. [16] J'M, SCHUMACHER, "Compensator synthesis using (C, A, B)-pairs," IEEE Trans. Aut. Contr., AC-25:1133-1138, 1980. [17] K.G. WAGNER, "Nonlinearnoninteraction with stabilityby dynamicstate feedback," SIAMJ. Contr. Optimiz., 29:609--622,1991. [18] lC. WILLEMS AND C. COMMAULT, "Disturbancedecouplingby measurement feedbackwith stabilityor pole placement," SIAMJ. Contr. Optimiz., 19:490-504, 1981. [19] J .C. WILLEMS, "Almost invariant subspaces: An approach to high gain feedback design-Part I: Almost controlled invariant subspaces, Part IT: Almost conditionally invariant subspaces," IEEE Trans. Aut. Contr., AC-26:235-252, 1981,and 27:1071-1085, 1982. [20] W.M. WONHAM, Linear Multivariable Control: A Geometric Approach, 1st edition, Lecture Notes in Economics and Mathematical Systems, Vol. 101, Springer-Verlag (New York), 1974; 2nd edition, Applications in Mathematics series, Volume 10, Springer-Verlag (New York), 1979.
A.I. &J.C.W
322
DECOUPLING AND POLE ASSIGNMENT IN LINEAR MULTIVARIABLE SYSTEMS: A GEOMETRIC APPROACH· W. M. WONHAMt AND A. S. MORSE:
1. Introduction. The current interest in linear multivariable control has led to several algebraic results with important applications to system synthesis. In particular, the problem of decoupling of individual system outputs by means of state variable feedback was studied by Rekasius [1], Falb and Wolovich [2] and Gilbert [3]; the problem of realizing arbitrary pole locations in the closed loop system transfer matrix was investigated by Wonham [4] and Heymann [5]. In the present article, new results are obtained along these lines. In § 3, the problem of neutralizing the effect of disturbances with respect to a specified group of output variables is solved. In § 4, the concept of a controllability subspace is introduced and its relation to pole assignability is investigated. This material is preliminary to the formulation of a general problem of output decoupling in § 5. In § 6 and § 7, necessary and sufficient conditions for decoupling are obtained in two specialcases; the results of§ 7 complement and extend those obtained previously in [1], [2J and [3]. In each case, the problem of pole assignment is solved completely. Our viewpoint is that such problems are usefully treated in a geometric framework in which both definitions and results become intuitively transparent. In this way, entanglement at the outset in a thicket of algebraic calculations is avoided. Of course, for applications, it is necessary to translate the geometric criteria into matrix operations suitable for computation. This matter will be considered in a future article. 2. Notation. The control system of interest is specified by the differential equation (2.1)
X(I)
= Ax(.t) + Bu(t'
with x an n-vector, u an m-vector and A, B constant matrices of dimension, respectively, n x nand n x m. Here and below, all vectors and matrices have realvalued elements. Script letters denote linear subspaces: 8" is real n-space; 'I '1 is the orthogonal complement of the subspace .'1"'; 0 denotes both the vector zero and the zero subspace. If K is a matrix, {K} or .K is the range of K, and. ·i' '(K) is the null space of K. If K is of dimension J.l x v and 'I.' ' c till, we write K- 1 1 . for the subspace
{z:zet8''',KzE 1"} C s: The controllable subspace of the pair (A, B), written {Alat}, is defined as {AI~}
= &4 + A~ + ... +
An-lYl.
• Received by the editors February 3~ 1969, and in revised form June 4. 1969. t Office of Control Theory and Application, NASA Electronics Research Center, Cambridge. Massachusetts 02139. The work of this author was supported by the National Aeronautics and Space Administration while he held an NRC postdoctoral resident research associateship. : Office of Control Theory and Application, NASA Electronics Research Center. Cambridge.. Massachusetts 02139. Reprinted with permission from SIAM Journal on Control, W. M. Wonham and A. S. Morse, "Decoupling and Pole Assignment in Linear Multivariable Systems: A Geometric Approach" Vol. 8, 1970, pp.1-18.
323
Thus, {AltI} is the largest subspace of 8" which the control u(·) in (2.1) can influence. Observe that {Alai} is an A-invariant subspace of In. With (2.1), we consider the auxiliary equation
(2.2)
y(t) = Hx(t),
where H is a constant q x n matrix. The vector y is the output. Equations (2.1) and (2.2) play no essentialrole but serve to guide the investigation. 3. Localization of disturbances. In place of (2.1), consider the perturbed system
= Ax(t) + Bu(t) + De(t), d matrix and c;( ·) is a disturbance input. If u(t) =
(3.1)
x(t)
where D is a constant n x Cx(t) + v(t) (where v( · ) is an external control input), then the output y( · ) will be unaffected by all possible c;( · ) if and only if {A + BCIP}} c ~K(H). This suggests the problem: given A, B, !J) c tI", .IV c tfn, under what conditions does there exist an m x n matrix C such that {A + BCI9}} c ..AI? If C exists, the effect of disturbances is, in an algebraic sense, localized to ./V: THEoREM 3.1. There exists C such that {A + BCI~} c ,iV· if and only if ~ c 1/;where l ' is the maximalsubspace such that
(3.2) Furthermore (3.3)
j/"
is given by "I/' =
1,'(0)
=
.,N'~
j"{lI),
j"(i)
=
where
rv: 1) n A - l(aJ +,,"(i-1}),
i= 1,2,···,v, and v = dim .A": Here and below, "maximal" ('''minimal'') mean l.u.b. (g.l.b.) with respect to the usual partial ordering of subspaces by inclusion. To prove the theorem we need two auxiliary facts. LEMMA 3.1. Let XiEIPI, UiEsm, i = 1, "', N, and write X = (Xl"'" XN), U = (U1'···' UN)' Thereexistsanm x n matrix Csuch that Cx, = Uhi = 1",·, N, if and only if %(X) c ..¥(U). C always exists if the Xi are linearly independent. The simple proof is omitted. LEMMA 3.2. Let "f'"' c tin. Thereexists an m x n matrix C such that (A + Be) 1, . c 1''"' if and only if Ar" c 91 + 'I': Proof. Necessity is clear. For sufficiency, let VI' ••• , VII be a basis of 1< Then AVi == BUi + Wi for some Ui E 8 m and Wi E r; Choose C, by Lemma 3.1, such that CVi = -Uh i = 1, ... , u; then (A + BC)v; = Wi' Proof of Theorem 3.1. For sufficiency, (3.2) implies l' C .. f' and A l' . c 91 + "f': By Lemma 3.2, there exists C such that (A + BC)"f' c l' ~ Then
{A
+ BCI,q)}
c
{A
+ Belt '}
= 1" c . ,~<
The maximal property of 'j' ' was not required. For necessity write {A + BC'~} = If: Then (3.4)
lI'~
c .'V:
A }f" c 324
d4
+
1/':
If '~is the class of all jf' c tin which satisfy (3.4), then clearly 0 E 1Y and 11' is closed under addition. Hence, 1f'''contains a (unique) maximal member 'II: Then ~ c 'II'" c 'I" and r: satisfies (3.2). To prove the second statement of the theorem, observe that 1"'(0):::> r; and if ; '(i- t) ~ 1~ then ,,'(0 ::> 1/ A - 1(~ + 1"') = '1/~ Thus, t"(i) ::':) '/.''' for all i; and since ;,'(i) c t ·(i-1). there is a least integer j such that ~.(;) = r '(j) if i ~ j. Since 'f '(j) :::> ;'- and 1,/"0) satisfies (3.4), tJ 'Cj) = 1- ~ Clearly, 0 ~ j ~ v: and if f} c 1-' we even have 0 ~ j ~ v - dim ~. Remark 1. Theorem 3.1 depends essentially on the fact that the class 11' determined by (3.4), or equivalently
n
n A - 1(~ + 11' ")} ~
'1Y = {lI": 1r c . J'
has a maximal element 1': Furthermore, 1" is defined constructively by means of (3.3). This fact will be used without special comment in the following sections. 4. CoatroUabiUty subspaces. In regard to the system (2.1), suppose that a subspace 91 c I" is selected and that it is desired to modify the system in such a way that ~, but no larger subspace, is completely controllable. This aim is to be realizedby feedback of state variablesand by formingsuitable linear combinations of control variables: that is, by setting u = ex + Kv, where K is an m x Ill' matrix for some m' ~ m. Then (2.1) becomes
x=
(A
{A
+ BCI{BK}}
+ BC)x + BKt'
and we require (4.1)
= .-Jt,
Condition (4.1) can be expressed more neatly by noting that {BK:· c tI and the following. LEMMA 4.1. If rJ c !M and {AI~} = 91, then {AI~ n JI} = ~jt. Conoersely, {!' {AliM n ~} = fJI, thereexists a matrix K such that {AI{ BK }} = iJt. Proof. {AI!f} = fJI implies rI c ~, so !f c 91 n ~, and thus ..Jt = l AI~: c {AlflI n ~}. Also, A9t c ~ implies A(aJ n at) c .~; by induction Ai(:jJ n .11) c ~,j = 1,2,··· ,and so {AI~ at} c YI. For the converse, let bi , i = 1, ... , m, be the ith column of B and let :"i~ j = I, · · . , m'} be a basis of 81 n 91. Then
n
m
rj
=
L k;jb;,
j
= I ~ , .... m',
i= 1
for suitable ki j , and we set K = [ki j ] . This completes the proof of the lemma. By Lemma 4.1, we can pose the synthesis problem as follows: Given A, Band rJI, find conditions for the existence of C such that (4.2)
{A
+ BCI~
n 9t} = JI.
If such a C exists, we call Yl a controllability subspace of the pair (A, B). Observe that at = 0 and fJl = {AI~} are controllability subspaces. Controllability subspaces can be characterized as follows. 325
THEOREM
4.1. Let A, B, .~ c I" be fixed. iJI. is a controllability subspace
(A, B) if and only if (4.3)
;~
A.JI c
~r
+ .Jt
and .~
(4.4)
= if,
where ~ is the minimal subspace such that ~ = .'if
(4.5) Furthermore, :i =
.\jI(P), where
p = dim :JI and
.~(O)
(4.6)
n (A~ + :M).
= 0, ;=1 . 2.···,1l.
WriteC for the classof matrices C such that (A + theorem we need two preliminary results. LEMMA 4.2. Let ~ c .rJl. For all C E C,
BC~ c
.'iI. To prove the
+ (A + BC).j = .:JI n (Abf + :JI). Proof. Let C E C. Then (A + BC).j c .rN and A.J + .M = (A + BC~.J + .~. Jl
n•
By the modular distributive rule for subspaces, .:Jt
LEMMA
n (A~ + ;M) = .if n [(A + Be).} + .A] = (A + BC~~ + .~ n ;JI,
4.3. I)' C E C then i
L (A + BCP-l(~ n :it) =
(4.7)
;.jf(i).
j= 1
where the sequence [JI(l) is defined by (4.6). Proof. Equation (4.7) is true for i = I, If it is true for i = k - I.. then by Lemma 4.2, Il
L (A
+ BCY- l(aI n ..JI) = ~ n .JI + (A + BC)af(Il-
I)
j= 1
= .JI n (A9r(k-l) + .rM) = .;J1(11.,. Proof of Theorem 4.1. By Lemma 3.2, C is nonempty if and only if (4.3) is true. Let .lJI = {A
(4.8)
+ BCltf nat},
Then C e C. By Lemma 4.3, .. jI =
L" (A + BCY-
I
(~
j= 1
326
n .:Jr.) = .Jttll) =
;:JI(P».
Conversely, if .'1t = ~(n). then (4.8) is true for every C E C. It remains to show that (4.5) has the minimal solution g,(p). By induction on i in (4.6), it is seen that Jt(i) c ~. i = 1,2•. · · • for every solution .j of (4.5), and that the sequence .jf(it is monotone nondecreasing. Hence, there is Ji ~ P such that .ljI(l) = ..Jt(Il) for ; ~ JL:
in particular, .tjf(p) c if and ;jf(II) satisfies (4.5), Remark 2. If:~ is a controllability subspace, then it was proved incidentally that ~~
= {A + BCI.qd n ~jf~
for every C such that (A + BC).~ c ~jf. This fact will be used later without special mention. Consider now the problem of assigning the eigenvalues of the restriction of A + Be to fJt. It will be shown that there is complete freedom of assignment and that simultaneously the control v introduced earlier can be made a scalar: i.e., in (4.1) K can be made an m-vector (111' = 1). For this, recall [4] that a subspace .1' is A-cyclic if there exists x E fl' such that {AI {x l} = PI'; that is. if .Uj' contains a generator x. Thus we can take m' = I if and only if.~ can be made (A + BC)-cyclic and ~ n PA contains a generator. THEOREM 4.2. Let (4.3) and (4,4) hold, and let ~1' ••• , ~/' be arbitrary real numbers (p = dim 91). Then C canbe chosensuchthat (4,2);s trueand.'1I is (A + Be)cyclic with characteristic polynomial I'
(4.9)
L ~iAi-l,
; ..p -
i=1
lf 0 ¢ b e fJI
n
fJI is arbitrary, C call be chosen so that, ill addition.. b generales .11. Proof', By Lemma 4.3 and Theorem 4.1, C is nonempty and
{A + BClbit
(4.10)
n .jf} = .jf
for every C E C, Choose C 1 E C arbitrarily and write A + Be 1 = A I ' Let b 1 = b E!JI ~ and let PI be the largest integer such that the vectors
n
bt,Atb l , are independent. Put r 1 = b, and rj
•••
'I
A/ll-lb l
= A1rj - 1 + b1,j = 2,···, Pl' Then riE.1f
and the ri are independent. If PI < p, choose b 2 E~:Jt
n .A such that r 1" • • • , '."."
b2 are independent; such a b2 exists by (4.7). Let P2 be the greatest integer such that
are independent, and define r p t + i = A1' PI + i -
1
+ h2 "
i
=
I ....... P2'
Then r. , , .. , 'Pl are independent and in~. Continuing thus, we obtain eventually r l ' • · · , rP independent and in .~, with the property ri+l
where hi E Jt
= Atri + bi ,
i = I....... p - I.
n 94. Now let C2 be chosen such that Be 2';'= hi' 327
; = l , ... " p,
where bp E;;t n JI is arbitrary. Since bi = BUi for suitable Ui' and the r, are independent, Lemma 3.1 guarantees that C2 exists. The situation now is that i=I~···.p-l
.
and
By independence of the r i'
+ Be 2H r 1 }-;. = .11; A + B((~l + e 2 ) with generator
.: A 1
E.*
that is, YI is cyclic relative to r, = b, n ;14. It is wellknown [4] that now an n-vector c can be found such that A + B(C1 + ('2) +b1c' (restricted to (1) has the characteristic polynomial (4•.9). Setting b, = Bg for suitable gEl"', it follows that the matrix C =
(~I
+ C 2 + gc'
has all the required properties. Remark 3. The result that any nonzero vector in dI ;~ can serveas generator is an extension of the useful lemma in [5]. Remark 4. If 9t = 8", (4.3) holds automatically and (4.4) amounts to {AI&4I} = tin, i.e., complete controllability of (A, B). Then Theorem 4.2 yields the known result [4] that controllability implies pole assignability. The construction just used furnishes a simpler proof of this fact than that in [4J. It will be necessary later to compute the maximal controllability subspace contained in a given subspace .C/. For this, let '1~ be the maximal subspace of .(/ which is (A + BC)-invariant for some C (recall Remark 1 following Theorem 3.1 ): and let C("Y)be the class of C for which (A + BC)"f~' c -1':' THEOREM 4.3. If C E C("Y), the subspace
n
(4.11)
jj
= {A + BCI31 n
17'}
is the maximal controllability subspace in .V}. Proof. By (4.2) and Lemma 4.1, ~ is a controllability subspace. Furthermore, by Lemma 4.3 with C(1,,4) in place of C. ~ is independent of C E C(T) and so is uniquely defined. Now suppose ;j = {A
+ BCI&it n di},
.lI c
i
.Y
•
fA is (A + Be)-invariant and 17" is maximal, there follows .iI c '1~. Let f" = ;j E9 f~ By the construction used in proving Lemma 3.2. a matrix C exists
Since
such that
cs c«. s
xE.i~
Then C E C(Jil, and ~ = {A c {A
+ BCI~ n ~} + BCI8I n 1~'}
= ~i;
that is.. ~ is maximal. 328
s. DecoupUng of output variables: Problem statement. Consider the output equation (2.2), with H=
(S.l)
where Hi is of dimension qi x n.. i = I, ... k, k (2.2) can be written 'l
~
2" til + .. , + 'It = 'I. Then i = 1, ... k .
(5.2)
'I
where Yl is a q,-vector. The vectors Yi may be regarded as physically significant groups of scalar output variables. It may therefore be desirable to control completelyeach of the output vectors Yi individually, without affecting the behavior of the remaining Yj, j =F i. This end is to be achieved by linear state-variable feedback togetherwith the assignment of a suitablegroup of control inputs to each Yi' That is, in (2.1) we set k
(5.3)
U
= ex + L
«»;
i= 1
For Vi to control Yi completely, we must have (5.4)
where Pltj is the range of Hi' Since the ith control r, is to leave the outputs Yj' j :#: i, unaffected, we require also (5.5)
Recalling the equivalence of (4.1) and (4.2), we can express conditions (5.4) and (5.S) more neatly as follows. Write tI" = I and (5.6)
i
Then our problem is: Given A, Band . ~ i , ···
bility subspaces ~ I '
'l
•
=
L··· k . 'I
ti, find a matrix C and controlla-
.Jt" with the properties:
··· , .
(5.7)
~~i = {A
(5.8)
.Jti
(5.9)
:Jt i c:
+ BC'~
+ . t·; = 8,
n . Jj~
n .Jtd- .
; = L··· . k . i = 1.. k. ; = 1. ., k .
i*i
Here (5.8) and (5.9) are equivalent, respectively, to (5.4) and (5.5). The relations (5.7}-(5.9) provide a geometric formulation of the problem of simultaneous decoupling and complete control of the output vectors ..vi" . · . " J'k' Thus stated, the problem definition is both natural and intuitively transparent. We observe that the output matrices Hi play no role beyond specification of thesubspaces. f'jOl Since the Hi need have no special structure, the.• "; are similarly 329
unrestricted. Nevertheless, we shall rule out trivialities by tacitly assuming: (i) . i·i #: I, i = 1, · · · , k, (ii] The subspaces ~ °t are mutually independent. 1 In particular, the .• i are distinct and 0
(5.10)
0
i :F 0,
l
; = 1~ · . · , k.
(iii) The pair (A, B) is completely controllable, i.e., {AI~~} = if. For if (i) fails, then for some i, . t i = I'; that is, Hi = 0 and Yi == O. If (ii] fails, then for some i, , t· 't
n L .r OJ =F 0 j~i
or, by taking orthogonal complements,
.•'i+
n.i'j~8 j~i
and (5.8) must fail. For (iii), if {AI~} = 8 1 #: 8 we can write I = 1'1
~
8 2 and
(2.1) as
Xl = A1x1
+ A 3 X 2 + Blu,
X2 = A 2 X 2 ,
"1'
where X;E"i, i = 1,2, and {Allbf!} = The problem is unrealistic unless A 2 isstable(i.e., the pair (A, B)is stabilizable [4]). Hence, we may assume X2(t) == 0 and take as starting point
The problem can then be reformulated with 8 1 in placeof I. We tum now to the determination of necessary and sufficient conditions for the existence of a solution to (5.7)--(5.9) in two special, but interesting, cases. In the following sections, ~i denotes the maximal controllability subspace such that (5.11 )
:li en
..Jj~
i
= 1,,·· ~ k.
j*i
The ~i are constructed according to Theorem 4.3. 6. Deeoupllng when rank (H) (6.1)
= n.
Our assumption is equivalent to
n" .i: = o. i= I
That is, there is a one-to-one mapping of state variables into output variables. THEOREM 6.1. If (6.1) holds, then the problem (5.7)-(5.9) has a solution {r and only if
; = 1, ..... k .
(6.2) 1
Equivalently.. the row spaces or the Hi are mutually independent.
330
Proof. If the problem has a solution :JI" i = 1, · .. , k, then by maximality of the ~i' i = 1, ... , k, there follows:A i c .ii' and (6.2) follows from (5.8). Conversely, suppose (6.2) holds. The jii are mutually independent; for.. by (5.11) and (6.1),
s, n L~Il [n .'j] n [L n "f\:] en. "j n " i =
O.
f
C
Il~i
j~i
Il:l=i
V~1l
j~i
Let C, be chosen such that ~j;i
= {A + BCi/:A n ~d"
i
=
L,·,. k .
Since the ~jii are independent there exists" by Lemma 3.1" a matrix C such that = C;r (reti i , i = 1. ..... k).. i.e,..
Cr
(A
+ BC)r =
+ Bei)r .
(A
re.ji"
i = I . ···. k .
Then
J; = -fA +
BCI~
n JJ i }· ..
i = L " .. k :
and C. together with the ~,' satisfy (5.7)-(5.9). Remark 5. By Theorem 4.2, the C, can be chosen so that A + BC;, restricted to ~i' has any desired spectrum. Hence, the same is true for A + BC. Furthermore. there exists b, E &f ~i such that
n
~i
= ·:A + BCI{b':·:'"
7. DecOlipUng when rank (8)
= k,
i =
L ., .. k .
Our assumption is equivalent to
dim:M = k,
(7.1)
Here the situation has been simplified by narrowing the choice of generating subspaces fJI s; The same assumption was made in [1], [2] and [3]" with the additional restriction that the outputs Yi be scalars. THEOREM 7.1. //(7.1) holds, then the problem (5.7>-(5.9) has a solution ~r and only ~r
n
(7.2)
·ji
+".i = 8,
; = L ., ... k .
and k
(7.3)
;;f
= L
~
n s;
i= 1
Furthermore, if C, Jt l '
·· . ,
·:Jtk is any solution" then
(7.4)
i
Proof", Part 1. Suppose C,
= L··· . k.
~ l ' ... " .il k
is a solution. The necessity of (7,2) follows, as in the proof of Theorem 6.1. To verify (7.3), write
~ n .II; = :iI; E9 [ !11 n .1; n j~j .lIi ] • 331
i
=
L···. k .
The ~i are mutually independent; in fact, HI,
n
L !/I
i_'
j C
:Mi
n
L ;M n ~j c
n (~ n .it;) n
.oA;
i-'
L·~.i =
o.
jei
Recall that the:A i are (A + Be)-invariant. Then
= ·~A + B('l:jfi~' + ·ji~
.11;
where
.Ai C{A +
Bcl L ilti}C L .Jt L n .1"~ c. t;. !j
j C
*' i
j ¢ i Il
j¢ ;
*j
Therefore. by (5.8), {A
+
BCI~':'
+ . t ; = A,
and since . Ji #: 4 there follows
~i
"# 0, i = 1....... k. Therefore
L ~i = L dim ~i = Ie
"
i=1
i= 1
dim
k:
so
(7.5)
.~
= ~A I ~
...
Ea :M"
"and dim..M;
= 1..
i
=
l .... · . k .
Since ~, c: fM n :iii C ~ n Jilt it follows that (7.3) is true. Proof. Part 2. To verify (7.4),it is enough to show that the subspaces ~M are independent. For then.. dim (fjI
n Ji
i)
n .iI;
; = 1....... k .
= 1..
and so
(7.6)
i
=
l.. ..... k .
Assuming (7.6) is true, let.ji = ~i G) ~i and choose Ci • by Lemma 3. t . such that (A
Then C, E C(aI,)
+ BCi).,ii C
'~i"
eir
= C,.,
re .if;,
n C(~,), so that ;it i
= {A + BCil~ n ·A'd" = ·fA + BCil;~ n ~d' =
.A;
which proves (7.4).
We proceed to show that the.-A
n ~ii are independent. Write
J;r = L s; j¢;
332
;
=
1.. ... " k .
It is even true that :An.lin~jr=O,
(7.7)
i=L···.k.
On the contrary, suppose (7.7) fails for, say, ; = 1. If dim ( j :J4n~l
(7.8)
If dim (at
n ~j 1) = L then
c.iT.
n ~1) ~ 2, and
(7.9)
;A
n ·ji ¢. ~r,
i = J,
0"
•
k,
then
Jl = 1, . · · , k - 1, that is,
dim
[.f,= ~ n ';;i] ~ 1
3;
and by induction dim
[.r ~ n !.Ii] ~ ,=1
+
k
I,
a contradiction. Thus (7.9) is false; combining this result with (7.8) there follows (7.10)
&f
n d(l. c fl,:
for some cxe(l,···, k). It will be shown below that there exists C2 such that (7.11)
(A
+ BCac)9l(l
C
~2;
(A
+ BC(l~:
c ;~:.
Assuming (7.11) is true, we have ~(I = {A
+ BC,JBI
n :i(l.}
{A + BC21~:} c ~: c .. i';.
C
and therefore (7.2) fails for i = ~. With this contradiction, (7.7) is established. It remains to verifythe existence of C2 For this we need the following result. LEMMA 7.1. Let 'IJ"~ 11" be arbitrary. There exists C such that 0
(A
+ Be)l-"
j ~
c
(A
+
Be) 'II' c
'JI'
if and only if' A1,'c9l+
1~
A11'cal+
'II~
A(1'
n ~"') c
:11
+
1"
n ~:
Proof. Necessity is obvious. For sufficiency, write l'
+
'H"
= tAO ~ (1 . n ';f") ffi 333
1i~
where 1-' c 1 ~ ,*~' c 11 ~ By the construction of Lemma 3.2, C can be chosen such that
n 'II ') c
(A
+
(A
+ BC){' c:
(A
+
BC)( t .
t'
1~
eci« c
'N~
This completes the proof of the lemma. ~:. Clearly A;j2 c J8 + Consider now fA = ~ n :12 + aJ n ~:, and so
s;
A(:i:l
n ;1:) c
n 'H ~
s; A-l: c
~11
+ .j:.
By (7.3)
.12 ) n (~ + ,~:) = 1~ + (;M + ~ 2) n .j: = 111 + (~ n ~j: + ·j2) n ~: = .eJI + ~2 n Ji:. (:M +
By applying Lemma 7.1, the existence of Cf/. is finally established. Proof. Part 3. We now prove that (7.2) and (7,3) are sufficient conditions for existence of a solution. Let fi be the maximal subspace such that 1~'
(7,12)
en. tj,
i = 1.. ..... k.
j:Fi
It is enough to check that the f; are compatible, in the sense that there exists C such that (A
+ Be')1;' c
t~·.
i=
L·,·~k.
;=
1.. ..• _k .
We show first that the subspaces
L 1~'
1~'" =
j'¢i
are compatible. From (7.12) there follows A
where ~i = 1M that
n c: at + '-:1 = :JI
n ~. +
= ~i
+
'f~1
(by (7.3))
t~,
n ·t~. By Lemma 3.2.. there exist B, with {Bd' = Jli't and
('i"
such
i = L··· . k .
Choosing a basis {VI' ... , vll }- for
1~'
+ .,. +
t~',
we define C such that
k
BCv,. =
L ec«; ;= I
334
\' = I . ··, .
}l.
Then
(A + BC)"f7 = (A + ec, + Ji BiCi) ",-,1' + BiC i )"f71 +
c (A
L B4j i*i
(7.13) c
'f-r
+
L 1j' j¢i
This proves compatibility of the
1li. Now define 1'; = n 'f"j,
i
= 1..','. k .
i
= 1.. .•.
't
k.
Clearly, 'I i :::> f;', i = I, · · · ~ k. By (7.13), (7.14)
(A
+ Be)'! i c i'i.
i = 1,,··· . k..
and, furthermore by the second condition of (7.12),
r,c
(7.15)
n L n j~i
,t~
= n . I j.l
2:1:j m¢'2
j*i
By (7.14) and (7.15), the j'j satisfy the conditions imposed on the '1~' in (7.12). Since the "fi are maximal, there results 'I i c j~, and, therefore, 1; = '1~' . i = 1, "., k. Remark 6. If the conditions of Theorem 7.1 are satisfied, then
(7.16)
.t
,-I
iii =
.± {A + BC/£f n
,= 1
= {A
;ji}
= {A + BCj.± £f n iii} ,=1
+ BClaI} = {Alai} = 8.
We turn now to the problem of pole assignment. In contrast to the situation of § 6, it is no longer possible, in general, to vary the spectrum of A + Be on each
:Ai independently. The following example shows that certain eigenvalues of A + Be may even be fixed for all admissible C. Let tf = 8 3 • k = 2 and
A =
I 0 0]I [001 1 I
~
B
= [:
~l
0] Ii = [~ . l.
2 This identity and its dual, rule for subspaces.
L n L = L, are readilyestablished by using the (modular)distributive 335
It is easily checked that (S.7H5.9) have the (unique) solution
and that C must have the form C
with arbitrary c 1 '
C2'
det (A
=
[('1 0 OJ
°°
Then
+ Be -
+
A./) = (1
"2
A..HI - A.)( 1 +
('1 -
('2 -
A).
Observe that the eigenvalue A. = I, belonging to the eigenvector (0'1 1'10)' of A
+
BC",
is fixed, To discuss the present case in general, we introduce a suitable decomposition of 8. Assume that the problem of (5.7H5.9) has a solution C, .j l ' ... , .11k .. and let C denote the class of matrices C for which (A + BC~~i C ~j;, ; = L .... k. We know that the spaces :l; are the unique solutions: for simplicity of notation" write Jt; for JI/. Define k
(7.17)
n ·*r,
tfo =
i= 1
and let tl i be any subspace such that .lJti=8i~(.'JtintfO)'
(7.18)
;= I.. ···,k.
In the following, J denotes the set of indices (1, ... , k), J o the set (0, 1, ... , k·). In intersections and summations involving 91's, the index ranges over J: in those involving 1'5, the index ranges over J 0 • LEMMA 7.2. The subspaces tl i have the properties (7.19)
4 o E9 8 1 ~
(7.20)
(A
••• ~8k
=I .
+ BC)tIi c li+ 4 o ,
ieJ o .. CECa
Proof. Assertion (7.20) is obvious by the fact that the :iii are (A invariant. For (7.19), observe first that 9l i
n tI0 = !JIi n n .~1 n fJtr = ·:Jli n ;Jt; j:;';
and so, if i e J,
8i
n (8
0
+
L tS
J
j)
c:
s, n (8 0 + .<Jtj)
j*O j:Fi
n ~r = G n :Jt n ,CJtr
= A'i
(7.21)
j
i
= o. Now for arbitrary subspaces ,Yi, i = 1, 2, 3, if ,y!
n ('~2 + ,Y3) =
.V1 n '2 336
+ ,Y~ n .3,
+ Be)-
then Applying this fact and using (7.21) we have 81
Ie
)
k
n ( 8 0 + i~2 8) = 0 = 8 1 n A'o + 8 1 n j~2 8'j.
and therefore ") Ion ( t!.+j~2'i =/o n / . + / o n JIi. 28j Ie
n L lSj.
= 10
j=2
Repetition of this argument yields, after k - 2 steps"
10 n (11 + jt2 Ii) = 1 0 n It = O.
(7.22)
Equations (7.21)" (7.22) state that the 8 i l E J 0' are independent. Finally" by (7.16) . 'l
Ie"
L Ii = L (8 i=O
k
i
+ 80)
~
i=1
L ~.;ti = I. i=1
Remark 7. If the 91; are independent, then 8 0 = 0 and G; = Jt;9; E J. For; E J o let Pi be the projection on 8; along 8 j and now let C E C be fixed. LEMMA 7.3. Let BI n '~i = {hi}' ieJ. Then
k¢;
tI; =
't
+ BC~{Pibd'}' i e J, Proof By (7.18) and (7.19), ti = Pial By (7.18) and (7.20),
(7.23)
{~(A
j•
n
P/Jti
= Pi L (A + BCy-l{b i } j= 1
n
= L [Pi(A + Bc)]j-l{Pibi:' j= I
= {~(A + BC~{~bd'}' LEMMA 7.4.
(7.24)
.4
n8
= O.
0
Proof. By (7.3) and (7.7),
~ n 40 = ( i
i= 1
Y4
n ::11;) n
= (iM n 91 1 + = (
n.:Jt1
j= 1
i ~ n rJl
i)
;=2
i ~ n .11;) n niN!
i=2
j=2
337
n .':1It n
n,1 lt
j=2
This completes the proof of Lemma 7.4. Next let C = C 1 be a fixed member ofC, let C2 E C, and write D = C 2 - C' l : thus A + Be 2 = A + BC I + BD. Now bie;~i c ~';i + ~IO (jEJ): and (7.20) yields BDtl i c Ii + 1 0 (ieJ o): therefore (7.25)
i.j e J;
i
r ).
Also, using (7.24) (7.26)
Write k
(7.27)
L h.jc/j~
BD =
j= 1
where, as before, .~ hj } = (7.28)
~!d
n ..;tj' Then
PiBDlj=Pibi,litrf'j=O,
ie i,
i#j,
We can now compute the spectrum 1\ of A + Ai = ~(A
(7.29)
+
8('2'
B(' d,
i E J 0,
+ Be 2)(1
- Po)
jeJ o .
Define
By (7.26) and (7.29), (7.30)
Po(A
+ Be 2) =
Po(A
+
AoPo,
E
J.
and by (7.28) and (7.29), k
PitA
+ Be 2 ) =
Ai
+ P;BD
(7.31)
LP
j
j=O
= Ai
+ Pib;d;P
j,
;
Suppose A. E A.. with corresponding (complex) eigenvector ~. A brief calculation from (7.30), (7.31) shows that either (i) for some i E J, Pi~ #- 0 and (Ai + Pibid~)Pi~ = APi~' or (ii) ~ = poe and Aoe = leo Conversely, if Ao~ = Ae for 0 -# ~ E &o - or (Ai + p;bida~ = A.~ for 0 '# ~ E IJj and some i E J, then AE 1\. Therefore k
1\ = U Ai' i=O
where Ai" ; E J o .. is the spectrum of the restriction of Pj(A + Be 2) to & i- By (7.30)" Ao is independent of the choice of C 2' i.e., is fixed uniquely by the requirement C E C. On the other hand, for; E J Lemma 7.3 states that 8:j is the controllability space of the pair (Ai' ~bi). Hence.. any choice of Ai can be realized by appropriate choice of d, : indeed, for any W E 8 there exists d, such that I
d·x = I
{W'X 0
for xeA'i'l for x E L 8'j. j*i
These results are summarized in the following theorem. THEOREM 7.2. Let the conditions oj' Theorem 7.1 be satisfied. (r C E C., the eigenvalues of A + Be can be partitioned into k + 1 disjoint sets
338
where
no =. dim
(.n ]11)' J= 1
n, = dim (.ji) - dim (~i
n J#r)~
iE J.
The set A o and the integers ni (i E J 0) are fixed for all C E C. The sets Ai (i E J) can be assigned freely (by suitable choice of C E C) subject only to the requirement that any Aii with 1m Aij =1= 0 occur in Ai in a conjugate pair. Remark 8. If basis vectors are chosen in the tS i , then the system differential equation can be put in a simple "normal" form. Let
and ~x ~
(A
+ BC 2 )x + Br .
Multiplying through by 1'; and using (7.30), (7.31), we obtain
= (Ai + PibidaZi + PiBv.. i E J , Zo = Po(A + Be 2)(Z 1 + ... + z,,) + Aozo + PoBt:. Let K be an m x m (= k x k) matrix such that BK = [b i ..• bk ] and put r = Zi
(7.32~
W == (WI'···' WIe)'.
K \V~
Since hiES; Et> tRIo, we have hi = Pb, + PObi ==
bi + b;o·
Adopting nrdimensional representations of the z., etc... we see that (7.32) can be written as i, =
(Ai
(7.33)
+ £'~)Zi + hiWi~
iE J ~
k
L AOj : j + Ao=o + Bolt\'.
to =
i'* 1
Equation (7.33) exhibits the system (2.1) as an array of k decoupled subsystems, each completely controllable by an independent scalar input "';'1 plus one additional subsystem which is driven by the others and by w, Finally, since ·;li n &'0 = .'Jti n c .f', it follows by (5.8) and (7.18) that s, +.~; = ef~ that is, H;&;
= .If;.
atr
Remark 9. The decoupled system is acceptable in practice only if the eigenvalues in the fixed set Ao are all stable. It is possible to check for stability of Ao as follows. Recall that Jt i C tl i + tf 0 (i E J) and note from (7.20) that A(&;i + a0) c Ii + tf0 + YI (i E J). Furthermore, ~
+ 80
en. 1j,
i E J.
j=/;i
It follows by Theorem 4.3 and the maximality of the .Jt; (='~i) that :Jt; = {A
+ BClat n (I'i + 1 0 ) :,
for any C with the property (7.20). That is, (7.20) is both necessary and sufficient that C e C. Thus, to compute Ao it is necessary only to compute the spectrum of A + Be0 (restricted to tf 0) where Co is any matrix such that (A + Be0 )($1 0 C ~ () • 339
Concluding remark. This article represents a preliminary investigation of the generaldecoupling problem formulated in § 5.The results for the specialcases of § 6 and § 7 suggest the possibility of a complete and detailed geometric theory of linear multivariable control, in which the concept of controllability subspace would playa central role. Specific problemsfor future study includenot only that of § 5 but also the problem of decoupling by adjunction of suitable dynamics (augmentation of the state space), and the problem of sensitivity. As formulated, decoupling represents a "hard" constraint, an all-or-nothing algebraic property. Of course, for applications a quantitative approach via "soft" constraints might also prove rewarding. It is clear that an adequate qualitative theory of large linear multivariable systems is currently lacking;and equallyclear that, with computers,such a theory would find wide application. REFERENCES [1] Z. V. REKASIUS. Decouplinx of multivariabte systems by means of state tariable feedback. Proc. Third AllertonConference on Circuitand SystemTheory, Urbana, Illinois,1965, pp. 439-447. [2] P. L. FALS AND W. A. WOLOVICH, Decoupling in the design and synthesis of multivariable control systems, IEEE Trans. Automatic Control, AC-12(1967), pp. 651-659. [3] E. G. GILBERT, The decoupling of multivariable systemsby statefeedback, this Journal. 7 (1969), pp.50-63. [4] W. M. WONHAM, On poleassignment in multi-input controllable linear systems, IEEE Trans. Automatic Control, AC-12 (1967), pp. 660-665. [5] M. HEYMANN, Pole assignment in multi-input linear systems, lbid., AC-13 (1968). pp. 748-749.
340
System Theory on Group Manifolds and Coset Spaces R. W. BROCKETT
T
HE 1960s witnessed important developments in controltheory revolving around the use of state space methods. Many of these developments were first reported in some of the papers reprinted in this volume. The principal resultswere revolution.. ary because they provided fundamentally new approaches to problems in optimization and estimation whichcouldbe solved in no otherway. The conceptof statespaceis absolutely central, for instance, to understanding recursive estimation and optimal controlof linear systems with whitenoise inputs and quadratic performance indices. As the decade of the 1960s gave way to the decade of the 1970s thereremained a greatdeal of excitement concerning the power of the state space point of view, and it is not surprising thattherewereattempts to extendthe statespacetheoryof linear control systems into the nonlinear domain. Several important papersappeared in the early 1970s, and two of these (thispaper by Brockett, and [14])as well as a later entry ([7])are reprinted in the presentvolume. The mostinteresting common threadwas the emerging general awareness that differentiable manifolds provided the right state space setting in which to study finite dimensional nonlinear control systems. (See the references in the paperby Brockett.) The two papersof 1972 appearing in this volume (this paper by Brockett, and the one by Sussmann and Jurdjevic [14] that follows it) were watershed efforts in exploring the essential differences between the linear and nonlinear theories of control. Although it is now widely recognized, it was a somewhat revolutionary observation made in the paperby Brockett that differentialgeometric methods arecentralto treating globalquestions involving strongly nonlinear systems (e.g.,in controlling rapid, large-angle slewing maneuvers in spacecraft). As the analytical tools from differential geometry became more refined, this observation became more widely appreciated, and the field of geometric nonlinear control theory was born. Therealization thatnonlinear systems shouldbe studied using the tools of differential geometry was singularly important becauseit motivated researchers to thinkaboutcontrolsystems in a different andfundamentally newway. Fromthegeometric viewpoint,a newanddeeperunderstanding of classical linearcontrol
theory emerged in which important classical results were seen as special cases of more general nonlinear results. It also led, veryearlyon, to the discovery thattherewereimportant partsof the nonlinear theoryfor whichthere wereno exactcounterparts in the linear setting. A few words contrasting Brockett's paper with that of Sussmann and Jurdjevic [14] may be in order. Sussmann and Jurdjevic developed theirtreatment withinthe verygeneralstate space framework of differentiable manifolds. Careful analysis within this very general setting showed how the topology of the reachable sets depended on the (Lie) algebraic structure of equations of motion. The groupand coset manifolds treatedby Brockett weremorespecialized buthadenoughadditional structure to facilitate a more wideranging studyof controltheoretic issues. Whilethe settingwas specialized, all the key features of very generalnonlinear systems werepresent. At the time when the paper appeared, the conceptof controllability for nonlinear systems was still not completely understood, and in Brockett's paper we find one of the earliest efforts to explain the important distinctions that must be drawnbetween systems with and without drift-in Brockett'snotation: X(t)
= (A +
t;
Uj(t)Bj) X(t)
(1)
and m
X(t)
= L u;(t)B;X(t)
(2)
;=1
respectively. The symmetric or drift-free case (2) is interesting both in the extentto whichit differs from (1) and also in that it admits a verycomplete characterization of controllability. More specifically, the system (2) is said to be controllable if for any two points Xo and Xl in the state space and any T > 0, there exist piecewise continuous (or merely integrable) controlfunctions Ul (.), ••• , um ( ·) suchthat the corresponding trajectory defined by (2) and X(0) = X0 satisfies X(T) = X1. This is a very strong notion of controllability that turns out to be natural for both symmetric systems, of the form (2),andclassical constantcoefficient finite dimensional linear systems, but that turns out
341
to be generally too restrictive for systems of the form (1). Indeed, systems with drift (1), where the evolution equation has the uncontrolledcomponent A, are almost never controllablein this sense. (See [14], for the complete details.) Both Brockettand Sussmann-Jurdjevic [14] have adoptedthe languageand styleof differentialgeometryand Lie theoryin developinga theoreticalframework for nonlinearcontrol systems. In Brockett's paper, controllability questions are posed in terms of transitive actionsof subgroupson orbits. The paper beginsby proving a versionof Chow's theoremin the settingof matrixLie algebras.The proof consistsof a sequenceof steps that compute Lie brackets and rescale time so as to show that for a system of the form (2) it is possible to steer a trajectory in the direction of any Lie bracket of the matrices B;. From this it is a short step, taken with the help of a basic result of Weiand Norman,to showthat (2) is controllable(in the senseof the abovedefinition) precisely on the orbit g · X o, where g = {exp Ala is the group generatedgeneratedby all finiteproductsof matricesof the form eC , whereC e A = the linear subspaceofR n x n spannedby the coefficientmatrices B 1, ... , Bm; X is a given initial condition of (2). The paper goes on to discuss the controllability of systems having the additional structure of the coefficient matrices Bi taking values in the classical semisimpleLie algebras.This was the first place in which semisimpleLie algebrasappearedin the control and systems engineering literature, and the importance of introducing these ideas was that they opened a new window on applicationsinvolvingcontrolledphysical systems.Brockett does not treat the nonsymmetric case (1) at all extensively, but he considers a special case of (1) in which the vanishingof certain Lie bracketsallowsus to concludethat the systemis controllable in the strong senseof the abovedefinition. This case further specializes to include constant coefficient finite dimensionallinear systems as a subcase. Observability is also treated only briefly (a page and a halt), but Brockett makes the important observation that in the nonlinear case, there is no duality between controllabilityand observability. It is interesting to observe that all the controllability results in this paper involve determining the dimension of the linear span of a finite number of matrices which are computed as Lie brackets of the coefficient matricesappearingin (1) and (2). All Lie bracket computations are explicitly prescribed and are finite in number. (The Lie algebras encountered in [14], on the otherhand, are typicallyinfinitedimensional.) The observability results involve a similar degree of simplicity and explicitness, and thus the control theory of systems defined on matrix Lie groups and coset spaces is comparable to that of finite dimensionallinear systems in terms of computational complexityand level of explicitness. Unfortunately, this degree of completeness is not found in more general nonlinear systems, and it is only when there is a special (e.g., polynomial,[1], [2]) structure
°
present that the theory rests on such explicit, finitely verifiable conditions. In the process of pointing out new research directions, the paper offeredthe field a definitive perspectiveon nonlinear control theory and its applications. After the appearance of this paper by Brockett and [14], geometric nonlinear control theory developed along two somewhat different but complementary lines. One line, which continues to grow and prosper, has been aimed at the development of new geometrically based principles of control.Prominentexamplescome from the literatureon nonholonomic motionplanning(e.g., [11]) and sub-Riemannian geometry (e.g., [5] and [4]). The other line has carried on the effort to extend important ideas rooted in classical control into the nonlinear domain. The paper [7], which also appears in this volume, was an early report of this work, and the textbooks [8] and [10] summarizethe larger history of the effort. The paper's influence on the control of mechanical systems has been especially deep, and it continues to be a primary source in this literature; see, e.g., [13], [9], [12], [6], and [3]. REFERENCES
[1] J. BAILLIEUL, "The geometryof homogeneous polynomialdynamicalsystems," Nonlinear Analysis: Theory, Methods and Applications, 4(5):879900, September 1980. [2] J. BAILLIEUL, "Controllability and observability of polynomial dynamical systems," Nonlinear Analysis: Theory, Methods and Applications, 5(5):543-552, April 1981. [3] J. BAILLIEUL, "The geometryof controlledmechanicalsystems,"in Mathematical Control Theory, 1. Baillieul and 1.C. Willems, Eds., SpringerVerlag(New York), 1998. [4] A.M. BLOCH, P.E. CROUCH, AND T.S. RATIU, Sub-Riemannian Optimal Control Problems, Fields Institute Communications, AMS (Providence, RI), 3:35--48, 1994. [5] R.W. BROCKETT, "Control theory and singular Riemannian geometry," in New Directions in AppliedMathematics, Springer-Verlag (New York), pp. 13-27,1982. [6] L.E. FAffiUSOVICH, "Collective Hamiltonian method in optimal control problems,"Cybernetics-i-, 25(2):230-237, March-April 1989. [7] R. HERMANN AND A.J. KRENER, "Nonlinear controllability and observability,"IEEETrans. Automat. Contr., AC-22:728-740, 1977. [8] A. ISIDORI, Nonlinear control systems, 3rd ed., Springer-Verlag (New York), 1995. [9] N.E. LEONARD, "Control synthesis and adaptation for an underactuated autonomousunderwater vehicle," IEEE J. OceanicEng., 20(3):211-220, July 1995. [10] H. NUMEUER AND A. VAN DER SCHAFf, Nonlinear Dynamical Control Systems, Springer-Verlag (New York), 1990. [11] R.M. MURRAY AND S.S. SASTRY, "Nonholonomicmotionplanning: steering using sinusoids,"IEEETrans. Autom. Control, 38(5):700-716, 1993. [12] V. RAMAKRISHNA, M.V. SALAPAKA, M. DAHLEH, H. RABITZ AND A. PEIRCE, "Controllabilityof molecular systems,"Phys. Rev. A, 51(2):960966, February 1995. [13] K. SPINDLER, "Optimal attitudecontrol of a rigid body," Appl.Math. Opt., 34(1):79-90, July-August 1996. [14] H.J.SUSSMANN AND V.JURDJEVIC, "Controllabilityof NonlinearSystems," J. Diff.Eqns., 12:95-116, 1972.
J.B.
342
SYSTEM THEORY ON GROUP MANIFOLDS AND COSET SPACES· R. W. BROCKETTt
Allltract. The purpose of this paper is to study questions regardingcontrollability,observabiJity, and realization theory for a particular class of systems for which the state space is a differentiable manifold whichis simultaneously a group 0', more generalJy, a cosetspace.We show that it is possible to give rather explicit expressions for the reachable set and the set of indistinguishable states in the case of autonomous systems. We also establish a type of state space isomorphism theorem. These results parallel. and in part specialize to, results available for the familiar case described by X(/) = Ax(I) + Bu(t), y(/) = Cx(t). Our objective is to reduce all questions about the system to questions about Liealpbras generated from the coefficient matricesentering in the description of the systemand in that wayarriveat conditions whichare easilyvisualized and tested.
I. Introduction.A standard assumption in modern control theory is that the state space is a vector space. This assumption is both valid and natural in many situations, but there is a significant class of problems for which it cannot be made. Typical or these are certain problems which arise in the control of the attitude of a rigid body. The state space in this case is not a vector space. Linearization often destroys the essence of the problem-even if one can work locally---and in any case new and different methods are needed for treating global questions. In this paper we substitute the following hypothesis for the usual vector space assumptions. We let /F and
(A +
it! u~t)Bi)X(t).
yet) =
~;X(t).
XE.F.
where A and B; belong to the Liealgebra associated with .F. the Ui are the controls. and the notation ~X(t) is to be interpreted as being a coset in ffi. We also study vector systems of a similar type in that we can view their evolution as occurring in a coset space. The results concern the explicit construction of the reachable set and a characterization of observability which is easily tested. OUf main point is that this classofsystemsis in many waysnot more difficultthan linear systemsof the usual type in R". There is a moderately large literature on the use of Chow's results [I] and related ideas to study controllability, including the work of Hermann. Kucera. Hermes. Haynes and Lobry (see [2]-[6]). This work is relevant here but we are directly interested in controllability only in so far as it contributes to the identification ofaframework in which we can study a fullrange of system theoretic questions, including observability and realization theory. Notice that it is impossible to pass * Received by the editors August 24. 1971. and in revised form October 27. J&J7J. Presented at the NSF Regional Conference on Control Theory. held al the University of Maryland Baltimore County, August 23-27. 1971. t Division of Engineering and Applied Physics. Harvard University. Cambridge. Massachusetts 02138. This work was supported in part by the U.S. Office or Naval Research under the Joint Services Electronics Program by Contract NOOOI4..67·A-0198·()()06 and by the National Aeronautics and Space Administration under Grant NGR 22 a007·172.
Reprinted with permission from SIAM Journal on Control, R. W.Brockett, "System Theory on Group Manifolds and Coset Spaces" Vol.10,1972,pp. 265-284. 343
directly from controllability results to observability results in the present setup because there is no clear notion ofduality. The main motivation for this work came from some work on Lie algebraic methods in differential equations (see [7]-[10]) and, above all, from being confronted with certain physical problems where linear theory was simply inadequate. Some unpublished work [18], [19] by Jurdjevic and Sussmann is related to this paper. In particular they give in [19] an alternate proof of our Theorem 5 and make a serious studyofthe unsymmetriccase (treated onlysuperficially inTheorem 7 here). We also mention a recent paper by Elliott (20]. 2. Examples. We postpone the development of the subject long enough to present a few simple examples which will help justify why the assumptions are set up the way they are. Example 1 (Control systemsdesign). Consider the problem of determining the gain, k, in the system x(t)
= A:«t) -
k(t)bc~"«l)
so as to achieve good performance relative to an index of the form
n»
{IX) x'(t)Mx(t)dt.
M = M'
~
o.
If a particular initial state is chosen and k( · ) is selected so as to minimize 'It then the performancemight be bad relativeto some other initial state. In caseswhere the initial state is not known, it is much more realistic to pick a collection of initial state vectors and to pick k in such a way as to minimizea weightedaverage of the individual performances. In fact.just to ensure stability it is necessary to average over at least n linearly independent initial states. If exactly n are chosen. then ~ should be regarded as controlling the evolution of the matrix equation cb(t)
= (A
- k(t)bc)<1>(t) ,
<1>(0)
= [x l ' x 2 '
••• ,
x n]
·
The state space is then the space of nonsingular n x n matrices, ~t(I1). Example 2 (Rigid body control). The orientation of a rigid body relative tc some fixed set of axes is described by a 3 x 3 orthogonal matrix A which satisfie: the differential equation
aZl (t)
aI 2(t ) a13(t ) a22(t ) a23(t)
a3 1(t)
Q32(t)
alt(t)
a33(t)
o =
w 3(t )
-w 3(t ) W2(t}
0 -WI(t)
- 0)2(1)
al1(t)
a 12(t)
U13(t)
w 1(t)
a21(t)
a2Z(t)
aZ3(t)
0
Q31(t)
a32(t)
Q3J(t)
The ro's themselves are usually controlled via the equations rol(t) = [(1 2 - 13)/l t ]w2(t)w3(t) + nl(t)/1 1 ,
= (1 3 ro3(t) = [(I I -
c.02(t)
I t)/l z]ln t(t)w3(t) + n2(t)/1 2
It
12)/l J]lJ)1(t)w2(t)
+ nJ(t)/1 J.
The state space for the first set of equations is YeJ(3)-the set of 3 x 30rthogona matrices. The state space for the second set of equations is H3-Cartesian 3-spacc 344
For our present purpose suppose that the center of mass of the body is fixed and suppose that the observed output of this system is a pencil beam of light generated by a light source which is mounted in the body along a line passing through the center of mass. In this case the output is Ci X(t), where ~ is a subgroup which corresponds to a rotation about the pencil beam (an undetectable motion).
FIG.
I. Illustration of the obseruability ofa rigid body
Example 3 (A model for DC to DC conversion).The electricalnetwork shown in Fig. 1contains switcheswhichare to be manipulated in such a wayas to transfer the energy stored on the capacitor 1 to capacitor 2. In order to have a sensible physical model we demand that there be exactly one path through the inductor at all times.
FIG.
2. An electricalnetworkfor which energy is conserved
The equations of motion are
x1(t )
0
x2(t)
-Sl(t)
..t J(t)
0
0
x.(t)
0
S2(t)
x 2(t )
-.s2(t)
0
xJ(t}
s.(t)
where Xl = l,.je";, X J = V3V/C~'t x:! = izV Land 51 and S2 are dependent on the switch positions and take on the values 1 or O. We have 51 = 1 and S2 = 0 if the switch on the left is closed, and we have S I = 0, 52 = I if the switch on the right is closed. 3. Lie algebras aad Lie groups. Let (R")(II denote the set of real n x n matrices; R")( II is a vector space of dimension n2 • By a Lie algebra !R in R")( II we understand a subset of R")(" which is a vector space and which has the property tha t if A and B belong to JR, thus so does [A, B] = AB - BA. If fRl and !f'2 are Lie algebras in An XII, and their intersection z", n !e2 is also a Lie algebra, then if A and B belong 345
to.fIJl and IRz and both are algebras, then [A, B] belongs to both ~1 and !f/2' The union.!l'J U ftJ2 of two Lie algebras, the sum!fll + !fJ2 of two Lie algebras and the commutator [9'1' ~2] of two Lie algebras are not necessarily Lie algebras. Given an arbitrary subset ofR"x" wecan add additional elements to it so as to imbed it in a Lie algebra. To obtain the smallest Lie algebra which contains a given set At"" we first add to . .¥ all linear combinations of elements in ./~,,. so as to get a real vector space ~ . Then we commute elements in .A"i to get ..'f/i = . f'; + [.0 -it; , .hi]; if this is not contained in .h~, then we form ..4-'3 = ./1'; + [~.~;, ,·+i] 't etc. Clearly this process stops in a finite number of steps since at each stage we increase the dimension of the vector space by at least one and the dimension is upper bounded by n2 • We call this Lie algebra the Lie algebra generated by ;4',.. and denote it by
{..,t'"L.·
If.A isa set of nonsingular matrices in Ria x n, welet {.,It} G denote the multiplicative matrix group generated by Ji, i.e., the smallest group in rR")(1I which contains J( and is closed under multiplication and inversion. If . A/' is a linear subspace of R")( ". then the set
contains no singular matrices since det (exp N i) = exp (tr N i ) > closed under multiplication and inversion and, in our notation,
o. Clearly .*" is
.~" = {exp t·¥"}G'
Let fR be a Lie algebra. At each point M in {exp .!t'}G there is a one-to-one map tPM from a neighborhood oro in !R onto a neighborhood of M in {exp !R}G which is defined by
rPM: ~
-+
{exp2'}G't
cPM(L)
= eLM.
This map has a smooth inverse which shows that {exp £P}G is a locally Euclidean space of dimension equal to the dimension of fR. We may check that the maps cP;,l satisfy the conditions for a Coc-manifold in the sense of [11, p. 97]. Thus we may give {exp ~}G the structure of a differentiable manifold. This justifies our referring to {exp .2'}G as a group manifold. If sf is a linear subspace of R")(" which is not necessarily a Lie algebra, we might inquire as to the relation between {exp d}G and {exp {d}A}G. Clearly the latter contains the former. The following theorem claims that they are identical, THEOREM 1. Let JIll , JJJf2 , • , • ., sip be a collection of linear subspaces of R")( ". Then
Before proving Theorem 1 it is appropriate to make a few remarks about its relationship to the controllability literature. (Perhaps a glance at Theorem 5 would help at this point.) In considering the equation
346
as a differential equation in <§t(n) it is clear from the theorem of Frobenius [12] that the solution passing through X E ~~/(n) lies in [exp -{ AiL-t }GX because at each point X in ~gf(n). The collection {A;XL.t is an involutive distribution which spans the tangent space of {exp {Ai}A}CiX at X. Wei and Norman [9]confirm this fact by giving (locally valid) formulas for the solution of this differential equation in terms of the functions u;( . ) and the structural constants of the Lie algebra generated by the Ai (without pointing out the differential geometric interpretation of the result). On the other hand, if we regard X(l) = (I~~ 1 u;(t)A;)X(t) as a control problem, then the natural question is not what manifold contains the solution, but rather what set can be attained from a given point, given freedom over the choice of (u•• U2'···' urn)' The results of Chow [1] (see also Hermann [2]) are applicable here. Chow showed under a suitable regularity condition that the set of points reachable for the vector system i(t) = 1 Ui(t).I~[x(t)] using piecewise constant controls is the same as the set reachable for
L7':
.i(t) =
I"
l'i(t)gi[X(t)],
;-:1
where {gj(x)} is a basis for the involutive distribution generated by {.I~(x)}. That is. {gi(X)} spans a vector space which includes C!;(x)} and is closed under the Lie bracket operation
.
[.I. g J =
~r
~- g
rx
-
2g . ~-.,./. ex
In our case the Lie bracket of A;X and AjX is [Ai' Aj]X. Thus we see that for the differential equation in question the reachable set includes {exp ':A i L.}G' and the theorem of Frobenius ensures that it includes nothing more. The proof of Theorem 1 given below could be shortened considerably by the use of these ideas. The reason for preferring the longer proof given here is that it is constructive, it is self-contained (nothing harder than the implicit function theorem is used) and it has the merit of proving a theorem about n x n matrices using the notation and tools natural to that subject. Proof We give a proof which relies on an implicit function theorem which, under suitable hypothesis, ensures the existence of a solution of :x equations in P < % unknowns. (See [13, pp. 29-30].) We also need the Baker- Hausdorffformuia which asserts that I 1 eA1Le- A I = L + [At.L] + 2[At. [At, LJ] + j![At.[At. [At. L]]) + ....
+ 1}th term in this series is less than ,IL /l2 n II A II In ! so that the series is majorized by the series iiLil [1 + 2:iAil + (2iIAdJ 1/1! + ... J = IILH . e2 i/AI II and hence is absolutely and uniformly convergent on - T ~ t T Note that the norm of the (n
for all T. Let
~.41'
rt
A:.! • . . . An} C .':It U "r:I2 U ... U -: be a basis for~/l + .':12 .... A,.: . Assume 'I
+ ... + ,.r:-/ and let !fJ be the Lie algebra generated by ~ A \ • A." p
347
this algebra is of dimension q. There exists a basis for !I) which consists of terms of the form
L, = A,. Lr + 1
= [A,,(,+ 1)' A,(,+ 1)],
L'+2
= [A,,(,+2).A U, .,.2)] '1
= [A"(r+s)' A'fr+s)]" L,+s+ 1 = [AIt(r+s+ u. [A'(r+s+ 1)'1 A mcr + s + 1)]]' L,+s
L q = [ ... [A".q,.. ..... Am,q)] ... ]. We are quite explicit here because at certain points in our proof it is necessary to regard these expressions as formal expressions as opposed to matrices. We introduce the following special notation. The operator EXP maps formal expressions into matrices. It is defined on Ai and its commutators (i.e... formal expressions such as Ai' [A j , A j]' [Ai' [A j , A,JJ, etc.) as follows: EXP Ail = eA "
'I
e·-1·JIeArl'e - A;.;te - Ajjl. EXP [Ai' Aj]t =
{ eAiJiil eA iJT1j e - A.;Jfii e - A.J'ti-,..
t < O.
The definition is completed by recursion. If A and B are formal expressions . then EXP [A. C]t =
{
(EXP AJ!)(EXP C ~/t)(EXP AJl)- t(EXP C ,,/1) - J.
1 ~
(EXP CJl~)(EXP AJjti)(EXP CJ'I)·- I(EXP A,i/n- I ~
I
0.
< O.
We now show that if B is a formal expression, then
EXP 81 = I + BI + o(t), where o(t)jt goes to zero as t goes to zero. We associate with each formal expression an integer ", called its degree, which is the largest number of times any particular entry of the formal expression is bracketed. To carry out this proof we use an induction on the degree of the expression. Suppose A and C are two formal expressions of degree n or less. Suppose that we have shown that for any formal expression of degree n or less.. EXP At = I
+ Al +
348
ott).
Now in view of the definition of EXP we can expand EXP At and EXP Ct in a convergent power series involving fractional powers of t. Let us say 1 + At + E(t) + Dt 1
+
2),
EXP At
=:;
EXP Ct
= I + C + G(t) + Ft 2 + 0((2),
0(t
where E and G involve powers of t between t 1 and (2 and E and F are the respective coefficients of t 2 • The power series expansions for the inverses are then (EXP At)- 1
= I - At -
(EXP Ct)-
=I
1
E(1) + (14 2 - D)(2
- Ct - G(t)
+
(e l
-
F)t 2
+
U«(2),
+ 0(t 1 ) ,
as is verified by multiplication with the corresponding expressions for EXP At and EXP Ct respectively. Now using these expressions we have for l nonnegative" EXP Bt = EXP [A, C]t = (I
+
Ajl + E(t) + De2 + o(t))(1 + C v/r + G(,/~) + Ft
(I - AJ{
+
E(t)
+ Dt + o(t))((1 + A 2 t/2 + A 2 t/2 2
= I + [A, C]t + Ft = I + [A, C]t + ott),
C
,/t 2
A t
G(\//t)
+ (C
2
-
+
+
C 2•
F)t -
ott)) F)t
-
+ oft))
e t + oft) 2
and the case t < 0 leads to the same result. Well-known properties of the matrix exponential function let one conclude that for Itt> 0, EXP Bt is continuously differentiable. The above argument shows that EXP Bt is differentiable with respect to t in a neighborhood of t = 0 and d
dt EXP Btft=o =
B.
Hence we have for the basis elements Lv'
Now consider a function of u = (U."lll'···" tl q ) and maps Rq x R' into RII x II and which is defined by F(U,l')
=
-I
= (l't,I'2'l ... , tel)
which
+ (EXP L1u.)(EXP L2U2) •.. (EXP Lqu q )
oexp(-L1Vl - L2 v2 Clearly F(O,O)
L'
..• -
Lqvq).
= O. Now the linear approximation of F at (u" 17) = (0.. 0) is given by
~U,I1)(U. v))(o.O)«()u" bV) = Ll~Ul
+ L 2l5U2 + ... +
- L 2 JV2
-
•.. -
Lq(}uq -
L1Jr t
Lqt>l'q
so that the range space of FeU,l')(U, v)l(o.o)(u, 0) is the q-dimensional subspace of 349
R" II spanned by {L i } • Now Fiu, v) + 1 is a finite product of exponentials which we write as It.
+ 1 = exp (Ai1uf.t exp (A i2Ufzl )
F(u .. v)
'CXp(-Lll~l -
L 2 l' 2
exp (Ai\'U~~')
-
-
Lqv q).
Since the Baker- Hausdorff formula lets one write eA"A" e- A " = AI(
+ [A
j •
A,Jt
+ ....
we see tha t eA,r A. e - Ai' belongs to the Lie algebra generated by the A's. Moreover, it is continuous with respect to t and at t = 0 takes on the value A". Using this result repeatedly we see that for each {Ui,,:· we can find Rile in !f' such that
exp (Ai.U~I) exp (Ai2U~2) ... exp (Aikl4tk)Aile cxp (Ai...
Ill::".·/) ... cxp (Ajvuf\~')
·exp(L 1 l ; 1 + L 2 l":?, + ... + Liltq ) = (F(u ~ (1)
+
I)RiJu i l ,
U i 2 •••• ~
u;)
simply by pushing A" past the exponentials one at a time. Clearly R;,JO.. 0" ... ) = Aile. Thus we see that for a and b small,
I)(J.
F(~.U)(u.v)II ••b)(t5U.t5V) = (F(a,b) +
Sj(a.b}t5uj
+ H i(a.bj(5 l 1;)
for some S~a, b) and H;(a, b) in IR. Since 8;(0,0) = Hi(O" 0) = L;and since S, and Hi depend continuously on their arguments, this establishes that the Jacobian of the map F: IR4 x IRq -+ R" x must have rank q in a neighborhood of (0.. 0) ; and hence" by the implicit function theorem cited earlier" there exists an t > 0 and a map ¢ : IRq -4 IRq such that if :1 r rl < e. then II
O.
F(l!J(v). v) =
Since Ftu, EXP
v)
= 0 implies that
LtUl
EXP
L2U2 ...
EXP Lquq = exp(Lt['l + L2r2
+ ... + Lqvq)
we conclude tha t there exists I; 1 > 0 such that if LEY and II L II < e'..
l; 1 • we can
write
= exp Lt, u. )exp(A· u,)··· exp Lt, {4.), 11
II
'2
'2
h.
Iv
Now for any L E!I' it follows that J1(l/rn)LIIII < tl for some integer nl" and thus we can express e': as fI--'nr e'!" ... e1.,'IPf. Likewise we can express e l. 1 eLl . . . e LfI in this form. Let ,K and !£ be Lie algebras in R It can happen that {exp .;r'}G is a bounded subset of R" x .. which is not closed"and it can happen that the closure of {exp .;tr}G equals {exp 9'}G with .K :F .!fJ. The skew-line on the torus [11] is an easy example. Also, {exp '~}G is not necessarily simply connected. Nonetheless.. we have the following result which we deduce from Theorem J rather then sending the reader to the literature. COROLLARY 1. If ,j{ and!R are Lie algebras in (RII )( II, then {exp .K"} G C {exp G if and only if .K c ff, and {exp'X"}G = {exp Y}o if and only if . J!' = Y. II
)( " .
Y:
350
Proof For both statements the sufficiency is obvious. To establish necessity in the first case notice that if {exp .K}G c {exp fil}a, then by Theorem 1,
{exp %}G = {exp %, exp ~}G
= {exp {.%, ~}A}G·
Suppose!R is of dimension n. To obtain a proof by contradiction, suppose that . 1"" is not contained in !.e. Then {Yl, ~}A is of dimension n + 1or greater. Then exp !l' is an a-dimensional manifold and {exp {Yr,!£}} is not, which contradicts {exp!fl} = {exp {!fI, %}}. To establish necessity in the second case repeat this argument verbatim but with "contained in" replaced by "equals" both verbally and symbolically. The notation ad~ B = B. ad] B = [A. B]. ad] B = [A. [A, B]], etc., is standard.lf.K and Z' are Lie algebras, we use the notation {adIt' f},.. to denote the Lie algebra generated by .K under commutation with elements of fiJ. That is, {ad~.K}A
=
{.f,[~.%],···.
(-5f,[9' ... [9',%] ... ]] ... }AO
This algebra may also be described as the intersection of all Lie algebras which contain .K and are closed under commutation with !fJ. If fF and f§ are groups. we introduce an analogous notation. The smallest group which contains .? and all products of the type GFG- 1 for Gin ~and F in~ will be denoted by {AD, ..F}Go This group may be described as the intersection of all groups which contain ;F and are closed under conjugation with elements of f§. If f§ is {exp ~}G and ./F is {exp~K}G' then clearly {AD• .F}G consists of products of terms of the form
THEOREM
2. Let .f and !t' be Lie algebras in Rft x ft. Then
U MfM- 1 } { Mefexp~lG
= {ad~.K}A A
and {AD{CXP~}G {exp%}G}G
= {exp{ad,i' .%"} AlG'
Proof From the Baker-Hausdorff formula we see at once that if L belongs to and K belongs .f{, then eLKe- L belongs to {ad!l' .K} A • Thus the right side of the first equality in question contains the left.On the other hand, expressions of the following type belong to the left side:
to
!{J
I
1
ef2 L - K e-~zl.. - --K
a
=
[L~
K]
+ o((X)~
~
e2L(~[K.L] + o(a»
e- 2/, -
~[K.L] + o(a) =
[L[L.K]]
+ ala).
etc.
Since IR")( is a finite-dimensional space and since {U Me(exp9'la M.t{ M - 1 ; ... is a linear subspace, it is closed. Thus [L, K], [L[L, K]] ... belong to this set and the first equality is seen to hold. The second statement is obtained by exponentiating the first. This gives II
p { ex {
U
M.!RM- 1}
Mefexp~}G
} A
351
G
= {exp{ad~.K}..t}(;.
but since
eL e" e:':
= exp (eLK e- L ) ., we see that
p U MYM- 1 } { ex { Me(exp,i'}G
=
}
A
{ADfcxPZ}G{exp.:fJ}G}G
G
so the result follows. The next theorem states a purely group theoretic result which, although easily proved, is stated formally becausewe need it in our study of observability. THEOREM 3. Let .Yf and ~ be subgroups of a group f§. Let ~ be the subset of f§ defined as 9 = {P: RPR- 1 e .j{~, all R E 9t}. Then ~ is a subgroup of :If, fJI~ is a subgroupof ~., and 9 is a normalsubgroup of A9. Thus Jt n ~ is a normal subgroup of:JI and ~9/~ is isomorphic to Jl/JI n ~. Proof, Suppose Pi and P2 belong to @. Then for each R in ~ there exist H I(R) and H 2(R) in .1f such that RP1R - l( RP"i 1 R- 1) = H l(R)· [H 2(R 2 )] - 1. Since .If isa group, this means RP 1 Pi 1 R -1 belongsto ~, and thus that ~ is a subgroup of ':f. Clearlyit isa subgroupof ,jf sincethe choice R = I is possible. To see thatBt~ is a group, note that if R I and R 2 belong to 91 and PI and P2 belong to 9, then
Since~ is a group and since r!J is a group whichhas the property that if P belongs to ~ then so does RPR- 1 for each R in !Jl, we see that this product belongs to ~.~. Clearly ~ isa normal subgroup of 91~ since RP~P- 1 R - 1 = ~ for each RP in 91.~. By the second isomorphism theorem (Rotman [14, p. 26]) ~ 9 is normal in :~ and ~9/~ ~ ~/ar n 9We now state and prove a Lie algebraic analogue of this theorem. Algebraic tests for observability will be derived from this result. THEOREM 4. Let ,1{ and !fJ be Lie algebras in R" Let .~ be definedas
n
X ".
9
= {P: RPR - I E {exp J'f}G, all R E {exp 9'}G}·
1f:f{ is a Lie algebra in A"XII, then {expf}G C 9 if and only if {adzf}A C .Yr. There exists a unique Lie algebra .~ such that {ad !I' ~ } A C J{ and .k'; contains all other Lie algebras having this property. Proof. Suppose {adz %}A c eK. Then for L, in!£ and K, in Y{ we see from Theorem 2 that {exp .K}G contains
By the hypothesis and Corollary 1, {exp {ad.!'f}A}G c {exp JfP}G. On the other hand, if for all L; in!£ and all K, in :f( we have
then since {exp Jf}G is a group, we see that {AD(cxpg)G{exp ·%}G}G c {exp .1t }G' and again from Theorem 2 and Corollary 1 we see that {ad~ %} A C .tI'. Finally, notice that if {ad~~}A c .Yr and {adZ~}A C .¥f', then {ad.v(.At~ + ~)}A C .tfJ. Thus there is a largest Lie algebra with this property. 352
4. ControUabiUty on group lDanifoids. The first question of a system theoretic character which we investigate is that of controllability. Since we want to emphasize global results, we work with the most elementary type of evolution equation appropriate to our present setting, namely, X(t)
= (A + LuJt)Bi)X(I).
The choice of control affects the direction in which X moves. However, A is a constant over which there is no control. This evolution equation has the property that the change of variable X -+ XP for P nonsingular leaves the equation unchanged. This invariance gives the vector field which a given choice of {u.~t)} establishes on ~t(n), a particularly simple form. THEOREM 5. Consider the linear dynamical system X(t) ==
(Jl
X = an n x n matrix.
U,{t)B;)X(t).
Given a time t. > 0 and giventwo nonsingular matrices X 1 and X 2' there exist piecewise continuouscontrols whichsteer the state from X 1 at t = 0 to X 2 at t = t. if and only if X 2 X l 1 belongs to {exp {Bi } A}G. Proof Sufficiency. Theorem 1 asserts that any matrix M in {exp {Bi } A}G can be written as a finite product.. say M = eBil«l eB·za. z ••• eB'm2m.
Suppose X 2X 1 1 = M. Divide the interval 0 ~ t ~ t" up into m equal intervals [tit t i + 1) whereby t i = i- tJm. Let tJm = p-l. On the interval [0.. t 1 ) all controls are zero except the imth control, which takes on the value a."./J. On the interval [t 1 ,t 2 ) all controls are zero except the i"'-1 th, which takes on the value cxm - 1P, etc., down to the last interval on which all controls are zero except the i 1st, which takes on the value a. 1 /J. Since the differential equation is linear and constant on each of the subintervals, the solution is a product of exponentials and the result follows. Necessity.' To show that X 2 cannot be reached from X 1 unlessX 2X~ 1 isof the form eLm eLm - 1 • • • eLI weassume the contrary and obtain a contradiction. Suppose that u 1( • ) .. • •• , Um( .. ) is a control which steers the system from X 1 at 1 == 0 to X 2 at t = t".. By Theorem I of [9] we know that there exists a sequence of times to, t 1 , t 2 , ••• , t m such that on each of the subintervals [ti' t., 1] the transition matrix of X(t)
can be written as
eRi(l)
=
for some H i( X(t a )
ttl
•)
U;(t)B;)X(t)
in .fR. Thus we can write
= eLm eLm -
I
•••
eLI X o.
which establishes the contradiction. As an application which emphasizes the ease with which we can study global questions using this theorem we observe the following results relating to the classical groups. Here J is given by J =
[_~ ~l
and a matrix is called symplectic if (J'JO
= J.
353
THEOREM 6. Consider the system oj' Theorem 5. Given a time ttl > 0 and given two nonsingular n x n matrices X I and X 2 with del X IX 2 > 0, there exists a piecewise continuous control which steers the state from X 1 at t = 0 to X 2 at t=t"if{Bi}A: (i) spans A" X", (ii) spans the (n2 - I)-dimensional subspace of Rft )( ft consisting of the zero trace matrices and det X 1 = det X 2' (iii) spans the (n(n + 1)/2)-dimensional subset of Rft x n consisting of the set oj' matrices which satisfy J A + A'j = 0 and X 2X 11 is symplectic, (iv) spans the (n(n - 1)/2)-dimensional subset of R")(" consisting of all skew symmetric matrices, and X 2X 11 is orthogonal. Proof As is well known, any nonsingular matrix can be written as OR with 8'0 = I and R = R' > O. Also, real orthogonal matrices with positive determinants and real symmetric positive definite matrices have real logarithms. Moreover.. in case (iii) the factors in the polar representation inherit the property of the group itself which is to say that the 0 and R in the polar representation of a symplectic
matrix are symplectic. To complete the proof we need only invoke Theorem 5 since the previous remarks justify our writing X 2X - 1 = eO eS with Q = - Q' and S = S' both in the appropriate Lie algebras. The results of Theorems 5 and 6 are somewhat unsatisfactory in that the A term is absent. The following theorem describes one way in which this can be relaxed. THEOREM
7. Consider the linear dynamical system
X(l)
= (A +
itl U~l)Bi)
X(t),
x = an n x
n matrix,
Suppose that [ad~ Bi , Bj ] = 0 for i. j = 1,2, ... , v and k = 0,1, , n 2 - 1. Let ~' be the linear subspace of IR" )( " spanned by ad~ B, for; = 1, 2, , \t and k = 0, 2 1, ... , n - 1. Then given a time t, > 0 and two n x n matrices X 1 and X 2 there exist continuous controls which steer the system from the state X 1 at t = 0 to the state X 2 at t = t. if and only if there exists H in J{J such that
Proof First of all, notice that
I
d: [eAIB i e- AI• Bj ] dt ,=0
= ddkk~ 11 [eAI[A. Bi ] e- At.Bj]1 t
'=0
d"-2 I = drk- 2 [eAI(ad~ Bj ) e- AI. Bj ] I: 0 = [eAr(ad~ Bi ) e- A t , Bj]l, = 0
= [ad~ Bi , B j ] . Thus [eA'Bie-A1,B j ] is identically zero if[ad~Bi,Bj]=O for k=O,I,2~···. However, ad, is a linear operator from an n 2-dimensional space into itself so that by the Cayley-Hamilton theorem all powers above n2 - 1 are lineary dependent 354
on the first n 2 - 1. Thus under the hypothesis of the theorem statement [eAt B, e- At, Hj ] vanishes identically. AIso Ar 0= eAIB.e-A1B. - B.eAIB.eI } J I
Now let t
+ (1 = fJ and)' = o, Thus for all fJ and "I, o = [e.4 11 e, «:". eAY e, «: A)I].
For the purpose of solving the differential equation we introduce Z(t)
= e-AtX(t) and observe that t(t) =
(.±•=
e, e'1I) Z(t) .
Ui(t) e - .11
1
But recall (see e.g. Martin [15]) that the solution of 2(t) = B(t)Z(t) is exp f~ B(u)d(j (I. Thus we can write
if [B(t) , B(u)] vanishes for all t and Z(t)
= exp ( ('
t
)0.=1
Ui{t) e-AIB i e·4 1
dt) 2(0).
It is a well-known and frequently used fact (e.g., [16, p. 79]) that the image space of the map taking continuous functions into IR P according to the rule x = L(u) = J~ eA«rbu(a) da, is spanned by the first p derivatives of eAtb evaluated at zero. Using this fact here we see that for each H in ~tf' and t a > 0 we have a continuous u defined on [0, ta] such that
Therefore in terms of X we see that we can reach at t a any X which can be expressed as eA ta eHx(O) with H in .1f'. As an application of this result we derive a familiar relationship. Example 4. Consider the system in IR". m
.x(t)
= Ax(t) +
L biuj(t):
x(O) is given.
i=1
Related to this is the matrix system in R(n + 1) '( 'n+ 1):
· == [.4o
X(i)
OJ X(t) + L Lli(t)[0 b0'J X(t). 0 m
0
1
i= 1
Let A and B, be the matrices appearing in this expression. In this case vanishes as required and so the reachable set from x(O) = I is
9t{t}
= exp [~t
~l {exp H; H E ,ff},
where .ff is the subspace spanned by ad~ Bi. A computation gives
[0
k AUk,b i,] Ad4B·:= .~ _0 I
355
[ad~4
Bi , B j ]
so that the reachable set at t is
~lHErange(B,AB,
[e;r
.rJf "'" {X:X =
where we have used the fact that
E!At,i"
~ ..K' for
' .. ,
all t and
.~.
An-IB)}, = range (B, AB,
...
'I
A"-lB).
5. Observability. In order to get a theory having a scope comparable to linear theory, it is necessary to treat observability. The choice of an appropriate form of the observational equation is critical for the success of the overall theory. As it turns out, the natural choice is indicated by the second example in § 2. Let fF be a matrix group and let ~ be a subgroup. Consider the system evolving in fF, X(t)
=
JI
(A + U~t)Bi)X(t),
y(t) == ~X(t)~
by which we mean that instead of observing X(t) directly, we observe to what 'equivalence class X(t) belongs with respect to the equivalence relation in §" defined by ({;. Thus y(t) takes on values in the coset space fF ICC which is generally not a group manifold (see § 7). We call two states X 1 and X 2 distinguishable if there exists some control which givesrise to different outputs for the two starting states. In general the zero control is not adequate to distinguish between all states which are distinguishable as contrasted with the situation one finds for linear systems. THEOREM 8. Let l€ be a matrix group and suppose that the set of points reachable .from the identity for the system
X(t) =
(A +
itl
U;(t)Bi ) X(t),
yet) =
.~
~'X(t).
is a group. Then the set a/initial states which are indistinguishable from the identity is given by {!jJ {!J
= {P: RPR - 1 E ce for a II R E .~} .
is a normal subgroup of ~91 and a subgroup of rt'. Proof Suppose that X is a starting state for the given equation which is
indistinguishable from the identity. That means that for each R in .fJI there is C(R) in «;' such that C(R)RX = R.
Since ~ and
C6J
are groups, we can take inverses to get
RXR
'-1
EC6'.
Thus the set [j' is exactly those states indistinguishable from the identity. The remainder of the conclusions come from Theorem 3. THEOREM 9. Let Jf and !R be Lie algebras in RtJ x '1. and suppose that all the
points reachable from the identity for X(t) =
(A +
it
UAt)Bi)X(t),
l
356
are {exp ..5f}G. Then the set of initial states /}J which are indistinguishable from the identity contains {exp %}G if' and only if {ad~ ,~·}.,t c -'II. Therefore a necessary condition jor all states to be distinguishable ,/;·011' the identity is thai .11 contain no subalgebraa: such that {adIr ·%}A C .ff. Proof Theorem 8 gives a characterization of :!I which permits one to bring
to bear Theorem 4. Theorem 4 immediately gives the desired result. One might be tempted to conclude that if there is no nontrivial algebra .~ meeting the requirements of Theorem 9, then all initial states are distinguishable. This is not true because &J can be a discrete subgroup and hence not trivial and yet not expressible as {exp .f('}G for any Lie algebra -or. The next example illustrates this. Example 5. In the numerical integration of the equations of motion of a rigid body one usually avoids Euler angle representation and uses instead quaternion or direction cosine representations. As is well known, the group of unit quaternions covers ..<1'(~)(3) twice. This causes an ambiguity in going from ,y:'((!{3) to the group of unit quaternions. This example illustrates this idea. Consider an equation in the group of unit quaternions .::2 which we parametrize in the usual way (a:! + h1 + c2 + d2 = 1): 0
II
-b
-c
-£I
d b
a
-£I
C
-lil
cit c
d
a
-h
-it:!
d
-('
h
II
--
-
-~
===
(6'
Ll:.
u]
1I
--h
-c
-d
0
llJ
-1l1
h
a
-d
('
UJ
0
II I
C
cI
a
--17
0
d
('
h
£I
+li 2
Ll_~
a
y(t)
Lit
·-h
--lit
-- ('
-
•.
_
····d
h
a
--.£1
c
d
1I
--/1
d
-('
h
a
('
where C6" is the subgroup given by 0
"J.
()
0
O
0
()
0
0
0
1.
()
0
-Y-
O
-Y-
(6
:exp.Klc;,
-It ==
Now it is true that {exp '#"}G includes I and - J~ and it is also true that this pair of elements forms a normal subgroup of lL. Thus I as an initial state cannot be distinguished from -I. Yet there is no nontrivial Lie algebra ,If" such that {ad~ ..t'~ L.t c
.e.
6. Realization theory. One of the central results in linear system theory is the fact that any two time-invariant, controllable and observable realizations of a given time-invariant input-output map are related to each other in a very simple way. Our purpose here is to establish a similar theorem in this context. 357
Suppose we have two systems X(l)
Z(t)
ttl = ttl =
U1{I)Bi)X(t),
y(t)
U;(t)G i)Z(t),
= .lrZ(t).
We assume that (i) the systems are observable in the sense that no two initial states give rise to the same response y for all piecewise continuous inputs, and (ii) there exist one-to-one maps, say c( . ) and h( . ), both mapping into a set S such that if each system starts at the identity state and if each system receives the same input, c(C€ X(t)) = h(.*,Z(t» for all future time. A pair of systems meeting these criteria will be said to be observable realizations of the same input-output map. We emphasize that X(t) and Z(t) are square matrices but not necessarily of the same dimension. Suppose we have two observable realizations of the same input-output map. Let u( · ) be a nonzero piecewise constant control defined on [0. 1] which when applied to the X system takes the state X(O) = 1 into the state X(l) = I. Then of course it must do the same for the Z system because they are observable realizations of the same input-output map. Thus we see that if iI' ;2' ... ., iq is a collection of integers with 1 ~ ilc ~ m and if (Xi are any real numbers such that
then
Let L 1 , L 2 , · · · ., L, be a set of commutator expressions in B 1 , B2 , · · · ~ Bm such that {Li } forms a basis for {B;} A. Let K 1 ~ K 2., ..• , K, be in an analogous expression obtained by replacing B 1 by G1 • B 2 by G2 , etc. Let S be an arbitrary commutator expression in B 1 , B 2 ., ..... Bm and let Tbe the analogous commutator expression in G.,G 2 , ••• , Gm • Then in the notation of the proof of Theorem I, there exist differentiable functions ex;(p) such that for Ipl small,
and
Since the exi are differentiable we can write (prime denotes derivative) 1ft
I
+P L
cxi(O)L i + O(p2) = I + pS + O(p2)
i==J
and m
I
+P
L ~~(O)Ki + O(p2) = i= J
358
I
+ pT +
O(p2).
Thus if r
s= I
l~iIJi'
i···1
then
,.
L ";'iKi
T=
o
i:;-: 1
From this we see that the algebra {G;} A is generated from {Bi } in exactly the same way as the algebra {Bi } A is generated from {Hi} and thus that the algebras are isomorphic. We summarize this discussion with a theorem. THEOREM to. Consider the two systems X(t)
= t~ U1{l)BiX(t)) ·
ZIt)
=
ttl u~t)GiZ(t)).
y(t) = (t X(t).
y(t)
= -:yt'Z(t),
where X and Z are n x nand q x q respectively. Suppose that these systems are observable realizations of the same input-output map. Then {BiL.. and {GdA are isomorphic as Lie algebras, and moreover lf L 1 , L 1 , ... , L, are commutator expressions in {B i } which form a basis for {Bi}A and if K 1 , K 2 • . . . , K, are the analogous expressions in G; obtained by replacing B, by G;, then K l ' K], . . . . . K,. is a basts"fior fG l ifl .4' and ~jO [L i , L j] =
I
lijkLk"
k--=l
then
[K j . K j ] =
I
'jj"K".
k:;;. 1
Of course this does not mean that the reachable sets from I. namely {exp {BiLot}G and {exp {G;} A}G' are isomorphic as groups. For example, the group of unit quaternions and the group of 3 x 3 orthogonal matrices have isomorphic Lie algebras, yet they are not isomorphic as groups. 7. System theory on coset spaces.. In this section we reinterpret our results in a somewhat different way. This interpretation leads to some facts about systems on
manifolds which do not admit a group structure. In particular we have in mind the n-sphere S" = {x: x'x = 1, x E IR n + 1} which, as is well known. does not admit a Lie group structure except for the cases n = 1 and 3. Let M c IR" be a manifold. Let C§ be a matrix group in IR")( ". We say that f§ acts on M if for every x E M and every G E~, Gx belongs to M. By the orbit of ~~ through x we mean the set of points C§X = {y: y = Gx, G E f§}. We say that f§ acts transitively on M if it acts on M and if for every pair of points x, y in M ~ there exists G in f§ such that Gx = y. If f§ acts transitively on M, then at any point x EM there will be a subset .ft);. c ~~ such that for each H E ,~, II x == x, Clearly if fl l E .11.'( and H 2 E.tfx ' then H 1H2 x = H,» = x and H-1x = x so that .ll~ is a subgroup. 359
We call .YI~ the isotropy group at x. Notice that if Gx = y.. then y = G·.f{xx = G.YI."G - 1 y, and thus G.tt."G -·1 is the isotropy group at y-all isotropy groups arc conjugate in t§. Now suppose M is a manifold for which there actualJy exists a group f§ acting transitively. Pick a point x E M. Define in f§ an equivalence relation whereby G1 ~ G2 ifand only ifG 1 = G2 f/:c forsome H, E .f{x' There isa one-to-one correspondence between this space of equivalence classes, q}/..I(y;, and .M. In this case we call M a coset space. We study systems in which the state is represented as an It-vector and the evolution is governed by (1)
x(t)
= (A +
Jl
J~t) = ~x(t).
Ui(t)Bi) x(t),
By ~'x(t) we mean an equivalence class of vectors, Xl being equivalent to X2 if and only if exI = X2 for some C in ~;. Let!f be a Lie algebra generated by {A, Bi } and let MeR" be a manifold such that {exp ~}G acts on it. Then the above equation can be thought of as evolving on the manifold M c [Rn, for jf x(O) EM, then regardless of the control.. x(t) E M for all t > o. If there exists a differentiable manifold M c IR" such that {exp ~}G acts on M, then we shall say that (1) is well..posed on M. Example 6. Consider the n-sphere, S". Let B I ' B 2 .. . . . .. Bm be (11 + )J x (n + 1) skew symmetric matrices. Clearly, the system i(t) =
Ltt
U.(t)Bi]X(t),
y(t)
=
et'x(t) .
is well-posed on S"since {exp ~}G consists of orthogonal matrices, and orthogonal transformations preserve norm. If we can observe only the first component of x, then we should let ({j be the subsets of .~(!i(n + 1)consisting of those matrices which have a 1 in the first column and first row. That is,
~. [~ Y·;'(1l)]. =
With respect to controllability we can say given any two vectors x I and x 2 in S" there exists a piecewise continuous control which steers the system from x 1 to X2 if and only if X 2 = RX1 for some R in {exp '?}G' where !f! is the Lie algebra generated by {Ai}. Also, an arbitrary point can be transferred to an arbitrary point if and only if {exp .!t'} G acts transitively on S". At the same time we might observe that any X o such that IIxoll = 1 can be transferred to any x 1 such that II x III = 1 if and only if {exp !t'} G acts transitively on S". This second point of view is useful because it puts the problem of controllability
on S" in contact with standard results in geometry. In particular a great deal is known about Lie groups which act transivitely on S". (See Samelson [17, p. 26].) As for observability, we note that two initial states x 1 and X 2 in S" give rise to the same y if and only if for all R in {exp ~}G there exists C(R) in f.6 such that RX 1 = C(R)Rx2' which is to say that R- 1C(R)Rx 2 = x.: We now abstract from this example the essential features and state formally a result which summarizes the development. 360
THEOREM
11. Consider the dynamical system (x(t) x(t)
=
ttl
U;(tIB;)Xlt J•
.I'll)
E !R"~
= [exp ,11 }c;x(C).
which is well-posed Oil the manifold A4 c ~n. Let ~ be the Lie algebra generated hy -{ B i } . A given state X2 is reachable from .x 1 ~l and Dilly if Xl = Nx 1 for .some N in {exp~} (i. Let ~'!/J = -( P: RPR - 1 E -{ exp.Jr.'} c; for all R E exp 2'):'. T~~'() states Xl and X2 are indistinguishable ij' and only ~l X2 = PX1 for some P in .:1'. III particular. two states XI and .'<2 are indistinguishable !I' x.2 ~ Px 1 for P ill ~ exp .if' : t: \\'ir!l;r' being lilly Lie algebra such that :ad YJ .k";' . . c 11: Example 7. Consider the submanifold tvl of ~n+ 1 consisting of those points whose last coordinate is 1. The evolution equation in ;\-1.. y(t)
0 ker CJ·)[xt('] ( [0 01
= exp
corresponds to the more familiar x(t) = Ax(t) + Bu(t) .. y(t) Theorem 7 we see that for the associated group equation ~¥(t)
=
A 0J[o 0 +
L m
lI i(t
)
;.-1
[0 0
= ('x(t).
Using
b'J X(t) I
0
the reachable set at time t consists of those matrices which can be written as ,11, =
{x: X
=
[e~t
';l
xErangeIB,AB.···, AII'IBJ},
Thus if B. AS.. . . . . A"- 1 B spans (R".. then the reachable group acts transitively on M and we have controllability. As for observability, we note that
ex
(0
p 0
kerC") ---: 0
(I
()
The subalgebras of .j{' which are closed under commutation with !/) correspond to the linear subspaces of ker C-' which are invariant under A.
Acknowledgment. Among the many people who made useful suggestions on earlier expositions I want to mention in particular D. Elliott V. Jurdjcvic.. H. Sussmann., Jacques Willems and Jan Willems. I also want to thank H. Rosenbrock
who some years ago acquainted me with work relating to differential equations and Lie algebras. REFERENCES [I] W. L, CHOW.. llber System ron linearen partiellen Differentialgleichungen erster Ordnung, Math. Ann.. 117(1939), pp. 98-105. [2] R. HERMANN, 0" the accessibility problem in control 11,,-'or.1'. Proc, International Symposium on Nonlinear Differential Equations and Nonlinear Mechanics, Academic Press. New York. 1963. pp. 325-332. [3] J. Kt:(,~RA, S01U1ioll ill large control problem: ,x .: (A( I - II) + But», Czech. Math. J. 16 (1966). no. 91. pp. 600-623.
or
361
Solution in large 0/ control problem: :i = (Au + BI!)X, Czech. Math. J.• 17 (967). no. 92, pp. 91-96. (5) C. LoaRY, Conlro/abilile dessystemesnunlinearies, this Journal, 8 (l970), pp. 573-605.
[4J J.
KUCERA,
[6] G. W. HAYNES AND H. HERMES. Nonlinear control/ability VUl Lie theory. this Journal. 8 (1970). pp. 450-460. [7] W. MAGNUS, On the exponential solution of differential equationsfor Q linear operator, Comm. Pure Appl. Math., 7 (1954). pp. 649-673.
(8] K. T. CHEN, Decomposition ofdifferentialequations. Math. Ann .. 146(1962). pp. 263 ·278. [9] J. WEI AND E. NORMAN, On global representations of the solutions of linear differentialequations as Q product of exponentials. Proc. Amer. Math. Soc .. 15 (1964). pp. 327-334. [10] - - , Lie algebraic Solul;on 0/ linear differentiDl equations, J. Mathematical Phys.• 4 (1963). pp.575··581. [11] I. M. SINGER AND J. A. THORPE. Lecture Notes on Elementary Topologv and Geometry, Scott.
Foresman and Co., Glenview, III., 1967. (12] S. STERNBERG, Lectures on Differential Geometry, Prentice-Hall. Englewood Cliffs. N.J.. 1964. (13] L. AUSLANDER AND R. MACKENZI~ Introduction to Differentiable MQn~fo/d~. McGraw-Hill. New York, 1963.
[14] J. J."ROTMANN, The Theory o/Groups, Allyn and Bacon, Boston, 1965. (IS] 1. F. P. MARTIN, Some results on matrices which commute with their deriratires, SIAM J. Appl. Math., IS (1967), pp. 1171-1183. [16] R. W. BROCKETT Finite Dimensional Linear Systems, John Wiley, New York. 1970. [17] H. SAMELSON, Topologyo[Lie groups, Bull. Amer. Math. Soc .. S8 (1952), pp. 2-37. (18] H. J. SUSSMANN AND V. JURDJEVIC.. Controllability ofnonlinear systems, J. DifferentialEquations. t
to appear.
[19] V. JURDJEVIC AND H. J. SUSSMANN. Control systems on Lie groups, J. Differential Equations. to appear. [20] D. L. ELLIOTT, A consequence ofcontrollabiiity, J. Differential Equations. 10 (1971). pp. 364-370.
362
Controllability of Nonlinear Systems HECTOR J. SUSSMANN AND VELIMIR JURDJEVIC
IN
the late 1950s and early 1960s, the introduction of statespace models in the analysis and design of linear control systems added enormous thrust to the development of systematic methods for the solution of major control problems such as stabilization by means of feedback, asymptotic regulation with minimal cost, disturbance rejection, and so on. Instrumental, in this expansion, were the notions of reachability and observability, introduced by Kalman to characterize interaction between input and state, on one hand, and state and output, on the other. An important feature of linear state-space methods was the possibility of obtaining information, from simple linear-algebraic tests on the model parameters, about either the set of all states that can be reached from a given one (using appropriate choices of inputs) or the set of all states that cannot be distinguished from a given one based on the observed output. Indeed, in view of the impact that these concepts had in the design of linear control systems, the problem of deriving similar characterizations for "nonlinear" systems soon became the focus of intense research effort. It must be stressed, in this respect, that the term "nonlinear" encompasses not just the consideration of phenomena such as saturation or dead-zone, whose influence on the stability and performance of an otherwise linear system was relatively well understood, but especially the kind of analytic nonlinearities found in the equations describing the motion of a rigid body, of a multi-link robot, of an electrical drive, and so on (typically, products between state variables or between input and state variables). Pioneering authors in this domain have been Hermann [7], Haynes and Hermes [6], Lobry [11], and Elliott [4], who realized that the appropriate language in which the study of the structural properties of a nonlinear system had to be cast was that of the differential geometry. Fundamental in this respect were the observations-based on an earlier result of Chow [3]-that the most natural geometric object into which the set of states reachable from a given point can be embedded is a differentiable manifold (i.e., a "surface," and not a subspace as in the case of linear systems) and the fact that this manifold can be "infinitesimally" characterized in terms of repeated derivatives, at the point of interest, of the vector fields that define the "admissible
velocities" at this point (namely, the velocity vectors that can be imposed at a given point of the state state by setting the control input). More specifically, the theorem of Chow (whose derivation predates by more than 20 years the birth of state-space methods in control theory, but which can easily be cast into its language) considers systems described by differential equations of the form oX
= II(X)UI
+ 12(X)U2 + ... + Im(x)u m
(1)
in which x E jRn and 11(X), ... , Im(x) are analytic complete! vector fields. Each of the "controls" U 1, ... , U m can take values in the set Q = {-I, 0, I}. Then, the theorem asserts that the set of states reachable from a given XO is a submanifold M of jRn, precisely the unique connected submanifold which contains XO and has, at any of its points, a tangent space spanned by the vectors of the set F = {II (x), ... , Im(x)} and by all vectors that can be expressed as (repeated) Lie brackets of vectors in F. 2 In other words, the vectors of F and their repeated Lie brackets provide a complete characterization of the set M of states reachable from xo, by determining the tangent space to M at any of its points. In the context of control theory, the main conceptual content of Chow's theorem is that the (repeated) Lie brackets of the vector fields 11(x), ... , Im(x) define the set of "directions" in which the state vector can be (infinitesimally) pushed. In the case of a linear system oX
= Ax
+ blUl + ... + bmum
this idea specializes to the familiar observation that the set of states reachable from x = 0 is not just the span of the vectors b l , ... , bm , but rather the span of all the vectors of the form Akbi • 1A vector field is complete if for any x ° the integral curve of.t = Ii(x) passing through XO at time t = 0 is maximally defined over (-00, +(0).
2We recall that the Lie bracket of two vector fields [(x), g(x), denoted [[, g](x), is the vector field
363
ag
[f, g](x) = -
a[
ax f(x) - -ax g(x).
Unfortunately, however, a basic hypothesis in Chow's result is that each of the inputs u 1, ... , U m take values in the set Q = {-I, 0, I}, which corresponds-in the control-theoretic framework-to the possibility of reversing the sign of the corresponding component of the velocity vector (or, equivalently, of reversing the sign of the flow of time) and/or turning-off every single component. Thus, this result cannot provide a satisfactory answer in the case of systems modeled by equations of the form
x=
fo(x)
+ fl(X)Ul + f2(X)U2 + ... + fm(x)u m
(2)
which, despite the formal similarity with (1), suffer from the basic limitation of having a component, namely fo(x), of the velocity vector (commonly referred to as the "drift term") which cannot be controlled. Many systems of practical interest in control applications fall into the class of systems modeled by equations of the form (2). A complete and satisfactory answer to a number of problems arising in the analysis of the geometric structure of the sets of reachable states for a control system of the form (2) (actually, for the more general class of systems modeled by equations of the form
x = F(x, u)
(3)
with F (x, 0) =1= 0) was provided by Sussmann and Jurdjevic in their 1972 paper. This paper can in full right be considered as the paper that laid the foundations of geometric nonlinear systems theory in a comprehensive and fully mathematically rigorous fashion. Sussmann and Jurdjevic have observed that, in the analysis of a system of the form (3), it is convenient to look separately at two different concepts of reachability: reachability with fixed end-time and reachability with free end-time. The former is used to define the set A(xO, T) of all states attainable from XO at time T (i.e., the set of all states x for which there exists some input, defined on the fixed interval [0, T], such that the integral curve x(t) of (3) satisfies x(O) = XO and x(T) = x), while the latter is used to define the set A(x) of all states attainable from XO (i.e., the set of all states x for which there exists some time interval [0, T] and some input, defined on the interval [0, T], such that the integral curve x(t) of (3) satisfies x(O) = XO and x(T) = x). The difference between these two concepts is present of course also in the case of linear systems, but nobody had pointed it out before. The reason was mainly because in the study of linear systems the initial state from which reachability is considered is the equilibrium state XO = 0, and in this case the two concepts coincide. But, if XO i= 0, it is readily seen that A(xO, T) = exp(At)xO
shown that A(xO) is contained in a manifold M for which a tangent space is determined, at each point, by the vector fields of F = {F (x, u) : U E Q} and their repeated Lie brackets, and that A(xO) has a nonempty interior in the topology of M (i.e., contains an open subset of M). In other words, A(xO) may not coincide with the whole of M (as it was in the particular situation addressed by Chow's theorem), but M has the least dimension among all manifolds which contain A(xO). Thus, M can be regarded, in a certain sense, as the "most appropriate" manifold into which the reachable set A(xO) can be embedded. Likewise, they prove that A(xO, T) is contained in a submanifold M r of M, of codimension at most 1 in M, and that A(xO) has a nonempty interior in the topology of MT. Again, the tangent space to M r is determined, at each point, by certain vector fields of F and their repeated Lie brackets. The concepts, the formalism, and the results contained in this paper soon became a standard reference for all authors working in that broad area of research which became known as "nonlinear control theory". The hypothesis of analyticity of the vector fields was soon removed by Sussmann himself [12] and an alternative proof of the property that A(xO) has a nonempty interior in M, due to Krener [10], also became available. The results were the point of departure for the work of Sussmann [12] on the theory of minimal realizations for nonlinear systems and the work of Hermann and Krener [8] (also included in this volume) on the theory of nonlinear observability, which these authors were able to develop in a dual-like fashion. The concept of Lie bracket played also an instrumental role in the work of Fliess on the characterization of the dimension of a minimal realization of a given nonlinear input-output map, through the concept of Lie rank of a Hankel matrix, introduced in [5]. The principles laid down in the 1972 paper by Sussman and Jurdjevic were instrumental in addressing many problems of "local" controllability and optimal control, a field of research that has been very active during the 1980s and the 1990s (see, e.g., [1], [14], and [15]). Finally, it is worth stressing that a specific outcome of the theory of nonlinear controllability developed in this paper was the possibility of characterizing (local) decompositions of nonlinear systems into "controllable/uncontrollable" parts, very similar in nature to the decompositions developed some 10 years earlier by Kalman (see the earlier preamble to [9], also included in this volume). The availability of these decompositions was one of the main thrusts for the development of systematic methods for solutions to problems of disturbance decoupling and noninteracting control, which occurred in the late 1970s and early 1980s (see the preamble to [8], later in this volume).
+R
where R is the (standard) set of states reachable from the origin, while A(xO) =
U
A(xO, T)
O~T
In other words, the family of sets A(xO, T), as T ranges over [0, (0), partitions the set A(xO). Assuming that for each fixed U E Q, with Q a fixed set, the vector field F(x, u) is analytic, Sussmann and Jurdjevic have
364
REFERENCES [1] J. BAILLIEUL, "Controllability and observabilityof polynomial systems," Nonlin. Anal., 5:543-552, 1981.
[2] R. W. BROCKETT, "System theory on group manifolds and coset spaces," SIAM J. Contr., 10:265-284,1972. [3] W. L. CHOW, "Uber Systemevon LinearenPartiellenDifferentialgleichungen Ester Ordnung," Math. Ann., 117:98-105,1938. [4] D. L. ELLIOTT, "A consequenceof controllability," J. Diff. Eqs., 10:364370,1970. [5] M. RIESS,"Realisationlocale des systemesnon lineaires,algebresde Lie
filtrees transitives et series generatricesnon commutatives," Invent. Math., 71:521-537, 1983. [6] G. w. HAYNES AND H. HERMES, "Nonlinearcontrollability via Lie theory," SIAM J. Contr., 8:450-460,1970. [7] R. HERMANN, "On the accessibilityproblemin controltheory," in Int.Symp. on NonlinearDifferential Equationsand NonlinearMechanics, Academic Press,pp.325-332, 1963. [8] R. HERMANN AND A. 1. KRENER, "Nonlinear controllability and observability," IEEE Trans. Aut. Contr., AC-22:728-740, 1977. [9] R.E. KALMAN, "Mathematical description of linear dynamical systems," SIAM J. Control, 1:152-192, 1963. [10] AJ. KRENER, "A generalization of Chow's theorem and the bang-
bang theorem to nonlinear control systems," SIAM J. Contr., 12:43-52, 1974. [11] C. LOBRY, "Controlabilitedes systemesnon lineaires,"SIAM J. Contr., 8: 573-605, 1970. [12] H. SUSSMANN, "Orbits of families of vector fields and integrability of distributions," Trans. Am. Math. Soc., 180:171-188,1973. [13] H. SUSSMANN, Existence and uniquenessof minimal realizationsof nonlinear system, Math. Syst. Theory, 10:263-284,1977. [14] H. SUSSMANN, Lie bracketsand local controllability, SIAM J. Contr. Optimiz., 21:686-713, 1983. [15] H. SUSSMANN, "On a general theorem on local controllability," SIAM J. Contr. Optimiz., 25:158-194,1987.
A.I.
365
Controllability of Nonlinear Systems* HECTOR
J.
SUSSMANN
Department 01Mathematics, University of Chicago, Chicago, Illinois 60637 AND VBLIMIR ]URDJEVIC
Departmmt01Mathematics, University of Toronto, Toronto, Canada Received April 26, 1971
1. INTRODUCTION In this article we study the controllability of nonlinear systems of the form
hldt
= F(x, u).
Our objective is to establish criteria in terms of F and its derivatives at a point x which will give qualitative information about the sets attainable from s. The studyis based primarily on the work of Chow [4] and Lobry [16], although it is similar in its approach to works by other authors in that it makes systematic use of differential geometry (for instance, see Hermann [8,9], Haynes and Hermes [6], Brockett [2], etc.), The state variable x is assumed to take values in an arbitrary real, analytic manifold M, rather than in R". We chose this generalization because it creates no essential new difficulties while, on the other hand, it allows for certain applications which are not commonly treated in control theory. For instance, when M is a Lie group, then the present results can be specialized to obtain more detailed controllability criteria. Control problems on Lie groups were first considered by Brockett in [2], and will be treated in a forthcoming paper by the authors. Most of the recent studies on controllability of nonlinear systems have essentially dealt with symmetric systems, i.e., systems of the form (*) with
* This work was performed while the authors were at Harvard University, Division of Engineering and Applied Physidl, Cambridge, Mass. The first author was supported by the U. S. Office of Naval Research under the Joint Electronics Program by Contract NOOOl4-67-A-0298-0006. The second author was supported by the National Aeronautics and Space Administration under ·Grant NGR 22-007-172. Reprinted with permission from Journal ofDifferential Equations, Hector 1.Sussmann and Velimir Jurdjevic, "Controllability of Nonlinear Systems" Vol. 12, 1972, pp. 95-116. 367
the property that F(x, -u) =::: -F(x, u) (Hermann [9], Haynes and Hermes [6], Lobry [16]). As remarked by Lobry in [16], the consideration of symmetric systems often excludes interesting situations arising from mechanics. In these cases the system is of the form
dxldt == A(x)
+ H(x) · u.
A notable exception is the work by Lobry [17]. Lobry stated (and proved for the case of two vector fields in R3) the result for nonsymmetric systems that appears here as Theorem 3.1. Our results apply to nonsymmetric systems. \Ve obtain some general information about the geometric structure of the attainable sets showing that they "practically"are submanifolds (see Theorems 4.4 and 4.5 for the precise statements). This information yields a complete answer to the problem of deciding when the sets attainable from a point x have a nonempty interior. The criteria obtained involve purely algebraic manipulations of F and its derivatives (of all orders) at the point x (see the Remark below). In particular, our results contain those of Kucera [14]. In this connection we observe that our proofs are of interest even for the case treated by Kucera (see Sussmann [21]). We have omitted the consideration of nonautonomous systems; they can be treated analogously by the familiar procedure of reduction to an autonomous system (i.'e., 'by considering the state variable to be defined in M X R). The 'erganization of the article is as follows: in Section 2 we introduce notations and basic concepts; in addition, we quote some well-known basic results which will be used later. In Section 3 we prove our main results in difFerentia~ geometric terminology. In Section 4 we apply these results to control systems. We derive the algebraic 'criteria mentioned above (Corollaries 4.6 and 4.7), and we prove two "global results": 'we show that, for a large class of manifolds, accessibility (i.e., the property that, for any given x, the set of points attainable from x has a nonempty interior) implies strong accessibility (i.e., that for any given x and 'any given fixed positive t, the set of points attainable from x at time t has a nonempty interior). We also show that, for a still larger class, including the Euclidean spaces, controllability implies strong accessibility. Finally, Section 5 contains examples. We show how our results can be used to derive the classical controllability criteria for the system dxjdt
== Ax + Bu.
We also derive the results of Kucera .and indicate some generalizations.
Remark. An assumption that is made throughout the article is that F is an analytic function of x. This guarantees that all the information about the 368
system is actually contained in F and its derivatives (of all orders) at a given point x. The analyticity assumption cannot be relaxed without destroying the theory (cf. Example 5.3). Another assumption that we make is that the trajectories of the system are everywhere defined. As opposed to the previous one, this assumption is not essential (except for the "global" Theorems 4.9 and 4.10). We use it, however, because it considerably simplifies all the proofs.
2.
PRELIMINARIES
We shall assume that the reader is familiar with the fundamental notions
of differential geometry. All the definitions and basic concepts utilized in this paper can be found in standard books, (for instance [1, 3, 7, 13 and 19]). The following notations will be used throughout:
R the set of real numbers. R"
n-dimensional Euclidean space.
M~
the tangent space to the manifold M at the point x.
TM the tangent bundle of the manifold M. V(M) the set of all analytic vector fields on the analytic manifold M. We will regard V(M) as a Lie algebra over the reals. For any X and Y in V(M), we will denote the Lie product by [X, Y] (Le., [X, Y] = XY - YX). All the manifolds will be assumed to be paracompact. Recall that a submanifold of a paracompact manifold is paracompact. Also, a connected paracompact manifold is a countable union of compact sets. These facts imply (cf. Lobry [16, p. 589]): LEMMA 2.1. Let M be a (paracompact) manifold of dimension n. Let S be a k-dimensional connected suhmanifold of M. If k < n, then the setof points of S has an empty interior in M.
A subset D of V(M) will be called infJo[utive if, whenever X and Y belong to D, then [X, Y] also belongs to D. A subalgebra of V(M) is an involutive subspace. Let D C V(M). An integral manifold of D is a connected submanifold S of M with the property that S~ = ~(D(x» for every XES, where D(x) = {X(x) : XED}, and where ~(D(x» is the subspace of M~ spanned by D(x). We state the following basic result about integral manifolds: LEMMA 2.2. Let D bean infJolutive subset of V(M), and let x E M. Then x is contained in a unique maximal integral manifold of D (here "maximal" me41lS "maximal with respect to inclusion").
369
This result is classical if the dimension of !l'(D(x») is the same for each x E M (Chevalley [3]). For a proof in the general case, see Lobry [16]. If D C V(M), we denote the smallest subalgebra of V(M) which contains D by !T(D), and the maximal integral manifold of !T(D) through x by lCD, x). Recall that. if X is a vector field on M, then ex is an integral curve of X if (X is a smooth mapping from a closed interval I, I C R, into M such that
da(t)/dt
=
X(a(t»
for all i«t.
2.3. If D is a subset of V(M), then an integral curve of D from a real interval [t, t'] into M such that there exist t = to < t1 < ... < t" = r, and elements Xl ,..., X/c of D with the property that the restriction of a; to [ti - 1 , til is an integral curve of Xi for each i = 1,2,... , k. We have the following elementary fact: DEFINITION
is a mapping
LEMMA
and let aCt)
c/,
2.4. Let D C V(M). Let ex : [to, t1] -+ M bean integral curve of D, = x for some t E [to, t l ]. Then, cx(s) E lCD, x)for all S E [to, t 1] .
Proof. It is sufficient to consider the case when 0: is an integral curve of X, XED. For each maximal integral manifold S of er(D), let J(8) be the set of all S E [to, tt] such that o:(s) E S. From the local existence and uniqueness of solutions of ordinary differential equations it follows that, if s E J(8), then there exists T > 0 such that (s - r s + r) (l [to, tt] C J(8). Thus, J(8) is
open relative to [to, til. Since the maximal integral manifolds of :T(D) are disjoint, we have that, for some maximal integral manifold S, [to, i 1] C J(8). But c¥(t) E I(D, x); therefore, our proof is complete. Chow's theorem provides a partial converse to the above lemma. If DC V(M), then D is symmetric if, whenever XED, -X also belongs to D. We can now state Chow's theorem as follows: LEMMA
2.5. Let DC V(M) be symmetric, and let x E M. Then, for every ex : [0, T] -+- M of D, w£tn T ~ 0,
y E I(D, x) there exists an integral curve such that (X(O) = x and ex(T) = y.
In other words, every point of the maximal integral manifold of ~(D) through x can be reached in positive time by following an integral curve of D having x as its initial point. DEFINITION 2.6. Let D C V(M), and let x E M. If T ~ 0, then, for any y E M, y is D-,eachable from x at time T if there exists an integral curve ex of D defined on [0, T] such that 0:(0) = x and a(T) = y. The set of all points D-reachable from x at time T is denoted by L~(D, T). The union of L~(D, t) for 0 t < 00 (respectively for 0 ~ t ~ T) is denoted by L~(D)
<
(respectively, LQ:(D, T». 370
3.
INTEGRABILITY OF FAMILIES OF ANALYTIC VECTOR FIELDS
As an introduction to the general situation we first consider the case when D is a symmetric subset of V(M). Chow's theorem can be utilized to obtain a necessary and sufficient condition for Lz(D) to have a nonempty interior in M. Let n = dim M = dim :T(D)(x). Then lCD, x) is an n-dimensional submanifold of M, and hence is open in M. By Chow's theorem we have that LJD) = I(D, x). We conclude that LJD) is open in M. Conversely (and without invoking the symmetry of D), if dim .r(D)(x) < n, then lCD, x) is a connected submanifold of M of dimension less than n; then from Lemma 2.1 it follows directly that lCD, x) has an empty interior in M. Since L:I'(D) C lCD, x), L.,(D) also has an empty interior. Thus, if D is symmetric, a necessary and sufficient condition for Lz(D) to have a nonempty interior in M is that dim.r(D)(x) = dim M. Moreover, this condition is necessary even in the nonsymmetric case (Lobry [16]). We shall show that it is also sufficient. For this purpose we shall assume that the elements of Dare complete-recall that a vector field X is complete if the integral curves of X are defined for all real t [13, p. 13]. THBOREM 3.1. Let M be an n-dimensional analytic manifold, and let D C V(M) be a family of complete 'Vector fields. A necessary and sufficient condition/or L~D) to htroe a nonempty interior in M is that dimcr(D)(x) = n. MoreOfJer, if this condition is satisfied, then for each T > 0, the interior of L(IJ{D, T) is dense in Lz(D, T) [thus, in particular, La:(D, T) has a ft01Umpty interior].
Proof. We already know that the condition of the theorem is necessary. So we assume that dim 5"'(D)(~) = n, and we prove the second statement. Clearly, this will imply that Lcc(D) has a nonempty interior in M. Without loss of generality we can assume that D is finite. Let D = {Xl ,..., X 1c} . For each i = I, 2,... , k, let ,pt(t, .) be the one-parameter group of diffeomorphisms induced by X~ (i.e., t -+ tP,(t, y) is the integral curve of Xi which passes through y at t = 0; the fact that it is defined for all real t follows from the completeness of Xi). If m is a natural number, t = (t1 , ••• , tm) is an element of Rm, and i = (~ ,..., im ) is an m-tuple of natural numbers between 1 and k, then we denote the element f/), (tl , fIJi (t2 , ••• , fIJi (tm , x) ...» by 1 I '" 4)t(t, x). Let ±D be the family of vector fields obtained from D by adjoining the vector fields -Xl ,..., -X1c to D. Then, ±D is symmetric, and dim9""(±D)(x) = n. From Chow's theorem we conclude that L(IJ(±D) is open in M. Clearly the elements of Lz(±D) are exactly those elements of M which are of the form (J).(t, x) for some m, some m-tuple i, and some tERm. For each I, and for each natural number N > 0, let A(i, N) be the set of all 371
+ ... +
points of M of the form <1>i(t, x), where II t II ~ N (here II til = I t1 I I t m I). Since A(i, N) is the image of the compact set {t : II t II ~ N} under the continuous mapping t -+
+ ... +
U = .Q n {t : II til
<
T - t} n {t : t1
> 0,•.., tm > O}.
U is open, and its closure contains the origin 0 of R'". Since dFt has rank n at each point t E U, it follows that F(U) is open. Let V = {fPJ(s, F(t) : t E U}. V is the image of F(U) under the diffeomorphism z ~ c1>J(s, z); therefore, V is open in M and, moreover, every element of V is D-reachable from x at time II s " II t II = t + II t!1 < T (here we use essentially the fact that t1 , ... , t m are nonnegative). It remains to be shown thaty belongs to the closure of V. Let {ttl} be a sequence of elements of U which converges to O. Then
+
This completes the proof of the theorem. We now want to state an analogous theorem for the sets Lz(D, T). For this purpose, we shall introduce a Lie subalgebra !To(D) of !T(D) which will be related to these sets in the same way as 9""(D) is related to the sets Lz(D, T). The aim of the following informal remarks is to motivate our definition of ~(D). We shall ignore the fact that time has to be positive. Moreover, we shall assume, for simplicity, that D consists of three vector fields Xl' X 2 and X s . Let 4>1 , 4>2 and 4>3 be the corresponding one-parameter groups. It is clear that .r(D) has the following "geometric interpretation". 9"'(D)(x) is, for each x E M, the set of all limiting directions of curves through x that are entirely contained in L~(D). Thus, for instance, if i = 1, 2, 3, then all the points in the curve t -+ cI>i(t, x) are attainable from x (recall that we are forgetting about positivity), and this is reflected in the fact that Xi(x) belongs to 9'"'(D)(x). Similarly, the curves (Xii(t) = ~i( -t, f/>;( -t, epi(t, et>j(t, x)) are 372
also contained in L:t(D). By the well-known geometric interpretation of the Lie bracket (cf. Helgason [7, p. 97]), the limiting direction of (Xij is [Xi' Xj](X) (after a reparametrization). Thus, it is clear why [Xi' Xi] belongs to .r(D). Obviously, a similar argument works for the brackets of higher order. The geometrical meaning of .r(D) is now obvious. If :J;,(D) is going to play the desired role it is clear that ~(D)(x) will have to be the set of all limiting directions of curves I' through x such that y(t) is "attainable from x in zero units of time" for all t, Notice that the curves CXij(t) of the preceding paragraph have this property. Indeed, cx(t) can be reached from x by "moving forward" in time 2t units, and then "backward" another 2t units. This shows that the vector fields [Xi , Xj] are reasonable candidates for membership in 9;,(D). A similar argument applies to higher order brackets, such as [Xi' [X; , X k ] ] , etc. On the other hand, a vector field such as Xi should not be included in ~(D) by definition, because we do not know whether the points j( -t,.epi(t, x». In other words, the subspace generated by the differences Xi - Xi will have to be included in !Jo(D). This subspace can A2X 2 + A3X3 also be defined as the set of all linear combinations AIXt such that Al + = 0 (that all the differences Xi - X, are linear combinations of this type is trivial; conversely, if Y = "tXt A2X2 A3X3 with Al "2 As = 0, then Y = AtXl A2X 2 (-AI - A2) X 3 , i.e., Y = "l(X l - X s) ~(X2 - X a)). We conclude that the reasonable candidates for membership in ffo(D) are: (i) all the brackets [Xi' Xi], [Xi' [X j , X k ] ] , etc., and (ii) all the sums "IX! A2X2 "aXa , where L Ai = O. Notice that the subspace generated by (i) is clearly the derived algebra of :T(D) (by definition, the derived algebra of a Lie algebra L is the subalgebra L' of L generated by all the brackets [X, Y], X EL, YeL; it is easy to check that L' is in fact an ideal of L; cf. Helgason [7, p. 133]). We now return to our formal development. Let ff'(D) denote the derived algebra of ~(D). Motivated by the previous remarks, we define 9;,(D) to be Y, where X is a linear combination :L:=1 AiXi with the set of all sums X Xl' ...' X1J ED and I: Ai = 0, and where Y E :T'(D). It is obvious that ~(D) is an ideal of .r(D). One shows easily that Y(D) is the set of all vector fields of the form L~-l ~Xi + Y, where Xl , , X p belong to D, Y belongs to er'(D), and Al ,.••, "1) are reals (but A1 A1) need not be zero). From this it follows immediately that ffo(D) is a subspace of ff(D) of codimension zero or one. The codimension will be zero if and only if some XED belongs to 9;,(D) (in which case every XED will belong to 9;,(D)). Similarly, for
+
+ "2 "3
+ +
+
+
+
+
+
+ +
373
+
+
+
each x E M, if k = dim 9'"(D)(x), then the dimension of 9;.(D)(x) will either be k or k - 1. We shall also be interested in associating to each D C V(M), a set D* of vector fields in the manifold M X R. Recall that the tangent space to M X R at a point (x, r) (x EM, r E R) is identified, in a natural way, to the direct sum M~ ffi R". If x e V(M) , Y E VCR), we define the vector field X E9 Y E V(M X R) by
(X (f) Y)(x, r)
= (X(x), Y(r».
EB 8/ot, where XED, and where a/at is the "canonical" vector field on R (i.e. (%t)f = f'). Using the identity [X <±) X', Y E9 Y'] = [X, Y] EB [X', Y'], one shows easily that /T'(D*) is the set of all vector fields of the form X E9 0, where X E~'(D) and 0 is the zero vector field. Therefore, $'(D*) is the set of vector fields of the form
The set D* is defined to be the set of all vector fields X
where Xl ,..., Xl) belong to D, Y E !T/(D), and AI'...' ,\2) are scalars. THEOREM 3.2. Let M be an analytic n-dimensional manifold, and let D be a family of complete analytic 'Vector fields on M. Let x e M, and let T > O. Then L[&(D, T) has a nonempty interior in M if and only if dim 9;.(D)(x) = n. Moreover, in thiscase, the interior ofLa;(D, T) is dense inL:x;(D, T).
Proof. The main idea in this proof is to modify our problem so that we can "keep track" of the time elapsed while we move along an integral curve of D. We shall then apply Theorem 3.1 to the modified system. We shall work in the manifold M X R. As in the preceding paragraphs, we let the family D* of vector fields on M X R he defined by D* = {X EE> (%t): XED}. It is clear that there is a one-to-one correspondence between integral curves <X of D such that cx(O) = x, and integral curves ~ of D* such that {3(O) = (x, 0). This correspondence is given by assigning to each curve ex the curve t -.. (cx(t), t). It follows thaty eL:x;(D, T) if and only if (y, T) E L(:x;.o)(D*, T). We show that Lz{D, T) has a nonempty interior in M if and only if L(~.o)(D*) has a nonempty interior in M X R. Assume that L~(D, T) has a nonempty interior in M, and let V be a nonempty open set such that V C L~(D, T). Let XED, and let 4> be the one-parameter group of diffeomorphisms of M generated by X. Consider the mapping F : V X R -+ M X R defined by F(v, t) = ((t, f), T + t). It is immediate that the differential of F has rank n + 1 everywhere. Therefore F maps open sets onto open sets. Since F(V X (0, C L(z.o)(D*), we conclude that L(~.o)(D*) has a nonempty interior in M X R.
(0»
374
To prove the converse, assume that L(z,o)(D*) has a nonempty interior in M X R. By Theorem 3.1, for each t with 0 < t < T, L(~.o)(D*, t) has a nonempty interior in M X R. Let V be a nonempty open subset of M, and let W be a nonempty open subset of R such that V X we LCm.o)(D*, t). Let s E W. Since V X {s} C L(z,o>{D*, t), we conclude that V C Lz(D, s). Let XED, and let ~ be the corresponding one-parameter group on M. Denote the mappingy -+ 4>(T - s,y) by G. Then G(V) is open. Since G(V) is contained in L~(D, T), it follows that L~(D, T) has a nonempty interior. We conclude from Theorem 3.1 that La:(D, T) has a non empty interior if and only if dim ~(D*)(x, 0) = n I. To complete the proof of the first part of our statement, we must show that this last condition holds if and only if dim 9;(D)(x) = ft. We recall, from the remarks preceding this proof, the fact that every X* E ~(D*) can be expressed as
+
(#)
where Xl ,..., X p belong to D and Y E ~'(D). Now assume that dim .r(D*)(x, 0) = n
+ I.
Let fJ E M z • Then (fJ,O) must belong to 9""(D*)(x, 0), so that (v,O) = X*(x, 0), where X" E .r(D*). Then formula (#) holds for suitable '\"
Xi' Y. Therefore,
and
The last equality implies that L ~ = 0, so that the vector field L ~Xi + y belongs to 9;(D). Thus !1 E 9;(D)(x). We have shown that M~ C 9;(D)(x). Therefore the dimension of ~(D)(x) is n. Conversely, let dim 9;(D)(x) = n. Let fJ e M~ . Then fJ E 9;(D)(x), so that
where the Xi belong to D, Y E er'(D) and L'\i (e, 0)
= O. Therefore,
((L ~Xi + Y) ffi (L >t.) :t) (x, 0) = (L~(X.ffi :t) + YffiO) (x, 0). =
375
This shows that (cv,O) belongs to ~(D*)(x, 0). Pick an XED. Then 6:> o!ot)(x, 0) belongs to D*(x, 0) by definition, and (X ~ O)(x, 0) belongs to er(D*)(x, O) by the previous remarks. Therefore (O,olot(O» belongs to ff(D*)(x, 0). We have thus shown that ff(D*)(x, 0) contains all the vectors (v,O), tJ E M~, and also the vector (O,olot(O». Therefore .r(D*)(x, O) = (M X R)(~.o) , so that dim .r-(D*)(x, 0) = n I as stated. We now prove the second part of the theorem. As we remarked earlier, there is no loss of generality in assuming that D is finite. Let y EL~(D, T). Using the notations of the proof of Theorem 3.1, let y = CPi(t, x), where i = (i 1 , ••• , i 11l) , and where tERm is such that t, > 0 for i = 1,... , m and 1/ t ] = T. Let {Sk} C (0, t'fYi) be such that limk-?co Sk = O. Since our condition for Lx(D, T) to have a nonempty interior is independent of T, we conclude that Lx(D, t) has a nonempty interior for all t > O. In particular, for each k > 0, there exists X k which belongs to the interior of Lx(D, Sk)' Let t k =: (t1 , ••• , t m - 1 , t m - Sk), and let Yk =:
(X
+
THEOREM 3.3. Let D C V(M) be a family of complete vector fields. Then, for each T > 0, the set Lx(D, T) is contained in I(D, x). Moreover, in the topology of I(D, x), the interior of L:z;(D, T) is dense on La:(D, T). Lx(D, T) has a nonempty interior in I(D, x) if and only if dim ~(D(x)) = dim S-(D)(x) and, in this case, the interior of Lx(D, T) is dense in Lx(D, T).
Proof. If X E :T(D), then X is tangent to I(D, x). Thus there is a welldefined restriction X# of X toI(D, x). We denote the set of all such restrictions of elements of D by D#. Since [X, Y]# = [X#, Y#], it follows that ff(D)# = :T(D#). Analogously, we have that .ro(D)# == ~(D#). If we now apply the previous theorems to the family D# of vector fields in I(D, x), we get all the conclusions of the theorem. COROLLARY 3.4. Let S he a maximal integral manifold of !T(D). Then the dimension of ~(D)(x) is the same for all XES.
Proof. If dim .r(D)(x) = k then, for each XES, the dimension of 9;.(D)(x) is either k or k - 1. We show that, if dim ~(D)(x) = k - 1 for some XES, then dim ~(D)(y) = k - 1 for all yES. Let fJ be a nonempty, open (relative to S) subset of La:(D) (this is possible by Theorem 3.3). We first show that) if y E Q, then dim ~(D)(y) = k - 1. If this were not the case, 376
then necessarily dim ~(D)(y) = k. Then L1J(D, t) would have a nonempty interior in S for all t > O. This would imply that Lz{D, t) has a nonempty interior in S. But by our assumption this is impossible. Thus, dim ~(D)(y) = k - 1 for all y E Q. Since S is connected, and Q is open in S, we have that dim ~(D)(y) = k - 1 for all yES: therefore, our statement is proved. We now proceed to study the case when dim ~(D)(x) = dim S"(D)(x) - 1. We begin by proving some preliminary lemmas. LEMMA 3.5. Let D C V(M) bea family of complete vector fields. If XED, let {t1>t} be the one-parameter group generated by X. Then, for every x E M, and every t E R the differential dtP t maps ~(D)(x) onto ~(D)(tPt(x».
Proof. We first show that for every y E M there is an T > 0 such that, if f) E ~(D)(y), then d(/)t(v) E 9;.(D)(CPt(Y» for all t with I t I < r. It is sufficient to show that for every y E M and every v E ~(D)(y) there exists an r > 0 such that dtPt(v) E ~(D)(
for all t in this neighborhood, where [XW), Y] = Y, and [x(n), Y] = [X, [x(n-l), Y]] for n = 1, 2,.... Since each term of the above series belongs to ~(D)(
Proof. Let X, Y, CPt, 'fit and S satisfy the conditions of the lemma. Let P be the maximal integral manifold of .r(D) which contains S. If dim P = dim S, then S = P, and cPt(S) = S = 'Pt(S). Assume that dim S = k = dim(P) - I. We first show that there is an r > 0 such that 377
= ",(8) whenever l t l < r. Let XES. The mapping (s, t) --. fPt(s) has rank k + 1 at (x, 0). Let D be a neighborhood of x in S, and let 8 > 0 be such that this mapping, restricted to {J X (-8, 8), is a diffeomorphism onto an open subset D'" of P. If yEaii', let s(y) and fey) be such that (f),(.)(s(y» = y. Clearly, I is analytic in QJIf, and fey) = 0 if and only if y e D. Moreover, XI I in {JfI'. For every t such that I t I < 8, the set cJ,(Q) is an integral manifold of !To(D). The vector field Y - X is tangent to cJt(Q) and, since f is constant on
=
XI
LEMMA 3.7. Let E be a locally convex 'Vector space, let K C E, and let C bea CMlfJex dense subset of K. Let F : K --. I(D, x) be a continuous mapping such that F(C) is contained in a maximal integral manifold S of ~(D). Then F(K) is contained in S, and F, as a mapping from K into S, is continuous.
Proof. If dim S = dim I(D, x), then S = I(D, x), and the conclusion follows trivially. Therefore, we shall assume that dim S = dim I(D, x) - 1. Let k E K, let XED, and let {tP t} be the one-parameter group induced by X. Then, as in the proof of Lemma 3.6, we can find a neighborhood of F(k) in lo(D, F(k», and a positive number 8, such that the mapping (s, t) --+ c1>t(s) is a diffeomorphism of {J X (-8, 8) onto an open subset Q# of I(D, x). Let U be an open convex neighborhood of k such that F( U fl K) C Q#. For each t E (-8, 8), the set cJt(Q) is an integral manifold of ~(D); therefore, if
n
378
4),(.0) intersects S, then tP,(O) is contained and open in S. Let A = {t : I t I < 8,
Remark 3.8. If is clear that the preceding lemma is valid under weaker assumptions about C and K. For instance, it is sufficient to assume that, for every k e K and for every neighborhood U of k, there exists a neighborhood V of k such that V C U and V n C is connected. We now state and prove the theorem towards which we have been aiming. THEOREM 3.9. Let D C V(M) be a set of complete oector fields, and let x E M. Then, for eaeh T > 0, L:e(D, T) C Io'(D, x) and, moreooer, the interior 01 L~(D, T) (relatifJe to Iot(D, x» is dense in L~(D, T) (and is, in particular, MIIefIIpty).
Proof. If dim 9;(D)(x) = dim :T(D)(x), then we have from Corollary 3A that 9O(D)(y) = 9"'(D)(y) for ally E lCD, x). Therefore,lo(D, x) = lCD, x) = I.'(D, x) and our conclusion follows from Theorem 3.3. Assume that dim 9O(D)(x) = k = dim .r(D)(x) - 1. It is clear from Lemma 3.6 that, if ex is an integral curve of D such that ex(O) = x, then ex(T) eloT(D, x); hence, LJD, T) C [.T(D, x). We now show that, if yeL:e(D, T), then y is the limit of points which belong to the interior ofL~(D, T). Let D = {Xl'...' X k } and lety = 4>.(T, x), where II T rI = T, and T, > 0 for i = 1, 2,..., m (the notations here are the same as in the proof of Theorem 3.1). Let j = (it ,... ,is) be an s-tuple of integers between 1 and k such that the rank of t ~ t/J.(t, x) is equal to dim 5"'(D)(x) for all t in an open dense subset D of R'. Let Q' = {t : t E R', t i > 0 for i = 1,..., s} n Q. Let {t2)} C (J be a sequence that converges to 0, and let T2) = (T1 , ••• , T m-I' T m -II t" JJ). We can assume that If tf) If < T m for all p > O. If we let YP = (f).(Tp , tl'J(tf) , x», then YP eLz{D, T). We next 379
show that Y'I' is in the interior of Lx(D~ T) relative to IoT(D, x). Since the mapping % --+- 4>j(TJJ , z) is a diffeomorphism from IBt~II(D, x) onto loT(D, x), it suffices to show that epJ(t p , x) is in the interior of L x( D, \I t, II). Let V'I' = {t : t E Rs, t1 > 0,..., i, > 0, II t II = II e, II}. Clearly, if t E VJ) , then eJ)j(t, x) ELz{D, II t'l'll). Let F1J : V p --+- l~tJlfl(D, x) be defined by F2)(t) = 4>.(t, x). We show thatF2) is analytic. SinceF2) is analytic as a map from V 1) into lCD, x), it suffices to show that it is continuous. But this follows from the previous lemma, because V1J is convex. The rank of t -+-
+
4.
ApPLICATIONS TO CONTROL SYSTEMS
We shall consider systems of the form dx(t)/dt = F(x(t), u(t» defined on an analytic manifold M. The functions u belong to a class d/I of "admissible controls". We make the following assumptions about '11 and the system function F: (i) The elements of tit arepiecewise continuous functions defined in [0, (0), having values in a locally path connected set D, D C Rm (D is locally path connected if, for every W E Q and every neighborhood V of (.0, there exists
a neighborhood U of (.0 such that U C V, and U (1 D is path connected). In addition, we assume that if! contains all the piecewise constant functions with values in D, and that efJery element of OIl has finite one-sided limits at each point of discontinuity. c:w is endowed with the topology of uniform convergence on compact intervals.
(ii) F: M X Q -+- TM is jointly continuously differentiable. For each u e Q, F(·, u) is a complete analytic oector field on M. We know that for each x E M, U E t¥I, the differential equation dx(t)/dt
= F(x(t), u(t», 380
x(O)
=
x,
(1)
has a solution defined for all t E [0, 8), where 8 > O. We denote such a solution by n(x, u, '), and 'We assume that n(x, u, t) is defined for all t E [0, (0). For the above-defined control system we now state the basic controllability concepts. We say that y E M is attainable from x E M at time t (t ~ 0), if there exists U E dIJ such that l1(x, u, t) = y. For each x E M, we let A(x, t) denote the set of all points attainable from x at time t. If 0 ~ t < 00, we define A(x, t) = Us
4.1. For each x EM, A(x) is contained in I(D, x).
The proof is identical to that of Lemma 2.4, and will therefore be omitted.
Remark 4.2. It is easy to see that the control system defined by restricting F to I(D, x) satisfies the same assumptions as the original system. Since it can be readily verified that the map u -+ II(x, u, t) is continuous as a map from t1I into M, it follows that this map is also continuous as a map from o/J into I(D, x).
We now want to obtain a result for A(x, t) which is similar to that of Lemma 4.1. It is here that the assumption about Q will be utilized. Let 9 be the class of piecewise constant .Q-valued functions defined on [0, (0). Clearly, fJ' is dense in ~. Moreover, the local connectedness of Q implies that the condition of Remark 3.8 is satisfied (this can be easily verified, and we omit the proof). Thus, we can apply Lemma 3.7, with C = f!} and K = 0/1, to obtain the following result: LEMMA
4.3. Let x E M. For each t
~
0, A(x, t) C Iot(D, x).
Proof. Since dIJ contains (!;, we have that L~(D, t) C A(x, t). Let G : ~ --..I(D, x) be defined by G(u) = I1(x, u, t). We have that G(~) = L~(D, t), and by Theorem 3.9, G(f!P) C Iot(D, x). Now our conclusion follows immediately from Lemma 3.7, and the proof is complete. The above lemmas combined with the theorems of the preceding section yields the following results; 381
THEOREM 4.4. Let x E M. Then A(x) C I{D, x). MoreOfJer,for every T > 0, tM interior of A(x, T)relatiw to I(D, x) is dense in A(x, T) (and, in particular, is ftOIIIfftJ'ty).
Proof. The first part is just the statement of Lemma 4.1. To prove the second part, we can assume that lCD, x) = M (if not, replace the original system by its restriction to lCD, x), cf. Remark 4.2). Since L~D, T) is dense in A(x, T), our conclusion follows immediately from Theorem 3.1. THBoREM 4.5. Let x e M. Then, for each t > 0, A(x, t) C Iot(D, x) and, ffUWlOfJer, the interior of A(x, t) relative to Io'(D, x) is dense in A(x, t) (and, in
particular, is ftOII8mpty).
Proof. The first part is just the statement of Lemma 4.3. To prove the second part, we apply Lemma 3.7 to the function G of Lemma 4.3, and we get that G is continuous as a map into [ot(D, x); therefore, L~(D, t) is dense in A(x, t) relative to lot(D, x). Our conclusion now follows immediately from Theorem 3.9, and the proof is complete. The following two controllability criteria follow immediately from Theorems 4.4 and 4.5, and from Lemma 2.1: 4.6. The system has the accessibility property from x if and only dim r(D)(x) = dim M. In this case A(x, T) has a nonempty interior for
COROLLARY
if
IfJery
T
> o.
COROLLARY
4.7. The system has the strong accessibility property from x = dim M. In this case A(x, T) has a nonempty
if and only if dim 9;(D)(x) ifttIrior for lfJery T > o.
The preceding results can be utilized to derive relationships between accessibility and strong accessibility. Even though the latter property seems much stronger than the former, we show that, for a very large class of manifolds (including the spheres Sft for n > 1, and all compact semisimple Lie groups, but not R"), it is in fact implied by it. On the other hand, for a still larger class of manifolds (including R") controllability (which trivially implies accessibility), is sufficient to guarantee strong accessibility (the fact that controllability implies that dim .r(D*)(x) = n 1 for all x was proved by Elliottin [5]). Consider a system on a connected n-dimensional analytic manifold M, having the accessibility property but not having the strong accessibility property. Let D be the family of associated vector fields. By Corollary 4.6, dim r(D)(~) = 11 for all x E M. By Corollary 3.4 the number dim 9;(D)(x) is independent of %. Since this number is either n or n - 1J Corollary 4.7
+
382
implies that dim r.,{D)(s) = n - 1 for all s E M. Choose a fixed X E D, and use (J'), to denote the one-parameter group generated by X (i.e., for every ye M, the integral curve of X that passes through y at t = 0 is the curve t -+ tI,(y». Define a mapping F from the manifold S X R into M by
F{s, t)
= t/J ,(s).
One shows easily that F is a local diffeomorphism onto M. Moreover, S X R is connected.In fact, we have (see [18] for the definition of a covering
projection): LEMMA 4.8.
The map F is a cO'lJering projection.
Before we prove Lemma 4.8, we show how the results mentioned above follow from it. THEOREM 4.9. Let M be a manifold whose universal CO'lJering space is compact. Then eoery system hcroing the accessibility property has the strong dCcessibility property.
Proof. If the universal covering space of M is compact, then every covering space of M is compact. If it were possible to have a system on M having the accessibility property but not the strong accessibility property, we could define, for such a system, S and F as above. It would follow that 8 X R is compact, which is clearlya contradiction. R.lmark. If n > 1, the sphere 8 ft is simply connected (and compact). Therefore Theorem 4.9 applies. Also, if M is a connectedcompactsemisimple Lie group (for instance SO(n), if n > 2), the universal covering group of M is also compact [7, p. 123] and, therefore, Theorem 4.9 applies in this case as well. THEOREM 4.10. Let M b, a mtmifold fJJhose fundammtal group has no elements of infinite order. Then efJery controllable system on M has thl strong aeceuibility property.
Proof. A controllable system obviously has the accessibility property. Assume it does not have the strong accessibility property. Define S and F as before. We show that F is one-to-one. Otherwise, there would exist So J 10' E S and aT#: 0 such thatF(so', T) = cj)T(SO') = F(so , 0) = so. Therefore clT(S) = S. Define H : 8 X R ~ S X R by H(s, t) = (4)T{S), t - T). Then H is well defined. because cPr (8 ) = S, and is a homeomorphism. Moreover, if (s, t) e S X R, F(H(s, t»
= (>"'T(cPr(S» = ePe(s) = 383
F(s, t).
Therefore H is a covering transformation [18, Chap. 2]. Moreover, if s is a point of S and if t belongs to R, then H'm(s, t) = (
t».
Remark. Theorem 4.10 applies, in particular, to any simply connected manifold, such as R". Proof of Lemma 4.8. We must show that every point of M has a neighborhood that is evenly covered by F. Let m E M. Since F is a local diffeomorphism onto, there exist s E S, t E R, E > 0 and a connected neighborhood U of sin S such thatF(s, t) = m and that the restriction ofF to U X (t - E, t + E) is a diffeomorphism onto an open subset V of M. We claim that V is evenly covered. Let A = {r : 4>.,.(8) = S}. For each TEA, let U., = tlJ.,( U). Since 4>., : S -+ S is a diffeomorphism, it follows that U., is open in S and connected for each TEA. We first show that, if 0 < IT - TJ I < 2£, TEA, 7J E A, then U., and UTI are disjoint. Assume they are not. Then
for each 'T E A, F maps W.,. diffeomorphically onto V, and
(c) the inverse image of V under F is the union of the sets W., . The first two assertions of (a) are obvious. If T and 1] belong to A, and I T - 1] I < 2E or I T - 1] I ~ 2£. In the first case W., and W'7) must be disjoint, because U., and U; are disjoint. In the second case, W., and W,., are also disjoint, because the intervals (t - T - e, t - T + E) and (t - 1] - E, t - 7J + f:) cannot have a point in common. To prove (b), take TEA. Define G : U X (t - e, t + E) ~ W'T by G(u, a) = (
T
=1= 1], then either
a
+
384
a»
<Pa(U) = F(u, a). Since the restriction of F to U X (t - E, t + E) is a diffeomorphism onto V, the same must be true for the restriction of F to W.,. . Finally, we prove (c). Let U E S, a E R be such that F(u, a) E V. Then there exist u' E U, a' E (t - E, t + E) such that F(u',a') === F(u, 0). Therefore u =
EXAMPLES
EXAMPLE 5.1. Let M = Rn, Q = Rm, and let F: M X Q -+ TM be defined by F(x, u) = Ax Bu, where A and B are, respectively, n X nand n X m real matrices. We have that D = {A(·) + Bu : U E Rm}. Let hi denote the i-th column of B. Then, as shown by Lobry [16], 9"'(D)(x) contains the vectors
+
i = l, ... ,m. It is not difficult to see that the above set of vectors forms a system of generators for !T(D)(x). From Corollary 4.6 we get that A(O, t) has a nonempty interior in Rn if and only if {±bi , ±Abi , ••• , ±An-1bi , i = 1, 2,..., m} has rank n; equivalently, A(O, t) has a nonempty interior in Rn if and only if rank[B, AB, ..., An-1B] = n. Since, obviously, ~(D)(O) = ..r{D)(O), we conclude that A(O, t) has a nonempty interior whenever A(O, t) does. The above statements, along with the fact that A(O, t) and A(O, t) are linear subspaces of R", imply that, if rank[B, AB,..., An-1B] = n, then for each t > 0 A(O, t) = A(O, t) = A(O) = R" (Kalman [12]). Thus, in this example, the accessibility property is equivalent to controllability. This is, of course, not true in general. EXAMPLE 5.2. Let M = R", D = {u E Rm : 0 ~ u, ~ 1, i = 1,..., m}, and letF(x, u) = (A o I:::l AiUi)X for all (x, u) ERn X Q, where A o ,••• , Am. are n X n real matrices. Then D is the set of all vector fields Xu where Xu(x) = (A o !:::1 UiAi)X. The set Mn of all n X n real matrices is a Lie algebra, with the bracket defined by [P, Q] = PQ - QP. To each matrix P there corresponds a vector field V(P) defined by V(P)(x) = Px, It is easy to check that V([Q, P]) = [V(P), V(Q)]. Using this fact, the spaces .r'(D)(x) and ~(D)(x) can be readily computed:
+
+
9""{D)(x) = {Px : PEL}, and ~(D)(x) =
{Px: PeL},
385
where L is the Lie algebra spanned by A o ,•••, Am , and L is the ideal of L spanned by AI'...' Am. We remark that for this example the theory of Section 4 is valid eoen if c1/ is the set of all bounded and mlasu,able {J-fJalued fametUms. This is so because the only properties of the class of admissible controls that were utilized in Section 4 were: (a) that the class of piecewise constant controls is dense in 'II (in the topology of uniform convergence), and (b) that, if {"eI} are elements of tpj that converge uniformly to u, then H(llo , X, t) converges to II(u, x, t). In our example, both (a) and (b) remain valid if the topology of uniform convergence is replaced by that of fJJeak convergence. This is easy to verify, and we shall not do it here (see Kucera [14]). Moreover, the set of .a-valued measurable functions defined in [0, T] is weakly compact. It follows that the sets A(x, T), A(x, T) are compact for each T > O. Denote their interiors (relative to lCD, x) and IoT(D, x), respectively) by int A(x, T), int A(x, T). It follows that A(x, T) is the closure of int A(x, T), and that A(x, T) is the closure of int A(x, T). Therefore, our results contain those of Kucera (in this connection, see also Sussmann [21]).
Remark. The result of the preceding example is a particular case of a more general situation. Let G be a Lie group, and let M be an analytic manifold on which G acts analytically to the left. Then there is a homomorphism Afrom the Lie algebra of G into V(M), defined by A(X)(m) = (dldt)[exp(tX) · m], the derivative being evaluated at t = O. If X o ,••• , X k belong to the Lie algebra of G, we can consider the control problem
where X/ = A(Xi ) . Example 5.2 results by letting G = GL(n, R) and = Rft.
M
EXAMPLE 5.3. This example shows that the analyticity assumptions are essential. Consider the following two systems defined in the (x, y) plane:
(8 1) X = fl(X, y, u),
y = gl(X, y, u), and
(St)
x = f,,(x, y, u), j = g,.(x, y, u). 386
=
Let /1 == /" = 1, g1 0, and g2(X, y, u) = ep{.t) where ep is a C«J function which vanishes for -00 < x < 1, and which is equal to 1 for x > 2. It is clear that for (81) the set A«O, 0» is the half line {(x, y) : y = 0, x ~ O} while, for (SI)' A«O, O}) has a non empty i~terior. However, both systems are identical in a neighborhood of (0, 0).
ACKNOWLEDGMENTS
We are grateful to Professor R. W. Brockett for his encouragement and advice. Also, we wish to thank Dr. E. H. Cattani, Dr. L. Marino and an anonymous referee for helpful suggestions.
REFERENcps 1. R. L.
BISHOP AND
R. I.
CRITTENDEN,
"Geometry of Manifolds," Academic Press,
New York, 1964.
2. R. W. BROCKETT, System theory on group manifolds and coset spaces, SIAM [. Control, to appear. 3. C. CHEVALLEY, "Theory of Lie Groups," Princeton University Press, Princeton, NI,1946. 4. W. L. CHOW, Uber Systeme von linearen partiellen Differential-gleichungen erster Ordnung, Math. Ann. 117 (1939),98-105. S. D. L. ELLIOTT, A consequence of control1ability, }. DiJJ. Eq,. 10 (1971), 364-370. 6. G. W. HAYNES AND H. HBRMBs, Nonlinear controllability via Lie theory, SIAM ]. Control 8 (1970), 450-460. 7. S. HELoASON, "Differential Geometry and Symmetric Spaces." Academic Preas, New York, 1962. 8. R. HERMANN, E. Cartan's geometric theory of partial differential equations, Advaneu in Math. 1 (1965), 265-315. 9. R. HERMANN, On the accessibility problem in control theoryt in "Intemational Symposium on Nonlinear Differential Equations and Nonlinear Mechanics," pp. 325-332, Academic Press, New York, 1963. 10. R. HBRMANN. The differential geometry of foliations II,]. Math. M,ch. II (1962), 30S-31S. 11. V. JURDJEVlC, Abstract control systems: controllability and observability, SIAM ]. Control 8 (1970), 424-439. 12. R. E. KALMAN, Y. C. HOt AND K. S. NARENDRA, Controllability of linear dynamical systems, Contrib. DifJ. Eqs. 1 (1963), 189-213. 13. S. KOBAYASHI AND K. NOMIZU, "Foundations of Differential Geometry," Vol. I, Interscience, New York, 1963. 14. J. Kui5BRA, Solution in large of control problem: x ::II (A(I - u) + Bu):c, C.eda. Math. ]. 16 (91) (1966), 600-623. IS. J. KuamA, Solution in large of control problem: = (Au + BfJ)'¥, Czech. Math. ]. 17 (92) (1967), 91-96. 16. C. LoBRY, ContralabUit~ des systemes non lin~aires~ SIAM J. Control 8 (1970),
x
,73-60S.
17. C. LOBRY, Une propriete de l'ensemble des etats accessibles d'un systeme guidable, C. R. Acad. Sci. Paris 272 Series A (1971), 153-156. 18. E. H. SPANIER, "Algebraic Topology," McGraw-Hill, New York, 1966. 19. S. STERNBERG, "Lectures on Differential Geometry," Prentice Hall, Englewood Cliffs, NJ, 1964. 20. H. J. SUSSMANN, The bang-bang problem for certain control systems in GL(n, R), SIAM ]. Control, to appear. 21. H. J. SUSSMANN, The control problem x = (A(1 - u) + Bu)x: a comment on an article by J. KuCera, Czech. Math. J., to appear.
388
Dissipative Dynamical Systems-Part I: General Theory JAN C. WILLEMS
EARLIER papers and discussions in this volume explained how the works of Popov [7], Yakubovich [13], and Kalman [4] led to the formulation of the Positive Real Lemma, through which passivity, a network theory concept, became useful for feedback control. In his input/output analysis, Zames [14] used passivity as a defining property of positive operators on extended inner product spaces and pursued a nonlinear analogy of Bode's gain-phase characteristics. In this analogy, passivity is viewed as a "phase" property behind both Popov and Circle Criteria. In Kalman's analysis of the inverse optimal Linear-Quadratic problem [5], a passivity property was shown to be necessary for optimality. In the mid-1960s, Jan Willems was among a select group of young researchers seeking to develop a general theory to unify newly discovered feedback properties. His theory of dissipative dynamical systems [12] was a major step in this direction. Only the general theory part of this two-part paper is reprinted here. The second part of the paper demonstrates the ability of this theory to unify previous stability and optimality results for linear systems. While most of the readers immediately saw the significance of the second part of the paper, the impact of the first part has been growing over a longer period. Jan Willems defines dissipativity as an input/output property but, in contrast to Zames, does not avoid the state-space and Lyapunov functions. Instead, he generalizes Lyapunov functions with his new concept of a storage function S(x). For an input/output pair u, y he introduces the notion of supply rate w(u, y) and defines a system to be dissipative if the increment of storage, S(XI) - S(xo), is not larger than the integral of the supply rate along any state-space trajectory x(t) from x(to) = Xo to X(tl) = Xl. A significant result of the paper is that input/output knowledge does not determine the storage function uniquely. Rather, there are two variational problems that can be posed, whose solutions determine two storage functions, the so-called available storage and required supply, and these two storage functions underbound and overbound any storage function that can be associated with the system. In the linear case, storage functions are associated with solutions of linear matrix inequalities. The extremal available storage and required supply functions correspond to extremal solutions of these inequalities
which are frequently also stabilizing or antistabilizing solutions of Riccati equations. Reviewing the results of Popov [8] and Zames [14] for feedback systems consisting of two nonlinear blocks, Willems points out that the small gain stability conditions are established with supply rate w = uTU - yT y, while w = uTY is the supply rate used to define passivity. The work of Willems stimulated Hill and Moylan [3] to derive a nonlinear analog of the PR Lemma and to quantify the "excess" or "shortage" of passivity using the supply rate w = uTY - vu TU - pyT Y where v and p are constants that can be negative or positive. In feedback interconnections, v and p for individual blocks can be changed by feedforward and feedback loop transformations. These results were instrumental for feedback passivation developed a decade later by Kokotovic and Sussmann [6], Byrnes et al. [2], and for nonlinear feedback design by Sepulchre et al. [9]. The concept of dissipativity was also crucial in recent extensions of linear robust control theory to nonlinear systems, as summarized by van der Schaft [11]. It also arose in the dynamic game approach to the disturbance attenuation problem by Basar and Bernhard [1], where the optimal (min-max) feedback control u(x) guarantees that the designed system is dissipative with respect to the disturbance input. For zT := [yT, uT ], the supply rate is of the form w = y 21dl2 - Iz1 2, and the L2-gain of the closed-loop system is ~ < y. IIdll3 As shown by Sontag and Wang [10], input-to-state stability, another currently popular concept, is directly related to dissipativity.
389
REFERENCES
[1] T. BA~AR AND P. BERNHARD, ?-loo OptimalControland RelatedMinimax
DesignProblems, Birkhauser(Boston),secondedition, 1995. [2] C.I. BYRNES, A. ISIDORI, AND J.C. WILLEMS, "Passivity, feedbackequivalence,and global stabilization of minimumphase systems," IEEE Trans. Aut. Contr., AC-36:1228-1240, 1991. [3] D. HILL AND P. MOYLAN, "The stabilityof nonlineardissipative systems," IEEE Trans. Aut. Contr., AC-21(5):708-711, 1976. [4] R.E. KALMAN, "Lyapunov functions for the problemof Lur'e in automatic control," Proceedings of the NationalAcademyof Sciencesof the United Statesof America, 49:201-205, 1963.
[5] R.E. KALMAN, "When is a linear control systemoptimal?" Transactions of the ASME, SeriesD, J. Basicengineering, 86:1-10, 1964. [6] P.V. KOKOTOVIC AND H.J. SUSSMANN, "A positive real condition for global stabilization of nonlinearsystems," Syst.Contr. Lett., 19:177-185, 1989. [7] V.M. PoPOv, "Absolute stabilityof nonlinearsystemsof automatic control,"Automation andRemoteControl, 22:857-875, 1962. Translated from Avtomatika i Telemekhanika, 22:961-979, 1961. [8] V.M. PoPOv, "Hyperstability and optimality of automatic systems with severalcontrolfunctions," Rev. Roumaine Sci.Tech. Electrotechn. etEnerg., 9:629-690, 1964. [9] R. SEPULCHRE, M.JANKOVIC, AND P.KOKOTOVIC, Constructive Nonlinear Control. Springer-Verlag (NewYork), 1997.
[10] E.D. SONTAG AND Y. WANG, "On characterizations of the input-to-statestabilityproperty," Syst.Contr. Lett., 24:351-359, 1995. [11] A.J. VAN DER SCHAFT, L2-Gain and Passivity Techniques in Nonlinear Control, Springer-Verlag (NewYork), 1996. [12] J.C. WILLEMS, "Dissipative dynamical systems Part I: General theory;
Part II: Linear systemswith quadratic supplyrates," Archive for Rational Mechanics andAnalysis, 45:321-393, 1972. [13] V.A. YAKUBOVICH, "The solution of certain matrix inequalities in automatic control theory," Doklady Akademii Nauk, 143:1304-1307, 1962. [14] G. ZAMES, "On the input-output stabilityof time-varying nonlinearfeedback systems-Parts I and II," IEEETrans. Aut. Contr., AC-ll:228-238 & 465-476, 1966.
390
P.V.K.
Dissipative Dynamical Systems Part I· General Theory JAN
C. WILLEMS
Communicated by C. TRUESDELL Contents Abstract . 1. Introduction . . . . . . . . . . 2. Dynamical Systems . 3. Dissipative I;>ynamica1 Systems . . . . . . . 4. Interconnected Systems . . . . . s. Stability . . . . . . . . . . . . . . . . 6. Non-Stationary Dynamical Systems . . . . . . 7. Applications . . . . . . . . . . . 7.1. Systems with a Finite Number of Degrees of Freedom .. 7.2. Stability of Feedback S¥stems 7.3. Electrical Networks . . 7.4. Thermodynamics . . 8. Conclusions . . . . . .
References . . . . . . . . . .
· . .. · · · .. · · · · · .. · ·
321 322 323 327 333 337 338 340 340 343 346 348 349 350
Abstract The first part of this two-part paper presents a general theory of dissipative dynamical systems. The mathematical model used is a state space model and dissipativeness is defined in terms of an inequality involving the storage function and the supply function. It is shown that the storage function satisfies an a priori inequality: it is bounded from below by the available storage and from above by the required supply. The available storage is the amount of internal storage which may be recovered from the systemand the required supply is the amount of supply which has to be delivered to the system in order to transfer it from the state of minimum storage to a given state. These functions are themselves possible storaae functions, i.e., they satisfy the dissipation inequality. Moreover, since the class of possible storage functions forms a convex set, there is thus a continuum of possible storage functions ranging from its lower bound, the available storage, to its upper bound, the required supply. The paper then considers interconnected systems. It is shown that dissipative systems which are interconnected via a neutral interconnection constraint define a new dissipative dynamical system and that the sum of the storage functions of the individual subsystems is a storage function for the interconnected system. The stability of dissipative systems is then investigated Reprinted with permission from Archive for RationalMechanics andAnalysis, Jan C. Willems, "Dissipative Dynamical Systems-Part I: General Theory" Vol. 45,1972, pp. 321-351.
391
and it is shown that a point in the state space where the storage function attains a local minimum defines a stable equilibrium and that the storage function is a Lyapunov function for this equilibrium. These results are then applied to several examples. These concepts and results will be applied to linear dynamical systems with quadratic supply rates in the second part of this paper. 1. Introduction Dissipative systems are of particular interest in engineering and physics. The dissipation hypothesis, which distinguishes such systems from general dynamical systems, results in a fundamental constraint on their dynamic behavior. Typical examples of dissipative systems are electrical networks in which part of the electrical energy is dissipated in the resistors in the form of heat, viscoelastic systems in which viscous friction is responsible for a similar loss in energy, and thermodynamic systems for which the second law postulates a form of dissipation leading to an increase in entropy. In the first part of this paper we hope to provide an axiomatic foundation for a general theory of dissipative systems. In the course of doing this we examine the concepts of an internal storage function and of a dissipation function. There will be an obvious search for generality in the theoretical discussion of the first part of this paper. This stems from a belief that in studying specialized classes of dynamical systems it is important to keep the axioms separated. Such a procedure has more than just an aesthetic appeal: it allows one to pinpoint clearly what is a consequence of what. My interest in dissipative systems stems from their implications on the stability of control systems. One of the main results in stability theory states that a feedback system consisting of a passive dynamical system in both the forward and the feedback loop is itself passive and thus stable. Moreover, the sum of the stored "energies" in the forward loop and in the feedback loop is a Lyapunov function for the closed loop system. The existence of a stored energy function is rather simple to eastablish since it is equivalent to the passivity assumption. It was in computing this stored energy function that we encountered some difficulties. It became clear that there is no uniqueness of the stored energy function, rather that there is a range of possible stored energy functions for a system with a prescribed input/output behavior. In this paper these concepts are studied in detail and generalized. The terminology dissipative will be used as a generalization of the concept of passivity and storage function as a generalization of the concept of stored energy or entropy. One of the main results obtained in this paper is that the storage function is as a rule not uniquely defined by the input/output behavior. It is shown that the storage function associated with a dissipative dynamical system satisfies an a priori inequality: it is bounded from below by the available storage and from above by the required supply. Moreover, and possibly more important, there is a continuum of possible storage functions between these upper and lower bounds. This situation has important consequences. To give but one example, consider the familiar area of linear viscoelasticity. This is a typical example of a situation where the internal physical mechanism which is responsible for a stress/strain 392
relationship is admittedly not completely understood. For many applications, one is, however, satisfied with an input/output description in terms of a relaxation function which may be obtained experimentally. Such an input/output description has, in fact, become the starting point of a general approach to the description of materials with memory. Nevertheless, the literature insists on postulating the knowledge of an internal energy function. It should be realized that this destroys some of the advantages of working with an input/output description since this knowledge of an internal energy function cannot be obtained from the relaxation function but requires additional information about the physical process. (In the present example one may often circumvent this difficulty by determining the heat production as well as the stress/strain relation, but this problem remains very fundamental in the context of thermodynamic systems where it is unclear what is being dissipated while the entropy increases.) There are several methods for further reducing the number of possible storage functions. One rather obvious method is to consider a system as an interconnection of dissipative subsystems. Another possibility is by assuming additional qualitative internal properties for the system. A typical example is by postulating internal symmetry conditions as the Onsager-Casimir reciprocal relations. These will be examined. in the second part of the paper. We shall use the state space formalism for representing systems with memory. This feature is felt to be essential and the absence of the state space formalism in continuum mechanics and thermodynamics is somewhat disturbing. It is indeed customary in these areas to assume that the functionals appearing in the constitutive equations of materials with memory may depend on the entire past history (see for example [1] and [2]). This approach, however, does not recognize the idea of "equivalent histories": two histories are said to be equivalent if they bring the system into the same state and are thus indistinguishable under future experiments. Hence, one should constrain a priori the constitutive relations of any internal function as, for example, the internal energy or the entropy to take on the same value for equivalent (but not necessarily identical) histories. The state space formalism is the natural way for incorporating this constraint. There has, in fact, been some recent work by ONAT [3, 4] which deals with the construction of state space models for continuum systems. We consider this paper as a contribution to mathematical system theory. The methods employed are those which have grown out of the modem developments of control theory; some of the auxiliary results, particularly in the second part of the paper, are drawn from network synthesis and optimal control theory. The implications of the results obtained and the methods used ought to be of interest to physicists, in particular those concerned with continuum mechanics and thermodynamics. We have tried to make the paper self-contained by being as explicit as possible whenever known results are being used.
2. Dynamical Systems A dynamical system is viewed as an abstract mathematical object which maps
inputs (causes, excitations) into outputs (effects, responses) via a set of intermediate variables, the state, which summarizes the influence of past inputs. The following 22*
393
lengthy definition is concerned with continuous systems (the time-interval of definition is the real line). In order to avoid unnecessary complications mainly of a notational nature, we will restrict ourselves to stationary ti.e., time-invariant, nonaging) systems. The time-varying case is briefly discussed in Section 6.
Definition 1. A (continuous stationary) dynamical system I is defined through the sets U, 'lI, Y, t!J, X and the maps c/J and r, These satisfy the following axioms: (i) fI is called the inputspaceand consists of a class of U-valuedfunctions on * R. The set U is called the set of inputvalues. The space tft is assumed to be closed under the shift operator, i.e., if uet:lt then the function UT defined by uT(t)=u(t+T) also belongs to tfI for any TeR; (ii) fJ is called the output space and consists of a class of Y-valued funcnons on R. The set Y is called the set of output values. The space ~ is assumed to be closed under the shift operator, i. e., if yetpj then the function Yr defined by YT(t)=y(t+T) belongs to l&f for any TeR;
(iii) X is an abstract set called the state space; (iv) q, is called the state transition function and is a map from Rt x X x dIJ into X. It obeys the following axioms: (iv), (consistency): cjJ(to, to, xo, u)=xo for all toER, XoEX, and uedlt; (iv), (determinism): q,(t1 , to, Xo, U1)=q,(t 1 , to, xo, U2) for all (11' to)eRt, xoeX, and "1' U2E~ satisfying U1(t)=U2 (t) for to~t~ t 1; (iv), (semi-group property): cfJ(t 2 , to, xo, u)=c/J(t 2 , t., cP(t 1 , to, xo, u), u) for all tO~tl~t2' XoEX, and uetJIt; (iv), (stationarity): cP(t 1 +T, lo+T, X O, UT)=q,(tt, to, xo, u) for all (t 1 , to)eRt, TeR, XoEX, and u, uTedIJ related by uT(t)=u(t+T) for all te R;
(v) r is called the read-out/unction and is a map from Xx Uinto Y; (vi) the Y-valued function r( cP (I, to, Xo, u), u(t) defined for t ~ to is, for all xoeX, toER and "e~, the restriction to [to, (0) of a function yedJJ. This means that there exists an element yetHI such that y(t)=r(t/J(t, to, xo, u), u(t)) for (~to.
A dynamical system thus generates outputs from inputs as follows: the system starts off in some initial state Xo at time to and an input u is applied toit. Then the state at time t 1 is given by q,(t1 , to, Xo, u). The output resulting from this experiment is given by y(t)=r(q,(t, to, Xo, u), U(I») and is defined for t~to. It is important (for applications to systems described by partial differential equations for example) to realize that state transitions, and thus outputs, need only be defined in the forward time direction.
We call 4>(t 1 , to, XO, u) "the state at time t 1 reached/rom the initialstate Xo at time to by applying the input u to the dynamical system E" and r(x, u) "the output
* We are using the following notation: R=the real numbers; R"=n-dimensional Euclidean space; R + = the nonnegative real numbers; Rt = the causal triangular sector of R 2 defined by Rt={(12, , t )e R21 t2~tl}; RC=the extended real number system= { - co} u R u{ + co}, 394
due to the presence of state x and the input-value u". We will denote the function r(t/>(t, tOt Xo, u), u(t») defined for t~to unambiguously by Y(l o, Xo, u). Definition I is precise and yet very general. By a suitable choice of the state space, the state transition function, and the read-out function, it includes all common deterministic models used in classical physics, in circuit theory, and control theory. The axiom of determinism is the crucial one. It expresses at the same time a fundamental property of the state and an important restriction on the class of systems which qualify for dynamical systems in the above sense. It states that the initial state summarizes the effect of past inputs in the sense that for future responses it does not matter how the system was brought into this state; it also implies that the state and thus the output before some time are not influenced by the values of the input after that time. We are hence in effect restricting our attention to systems in which future inputs do not affect past and present outputs. The idea is simple: since all experimental evidence indicates that physical systems indeed satisfy this property of causality, we require this to be preserved in the model. It should be emphasized that the read-out function is required to be a memory.. less map in the sense that the output only depends on the present value of the state and the input. All dynamical effects (i. e., those phenomena involving memory) are required to be taken care of by the state. The above definition is commonly used in mathematical system theory (see, for instance, references [5, 6]). Although physicists have been groping for a similar concept for a long time, it is only for systems in which the input space consists of only one element (i. e., the autonomous dynamical systems of classical mechanics) that such mathematical structures have been introduced in a formal way. In the framework of Definition I the state at every moment completely describes the present situation. It is, however, impossible to deduce a priori, in physical terms, what will be the state. This, indeed, is a very difficult problem even for relatively simple systems, and it appears to be the cause for much of the reluctance of introducing this concept in physics. The approach which has been taken for describing materials with memory is to allow the outputs to be a function of the whole past history of the input. This is particularly prominent in the pioneering work of TRUESDELL, COLEMAN, and NOLL [1, 2]. Another approach is that of ONAT [3, 4] where the state is constructed in terms of observables. These two extreme points of view are particular cases of Definition I, but we see DO compelling reason to adhere to either of them. The first approach does not recognize the idea of equivalent histories, and the second approach will lead to difficulties when we consider isolated systems for example. III view of this dichotomy, it would appear to be useful to allow some time discussing these state space concepts further. Let us take the point of view that all the information the experimenter may obtain about a system is a table of input functions in dIJ versus the corresponding output functions in d§. The so-called problem ofrealization is to define a state space X and the functions lP and r in such a way that the resulting dynamical system in state space form generates the given input/output pairs by a suitable choice for the initial state in each tabulated ex395
periment. This problem has attracted a great deal of attention in the literature. Both the questions, "Does a state space realization exist?" and "What are the maps cP and ,1", have been examined. For the first question we mention the work of ZADEH [7] and for the second question the work by YOULA [8] and, especially, KALMAN [5], among others. The existence question essentially only requires a determinism postulate on the input/output pairs. The construction of 4J and r is understandably much more intricate but has been satisfactorily resolved for large classes of systems. In particular, there exists a very elegant solution to this problem for linear systems with a finite number of degrees of freedom. This material is considered to be of prime importance and can be found in a number of recent texts (e.g., [10]). We now consider an important particular case of this realization problem. Assume that Fis a given map from dIJ into CTjj satisfying the postulate of determinism which states that inputs u 1 , U2EtJIJ satisfying Ut(t)=U2(t) for t~to yield outputs Yl=Fu l and Y2=Fu2 which similarly satisfy Yt(t)=Y2(t) for t~to. Assume in addition that this map is stationary i. e., two inputs u., U2 eqj related by U.1 (t)= U2 (t+n yield outputs Yl =Fu 1 and j , =Fu 2 which are similarly related by Y1 (t)= Y2(t+T). The question is to realize Fby a dynamical system in state space form. The solution to this problem is by no means unique. One possibility is to consider the function f: R+ -+ U defined by f(s)=u(t-s) for s~O as the state at time t resulting from the input u. It is clear how the state transition function and the read-out function may be defined from here [11]. This state space realization is of course completely inefficient: in trying to store sufficient information about the past inputs, we decided to store the whole past input. The most efficient and natural state space realization of F is the one obtained by considering as the state at time t the equivalence class of those inputs up to time t which yield the same output after time t regardless of how the input is continued after time t. More specifically, in this realization we start with the space of functions f: R+ --+ U satisfyingf(s)=u( -9), s~O, for some ued/l. We then group these functions into equivalence classes by letting j'; "'!2 if Yl =Fu 1 , and Y2= Fu2 satisfy Yl (t)=Y2(t) for t~O whenever U1(-t)=/1 (t), U2(-t)=f2(t), and U t (t)=U2(t) for t~O. The latter realization is sometimes called a "minimal realization" and plays a central role in control theory [5, 10]. A similar idea has been proposed by ONAT [3, 4] in a restricted context. The point of view taken in this paper is that the state space realization is given, i.e., it has been inferred from previous considerations what the state space is. We do not demand minimality since, in our opinion, there is no compelling reason for doing so: minimality is very much a function of the class of experiments and observations which are allowed, is sensitive to modelling, and is not necessarily a good physical assumption. Neither do we adhere to the idea that the state is the whole past input since this point of view leads to nonsensical situations. Consider for example an electrical RLC network which has a given set of charges on the C's and fluxes through the £'8. Does it make sense to allow the stored energy of such a system to depend on exactly how these charges and fluxes came about? The whole question of what the state space of a physical system is requires much consideration. In this paper we have taken the easy way out by assuming that this has already been decided. 396
3. Dissipative DynamicalSystems In this section the concepts, which will be the basis for the further developments, are introduced. Assume that a dynamical system E is given together with a realvalued function w defined on U x Y. This function will be called the supply rate. Weassume that for any (t 1 , toleRt, ue U, andye Y, the function w(t)= w(u(t),y(t)) satisfies *
tl
JIw(t)J dt< 00, i. e., w is locally integrable. to
Definition 2. A dynamical system I with supply rate w is said to be dissipative if there exists a nonnegative function S: X-+- R+, called the storage function, such that for all (t" to)eRt, xoeX, and UEU, t1
J
S(Xo) + w(t)dt~S(Xl) to
where
Xl
=4>(t 1, to, xo, u) and w(t)=w(u(t), y(t)), with y=y(to, Xo, u).
The above inequality will be called the dissipation inequality. Note that
t w(t) dt~O with f indicating that the dynamical system is taken from a particular initial state to the same terminal state along some path in state space. This condition is in itself inadequate as a definition for dissipativeness but dynamical systems which are dissipative in such cyclic motions only are of independent interest. The approach taken here proceeds from the knowledge, from physical considerations, that the dynamical system is dissipative and thus that the storage function exists. The fact that this storage function is "defined" via an inequality requires further analysis. Central in this analysis is the question: "In how far is S defined by the dissipation inequality?" (The question is not so much "Does a storage function exist?" but rather" What can it be?") A crucial role will be played in the sequel by a quantity termed the available storage: it is the maximum amount of storage which may at any time have been extracted from a dynamical system. The notion of available storage is a generalization of the concept of "available energy" [11, 12, 13] studied in control theory and of "recoverable work." encountered in the theory of viscoelasticity [14, 15].
Definition 3. The available storage, Sa' of a dynamical system rate is the function from X into Re defined by
r
with supply
II
Sa(x) = sup- Jw(t)dt x'"
tt~O
0
where the notation x -+ denotes the supremum over all motions starting in state x at time 0 and where the supremum is taken over all uedll.
The available storage is an essential function in determining whether or not a system is dissipative. This is shown in the following theorem:
* The shorthand notation w(t) for w(u(t), y(t)) will be used whenever it is obvious from the context what x o' to' and u are. 397
Theorem 1. The available storage, Sa' is finite for all xeX if and only if r is dissipative. Moreover, O~SG~S for dissipative dynamical systems and SQ is itself a possible storage function. Proof. Assume first that Sa< 00: it will be shown that I is then dissipative. It suffices therefore to show that Sa is a possible storage function. Notice that S,,~O since S.(x) is the supremum over a set of numbers which contains the zero tl
element
o. =0). Consider now the quantity Sa(Xo) + Jw(t) dt, We have to show '0
that this quantity is not less than S.(Xl) whenever w is evaluated along a trajectory generated by an input u which transfers the state from Xo at to to Xl at t 1. The proof of this is quite simple although writing out the details is somewhat laborious. The idea is the following: in extracting the available storage from I when it is in state Xo we could first take I along the path generated by u, thus transferring I to Xl' and then extract the available storage with X in state Xl. This combined process is clearly a suboptimal procedure for extracting the storage originally present with I in state Xo. Formalizing this idea immediately leads to the desired dissipation inequality for S._ Assume next that r is dissipative. Then
'1
S(xo) + Jw(t)dt~S(Xl)~O to
which shows that S(xo)~ sup xo~
'1
Jw(t) d t = Sa(XO). 0
tl~O
Hence Sa< ex> as claimed. This ends the proof of Theorem 1. Theorem 1 gives a method which in theory may be used to verify whether or not a dynamical system is dissipative and this procedure does not require knowledge of the storage functions. In this sense it is an input/output test. Note that the theorem only states that the available storage may be the storage function. Usually it will not be the actual storage function. In fact, under certain additional assumptions (e.g., the Onsager-Casimir reciprocal relations) it may be shown that it will not be the actual storage function. This fact should be kept in mind when interpreting the results of [12, 14, 15]. A dynamical system which has the available storage as its actual storage function has the interesting (and unusual) property that all of its internal storage is available to the outside via its external terminals. It is convenient to introduce at this point the concept of reachability. This notion .is related to controllability and plays a central role in mathematical systems theory. Definition 4. The state space of the dynamical system I is said to be reachable from X-t if for any xeX there exists a t: , ~O and uedIJ such that x=cP(O~ t-
It is said to be controllable to x 1 if for any that XI =q,(t 1, 0, x, u).
l' X-I' XE X
398
u).
there exists a I 1 ~ 0 and a ue OIJ such
Theorem 1 emphasizes what happens when the system starts off in a particular state. One may similarly examine what happens when the system ends up in a particular state. We will therefore introduce the concept of required supply. This is done by letting the system start in a given state and by bringing it to its present state in the most efficient manner, i. e., by using no more supply from the outside than is absolutely necessary. The notion of required supply has been introduced in [11]. Although one could choose any point in state space as the initial state, it is most logical to assume that the system starts in a state of minimum storage.
Assumption. It will be assumed that there exists a point x*eX such that S(x*) =min S(x) and that the storage function S has been normalized to S(x*}=O. xeX
Definition 5. The required supply, S" of a dissipative dynamical system I with supply rate w is the function from X into Re defined by o
J wet) d t
S,(x) = inf x*""x
t-l
t-l~O
where the notation inf denotes * the infimum over all uedIJ and x*-+%
1-1 ~O
such that
t-l~O
x=q,(O, t_ l , x*, u). Deorem 2. (i) Assume that the state space of E is reachable from x _ 1. Then r is dissipative if and only if there exists a constant K such that o
J w(t)dt~K
inf
for all
XEX.
X-l-t,X '-1
t-l~O
Moreover, o
Sa(X- t ) + inf X-l-tX
J w(t)dt t-J
t-l~O
is a possible storage function. (ii) Let E be a dissipative dynamical system and assume that S(x*)=O. Then S,(x*)=O and O~SQ~S~S,. Moreover, if the state space I is reachable from x* then S, < 00 and the required supply S, is a possible storage function. Proof. (i) By reachability and Theorem 1 we see that I is dissipative if and only if Sa(x, 1)< 00. Any K~ - Sa(x- 1 ) will thus yield the inequality in part (i) of the theorem statement. It remains to be shown that o
Sa(x-t)+ inf X-I-+'x
J w(t) d t t-t
l-l~O
is a possible storage function. This function is clearly nonnegative. To prove that it satisfies the dissipation inequality, consider the following idea: in taking the
system from X-t to Xl at t 1 , we can first take it to Xo at to while minimizing the supply and then take it from X o at to to x I at t 1 along the path for which we are to
* Thisnotation, along with the similarone introducedin Definition 3,will be usedthroughout. 399
demonstrate the dissipation inequality. This results in a suboptimal policy for taking the system to Xl and the formalization of this procedure leads to the desired dissipation inequality. (ii) That S,.(x*) = 0 is obvious. Moreover, any uedJt resulting in a transfer from o x* at t: t to x at 0 satisfies S(x) ~ w(t) dt by the dissipation inequality. The
J
t-l
inequality S,.(x)~S(x) follows by taking the infimum at the right-hand side. Assume now that the state space of I is reachable. Then clearly S,.~ 00. It remains to be shown that S,. is a possible storage function. This, however, follows from (i). It
It is an immediate consequence of the normalization S(x*)=O that for a tl
dissipative system any motion starting in x* at to satisfies
J }1J(t) dt~O
for all
to
uet1Jt and tl~tO' Thus the net supply flow is into the system. This idea has been proposed [16, 17, 18, 19] as a definition of passivity. It has the advantage of being an input/output concept which does not involve introduction of state space notions. However implicit in this approach is the fact that one knows the state of minimum internal storage. Note that the required supply is in general a function of Sand x*. Usually, however, the point of minimum storage is a unique a priori known equilibrium point which may thus be shown to be independent of S and this ambiguity does not arise. Remarks. 1. Under the assumptions of reachability from 'X- 1 and controllability to X t we always have the following inequalities for a dissipative system: 0
t1
S(x1)+sup - Jw(t)dt~S(x)~S(X-l)+ inf X .... Xl
0
X-l-X
tt~O
S w(t)dt. t-l
t-I~O
Note however that the lower bound on S thus obtained is itself in general not a possible storage function because it need not be nonnegative. 2. Often a state space model of a dynamical system is constructed on the basis of an input/output description. Particularly important realizations are the minimal realization mentioned earlier and the realization in which the state is the whole past history. It is quite simple to associate a storage function with these realizations when one has determined a storage function on a particular state space X. For example, defining S(u(-OOt O»)=S(x(O)) leads to a storage function on a state space which keeps track of the whole past input history. The available storage function of these realizations will in fact agree on that part of the state space which is reachable along some past history. Assuming that for t sufficiently small every element of dlI is equal to a fixed constant u* (typically the zero element of some o vector space) such that w(u*, y*)=O and that J w(u(t), Yet)) dt exists and is -ex;
nonnegative (thus the state at "t = - 00 " is assumed to be the state of minimal storage), then we may actually also evaluate the required supply for the realization in which the state keeps track of the whole past history. This does not require o
Jw(u(t),
any infimization and is simply equal to
-00
400
y(t)) dt. It may in principle
be different for every history. Moreover, the dissipation inequality holds with equality for this storage function. (This fact does not conflict with Theorem 4 since this realization will never be controllable.) If one works with the minimal realization then one may associate a storage function by defining S(XmiJ=S(X) where x is a state in the equivalence class Xmin. After elimination of the non-reachable states, one thus divides the state space X into equivalence classes and defines the storage to be the storage of an arbitrary element in this class. The available storage functions of these realizations again agrees on that part of the state space which is reachable along some past history. The required storage may now take on more values in X than in Xmin. An interesting consequence of the above reasoning is that the notion of available storage is defined purely as an input/output concept for states which are reachable. Thus, taking equivalence classes as the state or the whole past history as the state leads to the same value for the available storage function. This reemphasizes the importance of Theorem 1 as an input/output test for dissipativeness. There is an interesting paper by DAY [33] which has used the concept of available storage (or "useful work" as it is called in [33]) in setting up an axiomatic theory of thermodynamics. Although the technical details are quite different, the ideas exploited in that paper appear to be very much along the lines of those on which Theorem 1 is based. To summarize the above results, we have shown that the storage function of a dissipative dynamical system satisfies the a priori inequality Sa~S~Sr, i.e., a dissipative system can supply to the outside only a fraction of what it has stored and can store only a fraction of what has been supplied to it. The available storage always satisfies the dissipation inequality, as does the required supply for systems with a state space which is reachable from a point of minimum storage. (This show that the above inequality is the best of its type.) Of course not every function bounded by this a priori inequality will be a possible storage function. It appears to be difficult to state other general properties of the set of possible storage functions. One interesting property is its convexity:
Theorem 3. The set of possible storage functions of a dissipative dynamical system forms a convex set. Hence (XSa+(I-~) Sr, O~(X~ 1, is a possible storage function for a dissipative dynamical system whose state space is reachable from x*. Proof. This theorem is an immediate consequence of the dissipation inequality. Ii The ultimate test for a theory of dissipative systems is whether or not there exists a (possibly idealized) "physical" system which realizes the input/output exchange process and which has the desired storage function. Such a synthesis program based on interconnecting ideal elements may in fact be carried out for linear systems with a finite number of degrees of freedom and quadratic supply functions. Some results in this direction will be indicated in Part II. We now proceed with a few remarks regarding the evaluation of the available storage and the required supply:
(i) If the state of minimum storage x* is an equilibrium point corresponding to the constant input u*et¥J (i.e., lfJ(t, 0, x*, u*)=x* for all t~O)andif w(u*,y*) =0, 401
then
o
Sr(x)= lim
J w(t)dt;
inf x·.... x
t-J-+-OO
t-l
(ii) If for all xeX there exists a uet¥J such that w(u, y)~O ii.e., the external termination may always be adjusted so that the supply flows out of the system), then tl
SQ(x)= lim sup Jw(t)dt; x .... 0
tl-+<X)
(iii) The concept of required supply assumes that there exists a point x*eX such that S(x*) = min Sex). There need however not be a point of minimum x·.x storage. One may then define S,(x) by considering a sequence of states {X,.} with lim S(xll) = inf Sex) and define "-+<Xl .xeX o S,(x)= limSr.,.(x) where Sr,,.(x)= inf J w(t)dt. n-+co
Xn-+,x t-I t-t~O
We now show how to treat conservative systems as particular cases of dissipative dynamical systems. DefiDitioD 6. A dissipative dynamical system r with supply rate wand storage function S is said to be lossless if for all (iI' to)eRt, xoeX, and uetJIJ f1
S(Xo) + Jw(t)dt=S(Xt) to
where XJ =tjJ(t 1 , to, Xo, u). The following theorem is immediate:
Theorem 4. Let I be a lossless dissipative dynamical system and assume that o. If the state space is reachable from x* and controllable to x*, then Sa= S, and thus the storage function is unique andgiven by S(x*) =
o
S(x)=
J w(t)dt
'-1 with any t:
t ~O
and ue'W such that x = 4J (0, t _ j,x*, u), or S(x)= -
with any
tl~O
'1
Jwet) d t
o
and uetfl such that x*=l/J(t h 0, X, u).
The condition Sa=Sr which implies uniqueness of the storage function is in itself not sufficient to imply losslessness. We could call such systems quasi-lossless since they may be transferred between states without dissipation provided; however, this transfer is executed optimally. An arbitrary transfer instead is expected to involve dissipation. An interesting property of dissipative dynamical systems is the following: tl
(i) For dissipative dynamical systems with x(to)=x*, Jw(t) k(t) dt~O for bounded functions k with k(t)~O and k(t)~O; to 402
an
(ii) For lossless dynamical systems with x(to)=x* and 4>(t 1 , to, x·, u)=x*,
J'1 w(t) k(t) dt~O for
all bounded functions k with k(t)~o.
to
These inequalities formalize the idea that for a dissipative system with no initial storage the supply flows into the system before part of it is recovered whereas in addition all of it gets recovered in a lossless system. These expressions generalize similar inequalities obtained in [20, 21, 22]. We conclude this section with a discussion of the concept of a dissipation function.
Definition 7. A real-valued function d: X xU...., R is said to be the dissipation rate of a dissipative dynamical system I with supply rate wand storage function S if for all (/1' to)eRt, xoeX, and ueau It
S(x o)+ S(w(t)+d(t»)dt=S(x t ) to
where Xl =q,(tt, to, Xo, u). It is clear that dbeing nonnegative implies dissipativeness. Moreover, since the dynamical system E is lossless with respect to the new supply rate (w+d) it follows that the dissipation rate d uniquely determines the storage function S provided the appropriate reachability and controllability conditions are satisfied. The converse, i. e., that dissipativeness implies the existence of a nonnegative dissipation rate is also the case under some technical smoothness conditions. The set of dissipation rate functions for a given dissipative system forms a convex set.
Remarks. Note that if S(c/J(t, 0, x, u» is differentiable at t=O for all xeX and ueefl, then the dissipation inequality is equivalent to
si». u)~w(r(x,
u), u)
. This definition is t t=O more standard but slightly less general than the one proposed here. The dissipation function d is then given by d=S-w. for all xe X and
UE
U where
S denotes
dd s(q,(t, 0, x, U»)I
4. Interconnected Systems The main result obtained in the previous section yields an a priori bound on the storage function of a dissipative dynamical system. Moreover these bounds themselves define possible storage functions and the storage function is thus uniquely determined by the dissipation inequality if and only if the required supply equals the available storage. This situation is the exception and as a rule there are consequently many possible storage functions. If we consider the implications of this result to physical systems which dissipate energy or to thermodynamic systems, then we conclude that experiments on a physical system will usually only give bounds on the stored energy function or on the entropy function. This result is unexpected in the sense that in classical physical systems this ambiguity does not 403
arise: we thus expect that the additional structural assumptions implicit in such systems will greatly reduce the number of possible storage functions and often render it unique. In this section' we examine one such possibility: it will be shown that by considering a given dissipative system as an interconnection of dissipative subsystems the number of possible storage functions is greatly reduced. Other qualitative assumptions (linearity, reciprocity, etc.) on the system will be investigated m Part II. The idea of an interconnected system is actually quite simple, albeit somewhat difficult to formalize. We start with a collection of dynamical systems {E«} where rx ranges over some given index set A. For simplicity we will assume that A is a finite set. More general interconnections involve the introduction of a measure on A which would take us somewhat astray. Each Erl is determined, as in Definition 1, by a septuplet {UtI' dIIa, Ya, dJlrJ.' XC!, tPa, r(%}. We assume that the inputs and outputs of each dynamical system L« are divided into two groups, i.e., U«=U:xU~,
OJta=tJlt:xtW~,
Ya=Y:xY~,
and
'!!Ia.=~exdJI;
when the superscripts e and i stand for the adjectives external and interconnecting.
interconnecting system
Fig. 1. Illustrating the concept of an interconnected system
Next we introduce the notion of an interconnecting function which is simply a function* f: II (U;x y;)~ V where V is some vector space, and of the interaeA
connection constraint which states that I(
n
(U~ x Y;))=O: it is thus a relation
aeA
between the instantaneous values of the inputs and the outputs. The idea of an interconnected system is illustrated in Figure 1 and indicates that the external inputs u: are given but that the interconnecting inputs u~ are to be determined implicitly. More precisely, given any U:E~: and XaEXa, lXEA, we may attempt to solve the implicit equations
f( II (u~(t) x r~(cPa(t, to, x(%(u:, u~)), (u: (t), u;(t)))))=O,
t~ to,
(leA
for some set of the functions u~e~rl. This equation is of course not necessarily uniquely solvable and this fact needs to be assumed explicitly. Notice that only
* The notation n stands for the Cartesian set product. «eA
404
the values of u~(t) for t~O enter in the equations. Hence we can only hope these equations determine u~ on the half line t~to. However it may be expected that under reasonable assumptions there will be a map from the set of statesandexternal input values into the set of internal inputs and outputs which determines a solution to these equations. We will thus assume explicitly that the interconnected system, denoted by Ia//, is well-posed in the sense that it defines unique functions c/J and rand
n
ClSA
fjJ~: nX«x n'lU:~U~, cceA
CLEA,
l%eA
such that: (i) the septuplet
IIl U:, Il lilt:, Il Y:, Il qy:, Il XlI' 4>, r} defines a dynamical «eA
lIeA
«eA
«eA
l%eA
system in the sense of Definition 1; (ii) the function u~(t)=t/J~(4>(t, to,
nxcz' nu:), nu:(t») defined for czeA
l%eA
t~to is
«eA
the restriction to [to, (0) of a function in Olt~;
(iii)
iP(t, to,
n X«' nu:)= n
/ZeA
(iv)
(leA
«n-, no- nr:(x
a,
tleA
«eA
lIeA
Xa,
(leA
(u:, u~») with u~ as in (ii);
(U:, t/J~(nX/Z, «eA
ne»
«eA
(V) f(n(t/I~(nXcz' nU:),r~(x«,(u:,t/J~(nXa' nU:»))=O. tzeA
aeA
ilEA
ileA
aeA
It is easy to verify that the above conditions formalize the intuitive conditions one expects a well-posed interconnected dynamical system to satisfy. Examples of interconnected systems will be given in Section 7. Note that although the interconnected system may have many state space realizations we are insisting on using the one with state space the Cartesian product of the state spaces of the individual subsystems. This is indeed a natural thing to do since the interconnection itself introduces no memory, We now introduce the concept of dissipation in this framework. Assume therefore that each dynamical system I« has associated with it an externalsupply rate, defined on U: x Y: and an interconnecting supply rate, w~, defined on U~ x
w:,
y;.
Definition 8. Consider the dynamical systems I a with interconnecting supply rates w~_ Then the interconnection defined by the interconnection constraint f( (u~ x y~») = 0 is said to be neutralif all u~ and y~ satisfying this equality yield
n
aeA
, i i s: wa{u tl , YtI)=O.
~
tleA
In terms of Figure 1, an interconnection is thus said to be neutral if the interconnecting system itself is lossless with respect to the supply rate w~. Thus the
L
«eA.
mere interconnection does not introduce any new supply or dissipation. One thus expects the dissipativeness of the interconnected system to be a consequence of the dissipativeness of the individual subsystems. That this is indeed the case is shown in the following theorem:
Theorem 5. Let I a , ctEA, be a collection of dissipative dynamical systems with supply rates wtz= + w~ andstorage functions Sa. Let f be a neutral interconnection
w:
405
constraint. Then the interconnected system I ==
n
Iflll
is itself dissipative with
fleA
respect to the supply rate W=
L w: and S== L S(J is a storage function for
fleA
I.
SEA
Proof. Summing both sides of the inequality
'1
Scr(Xfl(tO»)+ J(w:(t) + w~(t»)d t~S(J(X«(tl») to
over fleA and using the assumption inequality for I. II
L w~(t)=O leads to the desired dissipation creA
The above theorem is intuitively obvious. Note however that by considering only storage functions which are additive in the sense that S( xcr)== Sa.(xa) one «eA a.eA obtains only part of the admissible storage functions for E. It is easy to see that since the interconnection introduces additional constraints on the inputs Urt == (u:, u~), we always have the inequality SQ~ L Srt,Q~ L Sfl,r~Sr, with equality holding
n
ClsA
L
(leA
exceptionally. Thus one obtains a unique additive storage function for the interconnected system if and only if Sa.,a=Sfl for each ~EA. In many physical systems encountered in practice, e.g., in lumped electrical networks or in continuum systems, one may postulate a priori that the system is an interconnection of dissipative systems and use this qualitative property to describe the system in terms of "local" states, i. e., to take the state space X = Xa '
t
n
nx )= aeA
and furthermore to require that the storage function be of the type S(
L Ss(x,J.
a
aeA
This natural requirement on the storage function of a dissipative
CI
dynamical system which consists of a family of dissipative systems interconnected by means of a neutral interconnection serves to reduce greatly the number of possible storage functions. This requirement leads to a unique storage function whenever it is possible to regard r as the interconnection of lossless systems with memory (capacitors and inductors, elastic systems) and a dissipative system without memory (resistors, friction elements). The lossless part possesses a unique storage function by Theorem 4 (under the additional hypothesis of reachability and controllability) whereas the dissipative part does not contribute to the storage since its state space is the empty set. The storage of the original system is thus given by the storage in the lossless subsystem and is consequently unique. In concluding this section we remark that the above method of considering interconnected systems is implicit in most treatments of dissipative systems. It is based on a qualitative assumption on the system (the idea of "simple" materials) and sometimes it results in the uniqueness of the storage function. This is however by no means always the case, and typical examples of areas where this nonuniqueness remains are linear viscoelasticity and the modern treatments of materials with memory in continuum mechanics and thermodynamics where the nonuniqueness of the storage function at the elementary particle level remains. In other words, one has to make more assumptions (or, equivalently, obtain more knowledge about the physics) in order to derive the storage function (internal energy, entropy, etc.) from the constitutive equations. 406
S. Stability
In this section we examine the stability of dissipative systems. As is to be expected, only some technical conditions are required in order for dissipativeness to imply stability of an equilibrium at a local minimum of the storage function. We shall be concerned with the stability of an equilibrium state, and in order to make this study meaningful weneed to isolate the system from its environment. Moreover, since stability is concerned with convergence the concept of a distance function on the state space needs to be introduced. Assume therefore that the following assumptions hold: (i) The system is isolated, i. e., the input space consists of one element only. In order to preserve stationarity we assume that this element is the constant function u(t)=u*; (ii) x*eX is an equilibrium point, i.e., 4>(t, to, x·, u*)=x* for all
t~to;
(iii) X is a subset of a Donned space and II /I denotes its norm; (iv) cf>(t, to, xo, u*) is continuous in t for (v)
t~ to;
w(u*, rex, u·»)~O for all xeXin a neighborhood of x*.
The following stability definition is a standard one in the context of Lyapunov stability theory [23]:
Definition 9. The equilibrium point x* of r is said to be stable if given 8>0 there exists a ~ (8)> 0 such that II Xo - x* II ~ ~ implies that
114J(t, to, Xo, u*)-x* II ~e
for all
(~to.
A very useful method for proving stability is by means of Lyapunov functions. The notion of a Lyapunov function is introduced in the following definition. It is a slight variation of the usual definition: Definition 10. A real-valued function V defined on the state space X of r is said to be a Lyapunov function in the neighborhood of the equilibrium point x* if (i) V is continuous at x*; (ii) V attains a strong local minimum at x*, i.e., there exists a continuous function a: R+-+R+ with a(u»O for 0->0 such that V(x)-V(x*)~ ~(II x-x*'D for all xeX in a neighborhood of x*; (iii) V is monotone nonincreasing along solutions in the neighborhood of x*, i. e., the real-valued function V(4> (t, to, Xo, u*») is monotone nonincreasing at t = to for all Xo in a neighborhood of x*.
It is a standard exercise in (8, b)-manipulations to show that an equilibrium point x*eX is stable if there exists a Lyapunovfunction in the neighborhood of x*. This leads to the following theorem:
Theorem 6. An equilibrium point x*eX of a dissipative dynamical system I is stable if the storage function S is continuous and attains a strong local minimum at ."(*. Moreover S is a Lyapunovfunction in the neighborhood of x*. 23 Arch. Rational Mech. Ana1., Vol. 45
407
Proof. It suffices to show that S(cP(t, to, xo, u*)) is monotone non-increasing at t=to if IIxo-x*U is sufficiently small. By the dissipation inequality S(t/J(t, to, xo, u*») is indeed monotone nonincreasing for all t ~ to and for all x*. II Note that S attains a strong local minimum at x* if Sa does-consequently this condition may usually be verified without explicit knowledge of S. Note also that the fact that x* is an equilibrium point itself follows if (i), (iii), (iv), (v) are satisfied and if S is continuous and attains a strong local minimum at x*. The consideration of the storage function is an extremely useful tool in stability investigations and by properly choosing the supply rates one may indeed obtain an interpretation for most of the existing stability criteria. In constructing a storage function it is natural to proceed to the evaluation of either the available storage or the required supply. These however lead to variational problems and it is only in exceptional circumstances that one may solve such problems, particularly if the dynamical system E is nonlinear. The concept of interconnected systems becomes in fact very useful in this context: it allows one to construct storage functions which correspond to neither the available storage nor the required supply, and which may be constructed by solving variational problems for-presumably less involved-subsystems of the original dynamical system I. This procedure will be illustrated in Section 7.2. One may refine the basic result of Theorem 6 in several directions. Some of these are briefly discussed below: (i) roughly speaking local minima of the storage function define stable equilibria and vice versa;
(ii) under appropriate additional hypotheses one may conclude that all trajectories actually approach the point of minimum storage. These additional hypotheses require the system to be strongly dissipative in the sense that no trajectory (other than the equilibrium) is free of dissipation. This strong form of dissipation is studied in [11]. We note here that one will usually not obtain S to be negative definite but merely semi-definite. The so-called invariance principles [24, 25] are thus very useful in establishing asymptotic stability in this context;
(iii) local maxima of the storage function will define an unstable equilibrium if all trajectories in its neighborhood involve some dissipation; (iv) if w(u*, y) = 0 for all OJ! E Y, and if the system is lossless, then local minima and maxima of the storage function define stable equilibra.
6. Nonstationary Dynamical Systems All of the above theory and results have been based on the hypothesis that the dynamical system under consideration is stationary. This stationarity has been postulated on two distinct levels:
(i) it has been assumed that the dynamical system I is itself stationary, i.e., the constitutive equations defined by the maps ljJ and r are invariant under shifts of the time axis; 408
(ii) the storage functions have been assumed to be time-invariant, i. e., the function S did not involve an explicit time dependence. Although it may not seem so at first glance, assumptions (i) and (ii) are separate since (i) mainly refers to input/output stationarity whereas (ii) supplements this with internal stationarity. There are important types of physical systems (e.g., rotating electrical machines) which are externally stationary but internally timevarying. We view stationarity postulates as an a priori qualitative assumption imposed on the mathematical model of the dynamical systems under consideration. In this section we will indicate the modifications required to extend the above definitions to time-varying systems. Once the conceptual framework is appropriately expanded, one may indeed generalize the results to the time-varying case without difficulty.
I. The following definition generalizes Definition 1. In contrast with most similar definitions which have appeared in the literature we allow for the state space itself to be time-varying. Definition 1'. A (continuous) dynamical system, L, is defined through the sets {Ut, eft, Yt,~, X,}, teR, and the maps 4> t l ' to' (t 1 , to)eRt, and r., teR. These satisfy the following axioms: (i) d/t is called the input space and consists of a class of functions u(t)'1 te R, taking their values at time t in the set of input values Ut ; (ii) 0/1 is called the output space and consists of a class of functions y (t), teR, taking their values at time t in the set of output values Yt ;
(iii) X t is called the state space at time t e R; (iv)
4>t, to is called the state transition function and maps Xto x Olt into Xli' 1t satisfies the analogous axioms of (iv)a' (iV)b' and (iv), of Definition 1;
(v) r t is called the read-out function and is a map from X, X U, into Yt ;
(vi) the function r,(4J(t, to, Xo, u), u(t)) defined for t~ to is the restriction to [to, (0) of an element of dJ/. The solution of the problem of state space realization in terms of equivalence classes goes through unchanged.
2. A (time-varying) dynamical system with supply rate at time t W,: U, X Yt ~ R is said to be dissipative if there exists a nonnegative function Sf: X t -+ R +, called the storage junction, such that tl
Sto(X o) + SWt(t)dt~Stl(X1)· to
The available storage is defined by t1
Sto,a(x)= sup x'-'
Jwt(t)dt, to
t\~to
whereas the definition of required supply necessitates again the notion of a point of minimal storage. Assume then that x~eXt minimizes St(x) over xeXt and as23*
409
sume in addition that S,(x~)=O (this postulate now involves more than simply adjusting an additive constant). The required supply then becomes Sto,,(x)= inf
'0
J wt(t)dt.
X·(t1)-+X t-1 t-l~to
The results of Theorems 1and 2 follow with the obvious modifications in notation. The available storage and the required supply are thus bounds on the storage functions and are themselves possible storage functions. 7. Applications In this section we shall present a series of applications whichserveto illustrate the previous theoretical developments. 7.1. Systems with a Finite Number of Degrees of Freedom Consider the dynamical system described by the set of first order ordinary differential equations %=/(x, u), y=g(x, u)
and assume that the supply function is given by w=(u, y)=u' y
(prime denotes transposition).
Here, xe k", u, yeR"', and it is assumed that / and g are Lipschitz continuous in oX and u jointly. It is well known that this implies that the above differential equation has a unique solution for any x (to)eR" and any locally square integrable u(t). Moreover the resulting functions x(t) and y(t) are themselves also locally square integrable. The above differential equation thus describes a dynamical system in the sense of Definition 1 with U= Y=R"', X=R", and ft=~ the locally square integrable R'ft-valued functions defined on R. The differential equation itself defines the state transition map 4J whereas the relation y=g(x, u) describes the
read-out function r. Note also that the supply function is locally integrable for uec1Jt and ye~. The problem at hand is (i) to determine conditions on I and g which make the dynamical system under consideration dissipative with respect to the given supply function and (ii) to discoverthe possiblestorage functions. If we restrict ourselves to sufficiently smooth storage functions then we are asking to find those func.. tions S: R"~ R + satisfying d
lit S(x) = 17. 8 (x) ·/(x, ")~(1I, y)=(u, g(x, u) for all xeK' and ueR"'. BROCKETT [26] has in fact proposed this as a definition of passivity. This equivalent statement hardly solves the problem. The question of dissipativeness is by Theorem 1 equivalent to whether or not ex;
inf
J(u, st». u»
lIe4l'0 0
410
d t,
subject to the constraints %==/(%, u); x(O)==xo, is finite for all %oeK'. The value of this infimum(whichis seen to be nonpositive by taking_ == 0) yields the negative of the available storage function. This variational problem and the analogous one involved in the computation of the required supply are standard problems in optimal control and these techniques will be used in Part II to obtain some specific answers to the above questions. At the level of generality posed here it is impossible to obtain necessary and sufficient conditions on f and g for dissipativeness, but some interesting specialcases offer a great deal of further insight: (i) Consider the particular case (corresponding to elastic systems and to capacitive networks):
x=u;
y=g(x).
In this case it is convenient to derive the conditions for dissipativeness directly from the dissipation inequality. Restricting ourselves again to sufficiently smooth storage functions (the available storage will in fact be smooth as a result of the assumption onl and g madeearlier),weseethat the dissipationinequalitydemands that there exists an S: g' -+ R+ such that
for all ueR'" and xeJr'. This is the case if and only if the function g'(x) is the gradient of a nonnegative function. It is well known that this requires iJ:' (x) iJ
~J(x) • This condition may be obtained in a different manner by nOtici~ that Xi
t
&1
Iw(t) dt= Jg' (x) dx.
The integral on the right is bounded from below for a o -. given %0 and Xl only if it is path independent which in tum requires g' (x) to be the gradient of a real-valued function. The necessary and sufficient conditions for dissipativeness may thus be expressed in terms of ,(x) by: (i) iJ g,(X)
oX j
s
(ii) the path integral P(x)=
J,'(x) dx is bounded from below. %.
Here, x* is arbitrary and the function P differs from S .only by an additive constant. It thus follows that the system is dissipative if and only if it is lossless. The storage function is thus unique and plays the role of a potential function since it determines the dynamical equations by
x=,,;
y=J7;S(x).
Note also that in this case one obtains reciprocity (condition (i») as a result of dissipativeness. This is by no means a general property of dissipative systems
however. 411
(ii) If we add "resistive" terms to the equations of motion studied in (i) then we obtain the dynamical system
x=u;
1=&'1(X)+&'2(8).
If we assume (without loss of generality) that g2 (0)=0 and concentrate again on sufficiently smooth storage functions then the dissipation inequality demands that there exists an S:K'-+R+ such that J7zS(x)·"~
for all xeR" and ueRm.
This inequality is satisfied if and only if for all xeRn
J7z S(x) = gl (x)
and
for all ueR"'.
The conditions for dissipativeness are then those obtained in (i) augmented by the additional requirement
(,,»
%
J11 (x) d x. The
by the path integral
dissipation function (also unique) is given
~
by dtx, u)= (a, 12 (a». Notice that as a consequence of dissipativeness, we obtain reciprocity of the "elastic" terms (81 (x» but not of the "resistive" terms (g2(8». The system studied here may be considered as the interconnection of the three systems: II: %1="1;
11=gl(Xl)
Z'2: 12=12("2)
L3: 13=14= -,,;
with
w1 = ( " t, 11) '
with
W2=(U2,Y2),
Y="3 +"4 with w3 =("3' )'3)+("4, '4)+(U, y)
with the neutral interconnection constraint
The storage function (but not its uniqueness) follows directly from here. The variables with subscripts represent the interconnecting inputs, outputs, and supply rates. (iii) Consider the system described by the equations (uz is scalar-valued):
Xl ="1;
11 = gl (X1, X2)
%2=12(X1,X2,U2);
12=g2(Xl,X2,U2)·
A simple calculation shows that dissipativeness implies that V%t
S(X1' X2)=&'~ (Xl' %z)
and
V%2 S(X1' %2) -/2(Xl' X2' U2)~U2 g2(Xl, X2' u 2 ) · for all Xl' X2' and Uz. Thus only part of the dependence of the storage function on the state vector is determined by the dynamical equations. The above dynamical system is a particular case of the one studied by COLEMAN [27] and GURTIN [28] (see [2], Chapter 3). These authors obtained in fact 412
very similar results. It should be realized however that for dynamical systems described in this amount of generality one needs a lot more information about the physics of the situation in order to obtain a unique storage function. 7.2. Stability of Feedback Systems
Consider the dynamical systems Z, and X2 and assume that U1 = U2 = Yt = Y2 are inner product spaces. Assume now that :E 1 and ~ 2 are interconnected via the constraint U2 = Y1 and Ul = - Y2. This results in the feedback system shown in Figure 2.
Fig.2. Feedbacksystern
The theory of dissipative systems discussed above offers a powerful method for investigating the stability of this feedback system. Assume that we associate the supply rate Wl(U 1 , Yt) with II and the supply rate W2(U2'Y2) with I 2 If W t and W2 are such that Wl(U,Y)+W2(Y' -u)=O for all u and Y, then the feedback system may be considered as an interconnected system with the neutral interconnection constraint: U2=Yl' U1 = -Y2' Thus in order to prove stability it then suffices to show that It is dissipative with respect to Wt and that E2 is dissipative with respect to W2 (or, equivalently, with respect to CXW 2 for some cx>O). It may be verifiedthat essentially all of the frequency-domain stability criteria which have recently appeared in the literature [17, 18] are based on this principle. Particularly important choices of the supply rates are WI = II U 1 1/2 -II Yl 11 2 , w2 = ll u21 12 - II Y2112 ; w1 = ( uJ, Yt ) , W2=(U2'Y2); and wt = ( u t + aYt , u 1 + bYl ) ' 1
W2 = -ab ( U2 -
~ Y2' U2 - ~ Y2)'
The stability theorems resulting from these
choices of the supply rates are known as the small loop gain theorem, the positive operator theorem, and the conic operator theorem. The interpretation of these stability principles in terms of dissipative systems gives further insight in these results and unifies the existingconditions. As an example, consider the autonomous dynamical system described by the set of first order ordinary differential equations: I: x=Ax-B/(Cx) m, where xeK',!:Rm--.R and A, B, and C are matrices of appropriate dimensions. We assume again thatj is Lipschitz continuous. Letf(O)=O; then the trajectory x(t}=O is a solution to this differential equation and the stability properties of this solution have been the subject of a number of recent papers in the control theory literature. Particularly the construction of Lyapunov functions is a matter of great practical importance. The best known result in this area is the so-called 413
Popov criterion [29, 30] which answered a long-standing question known as the "Lur'e problem". We will reproduce this result for a representative case using the theory of dissipative systems. In doing so we obtain a wholeclass of Lyapunov functions in a systematic manner and extend the results recently obtained in [11]. We begin with viewing this dynamical system as the feedback interconnection of two dynamical systems, namely: II: xt=Axt+BUt;
Yl=CAXt+CQXt+CBul
x2 :
Y2=/(X2)
and
X2=-Q X 2 +
U2 ;
with the interconnectionconstraint III = - '2;"2 =Yl. The matrix Q is an arbitrary (n x n) matrix which features in the conditions for stability. It is clear that the above interconnection constraint defines a neutral interconnection with respect to the supply function WI = <"1' '1) and W2 = ("2' '2). This interconnection leads to a "closed" system since we only have interconnecting, but no external, inputs and outputs. It is a simple matter involving only algebraic manipulations to show that the interconnection of Xl and X2 via the given interconnection con.. straint yield Xl(t)=X(t) and X2(t)=CX(t) provided the initial conditions are chosen as %1 (0) =x (0) and X2(O)=CX(O). The philosophy behind this equivalence is shown in Figure 3.
, ~2(o)
Fig. 3. Illustrating the interconnected system studied in Section 7.2
We now postulate the conditions for stability. These are:* (i) {A, B, C} is a minimal realization of G(s)=C(ls-A)-l B
* Re A.{A} denotes the real part of an arbitrary eigenvalue of A, the matrix inequality p~ 0 indicates that the Hermitian matrix P is nonnegative definite, and "minimal realization" is system theory jargon which will be explained in detail in Part II. Positive real functions have been studied, particularly in the context of electrical network synthesis, and will be discussed in Part II.
414
(ii) (ls+Q)G(a) is a positive real function of s
(iii) I is the gradient of a nonnegative function, i. e.,
11' (e) dtJ'~O for all zeR'" s
and the path integral
(iv) j'(a) Qcr~O for all tieR"'. This stabilityclaim will be verified with the aid of a suitable Lyapunov function. The idea behind the above conditions is that they make, both Eland I 2 into dissipative systems. In fact, conditions (i) and (ii) ensure that II is dissipative with respect to Wt =("1' )'1). The available storage x' Qax and the required supply x' Q,x are positive definite, quadratic forms. These functions and the other possible storage functions which are quadratic will be the subject of study in Part II of this paper. In order to verify that E2 is dissipative with respect to W2 = ("2' '2)' we shall use conditions (iii) and (iv), Consider thus T'1.
J
'1= -inC ("2,Y2)dt T1
subject to %2= -(2%2+"2; 12=/(%2). We may eliminate integral in terms of X2. This yields
U2
and 12 from this
Note that the first integral is independentof path by (iii) and the second one has an integrand which is always nonnegative by (iv). This last integral may be made arbitrarily small in the evaluation of the available storage and the required supply. The dynamical system X2 is thus dissipative and its storage function is uniquely given by Z'1.
S2(X2)= Jf'(tI)dtl~O. o
The interconnected system is hence dissipative and has Sl(Xt)+S2(X2) as an admissible storage function. Restricted to initial conditions %2(O)=CX 1 (0), this
1/' (tJ') dtJ' is nonincreasing along Cz
statement implies that the function x' Q.x+
solutions of I whenever Q.=Q; defines a quadratic storage function of Et • Since Q.==Q;~Q.=Q~~81 for some 8~O it follows that this Lyapunovfunction establishes the stability of the solution %(t)==O. By strengthening condition (ii) to include Re A {A} <0 we may in fact show, using this Lyapunov function,that all solutions approach their equilibrium solution s == 0 as t ... 00. 415
7.3. Electrical Networks
Consider an electrical network with n external ports and with a number of internal nodes and branches. We shall denote the external voltages and currents by the n-vectors Vand 1, respectively. Assume that the network contains resistors, capacitors, inductors, and lossless memoryless elements (e.g., transformers, gyrators, etc.) which need not be specified further. Let nR' nc, and nL denote the number of resistors, capacitors, and inductors, and denote the voltage across the elements and the current into these elements by the nR-vectors VR , I R' the nc: vectors Vc, Ie, and the nL-vectors VL , IL respectively. We take the sign conventions shown in Figure 4.
lossless network without memory
Ik+
v; _
n
}
external ports
'----r--' nL inductor ports ,----A---.
~
+
Vl k
-I
llk~ Lk
Fig. 4. The electrical n-port considered in Section 7.3
We now turn to the question of what should be considered inputs and outputs. This is a somewhat annoying question since what are the most convenient variables to work with depends on the network under consideration. In fact it has recently become apparent that the so-called scattering representation [31] (input = v+pi; output=v-pi, p>O) is by and large the most convenient model to consider. We shall consider here a somewhat simpler case and assume that [34] (i) V is the input and I is the output; 416
(ii) the characteristic of the k th resistor is given by
This leads to the relation VR=R(IR)IR. (iii) the characteristic of the k th capacitor is given by lc,,=Ck(Vc,)
d;~"
with
This leads to the relation Ic=C(Vc)
Cl~8>O.
d~c.
(iv) the characteristic of the kth inductor is given by VL,,=Lk(IL,,)
d~~"
with
This leads to the relation VL=L(IJ
Lk~Il>O.
~:.
(v) the part of the network which neither involves dissipation nor memory defines an instantaneous relation from the voltages across the external ports, V, the currents into the resistor ports, l~, the voltages across the capacitor ports, V~, and the currents into the inductor ports, into the currents into the network at the external ports, I, the voltages across the resistor ports, V~, the currents into the capacitor ports, V~, and the voltages across the inductor parts, Y1. It is also assumed that Jt: I) + i I " .• (IR' VR)+(V C , 'c)+(/i., Vl)=O.
It,
<
The interconnected network may thus be considered a dynamical system with input V, output I, and state
[·r:l
It is a neutral interconnection of dissipative
v:
systems with interconnection constraint VR = V~, Vc = V~, VL = IR= -lk, lc= -I~, and I L = -It. The external supply rate is <~ I) and the internal supply rates are (lR' VR) , (Vc, Ie), and (/ L , VL ) . The stored energy in the capacitors and inductors is uniquely defined by lie
nL
E(Vc, I L ) = L:E,,(V,,)+ L:E,(I,) 1
1
with
Thus in standard electrical networks no ambiguity in the stored energy function arises. This is an immediate consequence of the fact that we are able to consider these systems as an interconnection of very simple subsystems in which the ele417
ments with memory are lossless. The individual dynamical subsystems involving memory are in fact described by first order scalar differential equations. 7.4. Thermodynamics Consider a thermodynamic system at a uniform temperature. Assume that the mathematical model used for describing this system is in the form of a dynamical system in the sense of Definition 1 and that the outputs to this dynamical system X contain (possibly among other things) w, the rate of work done on the system, q, the rate of heat delivered to the system, and qlT where T denotes the temperature of the system. We assume that every admissible input and every initial state yield functions w, q, and qlT which are locally integrable. The first and the second laws of thermodynamics may then be formulated by stating that a thermodynamic system is dissipative and lossless with respect to
the supply rate (w+q) and dissipative with respect to the supply rate - ~. In terms of our definitions this implies the existence of two nonnegative functions E and - S defined on the state space X of I such that every motion with Xl = t/J(tt, to, xo, u) yields tl
I
E(xo) + (w(t) + q(t»dt==E(Xt)
(Conservation of Energy)
to tl q(t) S(xo) + T(t) dt~S(Xl)
t!
(Clausius' inequality).*
The function E is called the internal energy and S is called the entropy. It follows from the results obtained earlier that E is uniquelydefined once the equations of the dynamical system are given but that in general there will be many possible entropy functions. Two particular possibilities, Sa and Sr, may be computed a priori via the variational problems S.(x)== - sup x...
'1 q(t)
J T()t
0
dt
fl~O
and •
S (x)= - InC ,
x....x
q(t) J -T(t) -dt 0
t-I
t-l~O
where x· is a point of maximal entropy normalized to S(x*)=O. We may also conclude that, whatever the actual entropy may be, it satisfies the a priori inequality S,~S~Sa~O. For reversible thermodynamic systems, i.e., when I is
* DAY [3S) has recently written a paper in which he shows how to replace this axiom by one involving the heat delivered and absorbed and the maximum and minimum temperature attained. 418
lossless with respect to -
i
as well, the entropy is given unambiguously by
S=S,,=Sr provided the state space of E is reachable from x* and controllable to x·. It should be emphasized that this ambiguity in the entropy function for irreversible thermodynamic systems is fundamental: the dynamical equations do not provide enough information to define the entropy uniquely. This difficulty has long been advertised by MEIXNER [32].
8. Conclusions
In the first part of this paper we have attempted to outline a general theory of dissipative dynamical systems. The mathematical model employed is a so-
called state space model in which the map which generates outputs from inputs is viewed as the composition of a state transition map and a memoryless read-out function. This type of model is standard in control theory and dynamic estimation theory and it is argued that this model offers conceptual advantages for describing general physical systems with memory. The definition of a dissipative dynamical system postulates the existence of a storage function which satisfies a dissipation inequality involving a given function called the supply rate. In many applications one knows from physical considerations that a storage function exists but it is often a difficult task to determine it. It is then shown that this difficulty is genuine and that the dynamical equations are insufficient to specify the storage function uniquely. However, the storage function satisfies an a priori bound, i.e., it is bounded from below by the available storage and from above by the required supply. The available storage is the amount of internal storage which may be recovered from the system and the required supply is the amount of supply which has to be delivered to the system in order to transfer it from a state of minimum storage to a given state. Both these functions are themselves possible storage functions and their. evaluation may be posed as variational problems. These ideas were then applied to interconnected systems and it was established that for interconnected systems with interconnections which instantaneously redistribute the supply (the so-called neutral interconnections), the sum of the storage functions of the individual subsystems is a possible storage function for the interconnected system. The stability properties of dissipative systems were then investigated and it was shown that states for which the storage function attains a local minimum are locally stable and that the storage function is a suitable Lyapunov function. Part II of this paper will be devoted to an examination of linear systems with quadratic supply rates. This research was supported in part by the National Science Foundation under Grant No. OK-25781 and in part by the U.K. Science Research Council. This paper was prepared while the author was a SeniorVisitins Fellowat the Departmentof Applied Mathematics and Theoretical Physics of the University of Cambridge, Cambridge, England. 419
References 1. TRUESDELL, C., Elements of Continuum Mechanics. Berlin-Heidelberg-New York: Springer 1966. 2. TRUESDELL, C., Rational Thermodynamics. New York: McGraw-Hi111969. 3. ONAT, E. T., The Notion of State and Its Implications in Thermodynamics of Inelastic Solids, pp. 292-314 in: Proc, 1966 IUTAM Symposium, H. PARXUS & L. I. SSDOV, Ed. Berlin..Heidelberg-New York: Springer 1968. 4. ONAT, E. T., Representation of Inelastic Mechanical Behavior by Means of State Variables, pp.213-224 in: Proc, 1968 IUTAM Symposium, B. A. BOLEY, Ed. Berlin-HeidelbergNew York: Springer 1970. 5. KALMAN, R. E., P. L. FALB, & M. A. ARDIB, Topics in Mathematical System Theory. New York: McGraw-Hill 1969. 6. DESOER, C. A., Notes for a Second Course on Linear Systems. New York: Nostrand Reinhold 1970. 7. ZADEH, L. A., The Concepts of System, Aggregate, and State in System Theory, in: System Theory, L. A. ZADEH & E. POLAK, Ed. New York: McGraw-Hill 1969. 8. YOULA, D. C., The synthesis of linear dynamical systems from prescribed weighting patterns. J. SIAM Appl. Math. 14, No.3, 527-549 (1966). 9. BALAKRISHNAN, A. V., Foundations of the state-space theory of continuous systems I. J. Computer and System Sci. 1, 91-116 (1967). 10. BROCKEIT, R. W., Finite Dimentional Linear Systesm, New York: Wiley 1970. 11. WILLEMS, J. C., The construction of Lyapunov functions for input-output stable systems. SIAM J. Control 9, No.1, 105-134 (1971). 12. BAKER, R. A., & A. R. BERGEN, Lyapunov stability and Lyapunov functions of infinite dimensional systems. IEEE Trans. Automatic Control AC-14, 325-334 (1969). 13. ESTRADA, R. F., & C. A. DESOER, Passivity and stability of systems with a state representation. Int. J. Control 13, No.1, 1-26 (1971). 14. BREUER, S., & E. T. ONAT, On recoverable work in linear viscoelasticity. ZAMP 15, No. 1. 12-21 (1964). 15. BREUER, S., & E. T. ONAT, On the determ.ination of free energy in linear viscoelastic solids. ZAMP 15, No.1, 184-191 (1964). 16. YOULA, D. C., L. I. CASTRIOTA, & H. J. CARLIN, Bounded real scattering matrices and the foundations of linear passive network theory. IRE Trans. on Circuit Theory CT-6, 102-124 (1959). 17. ZAMES, G., On the input-output stability of time-varying nonlinear feedback systems. Part I: Conditions derived using concepts of loop gain, conicity, and positivity; Part II: Conditions involving circles in the frequency plane and sector nonJinearities. IEEE Trans. Automatic Control AC..l1, 228-238, 465-476 (1966). 18. WILLEMS, J. C., The Analysis of Feedback Systems. Cambridge, Mass. : The M.I.T. Press 1971. 19. MEIXNER, J., On the theory of linear passive systems. Arch. Rational Mech. Anal. 17, 278-296 (1964). 20. CARLIN, H. J., Network theory without circuit elements. Proe. IEEE 5S, No.4., 482-497 (1967). 21. GRUBER, M., & J. L. WILLEMS, On a Generalization of the Circle Criterion. Proe. of the Fourth Annual Allerton Conf. on Circuit and System Theory, pp. 827-835, 1966. 22. FREEDMAN, M., & G. ZAMES, Logarithmic variation criteria for the stability of systems with time-varying gains. SIAM J. on Control 6, 487-507 (1968). 23. WILLEMS, J. L., Stability Theory of Dynamical Systems. Nelson, 1970. 24. YOSHlZAWA, T., The Stability Theory by Liapunov's Second Method. Math. Soc. Japan,
1966. 25. LA SALLE, J. P., An Invariance Principle in the Theory of Stability, pp.277-286, in: Differential Equations and Dynamical Systems, J. K. HALE & J. P. LA SALLE, Ed. New York: Academic Press 1967. 26. BROCKETT, R. W., Path Integrals, Liapunov Functions and Quadratic Minimization. Proc, 4th Annual Allerton Conf. on Circuit and System Theory, Monticello, 111., pp. 685-698, 1966.
420
27. CoLEMAN, B. D., Thermodynamics of materials with memory. Arch. Rational Mech. Anal. 17, 1-46 (1964). 28. GURTIN, M., On the thermodynamics of materialswith memory.Arch. Rational Mech. Anal. 28, 40-50 (1968). 29. Popov, V. M., Hyperstability and optimality of automatic systems with several control functions. Rev. Roumaine Sci. Tech. Electrotechn. et Energ. 9, No.4, 629-690 (1964). 30. KALMAN, R. E., Lyapunov functions for the problem of Lur'e in automatic control. Proc. Nat. Acad. Sci. U.S.A. 49, 201-205 (1963). 31. NEWCOMB, R. W., Linear Multiport Synthesis. New York: McGraw-Hill 1966. 32. MEIXNER, J., On the Foundation of Thermodynamics of Processes, pp. 37-47 in: A Critical Review of Thermodynamics, E. B. STUART, B. GAL-OR & A. J. BRAINARD, Ed. Mono Book Corp. 1970. 33. DAY, W. A., Thermodynamics based on a work axiom. Arch. Rational Mech. Anal. 31, 1-34 (1968). 34. BRAYTON, R. K., & J. K. MOSER, A theory of nonlinear networks. Quart. Appl, Math. 22, 1-33, 81-104(1964). 35. DAY, W. A., A theory of thermodynamics for materials with memory.Arch. Rational Mech. Anal. 34, 86-96 (1969). Decision and Control Sciences Group Electronic Systems Laboratory
Department of Electrical Engineering Massachusetts Institute of Technology (ReceivedJanuary 19, 1972)
421
On Self Tuning Regulators K. J. AsTROM AND B. WITTENMARK
IN
1973 adaptive control was in one sense an old subject. The 1950s had seen a vigorous development, where virtually all the basic concepts were born. At the beginning of the 1960s the fundamental theoretical tools and formulations were laid out: Bellman, [3], had shown how Dynamic Programming was a natural tool for decision making under uncertainty and for showing how to employ optimally the increasing process knowledge that observations provided. Feldbaum, [5], had with Dual Control pinpointed the basic conflict between applying control and probing the system to secure more information (for better future control). Kalman, [8], had outlined the precursor to the self-tuning regulator: least-squares identification of a difference equation model of the process coupled with a regulator that was optimal (in some sense) for the current model estimate. Self-oscillating adaptive systems (SOAS) and Model Reference Adaptive Control (MRAC) had been devised and also tested in various applications. An excellent description of the status of adaptive control around this time can be found in [2]. It can perhaps be said that enthusiasm for adaptive control declined from the mid-1960s. There were a number of setbacks: flight tests of aircaraft using adaptive control to handle varying dynamics had not proven to be successful and model reference control had convergence difficulties when used for more than gain adjustment. Optimal (dual) control proved to be impossibly complex for use in practice. A number of papers appeared around 1970 describing approximations to dual control, but they merely emphasized the complexity of the issues involved. It is in this light that the impact of this paper by AstromWittenmark should be seen. Like many important results, the paper's impact comes from the simplicity of the main idea it put forward. The approach is to combine two methods well known at the time: recursive estimation of the parameters of a linear discrete-time model of the system
y(t) + aty(t - 1) =btu(t - nk)
+ +
+ anay(t + bnbu(t -
na ) nk - nb + 1)
(1)
and application of dead-beat control using the estimated model.
When the delay, ni, is 1, the dead-beat control is u(t)
=
:1
(Oly(t)
+ ···+ On.y(t -
na
+ 1)
-b2u(t - 1) - · · · - bnbu(t - nb + 1)
(2)
A useful reparametrization of the system that permits direct estimation of the regulator parameters when ni > 1 is described in the paper. The main technical contribution of the paper is the following:
(under certain conditions) if theself-tuning regulator converges, it converges to the minimum variance controller. The surprising aspect of this result is that it also applies when (1) is subject to colored noise, in which case the true dead-beat controller gives bad control, and the least-squares parameter estimates are biased. These two errors cancel each other, so to speak. This result can be understood by observing that the parameter estimates are driven by the correlation between the control error and the past inputs and outputs. Theorem 5.1 in the paper shows that if the algorithm converges, then the control error becomes uncorrelated with all past data, so that minimum variance control is obtained. Convergence itself is not claimed or proven in the paper, but examples and approximate analysis establish that the regulator in fact "often" converges to the optimal regulator. It later turned out that there are cases where convergence does not occur [9]. The paper also describes important extensions, such as feedforward, nonminimum phase systems, and nonlinear systemsproblems that would keep the community busy for the next few decades. The paper by Astrom and Wittenmark had an immediate impact. Literally thousands of papers on self-tuning regulation, both theoretical and applied, appeared in the next decade. On the theoretical front, the paper left open the questions of convergence and stability and this inspired much subsequent research. A pioneering paper by Monopoli [10], although not complete, pointed the way, but the first proof of convergence of deterministic, continuous-time, adaptive systems systems, due to Feuer and Morse [6], did not appear until 1978; convergence was achieved by an ingenious but complex controller, not of the certainty
423
equivalence type, that employed nonlinear damping to dominate an error (arising in all adaptive controllers) due to the rate of change of parameter error. A second proof (in a deterministic, discrete-time context), closer in spirit to the Astrom-Wittenmark paper and resulting in a much simpler controller, was provided in [7], also included in this volume. The approach in this paper (also included in this volume) used certainty equivalence, as does Astrom- Wittenmark, and achieved convergence by slowing down the identifier (by using error normalization, also proposed in [4]) thus reducing the error due to rate of change of the parameter error. Further details on this, and a brief discussion on subsequent work on this topic, can be found later in this volume, in the preamble to [7]. At the applied end of the spectrum, the paper (by Astrom and Wittenmark) inspired, almost uniquely, many industrial applications. Providers of control equipment developed many hardware and software products that implement variants of the selftuning regulator. Some of these developments are described in [1]. The lasting influence of the paper is perhaps best judged by the fact that today there are many thousands of control loops in practical use that have been designed using the self-tuning concept. One may wonder why the impact of the self-tuning regulator was so massive and immediate. I believe there are several reasons: its simplicity, an almost "magical" robustness property (described in the paper), some early successful industrial applications, and the newly coined term self-tuning, all combined to
attract a considerable amount of attention. The self-tuning regulator revitalized the field of adaptive control that had lost, in the early 1970s, some of its earlier lustre. REFERENCES [1] K.J. AsTROM AND B. WITIENMARK, Adaptive Control, Addison-Wesley
(Reading,MA), 2nd edition, 1995. [2] K.J. ASTROM, "Adaptive control around 1960," IEEE Control Systems (Magazine), AC-16(3):44-49, 1996. [3] R. BELLMAN, AdaptiveControlProcesses- A GuidedTour, PrincetonUniv. Press (Princeton,NJ), 1961. [4] B. EGARDT, Stability of Model Reference Adaptive and SelfTuning Regulators, Lund Inst. Techn. Tech. Report (Lund, Sweden), 1978. [5] A.A. FELDBAUM,"Dual control theory I," Avtomatica i Telemekhanika, 21(9):1240-1249, 1960. (English translation: Automat. Remote Contr., 21(9):874-880, 1961.) [6] A. FEUER AND A. S. MORSE, "Adaptive control of single-input singleoutput linear systems," IEEE Trans. Automat.Contr., AC-23(4):557-569, 1978. [7] G. C. GOODWIN, P.J. RAMADGE, AND P. E. CAINES, "Discrete time multivariableadaptivecontrol,"IEEE Trans. Automat.Contr., AC-2S(3):449456,1980. [8] R. E. KALMAN, "Designof a self-optimizing controlsystem,"Trans. ASME, J. Basic Engineering, 80:468-478,1958. [9] L. LJUNG, "On positive real transfer functions and the convergence of some recursive schemes," IEEE Trans. Automat.Contr., AC-22:539-551, 1977. [10] R.V. MONOPOLI, "Model reference adaptive control with an augmented error," IEEE Trans. Automat.Contr., AC-19:474-484, 1974.
L.L.
424
On Self Tuning Regulators' Sur les Regulateurs Auto-Syntonisants tiber selbsteinstellende Regier
o CaMOHaCTpaHBalOI.Q.HXCg peryJIJITOPax K. J. ASTROM and B. WIITENMARK
Control laws obtained by combining a leastsquares parameter estimator anda minimum variance strategy basedon the estimated parameters have asymptotically optimal performance. S..-ry-The problem of controlling a system with constaDt but unknown parameters is considered. The analysis is restricted to dilcrete time single-input sinsJe-output systems. An allorithm obtained by combining a least
From a practicalpoint of view it is thus meiDingful to consider the control of systems with constant but unknown parameters. Optimal control problemsfor such systems can beCormulated and solved using non-linear stochastic control theory. The solutions obtained are extremely impractical since even very simpleproblemswillrequire computations far exceeding the capabilities of todays computers. For systems with constant but unknown parameters it thus seems reasonable to look for strategiesthat will converge to the optimal strategies that could be derived if the system characteristics were known, Such algorithms will be called se/j:'tuning or self-adjusting strategies. The word adaptive is not used since adaptive, although never rigorously defined, usually implies that the characteristics of the process are changing. The problem to be discussed is thus simplerthan the adaptive problem in the sense that the system to be controlled is assumed to have constant parameters. The purpose of the paper is to analyse one class of self-adjusting regulators, The analysis is restricted to single-input single-output systems. It is assumed that the disturbances can be characterized as filtered white noise. The criterion considered is the minimization of the varianceoftbe output. The algorithmsanalysed are those obtained OD the basis of a separation of identification and control. To obtain a simple algorithm the identification is simply a least squares parameter estimator. The main result is a characterizationof the closed loop systems obtained whenthe algorithmis applied to a general class of systems. It is shown in Theorem 5.1 that if the parameter estimates converge the closed loop loop system obtained will be such that certain covariances of the inputs and the outputs of the closed loop system are zero. This is shown under weakassumptionson the system to be
squues estimator with a minimum variance reauJator computed from the estimated model is analysed. The main
results are two theorems which characterize the closed loop system obtained under the assumption that the parameter
estimates conYerae. The fint theorem states that certain covarianc:es of the output and certain croa-covariances of the control variable aDd the output will vanish under weak assumptions on the system to be coatrolled. In the second theorem it is assumed that the system to be controlled is a amerallinear stocbutic 11th order system. It is shown that if the parameter estimates converp the control law obtaiaed is in fact the minimum variance control law that could be computed if the parameten of tbe system were Imown. This is somewhatsurprisiol since the least squares estimate is biased. Some practical implications of the results are discussed. In particular it is shown that the algorithm can be feasibly implemented OD a small process computer.
1.INTRODUcnON
IT HAS been shown in several cases that linear stochastic control theory can be used successfully to design regulators for the steady state control of industrial processes. See Ref. [1]. To use this theory it is necessary to have mathematical models of the systemdynamics and of the disturbances. In practice it is thus necessary to go through the steps of plant experiments, parameter estimation, computation of control strategies and implementation. This procedure can be quite time consuming in particular if the computations are made ofT-line. It might also be necessary to repeat the procedure if the system dynamics or the characteristics of the disturbances are changing as is often the case for industrial processes. • Received 2 March J972; revised 12 September 1972. The oriJinai versiOD of this paper was presented at the 5th IFAC Congress which was held in Paris. France during lune 1972. It was recommended for publication in revised form by AssociateEditor A. Sap.
Reprinted with permission from Automatica, K. 1.Astrom and B.Wittenmark, "On Self Thning Regulators" Vol. 9, 1973, pp.185-199.
425
controlled.. Moreover if it is assumed that the system to be controlled is a sampled finite dimensional linear stochastic system with a time delay in the control signal it is demonstrated in Theorem 5.2 that, if the parameter estimatesconverge, the corresponding regulator will actually converge to the minimum variance regulator. This is true, in spite of the fact that the least squares estimate is biased. The major assumptions are that the system is minimum phase, that the time delay is known and that a bound can be given to the order of the system. The first two assumptions can be removed at the price of a morecomplicated algorithm. The paper is organized as follows: sections 2 and 3 provide background and a motivation. The algorithm is given in section 4. Control strategies for systemswith known parameters are givenin section 2. Least squares parameter estimation is briefly reviewed in section 3. Some aspects on the notion of identifiability is also given in section 3. The algorithm presented in section 4 is obtained simply by fitting the parameters of a least squares structure as was described in section 3 and computing the corresponding minimum variance control strategy AS was described in section 2. The possible difficulty with non-identifiability due to the feedback is avoided by fixing one parameter. The main result is given as two theorems in section 5. We have not yet been able to prove that the algorithm converges in general. In section 6 it is, however, shown that a modified version of the algorithm converges for a first order system. The convergence properties of the algorithm are further illustrated by the examples in section 7. Some practical aspects of the algorithm as well as some problems which remain to be solved are given in section'S. In particular it is shown that the algorithm is easily implemented on a minicomputer.
and the polynomials
are introduced, the equation (2.1) describing the system can be written in the following compact form: A(q)y(t) =B(q)u(t - k) +). C(q)e(t).
(2.2)
It is well known that (2.1) or (2.2) is a canonical representation of a sampled finite dimensional single-input single-output dynamical system with time delays in the output and disturbances that are gaussian random processes with rational spectral densities. The model (2.1) also admits a time delay T in the system input which need not be a multiple of the sampling interval. The number k corresponds to the integral part of tjh, where h is the sampling interval. Let the criterion be (2.3)
or 1
N
Ly2(t) . Nl
V 2=E-
(2.4)
The optimal strategy is then u(t)= _ qkG(q) yet) B(q)F(q)
(2.5)
where F and G are polynomials 2. MINIMUM VARIANCE CONTROL
F(Z)=Z"+!tZ"-l+ ... +J~
This section gives the minimum variance strategy for a system with known parameters. Consider a system described by
determined from the identity
y(t)+a t y(t - l) + .... +D,.y(t-n)=b1u(t-k-l)
+ ... + ...
q"C(q) = A{q)F(q) + G(q).
+bnu(t-k-n)+A.[e(t)+cte(t-l)
(2.8)
Proofs of these statements are given in [2]. The following conditions are necessary:
+clle(t-n)],
1=0, ±l, ±2, ...
(2.6)
(2.1)
-The polynomial B has all zeroes inside the unit circle. Thus the system (2.1) is minimum phase. -The polynomial C bas all zeroes inside the unit circle.
where u is the control variable, y is the output and {e(/), 1=0, ± I, ±2, ... } is a sequence of independent normal (0, I) random variables. If the forward shift operator q, defined by
These conditions are discussed at length in [1). Let it suffice to mention here that if the system (2.1 )
qy(t)=y(t+ I) 426
{e(t)} it may not be possible to determine all parameters. When the inputs are generated by a feedback they are correlated with the disturbances and it is not obvious that all the parameters of the model can be determined. Neither is it obvious that the input generated in this way is persistentlyexciting of sufficiently.high order. A simpJe example illustrates the point.
is non..mimmum phase the control strategy (2.5) will still be a minimum variance strategy. This strategy will, however, be so sensitive that the slightest variation in the parameters will create an unstable closed loop system. Suboptimal strategies which are less sensitive to parameter variations are also well known. This paper will, however, be limited to minimum phase systems.
Example 3.1 Consider the first order model
3. PARAMETER ESTIMATION
For a system described by (2.1) it is thus straight forward to obtain the minimum variance regulator, if the parameters of the model are known. If the parameters are not known it might be a possibility to try to determine the parameters of (2.1) using some identification scheme and then use the control law (2.5) with the true parameters substituted by their estimates. A suitable identificationalgonthm is the maximum likelihood method which will give unbiased estimates of the coefficients of the A, B and C polynomials. The maximum likelihood estimates of the parameters of (2.1) are, however, strongly non-linear functions of the inputs and the outputs. Since finite dimensional sufficient statistics are not known it is not possible to compute the maximum likelihood estimate of the parameters of (2.1) recursively as the process develops. Simpler identification schemes are therefore considered.
y(t)+ay(t-l)=bu(t-l)+e(t).
Assume that a linear regulator with constant gain
u(t)=t%y(t)
(3.3)
is used. If the parameters a and b are known the gain (X=a/h would obviously correspond to a minimum variance regulator. If the parameters are not known the.gain cx=a/6 where and 6 are the least squares estimates of a and b could be attempted. The least squares parameter estimates are deter .. mined in such a way that the loss function
a
N
J/(a, b)=
I: [y(t+ 1)+aY(l)-bu(t)]2
(3.4)
1
is minimal with respect to a and b. If the feedback control (3.3) is used the inputs and outputs are linearly related through
The least squares structure The problem of determining the parameters of the model (2.1) is significantly simplified if it is assumed that c,=O for ;= 1, 2, ... ,n. The model is then given by
A(q)y(t) = B(q)u(t-k)+ Ae(t+n).
(3.2)
u(t) - a;y( t) = 0 .
(3.5)
Multiply (3.5) by -1 and add to the expression within brackets in (3.4). Hence
(3.1)
The parameters of this model can be determined simply by the least squares method [3]. The model (3.1) is therefore referred to as a least squares model. The least squares estimate has several attractive properties. It can easily be evaluated recursively. The estimator can be modified to take different model structures, e.g. known parameters, into account. The least squares estimates will converge to the true parameters e.g, under the following conditions.
N
V(a, b)= L[Y(t+l)+(a+~}')y(t)-(b+i')U(I)]2 1
=V(a+~'Y,
b+l').
The loss function will thus have the same value for all estimates a and b on a linear manifold. Thus the two parameters a and b of the model (3.2) are not identifiable when the feedback (3.3) is used. The simple exampleshows that it is in general not possible to estimate all the parameters of the model (3.1) when the input is generated by a feedback. Notice, however, that all parameters can be estimated if the control law is changed. In the particular example it is possible to estimate both parameters of the modeJ, if the control law (3.3) is replaced by
-The output {y(t)} is actually generated from a model (3.1). -The residuals {e(t)} are independent.
-The input is persistentlyexciting, see Ref. [3]. -The input sequence {u(t)} is independent of the disturbance sequence {e(t)}. These conditions are important. If the residuals are correlated the least squares estimate will be biased. If the input sequence {u(t)} depends on
U(l)=~y(t-l)
427
Step 1 parameter estimation
or
At the sampling interval t determine the para-
meters ex 1 t
or if a time varying gain is used.
• • •
Pit ... , fJ,
of the model
y(t)+ cx ly(t - k - l )+ ... +cxllly(t-k-m)
=fJo[u(t-k-l)+ fJtu(t-k-2)
4. THE ALGORITHM
In order to control a system with constant but unknown parameters the following procedure could be attempted. At each step of timedetermine the parameters of the system (3.1) using least squaresestimation based on all previously observed inputs and outputs as was described in section 3. Then determine a control law by calculating the minimum variance strategy for the modelobtained. To computethe control lawthe identity (2.8) must beresolved in eachstep. The problem of computing the minimum variance regulator is simplified if it is observed that by usingthe identity (2.8) the system (3.1) can be written as
+ · · · +p,u(t-k-i-l)]+e(t)
using least squares estimation based on all data available at time t, i.e.
minimum. The parameter flo is assumed known. Step 2 control At each sampling interval determine the control variable from
1
y(t+k+l)+txty(t)+ ... +«m(t-m+ 1)
U(I)=- [tXly(t) +
Po
=Pu[u(t) + Plu(t-l}+ · · · + P,u(t-l)]
... +«",y(I-".+ 1)]
- fJtu(t-l)- · .. -fJ,u(t-l)
+s(t+k+ I)
(4.2)
(4.1)
where the parameters !x, and fJ, are those obtained in Step I. The control law (4.2) corresponds to
wherem==n and l=n+k-l. The coeffieients e, and
Pi are computed from the parameters Q, and b, in (3.1) using the identity (2.8). The disturbance s(t) is a movingaverage of order k of the driving noise e(t). The minimum variance strategy is then simply
qf-"'+~(q)
= 1
u(t)==-[«I1
-fJl u(t -
(4.1)
l ) - ... -P,u(t-l).
a(q}
y(t).
(4.3)
Since the least squares estimate can be computed recursively the algorithm requires only moderate
(4.2)
computations.
It should be emphasized that the algorithm is not optimal in the sense that it minimizes the criterion (2.3), or the criterion (2.4). It fails to minimize (2.3) because it is not taken into account that the parameter estimates are inaccurate and it fails to minimize (2.4) because it is not dual in FELDBAUM'S sense [4]. Thesemattersare discussed in [2]. It will however, be shown in section 5 that the algorithm has nice asymptotic properties. The idea of obtaining algorithms by a combination of least squares identification and control is old. An early reference is KALMAN [5]. The particularalgorithmused here is in essence the sameas the one presented by PETERKA [6]. A similar algorithm where the uncertainties of the parameters are also considered is given in WIESLANDERWITrENMARK [7].
In order to obtain simple computation of the control strategy it could thus be attempted to use
the model structure (4.1) which also admits least squares estimation. The trade-off for the simple calculation of the control law is that k more para.. meters have to be estimated. As was shown in Example 3.1 all parameters of the model (4.1) can not necessarily be determined from input-output observations if the input is generated by a feedback (4.2) with constant parameters. In order to avoid a possible difficulty it is therefore assumed that the parameter flo is given. It will be shown in section 6 that the choice of Po is not crucial. Summing up, the algorithm can be described as follows. 428
s.
Assume that the parametersconverge, For sufficiently large No the coefficients of control law (4.2) willthen converge to constant values. Introduction of (4.2) into (5.3) gives
MAIN RESULTS
The properties of the algorithm given in the previous section will now be analysed. We have
Iy(t+k+ 1)y(t)== 0
Theorem S.t
Iy(t+k+ 1)](t-l)=O
Assume that the parameter estimates (1,(1), ;= 1, ..• ,m, fl,(t), j= 1, ... I converge as t-+oo and that the closed loop system is such that the output is ergodic (in the second moments). Then the closed loop system has the properties EY(t+tlv(t)==r,(-r)==O
T==k+ 1, ... k-s-m
!y(t+k+ l)Y(t-m+ 1)=0
Iy(t+k+ l)u(t-l)=O
(5.1)
l:y(t+k+ 1)14(/-1)=0. EY(t+T)u(t)==r,.(T)=O
t=k+ I, ... , k+l+ 1.
Using the control law (4.2) it also follows that
(5.2)
Iy(t+k+ l)u(t) =0.
Proof
Under the ergodicity assumption the sums can furthermore be replaced by mathematical expectations and the theorem is proven.
The least squares estimates of the parameters «I' (X2, • • . , fl.., PI' fJ2, ••• , P, is given by the equations I
Ey(t)2
1:y(I)y(I-l)
l:y(t)y(t-m+l)
i -fJoI.y(t)u
- /lor,y(t)u(t-I)
I
Iy(t-l)y(t-m + 1)1
Iy(t)y(t -1)
I I
I
1
i
Ii -P oEy(, - m + l )u(t - l ) . . .
1:y(I»'(t-m+l)
-PoIy(t-m+ l)u(t-1)
--•••••••••••••••••••••..•.•..••••.•..•••••.•.•....•..·•••••·•·••·••·•··•···•·••••••••••·•••..······fI
- Pol y(t)u(t -
t
t P~Iu2(t-l)
l}
.•
.... fJ~tu(t-l)u(t-l)
i i
I I !
j
- fJor.y(t)u(t-l)
i
+ Jlol:u(t)y(t)
Cll
-~y(t+k+ l)y(t)
fl2
1:)'(t+ k+ l)y(t-l)+ Potu(t)y(t-l)
-Iy(t+k+ l)y(t-m+ 1)+/loEu(t)y(t - m + 1)
fl.
PI
p,
=
- /101:1<'+ k+ l)u(t-l}-P~Iu(t)u(t-l)
JloIy(t+ k+ l)u(t-l)- P~u(t)u(t-l)
where the sums are taken over No values. See Ref. [3]. 429
(5.3)
Remark 1
Proof'
Notice that theassumptions on the system to be controlled are very weak. In particular it is not necessary to assume that the system is governed by an equation like (2.1) or (3.1).
Assume that the least squares parameter estimates converge. The regulator is then given by (4.3) i.e,
Remark 2
u(t)=
It is sufficient for ergodicity that the system to be controlled is governed by a difference equation of finite order, e.g. like (2.1), and that the closed loop system obtained by introducing the feedback law into (2.1) gives a stable closed loop system. Remark 3 The self tuning algorithm can be compared with a PI-regulator. If the state variables of a deterministic system with a PI-regulator converge to steady state values, the control error must be zero irrespective of the properties of the system. Analogously theorem 5.1 implies that if the parameter estimates of the self tuning algorithm converge the covariances (5.1)and (5.2) are zero. Remark 4 Theorem 5.1 holds even if the algorithm is modified in such a way that the parameter estimation (Step 1) is not done in every step. If it is assumed that the system to be controlled is governed by an equation like (2.1) it is possible to show that the conditions (5.1) and (5.2) in essence imply that the self tuning regulator will converge to a minimum variance regulator. We have
q'-"'+ ld(q) q".rI(q) y(t)=-tM(q) dI(q)
where the coefficients of .RI and £f are constant. Since the system to be controlled is governed by (2.1) the closed loop system becomes [A(q)SI(q) - B(q ).rI(q)]y(t) =AC (q) If(q)e(t) .
(5.4)
The closed loop system is of order r=n+/. Introduce the stochastic process {,(t)} defined by o(t)=l q'C(q) e(t). A(q) fM(q) - B(q)JiI(q)
(5.5)
Then y(t) =q -'aJ(q)v(t)
(5.6)
U(l)=q-"+ 1d(q)v(t).
(5.7)
and
Multiplying (5.6) and (5.7) by y(t+t) and taking mathematical expectations gives
Theorem 5.2 Let the system to be controlled be governed by
(5.9)
the equation (2.1). Assume that the self tuning algorithm is used with m =n and 1=n + k - I. If the parameter estimates converge to values such that the corresponding polynomials .91 and eI have no common factors, then the corresponding regulator (4.2) will converge to a minimum variance regulator.
1
P,
PI
0 1
Furthermore it follows from Theorem 5.1, equations (5.1) and (5.2), that the left member of (5.8) vanishes for t = k + I, ... , k + m and that the left member of(5.9) vanishes for t=k+ I, ... ,k+/+ 1. We thus obtain the following equation for r )'p(t).
0
0
Pi
r,,,(k+l)
0
r,,,(k+2)
0
0 0 1
0 0
(Xl
p,
PI ex".
=
(5.10)
0
0
o o
o 430
Sincethe polynomials d and a have no common factor it follows from an elementary result in the theory of equations [8, p. 145] that the (/+m)
Remark The conditions m = nand / = n +k - 1 mean that there are precisely the number of parameters in the model that are required in order to obtain the correct regulator for the process (2.1). Theorem 5.2 still holds if there are more parameters in the
x (/+m)-matrix of the left member of (5.10) is non-
singular. Hence
t=k+ I, ... , k+l+m.
regulator in the following cases. (i) Theorem 5.2 still holds if m=n and /~n+k-l. In this case the order of the system is r=n+1 and since m=n the equation (5.10) implies (5.11)-(5.13) and the equation (5.16) is changed to
(5.11)
Since v is the output of a dynamical system of
order r == n +/= m + 1driven by white noiseit follows from (5.11) and the Yule-Walker equation that
'11J(t)=O
'r~k+
1.
(5.12)
1.
(5.13)
G=-
q,-a:-m+ 1BFd 91
.
(5.16')
The equation (5.8) then implies
r,(r)=0
't~k+
The rest of the proof remains unchanged. (ii) If m~n and /=n+k-l the theorem will also hold. The closed loop system is of order r ~ m + I but the equation (5.10) will still imply (5.11) and (5.16) is changed to (5.16'). The rest of the proof remains unchanged. (iii) If mz-n and />n+k-l the theorem does not hold because .JJI and 111 must have a common factor if the parameter estimates converge. It can, however, be shown that if the algorithm is modified in such a way that common factors of ~ and £f are eliminated before the control signal is computed, Theorem 5.2 win still hold for the modified algorithm.
The output of the closed loop system is thus a moving average of white noise of order k, Denote this moving average by
(5.14) where F is a polynomial of degree k. It follows from (5.4) and (5.14) that
q"C=FA_
BFd. ~
Since qkC and FA are polynomials BF.9I/!f must also be a polynomial.
6. CONVERGENCE OF THE ALGORITHM
It would be highly desirable to have general results giving conditions for convergence of the parameter estimates. Since the system (2.1) with the regulator (4.2) and the least squares estimator is described by a set of nonlinear time dependent stochastic difference equations the problem of a general convergence proof is difficult. So far we have not been able to obtain a general result. It has,
Hence
q"C=FA+G
(5.15)
G=_BFd dI
(5.16)
where
however, been verified by extensive numerical simulations that the algorithm does in fact converge in
is of degree n-l and F of degree k, A comparison with (2.8) shows, however, that (5.15) is the identity which is used to derive the minimum variance strategy. It thus follows from (2.5) that the minimum variance strategy is given by
many cases. The numerical simulations as well as analysis of simple examples have given insight into some of the conditions that must be imposed in order to ensure that the algorithm will converge. A significant simplification of the analysis is obtained if the algorithm is modified in such a way that the parameter estimates are kept constant over long periods of time. To be specific a simple example is considered.
lG
u(t)= -Ly(t). BF
The equation (5.16) then implies that
Example 6.1 Let the system be described by
(5.17) and it has thus been shown that (4.2) is a minimum variance strategy.
y(t)+ay(t-l)==bu(t-I)+e(t)+ce(t-l)
lei < I. (6.1)
431
The point x =0 is a fixed point of the mapping 9 which corresponds to the optimal value of gain of the feedback loop, i.e. cx=(a-c)/b. The problem is thus to determine if this fixed point is stable. Since the closed loop system is assumed to be stable it is sufficient to consider
Assume that the control law u(l) = tX,.y(I)
(6.2)
is used in the time interval 1,,
c-l c+l -<X<-. b b
(6.3)
y(t+ l)+cxy(t)=u(t)+s(t+ 1)
to the data {u(t), y(t), 1=1..- 1 , ••• , t,,-I)}. The least squares estimate is given by
Three cases have to be investigated 1. c=O
2. c>O or c
',._2
~ y(t)[y(t + 1)- u(t)] ex,,- - _,=-.0'"-..-......1 _ _- - - t.. -2
L
'=',,-1 '"-2
L
'='11-1
(6.11)
For all cases g(x)~(l-b)x if x is small. This implies that solutions close to 0 converge to x=O irO
y2(t)
y(t + 1)y(t)
Case 1 The equation (6.10) reduces to
(6.4)
X"+l =(l-b)xn
where the last-equality follows from (6.2). Assume that tIl - ',.-1-. co and that
and the fixed point x = 0 is stable if 11 - bl < 1 i.e. O
la-ball_I' < 1
ease 2 The pincipal behavior of g(x) when c>O and O
(6.5)
which means that the closed loop system used during the time interval t"_1 < t< til is stable then r,(I)
(6.6)
CX,,==Cl"-l--
',(0)
where r ,('r) is the covariance function of the stochastic process {y(t)} defined by y(t)+(a-bCZIl_t)y(t-l)=e(t)+ce(t-l). (6.7) 0 -+--~~:-------::IIIi~--------t
Straightforward algebraic manipulations now give
«'=1%.-1 (c-a + b~II_I)(l-ac+ bCII.-I). (6.8) l+c -2ac+2bcex._l
The problem is thus reduced to the analysis of the nonlinear differerence equation given by (6.8).
-1
Introduce
-1
a-c
X..=CX,,--
b
(6.9)
FIG. 1. Graph of the function 9 when O
(6.10)
Case 3 The function g(x) for this case is shown in Fig. 2. It is not obvious that x will converge to zero, because there might exist a "limit cycle".
the equation (6.8) then becomes X"+l=g(x,.)=(l-b)x,,+
b2c X 2 2
"
l-c +2bcxlI
•
o
432
The analysispresented in the simple example can be extended to give stability conditions for the modified algorithm in more complex cases. The analysis required is tedious. 7. SIMULATIONS
The results in section 5 are given under the assumption that the least squares estimator really converges, but yet we have not been able to give generalconditionsfor convergence. But simulation of numerous examples have shown that the algorithm has nice convergence properties. This section presents a number of simulated examples which illustrate the properties of the selftuning algorithm.
-1-F-----..a.--+--------1
-t
0
Example 7.1 Let the system be
Flo. 2. Graph of the function g when 1 < b < 2. The figure is drawn with the parameter values a=-O·S, 6=1·.5 and c==O·7.
y(t)+ay(t-l)=bu(t-l)+e(t)+ce(t-l) If c>O and starting with x.>O then if g(xo) <0 it can be shown that after two iterations the new
with a= -0-5, b=3 and c=O·7. The minimum variance regulator for the system is
value x, will satisfy O<xz -cx, i.e.
o
a-c
u(t)=-y(t) = -0·4y(1). b
if x>O, g(x)
If(c-I)/b<x
A regulator with this structure can be obtained
y(t+ 1)+ tzy(t) =Pou(t) + s(t+ 1).
Summary From the analysis above we can conclude that
-l
and (6.12)
The example shows that under the condition (6.12) the version of the self.tuning algorithm where the parameters of the control law are kept constant over long intervals will in fact converge. In the analysis above fl. == 1 was chosen. If P. #: 1 then the condition (6.12) is replaced by (6.12')
O.5b
(6.13)
(7.3)
Figure 3 shows for the case Po= 1 how the parameter estimate converges to the value (X = - 0·4 which corresponds to the minimum variance strategy (7.2). In Figure 4 is shown the expected variance of the output if the current value of tz should be used for all future steps of time. Notice that the algorithm has practically adjusted over 50 steps.
x =0 is a stable fixed point if
O
(7.2)
by using the self-tuning algorithm based on the model
dition above will be satisfied. If c < 0 then it can be shown that g(g(x» > x if x
O
(7.1)
0.0 ..........- - - - - - - - - - - - - - - - ,
or ~1.0_+__---......._--__r---~---"""'1
o
200
Ti",•
.The condition (6.13) implies that it is necessary to pick the parameter Po in a correct manner. The algorithm will always converge if P. is greater than b. Under-estimation may be serious and the value fJ.
FIG. 3. Parameter estimate a(l) obtained when the self tuning algorithm based on the model (7.3) is applied to the system given by (7.1). The minimum variance regulator corresponds to a= -0·4 and is indicated by the dashed line.
433
I•
.!
i
iI 1
2-
~
, --,,-"._..._:: . .-.: .:-. . -..,~----...- . .------....,j r~"i\,,_
_
......----------------~
i
dl o-l----.....----...,-.----ro-------t J I I
400
200
..
FlO. 4. Expected variance of the output of Example 7.1 if the control law obtained at time 1 is kept constant for aU future times. Notice that the estimate at time 1==26
would live an unstablesystem.
..
••
The analysis of Example 6.1 shows that, since b>2, and Po== 1 the modified self-tuning algorithm obtained when the parameters of the controller are kept constant over long intervals is unstable. The simulation in Example 7.1 shows that at least in the special case a conservative estimate of the convergenceregion is obtained by analysing the modified algorithm. If the value of b is increased further it has been shown that the algorithm is unstable. Unstable realizations have been found for b= s. In such cases it is of course easy to obtain a stable algorithm by increasing fl•. This requires, however, a knowledge of the magnitude of b. The system of Example 7.1 is very simple. For instance, if no control is used the variance will still be reasonably small. The next example is more realistic in this aspect.
E
! l. . ) - t - - - - , . . . . - - - - , . . - - - - - , . . . . - - - - - I o
FIG. S. Parameter estimates aI, 4 2. fJl and 62 obtained when the self tuning alloritbm based on (7.6) is applied to the system given by (7.4). The thin lines indicate the parameter values of the minimum variance strategy.
1000
: 500-,
I
.!
!
Example 7.2
200
Ti....
i
D
~
Consider the system
E ~
C
y(t)-1.9y(t-l)+O-9y(/-2)=u(t-2)
0
--F----...------r------r----~
o
200
coo
Tim.
+1I(t-3)+e(t)-O-Se(t-l).
FIG.
(7.4)
,
L y2(S) ,=t
If no control is used the variance of the output is infinite. Also notice that B(z)=:=z-l. The assumption that B has all zeroes inside the unit circle is thus violated. The minimum variance strategy for
for a simulation of the system (7.4) when usinl the self tuning algorithm (thick line) and when USiDg the optimal minimum variance regulator (7.S) (thin line).
the system is
using the self-tuning algorithm and when using the optimal minimum variance regulator (7.5). In both examples above, the models in the selftuning algorithm have had enough parameters so it could converge to the optimal minimum variance regulator. The next example shows what happens when the regulator has not enough parameters.
U(/)== -1-76y(/)+ t-26y(t-l)-O·4u(t-l)
+ }·4u(t-2).
6. Accumulated loss
(7.5)
A regulator with this structure is obtained by using the self-tuning algorithm with the model
Example 7.3 Consider the system ::u(t)+fJtu(/-l)+P2u(t-2)+s(t+2).
y(t)-t·60y(t-l)+ 1-61y(t-2)-O·776y(t-3}
(7.6)
== I·2u(t-I)-O·9Su(t-2)+O·2u{t-3)+e(t)
The convergence of the parameters is shown in Fig. 5. Figure 6 showsthe accumulated losses when
+O-1e(t-l)+O·2Se(t·-2)+O·S7e(t-3). (7.7) 434
shows that the algorithm does in fact converge and that the sample covariance Py( I) does not differ significantly from zero. SeeFig. 7(c). When using regulators of lower order than the optimal minimum variance regulator, the parameters in the controller will not converge to values which for' the given structure gives minimum variance of the output. In Table I is shown the variance of the output for the system above when using different regulators. The loss when using the self-adjusting regulator is obtained through simulations. The optimal regulator is found by minimizing ',(0) with respect to the parameters in the controller.
The polynomial A(z) has two complex zeroes near the unit circle ( +O·4±O·9i) and one realzero equal to 0-8.
If a self-tuning regulator is determined based on a model with m =3 and 1=2 it will converge to the minimum variance regulator as expected. Figure 7(a) shows a small sample of the output together with the sample covariance of the output, , ,(t).
1
•~
.!
i
8
i.
'i
~
O~
O~""""~"""--f
- 5 -+---
o
--..-----1 2S
so
7S
I
1 -l~----~ o
10
Tinte
TABLE
1
s-....---------,
N
:c
Loss 1 I HY2(t)
.! ~
..
•
";
; 0
Nt=1
D 8
0
(It)
m
"& E
-5-+------,-------4 so o 2S
II
1
r ·-------·--l
.; oAAAlI
te)
{
j 2S
so
I
•
Optimal
.L.'. .. o
3
2
1-0
1·0
2
1
2·S
1·9
0
4·8
3·4
The previous examples are all designed to illustrate various properties of the algorithm. The following example is a summary of a feasibility study which indicates the practicality of the algorithm for application to basis weight control of a paper machine.
II
10
TiM.
FlO. 7.
Self-adjusting
D
IA
Output or the system (7.7)and samplecovariance
of the output ;-<1') when controllilll with self tuning regulator is having different Dumber of parameters
Example 7.4 The applicability of minimum variance strategies to basis weight control on a paper machine was demonstrated in [9]. In this application the control loop is a feedback from a wet basis weight signal to thick stock flow. The models used in [9] were obtained by estimating the parameters of (2.1) using the maximum likelihood method. In one particular case the following model was obtained.
If the self-tuning algorithm instead is based OD a model with m==2 and 1= 1 it is no longer possible to obtain the minimum variance regulator for the system since there are not parameters enough in the self-tuning regulator. Theorem 5.1 indicates, however, that if the self-tuning regulator converges, its parameters will be such that the covariances "( I), r,(2), r1.( 1) and ry,,(2) are all zero. The simulation shows that the algorithm does in fact converge with Po == 1-0. The covariance function of the output is shown in Fig. 7(b). It is seen that the sample covariances P,( I) and , ,(2) are within the 5 per cent confidence interval while P,(3) is not as would be expected from Theorem 5.1. If a self-tuning algorithm is designed based on a model with m= 1, 1=0 then Theorem 5.1 indicates that ',(1) should vanish. Again the simulation
where the output y is basis weight in gjm2 and the control variable is thick stock ftow in g/m2 • The disturbance {v(t)} was a drifting stochastic process which could be modelled as
(7.9) 435
where {e(t)} is white noise. The sampling interval was 36 sec and the numerical values of the parameters obtained through identification were as follows
150~-------------'
1 ' 00
:!
.!-1 50
i
";
e
e
C O.......-~---__r_------------4
o
60
30
90
TiMe [",in)
Cl
FIG. 9. Accumulated loss for Example 7.4 when using the self tUDiq rep)ator (thick line) and when using the minimum variance replator (thin line).
= -1·438
O--.C""'2'""------------i' ).,=0·382.
~-~L,-----,.------
"
To investigate the feasability of the self-tuning algorithm for basis weight control, the algorithm was simulated using the model (7.8) where the disturbance , was the actual realization obtained from measurements OD the paper machine. The parameten of the regulator were chosen as k = I, 1=3, m=4 and fJ.=2·S and the initial estimates were set to zero. The algorithm is thus tuning 7
o
Ji .... [mift]
The results of the simulation are shown in Figs. 8-·10. Figure 8 compares the output obtained when using the self.tuning algorithm with the result obtained when using the minimum variance regulator computed from the process model (7.8) with the disturbance given by (7.9). The reference value was 70 gfm2 • In the worst case the self-tuning regulator gives a control error which is about 181m2 greater than the minimum variance regulator. This happens only at two sampling intervals. After about 7S sampling intervals (4S min) the output of the system is very close to the output obtained with the minimumvarianceregulator.
~
90
Ii""
,
Jl(t)=
L y2(n) a=O
obtained with the minimum varianceregulator and the self-tuning regulator. Notice that io the time interval (21, 24) minutes there is a rapid increase in the accumulated lossof the self-tuning regulator of about 17 units. The largest control error during this interval is 2·7 81m 2 while the largest error of the minimum variance regulator is 1 g/m2 • The accumulatedlossesover the last hour is 60 units for the self-tuning regulator and S9 units for the minimum variance regulator. The control signal generated by the self-tuning algorithm is compared with that of the minimum variance regulator in Fig. 10. There are differences in the generated control signals. The minimum variance regulator generates an output which has more rapid variations than the output of the selftuning regulator. The parameter estimates obtained have not converged in 100 sampling intervals. In spite of this the regulator obtained will have good performance as has just been illustrated. The examplethus indicatesthat the self-tuning algorithmcould befeasible as a basis weightregulator,
-w------------------
•
i
4
••;
-i
j
I 60
I
Figure 9 compares the accumulated losses
/;'70 ......~......I_+_tI~.........f+A~........~...___w_..... '. ..............
J
,
30
I I
FIG. 10. The control in IIm1 for Example 7.4 when usina the self tUDiDa replator (thick line) and when USiDI the minimum variance repJlator (thin line).
parameters.
75
I
ss-+------r-------------. ... 60 30 o '1-
90
[..a,,]
FIG. 8. Wet basis weiaht when using the self tuning resulator (thick line) and when using the minimum
variance Raulator based on maximum likelihood identiftcation (thin line). The reference value for the con-
troller was701lm2• 436
8. PRACTICAL ASPEcrs
that have the correct asymptotic properties. Apart from the algorithm given in section 4 we have the algorithm which minimizes (2.4). But that algorithm is impossible to use due to the computational requirements. It is of interest to investigate if other possible algorithms have better convergence rates than the algorithm of section 4. No complete answer to this problem is yet known. A few possibilities will be indicated. It could be attempted to take into account that the parameter estimates are uncertain. SeeRefs. [2, 7 and 10]. The least squares identifier can be improved upon by introducing exponential weighting of past data. This has in some cases shown to be advantageous in simulations. Algorithms of this type have in simulations been shown to handle slowly drifting parameters. Another possibility is to assume that the parameters are Wiener processes, which also can be incorporated fairly easily [2, 7]. It has been verified by simulation that the region of convergence can be improved by introducing a bound on the control signal.
A few practical aspects on the algorithm given in section 4 are presented in this section which also covers some possible extensions of the results A prior; knowledge
The only parameters that must be known aprior;' are k, I, m and Po. If the algorithm converges it is easy to find out if the a prior; guesses of the parameters are correct simply by analyzing the sample covariance of the output. Compare Example 7.3. The parameter Po should be an estimate of the corresponding parameter of the system to be controlled. The choice of Po is Dot critical as was shown in the Examples 6.1 and 7.1. In the special cases studied in the examples an under-estimate led to a divergingalgorithm while an over-estimate was safe.
Implementation on process computers It is our belief that the self-tuning algorithm can beconvenientlyused in process control applications. There are many possibilities. The algorithm can be used as a tool to tune regulators when they are installed. It can be installed among the systems programs and cycled through different control loops repetitively to ensure that the regulators are always properly tuned. For critical loops where the parameters are changing it is also possible to use a dedicated version which allows slowly drifting parameters. A general self-tuningalgorithm requires about 40 FORTRAN statements. When compiled using tbe standard PDP 15/30FORTRAN compiler the code consists of 450 memory locations. The number of memory locations required to store the data is (1-1 +m)2+3(/-l +m)+2k+4. Execution times on a typical process computer (PDP 15) without floating point hardware are given in the table below. The major part of the computing is to update the least squares estimate.
Feedforward In many industrial applications the control can be improved considerable if feed forward is used. The self tuning regulators in this paper can include feed forward control by changing the process model (4.1) to y(t+k+ 1)+(Xty(t)+ •..
== Polu(t) + P1u(1- 1) + .. . + P,u(t -I)] + 'Y 1s(t)
+ ...
u(t)=-[cxty(t)+ .•.
Po
Execution time ms
I-s-m
16
S
34
8
69
(8.1)
+ (.(.y(t- m + 1)-')'1s(t)
- . . . -}'"s( t - p + 1)] - fll u(t -1 )
5
3
+'Y"s(l-p+ l)+e(t+k+ 1)
where S(I) is a known disturbance. The parameters ai' Pi and "Ii can be identified as before and the cootrollaw (4.2) will be changed to 1
Number of parameters
+~mY(t-m+1)
- • • • -fJ.u(t -I).
(8.2)
Nonminium phase systems Difficulties have been found by a straightforward application of the algorithm to nonminimum phase systems, i.e, systems where the polynomial B has zeroes outside the unit circle. Several ways to get around the difficulty have been found. By using a model with B(z) = fJo it has in many cases been possible to obtain stable algorithms at the sacrifice of variance.
Improved convergence rates The results of this paper only shows that if the parameters converge the regulator obtained will tend to a minimum variance regulator. Nothing is said about convergence rates, which of course is of great interest from practical as well as theoretical points of view. There are in fact many algorithms 437
[3] K. J. ASTROM and P. EVKHOfF: System identificationA survey. Automatica 7, 123-162 (1971). (4) A. A. FELDBAUM: Dual control theory I-IV. Aut. Remote Control 21, 874-880, )033-1 039 (1961); 22, 1-12, 109-121 (1962). (5) R. E. KALMAN: Design of a self optimizing control System. Trans. ASME 80, 468-478 (1958). [6] v. PE1'mucA: Adaptive Digital Regulation of Noisy Systems, 2nd Praaue IFAC Symposium on Identification and Process Parameter Estimation (1970). [7] J. WIESLANDER and B. WITTENMARK: An approach to adaptive control using real time identification. Auto.. mattca 7, 211-217 (1971). [8] L. E. DICKSON: First Coursein the Theory ofEquations. Wiley, New York (1922). [9] K. J. AsmOM: Computer control of a paper machineAn application of linear stochastic control theory. IBM J. Res. Development 11. 389-40S(1967). (10] J. B. FAJUSON et a1.: Identification and control of linear discrete systems. IEEE Trans. Aut. Co"trol12, 438-442
It is well-known that the minimum variance regulators are extremely sensitive to parameter variations for nonminimum phase systems [I]. This is usually overcome by using suboptimal strategies which are less sensitive [1]. The same idea can be used for the self-tuning algorithms as well. The drawback is that the computations increase because the polynomials F and G of an identity similar to (2.8) must be determined at each step of the iteration. An alternative is to solve a Riccati-equation 'at each step.
Multivariable and nonlinear systems It is possible to construct algorithms that are similar to the one described in section 4 for multivariable and nonlinear systems as long as a model structure which is linear in the parameters [3, p, 131] is chosen. For multivariable systems the structure given in equation (3.2) of Ref. [3] can thus be attempted. Analyses of the properties of the algorithm obtained when applied to a multivariable or a nonlinear system are not yet done.
(1967).
Resume-on
considereIeproblemedu controle d'un systeme avec parametres constants mais inconnus. L'analyse se limite aux systemes discrets Aentree unique et sortie unique. Un algorithme obtenu eo combinant un estimateur a carres
minimum avec un r6gulateur a variance minimum caleulee du modelecstim6,est analyse. Les resultats principaux sont deux theoremes qui caractCrisent le systeme a boucle fermee obtenu en supposant que les estimations des parametres converaent. Le premier theoreme dit que certaines covariances de la sortie et certaines covarianee5 transverses de la variable de contr6le et de la sortie disparaitront avec des suppositions faibJes du systeme contr61e. Dans Ie second th60reme il est supp0s6 que Ie systeme a contr61er est un systeme general line-ire stochastique du neme ordre. II est montre que si les estimatioDS des parametres convergent. la loi de contraJe obtenue est en fait 1a loi de contrele de
9. CONCLUSIONS
The paper has been concerned with control of systems with constant but unknown parameters. The analysis has been limited to single-input singleoutput systems with disturbances in terms offiltered white noise. A control algorithm based on least squares parameter estimation and a minimum variance regulator computed from the estimated parameters has been analysed. Assuming that the parameter estimates converge the closed loop system has been analysed. A characterization of the closed loop system has been given under weak assumption on the system to be controlled. Under stronger assumptions on the system to be controlled it has been shown that the regulator obtained will actually converge to the minimum variance regulator if the estimates converge. Since the closed loop system is characterized as a nonlinear stochastic system it is very difficult to give general conditions that guarantee that the estimates converge. The convergence has only been treated for simple examples and under further assumptions as in section 6. But simulations of numerous examples indicate that the algorithm has nice convergence properties. The simplicity of the algrorithm in combination with its asymptotic properties indicate that it can be useful for industrial process control. The feasibility has also been demonstrated by experiments on real processes in the paper and mining indust-
variance rninimale qui pourrait etre calculee si les paradu systeme 6taient connus, Ceci est q uelque peu surprenant car l'estimatioD des canes minimum est partiale. On discute certaines des implications pratiques des resultats, II est montre en particulier qu'il est possible d'appliquer l'algorithme a un petit ordinateur.
metres
Zusammenfassung-Betrachtet wird das Problem der Steuerung eines Systems mit konstanten, aber unbekannten Parametem. Die Analyse wird auf Systeme mit Diskrctzeit und einem Binpng bzw. einem Ausgang beschrinkt. Ein durch Kombination einer Schitzeinrichtung und der Methode der k1einsten Quadrate mit einer aus dem Schatzmodell berechneten Regeleinrichtung erhaltener Algorithmus wird analysiert. Die Hauptergebnisse sind zwei Theoreme, die unter der Annahme, daB die Parameterschatzungen konvergieren, den erhaltenen geschlossenen Regelkreis charakterisieren. Das erste Theorem konstatiert, daB bestimmte Kovarianzen des Ausganges und bestimmte Kreuz-Kovarianzen der Steuervariablen und des Ausgangs unter schwachen Annahmen ii her das zu regelnde System verschwinden. 1m zweiten Theorem wird angenommen, daB daszu regelnde System ein allgemeines lineares stochastisches System n-ter Ordnung ist. Gezeigt wird, daB bei Konvergenz der Parameterschitzung des erhaltenen Steuergesetzes in der Tat das Steuergesetz bei minimaler Varianz ist, das berechnet werden kann, wenn die Parameter des Systems bekannt waren. Das ist etwa iiberraschend, weil die Schitzung nach den kleinsten Quadraten angesteuert wird. Einige praktische Foigerungen aus den Ergebnissen werden diskutiert. Speziell wird gezeigt, daB der Algorithmus auf einem kJeinen ProzeBrechner leicht verwirklicht werden kann.
ries. 10. REFERENCES [l] K. J. AsTROM: Introduction to Stochastic Control Theory. Academic Press. New York (1970). [2] K. J. AsmOM and B. WlTTENMARK: Problems of identification and control. J. Math. Analysis App/ic. 34,
Pe3IOMe-PaccMaTpHBaeTCB CHCTeM
npH nOMoIllH
npo6neMa
perynHpOBaHHJI
nOCTOJlHHWX. HO HeH3BeCTHbIX napa-
MeTpOB. AHaJIH3 orp8HH'lHBaeTCJI CHCTeMaMR AHClCpeTHOro &peMeHH c O,llHHM BBOllOM He O,llHHM BWBO.llOM. A HaJtH3H"
90-113 (1971).
PYCTCJI anropHTM, DOny'lcHB.bIit nyTCM KOM6HHBPOBaHHJI
438
ICOaapBaHTHOCTIa BWBOJnl B HeXOTOpu ncpec:eKalOD,lU ItOaapRaHTHOCTIa XOHTPOnldlhlx nepeMeHBWX BbI8O,l.l& BC'le3" JIYT Upll HeOOJII»lDRX.llonyweRBJIX B peryJIHPyeMol CBCTeMC.
cHCTeMol: nopJI):tKa n.th, nOD3aHO, 'ITO, ecna OUCHXH napasserpoa CXO,IUITCJI, TO nOJIY'leBBwA 38JCOH pery 1IHpOBJCJI HBJIJIeTCJI ~eCJCII 3UOHOM peryJIBPOBKH MHIIIIMam.aoro XOJle6aHml, ICOTOpoe Morna 6.. 61a1TL BhI'DICJleHO npa B3BeCTIIYX oapaMerpax CBCTeMW. 3To JJ;O HClCOTOpol CTcnCHH HeoDll8HHOt Tax 1CU OUeBD HallMCllblDRX J(Ba)IpaTOS DOeTCJI CMe~eHHol. 06cyQ8.lOTCJI HelCoTopwe lJPUTB1Iecme BWBO,llW H3 PC3ym.TaTOB. B 'laCTROCTB, DOD3allO, lITO nOT anropBTM MOKCT 6Ja1Th npHMeBeH aa
So BTopol TeOpeMC npennonaraercs, 'ITO lCOHTpOJIBpyeMU CRCTeMa JlBJUleTCJI 06111el JlHHeitHoil croxacrasecsoa
COB.
OQeHOIC HaHMeHI»WHX KBa.llpaTOB c peryJIHpoBKOI MHHHMaJU.HLlX lCone6aHHl, BbJ'IHcnClfHhlX H3 paC1iCTHOI MOllC1tH. OcHOBHWMB
pC3YJllaTaTaMH
JlBJUJIOTCSI
ABe
TeOpeMW,
xaparrcPB3)'1ODJ.HC CHCTeMy 3aMlCHyToro KOHTypa, DOJl)"leHsoro B npe;uJOJlOZCHHH, 'ITO oueHKH napaMeTpOB CXO,lt-
JlTCJI.
DepBaJI
TeOpeMa
ymepXCJl'leT,
sro
HeKOTOpU
He6om.mol CtleTRO-8W'UICJIIITeJILBOI
439
MamHHe .lXJJJI
nponec-
Nonlinear Controllability and Observability ROBERT HERMANN AND ARTHUR J. KRENER
AT
the time of its publicationin 1977,this paperrepresented,
for the analysis of the "observability" properties of a nonlinear system, the same kind of breakthrough that the earlier 1972 paper by Sussmann and Jurdjevic (also included in this volume) had represented for the analysis of the reachability properties of those systems. In particular, the paper developed the appropriate language and mathematical tools for the analysis of certain properties of "local" observability, which lend themselves to an appealing algebraic test as well as the possibility of finding decompositions of the system into "observable/unobservable" parts. In an abstract system-theoretic setup, a system is considered to be "observable" if in its state-space it is not possible to find pairs of indistinguishable states, i.e., pairs of initial states that induce two identical input-output maps. This notion yields a very simple characterization in the case of linear systems, where the output is the sum of a function of the input (the zero-state response) and a function of the initial state (the zero-input response), which in tum is a linear function of its argument. In fact, in this case, two states are distinguishable if and only if the associated zero-input responses are different, Le., if their difference produces a zeroinput response that is not identically zero. In a nonlinear system, on the other hand, two states are distinguishable if there is at least one input function (not necessarily the zero-input) for which the output responses are different, and this input may in principle depend on the particular pair considered. In other words, the study of observability in a nonlinear system must explicitly take into account the way in which inputs affect the response, and this makes the analysis substantially more involved. Pioneering authors in the study of observability for nonlinear systems were Brockett [2], d' Alessandro et ale [3], Sussmann [11], [12], and Sontag [10]. In particular, Sussmann, generalizing the work of Brockett, [2], on Lie groups, showed that under appropriate hypotheses the relation of indistinguishability between states is a closed regular equivalence relation, which can be used to define a reduced system (with state defined on the associated "quotient" manifold) having the same inputoutput behavior as the original system but no pairs of indistinguishable states. Sontag studied the observability of polynomial
discrete-time nonlinear systems, and introduced the notion of "observation space." The approach of Hermann and Krener to the analysis of the observability properties of nonlinear systems was similar to the one that had proved very successful a few years earlier in the analysis of the controllability properties, namely the approach of differential geometry. In particular, considering systems described by equations of the form oX
y = h(x)
= F(x, u)
(1)
in which x E ~n, F (x, u) is an analytic vector field for every u E Q, with Q a fixed set, and h(x) is an analytic ~P-valued map, they realized that the appropriate geometric object to look at was the subspace (of differential forms) spanned by the differentials of the p components of the vector h(x) and their (repeated) Lie derivatives' along the vector fields of the set F= {F(x,u): U E Q}. In particular they proved that if, at a given point xo, this subspace has dimension equal to that of the state-space, then all states in a neighborhood. of this point are distinguishable form xo. In other words, the computation of (repeated) Lie derivatives of the differentials of the elements of h(x) along the vector fields of F yields a simple test for observability, very much in the same way the computation of the (repeated) Lie brackets of the vector fields of F yields a simple test for reachability. This test amounts to checking the rank of a matrix having n columns, whose rows consist of the differentials of the elements of h(x) as well as their (repeated) derivatives along the vector fields of F and appears as a natural extension of the familiar test which determines the observability of a linear 1The
differential of a real-valued function :A(x) is the one-form d:A(x) =
(~ ... ~) aXl
aX n
and the Lie derivative of a one-form w(x) along a vector field !(x), denoted L f w(x) is the one-form
441
system
x=Ax+Bu
y =Cx
by looking at the rank of a matrix whose rows are the rows of C (the differentials of the elements of C x) and rows of the form C A k (the differentials of the kth Lie derivative of C x along the vector field Ax). In case the subspace in question does not have dimension n at xo, but its dimension is constant, say d, then Hermann and Krener also pointed out that, in a neighborhood of xo, it is possible to find d real-valued functions ~l (x), ... , ~d(X), whose differentials span that subspace, and these can be used to define, locally near x", a reduced (d-dimensional) system having the same input-output behavior as the original one, but no pairs of indistinguishable states. This observation had a significant fallout in subsequent years, as it became one of the main ingredients for understanding the solution of the problems of decoupling and non-interacting control via state-feedback, thus paving the way to the systematic development of geometric methods for nonlinear feedback design. In particular, this observation and the results of [6], suggested to Krener and to the author of this preamble (in the course of the joint work which led to the paper [5]) that the set of functions ~l (x), ... , ~d(X), augmented with a complementary set of n - d functions having independent differentials, could be used for a local change of coordinates which highlights a decomposition into observable/unobservable parts. For example, in a singleinput single-output system modeled by equations of the form
x
= I(x) + g(x)u
y
= h(x)
(2)
the change of coordinates in question yields the decomposition
= 11 (Xl, X2) + gl (Xl, X2)U X2 = 12(X2) + g2(X2)U y = h2(X2)
Xl
where Xl is of dimension d and X2 has dimension n - d, and where it is clear that, from an input -output point of view, it is only the subsystem formed by the second and third equation (i.e., the "x2-subsystem") that matters (extensive discussions of these decompostions can be found in [4]). Now, it is immediately clear from this that if the state equation of (2) is perturbed by an additive term of the form p(x)w in which w represents an extra input channel and, in the new coordinates, the vector field p(x) transforms into a vector of the form
(Pl(X~' X2») then the input w has no effect on the output y (whatever the initial state might be). In other words, the input w is decoupled from
the output y. From this to the next stage, i.e., understanding how to use feedback so as to enforce decompositions of this kind, the step was short and the geometric theory of nonlinear feedback design was born. Another relevant offspring of Hermann and Krener's theory of nonlinear observability, which appeared a few years later, was the identification of conditions under which there exists a local change of coordinates z = (x) transforming a system of the form
x = f(x)
y
= h(x)
(3)
to a system of the form i = Az
+ ¢(y)
y =Cz
with (A, C) an observable pair (see [1] and [7]). The importance of this result is that for a system having this property it is very easy to construct a state observer, in the form ~ = (A - GC)~ - ¢(y) - Gy
which yields an observation error e = (x ) -
~
satisfying
e=
(A - GC)e. The kind of "observability" property thus charac-
terized was used in many applications and has led, for instance, to the development of the nonlinear adaptive controllers studied by Marino and Tomei [9] and by Krstic et al. [8]. REFERENCES
[1] D. BESTLE AND M. ZEITZ, "Canonical form observer design for nonlinear time-variable systems," Int. J. Contr., 38:419-431, 1983. [2] R.W. BROCKETT, "System theory on group manifolds and coset spaces," SIAM J. Contr., 10:265-284,1972. [3] P. D'ALESSANDRO, A. ISIDORI, AND A. RUBERTI, "Realization and structure theory of bilinear dynamical systems," SIAM J. Contr., 12:517-535, 1974. [4] A. ISIDORI, Nonlinear Control Systems: An Introduction, 3rd ed., SpringerVerlag (New York), 1995. [5] A. ISIDORI, AJ. KRENER, C. GORI GIORGI, AND S. MONACO, "Nonlinear decoupling via feedback: a differential geometric approach," IEEE Trans. Aut. Contr., AC-26:331-345, 1981. [6] AJ. KRENER, "A decomposition theory for differentiable systems," SIAM J. Contr. Optimiz., 15:289-297,1977. [7] AJ. KRENER AND A. ISIDORI, "Linearization by output injection and nonlinear observers," Syst. Contr. Lett., 3:47-52, 1983. [8] M. KRSTIC, I. KANELLAKOPOULOS, AND P. KOKOTOVIC, Nonlinear Adaptive Control Design, Wiley (New York), 1995. [9] R. MARINO AND P. TOMEI, "Global adaptive output feedback control of nonlinear systems," IEEE Trans. Aut. Contr., AC-38:17-48, 1993. [10] E. SONTAG, Polynomial Response Maps, Springer-Verlag (New York), 1979. [11] H. SUSSMANN, "Existence and uniqueness of minimal realizations of nonlinear system," Math. Syst. Theory, 10:263-284,1977. [12] H. SUSSMANN, "Single input observability of continuous time systems," Math. Syst. Theory, 12:371-393, 1979.
A.I.
442
Nonlinear Controllability and Observability ROBERT HERMANN AND ARTHUR J. KRENER, MEMBER,
IEEE
AIMtrtId-11ae properties of coatroUabllity, observability, aDd the theory 01 . . . . . reaIizadoB 'or Ibaear systeDas _. weU-understooci and have beeD 'elY Ulelul iD ...tyz...... systems. Ttus paper deals widl JUIIIIo.
the state space M is still too large, then one would like to apply a systematic technique to reduce M and still preserve the input-output structure of the model. . . . ~ for DOIIIIaear syste8. Using the ideas of controllability and observability, in the early 1960's Kalman and others carried out this pro.. I. INTRODUCTION gram for linear systems. The similar questions for nonlinFREQUENTLY, control systems of the following form ear systems were not effectively treated until the early are used to model the behavior of physical, biological, 1970's. Based on the work of Chow [5], Hermann [9], Haynes-Hermes [8], and Brockett (9] and working indeor social systems, pendently, Lobry [21], (22), Sussman-Jurdjevic [25], and x= f(x,u) };: ( 1) Krener [19], [20] developed the nonlinear analog of linear y=g(x) controllability in terms of the Lie algebra ~ of vector fields on M genera ted by the vector fields f (., u) correwhere u E {l, a subset of R I, x E M, a C-' 00 connected sponding to constant controls u E Sl. manifold of dimension m, y E R" and f and g are COO It was shown that if the dimension of ~ is constant or if functions. the system ~ is analytic, then there exists a unique maxiThe control variable u represents the externally applied which carries all the controls or exogenous inputs to the system and the output mal submanifold M' of M through xO O variable y represents the observable parameters of the trajectories of I passing through x such that any point system. The state variable x mayor may not be directly on this submanifold can be reached from xO going for.. measurable and is used to represent the memory of the ward and backward along the trajectories of the system. system. The past history of ~ affects its future evolution In particular if the dimension of §7(x~ is m then M = M' and so the system is "controllable" in some sense, to be only through information convey·ed by this variable. The state space M may be deficient in one of two ways. made precise in Section II. If it is less than m then ~ can It may be too small to adequately represent the full be restricted to M' where it is "controllable." This is one variety of memory states, i.e., it may fail to distinguish half of reducing the state space M. For linear systems between real states where some control exictes different observable effects. If this is the case, then the mathematical system ~ fails to adequately model the real system and hence, must be revised.
x=Ax+Bu y=Cx
(I)
On the other hand, the state space may be too large. where uER', xER m , yERn, this reduces to the wellThe system may not be controllable, i.e., if ~ is known to known criterion that the rank of the matrix be in a given state xO E M at some time there may be (B:AB:··· :Am-1B) other states x E M where the system cannot possibly get to or have come from using the given set of inputs. Or there bem. may be distinct states which are indistinguishable from an The other half of the program of reducing the state input-output point of view, i.e., if the same input is space was supplied by Sussmann {28] for analytic or applied in either of these states then the same output symmetric systems. This generalized the work of Brockett results. These problems can be caused by ignoring possi[3] on Lie groups. Sussmann noted that indistinguishabilble real inputs or real observables, in which case these ity is an equivalence relation on. M and showed that for must be added to the model. If, after all the significant analytic "controllable" or symmetric "controllable" sysinput and output variables have been incorporated into I, tems, indistinguishability is a closed regular equivalence relation so that the quotient is another manifold. He also Manuscript received December 16 1976; revised June 7, 1977. R. showed that the quotient inherits a system which has the Hermann's work was supported in part by a National ResearchCouncil Senior Postdoctoral Research Associateship at Ames Research Center same input-output behavior as ~ but is "controllable" (NASA), and in part by the National Science Foundation under Grant and observable, i.e., has no indistinguishable states. MCS7s.07993. A. J. Krener's work was supported in part by the Na.. In this paper we take a different approach to nonlinear tioual Science Foundation under Grant MPS75-0S248. Paper recommended by J. Davis, Chairman of the Stability, Nonlinear, and Distrib- observability, related to that of previous authors [16H18] uted Systems Committee. R. Hermann is with the Department of Mathematics, Rutgers Univer- but which is more in the spirit of the approach to nonlinsity, New Brunswick, NJ 08903. ear controllability described above. In particular we deA. J. Krener is with the Department of Mathematics, University of velop results involving observability which are analogous California, Davis, CA 95616. t
Reprinted from IEEE Transactions on Automatic Control, Vol. AC-22, October 1977, pp. 728-740.
443
to the controllability results depending on the dimension of ~(x) . The relevant objects in this study are §, the smallest linear space of functions on M which contains the observations gl(X),··· ,gll(x) and which is closed with respect to Lie differentiation by £?f, and the differentials of § denoted by d§. If the dimension of drJ is constant over M then indistinguishability need not be regular but there is a related regular equivalence relation which we can use to factor M. On the quotient there exists a system with the same input-output behavior as I but which is "observable" in the sense that neighboring points are distinguishable. In particular if the dimension of d§ (x) is always m then ~ has this property. For linear systems this reduces to the well-known criterion that the matrix
c
--- - - -CA
be of rank m. In order to bring out the "duality" between "controllability" and "observability" (which is, mathematically, just
It is with no loss of generality that we assumeyERn. If y EN, a Coo manifold, then by the Whitney Imbedding Theorem, N can be imbedded in R" for some n which then can be taken as the range of g. The assumption of infinite differentiability for M,!, and g is not essential but is only invoked to avoid counting the degree of differentiability needed in a particular argument. Occasionally we consider analytic systems where these are assumed to be an analytic manifold and analytic mappings.. Nonautonomous systems are handled in the standard fashion by assuming time is one of the state variables. We also assume the system is complete, i.e., for every bounded measurable control u(t) and every xO EM there exists a solution of the differential equation .x =!(x(t), u(t» satisfying x(t~=xo and x(t)EM for all fER. We use the notation (u(t),[tO,t l ] ) to denote functions defined on [to,t l ] . Given a subset U ~ M, Xl is V-accessible from XO (denoted by xlAux~ if there exists a bounded measurable control (u(t),[tO,t l ]) satisfying-u(t)ESl for tE[tO,t l ] such that the corresponding solution (x(t),[tO,t l ] ) of the differential equation (I) satisfies x(t~=xo, x(tl)=X 1 and x(t)E U for all tEltO,t']. M-accessibility and AM are simply referred to as accessibility and A. Given any rela.. tion R on M we use the notation
the duality between vector fields and differential forms), R (xO)= {Xl EM = xIRxO}. we first review in Section II the known facts concerning nonlinear controllability and then in Section III we dis- For example A (x~ is the set of points accessible from xO. cuss our approach to nonlinear observability. Finally Sec- The system}; is said to be control/able at XO if A(x~= M tion IV deals with the question of minimality for nonlin- and contro/lable if A (x) = M for every x EM. ear systems. If ~ is controllable at xO it still may be necessary to travel a considerable distance or for a long time to reach points near to xo. As a result this type of controllability is II. NONLINEAR CONTROLLABILITY not always of use and so we introduce a local version of We consider the system ~ described in Section I. Recall this concept. ~ is locally controllable at x 0 if for every a C~ (analytic) m-dimensionalmanifold M is a Hausdorff neighborhood U of xo, Au(x~ is also a neighborhood of topological space with 'a CCXJ (analytic) structure, i.e., a xo; ~ is locally controllable if it is locally controllable at countable cover of coordinate charts (U", x a ) where every x EM. (This is called local-local controllability by Haynes and Hermes [8].) 1) U" is an open set in M, Accessibility is a reflexive and transitive relation but for 2) xa==col(xal'··· ,xam ) : ua~Rm is a homeomorphism nonlinear systems it need not be symmetric. For this onto its range, and reason we need a weaker relation. Given an open set 3) if (UG,xa ) and (UfJ,xp ) are two such coordinate U M there is a unique smallest equivalence relation on charts then the change of coordinates x{J 0 xa- I : xa ( U" n U which contains all V-accessible pairs. We call this UfJ)~xp( U" n U/l) is C ~ (analytic). relation weak Usaccessibility and denote it by WA u' It is For further details we refer the reader to Hermann [10] or easy to see that x'WAux" iff there exists xo,... . x k such Boothby [2]. A non-Hausdorff manifold is a non-Haus- that xO=x', xk=x" and either xiAuX i - 1 or Xi--1Aux i for dorff topological space with such a structure. If M is not i= 1,'·· .k, Weak M-accessibility and WA M are simply Hausdorff then solutions of ordinary differential equa- referred to as weak accessibility and WA. L is weakly control/able at XO if WA(x<)=M in which case WA(x)= tions on M need not be unique. Throughout this section we simplify notation by assum- M for all x and so L is weakly controllable. Notice that weak controllability is a global concept and ing M admits globally defined coordinates x = col (XI' · · · ,xm ). This allows us to identify the points of M does not reflect the behavior of L restricted to a neighbor.. with their coordinate representations and to describe con- hood of xo. So again we introduce a local concept. L is trol systems in the familiar fashion (1). However when locally weakly controllable at XO if for every neighborhood dealing with nonlinear observability in-the next section we U of xo, WAu(x~ is a neighborhood of xu. L is locally weakly controllable if it is locally weakly controllable at shall be forced to consider several coordinate charts.
c:
444
~ y/(x,tO)= h( y,(x, to))
every x EM. Clearly (local) controllability implies (local) weak controllability and it is not hard to show using the transitivity of (weak) accessibility and the connectivity of M that local (weak) controllability implies (weak) controllability, i.e., we have the following implications: ~
Yto(x, to) = x. l
Then 1to(·,t ) is a diffeomorphism of a neighborhood Vof x 1 onto a neighborhood of xo. Moreover we can choose V ~Au(x~ sufficiently small so that ytJ
=> ~ controllable ~ ~ locally weakly controllable ~ }: weakly controllable.
locally controllable
o
~
The advantage of local weak controllability over the other forms of controllability discussed above is that it lends itself to a simple algebraic test.. First we introduce some additional mathematical concepts. The set of all COO vector fields on M is an infinite dimensional real vector space denoted by ~(M) and also a Lie algebra under the multiplication defined by the Jacobi bracket [h l,h 2l given by
In general there are no other implications but for autono-
mous linear systems it can be shown that all four concepts are equivalent. The following result gives an intuitive interpretation of local weak controllability. Loosely put it shows that}: is locally weakly controllable iff one needs local coordinates of dimension m to distinguish the trajectories of ~ from any initial point. Theorem 2.1: ~ is locally weakly controllable iff for every x E M and every neighborhood U of x the interior of A u (x ) :F0 . Proof: Suppose ~ is weakly controllable. Given any XO E M and any neighborhood U of XO we can choose u l en such thatfl(x)== f(x,u l ) is not zero at XO (assuming m>O, if m=O then the result is immediate). Let s~y}(x) denote the flow of 11, i.e., the family of solutions of the differential equation
ah2
where hI' h2 and [hl,h2]E~'t(M) . Elements of ~(M) are represented by column m-vector valued functions of x. For any fixed hi E ~,(M) the real linear transformation from CX(M) into itself which sends h2~[hl,h2] is called Lie differentiation with respect to hi and is denoted by
i;
Each constant control u EO defines a vector fieldj(x,u) E ~ (M), we denote by ~ 0 the subset of all these vector fields. Let §' .denote the smallest subalgebra of ~(M) which contains ~ 0. A typical element of ~ is a finite linear combination of elements of the form
d ds y;(x)= P (y;(x»
satisfying the initial conditions
[Ik - I ,Jk] ." ]]] wherej'{x) = f(x,u i ) for some constant u' En. We denote
yJ(x)=x.
[ / I [ /2 [ . • .
For some £>0 the set V 1 = { Ysl(XO:> :O< s < E: } is a submanifold of U of dimension 1. Suppose inductively Vj - ·I is a j -1 dimensional submanifold of U defined by
Vj - I == {.. Jr ) 0 r~_1
• ••
0
",I
'''I
ah,
[h.,h 2](x)= ax(x)h1(x)- a; (x)h 2(x)
(Xo) :(s I" ... s,J-I )
in some open subset of the positive orthant of Hj-I }
by ~(x) the space of tangent vectors spanned by the vector fields of GJ at x. ~ is said to satisfy the control/ability rank condition at XO if the dimension of '.J(x~ is m; }: satisfies the controllability rank condition if this is true for every xEM. Theorem 2.2: If}: satisfies the controllability rank conXO then ~ is locally weakly controllable at xo. Proof: The proof is very similar to Theorem 2.1. We start by choosing a neighborhood of U of XO small enough
dition at
where y;(x) is the flow of ji(X) = f(x,u i ) for some u' EQ. Clearly Vi-I ~Au(x~. Ifj <; m we construct Vi by choosing a II E U and x j - 1 E Vi - I. This is always possible for if not then every trajectory of ~ starting on Vi - 1 would remain on Vj - 1 for a while. This contradicts the local weak controllability of ~. It follows that we can choose an open subset of the p
so that ~ satisfies the controllability rank condition at every x E U. We construct a sequence of submanifolds as before but this time there is a different reason why one can always choose u' En and x J - 1 E Vj - I such thatfJ(x) =f(x,rl) is not tangent to V j - l at xj-I. If this is not possible then §" restricted to Vi -I is a subalgebra of X( Vi - I) which implies the dimension of ~(x) ~ j - I < m on Vj - I U which is a contradiction. It follows that for every neighborhood U of xo, the interior of Au(x~ is not empty. The second 'half of the proof of Theorem 2.1 implies WAu(x~ is a neighborhood
c:
of xo. 0 From the above we see that if ~ satisfies the controllability rank condition then it is locally weakly controllable. The converse is almost true as we shall see later on in Theorems 2.5 and 2.6.
445
Suppose the trajectories of I are required to satisfy the
been shown by Sussmann [29] that for any C 00 system, WA (x~ can be given the structure of a C 00 manifold, in general it will not be an integral submanifold of ~. Theorem 2.4 (Sussmann [29]): Given any realization (~, xO:>, there exists a weakly controllable realization (~',x~ of the same input-output map on M'= WA(x~. Notice that unlike Theorem 2.3 the above does not guarantee that If is locally weakly controllable or satisfies the controllability rank condition. Locally weakly controllable systems "almost" satisfy the controllability rank condition. Theorem 2.5: If ~ is locally weakly controllable then the controllability rank condition is satisfied genericaJJy, i.e., on an open dense subset of M.
initial condition
x(tO)==XO, then I defines a map from inputs to outputs as follows. Each admissible input (U(/),[/O, II]) gives rise to a" solution (x (t), (1°, II]) of i == f(x, u(t» satisfying the initial condition. This, in turn, defines an output (y( t), [to, t I) by y(t)==g(x(t». We denote this map by
Ixo:(u(t), [to,tl])~(Y(/), [to,t1 ] ) and call it the input-output map of }: at xo. Given a map P:(U(/), [/°,1 1])t-+(Y(/), [/O,t l])
Proof: For any system the controllability rank condi-
from inputs to outputs, the pair (~, x~ is said to be a realization of {J if ~JC0:= fJ· Now suppose I is neither locally weakly controllable nor weakly controllable, Given the input-output map I xo we would like to find another realization (~',z~ of this map which is weakly controllable in some sense. The obvious way to proceed is to find a submanifold M' of M which contains xO and all the trajectories of I passing through XO then let ~' be the restriction of ~ to M' and zo::: xo. If M' is chosen small enough then hopefully ~' will be weakly controllable in some sense. To carry out this program we introduce some mathematical tools. A connected submanifold M' of M is an integral submanifold 'of Cff if at each x EM' the tangent space to M' at x is contained in ~(x). M' is a maximal integral submanifold of Cff if it is not properly contained in any other integral submanifold of ~. There are two important cases when f!f has maximal integral submanifolds. Frobemus Theorem [10]: If the dimension of lff(x) = k for every x EM, then there exists a partition of Minto maximal integral submanifolds of Gf all of dimension k. Hermann-Nagana Theorem [11J, [J2], [14], [23]: If the system is analytic then there exists a partition of M into maximal integral submanifolds of ~ of varying dimensions. The dimension of §"(x) can vary but it will be constant on each submanifold of the partition and equal to the dimension of that submanifold. The relationship of these two theorems with controllability is given by the following. Chow Theorem [5]: If either of the above is satisfied and M' is the maximal integral submanifold of OJ containing xO then M'== WA(x~. This leads to the following. Theorem 2.3: Suppose (I,x~ is a realization of an input-output map such that either 1) ~(x) is of constant dimension, or 2) ~ is analytic, then there is a locally weakly controllable realization (~',x~ of the same input-output map on the maximal integral submanifold M' of ~ containing xo. In fact, }:' satisfies the controllability rank condition. For completeness we mention the case when}: does not satisfy either of the hypotheses of Theorem 2.3. It has
tion is satisfied on an open subset of M, possibly empty. To see that it is dense for locally weakly controllable systems suppose there exists an open subset U of M where the dimension of §="(x) < m. Without loss of generality we can assume dim'?J(x)=k<m for all xE U. For some xO E U, let V' denote the maximal integral submanifold of ~(x) in U given by the Frobenius Theorem, then Au(x~ k U' and so using Theorem 2.1 we see ~ is not locally
weakly controllable at xo. 0 For analytic systems we can strengthen the above, in fact, weak controllability, local weak controllability, and the controllability rank condition are equivalent. Theorem 2.6: If ~ is analytic then L is weakly controllable iff it is locally weakly controllable iff the controllability rank condition is satisfied. Proof: We have already shown that for Coo systems the controllability rank condition implies local weak controllability which implies weak controllability. The reverse implications follow for analytic systems if we show that xOWAx1 implies that the dimension of ~t(x~ and ~(Xl) are the same. For then, if ~ is weakly controllable, the dimension of ~(x) must be constant and hence equal to m by the Frobenius and Chow Theorems. To show that xOWAx l implies the dimension of ~(x~ and GJ(x l ) are the same, it suffices to consider the case x I = 'Y(x~ where y is the family of solutions of the vector fieldf(x)=f(x,u) for some constant uE1l and x>O. The map (o/ax)'Y _s(x 1) is a linear isomorphism from the tangent space at x I to the tangent space at xo. For any h E ~(M), the Campbell-Baker-Hausdorff formula allows us to expand (a / ax)y _s(xl)h(x 1) in a convergent senes
In particular, if h E ~f then the right side of the above is a vector in £j'(xO:> so (%x)'Y_s(x carries ~J(XI) into ~(x~. Therefore, the dimension of LJ(x I) c dimension of '?f(x~. Reversing the argument shows that the inverse map (a / ax)ys(xo) carries §"(x~ into ~f(x I) and the result J
)
follows.
446
0
Example 2.7: Consider the linear system (2). In this
case ~ 0 == {Ax + Bu : u EO}, so the Lie algebra is generated by the vector fields {Ax,B• .,··· ,B./} where B*j denotes the jth column of B considered as a constant
From this it can be shown that the controllability rank condition is equivalent to the familiar requirement that
vector field. Computing brackets yields
[Ax,B.j ] = -AB.j [Ax[ Ax,B.j
] ]
for every t E R where
[B.j,B. k ] =0
=A2B.j
f:.cB(t)=
[ B. j [ Ax,B.k ] ] =0
and so OD. The Cayley-Hamilton Theorem implies that Cff is spanned by the linear vector fields Ax and the constant vector fields AiB.j where ;=0,··· ,m-l andj= I,··· ,I. This system is analytic so the Hermann-Nagano Theorem guarantees the existence of maximal integral submanifolds of ~ through each x E R "', In particular if we let M' denote the maximal integral submanifold through x'=O then M' must contain the integral curves of all linear combinations of the constant vector fields {A iB. j } starting at 0 and hence must include the linear subspace spanned by these vectors. In fact M' is precisely this subspace because at each x EM' the tangent space to M' contains Ax also. Notice that in this context the controllability rank condition reduces to the well-known linear controllability condition that
~B(t)-A(t)B(t),
f:.~B (t) = (~ - A (t) )f:.~-lB (t). See (4), [6], or [15J. Example 2.9: Consider the bilinear system I
x=Ax+
L uiBix, i=l
where uEU=R I and M is either R'" or an m-dimensional subgroup of GI(h,R), the group of invertible h X h matrices. The Jacobi brackets of linear vector fields Dx, Ex is seen to be
[Dx,Ex] = [D,E]x where
rank(B:AB:··· :Am-JB)=m.
[D,E] =ED- DE
The controllability rank condition only implies local weak controllability but for linear systems it also implies con- is the commutator of the matrices D and E. ~ is isomorphic to a Lie subalgebra of g/(mR) if M=Rm and trollability (see [4] for details). gJ(h,R) if M is a subgroup of G/(h,R}. (g/(m,R) is the If the controllability rank condition fails and the rank Lie algebra of aU m X m matrices with commutation as the of the above matrix is m' < m, then by restricting the bracket). Let F denote the group system to M '= Rm', the maximal integral submanifold of ~ through 0, we obtain a controllable linear system. F = { exp X : X E ~}. Notice that in this case the dimensions of the maximal integral manifolds of ~ can vary, for if Ax ~ span {A iB*j} The maximal integral manifold M' of OJ through Xo is just then the maximal integral manifold of §" through x will be the orbit of xO under F of dimension> m', It cannot be a linear subspace of R'" M'=Fx O because it does not contain 0 which must lie in a maximal integral manifold of dimension m'. and ~ satisfies the controllability rank condition iff Facts Example 2.8: Consider the linear system (2) but where transitively on M. A (I), B(t), and C(t) are Coo matrix valued functions ,of time. We adjoin another state variable Xo= 1 and rewrite III.
the system equations
xo=l
i=A(xo)x+B(xo)u Yo=x o
(3)
y:=C(xo)x.
The structure of ~ is similar to that for autonomous
systems, for example,
[
(~(xO)x ~.J{xo) ),(
NONLINEAR OBSERVABILITY
We consider the system ~ as described in Section I and the input-output map of the pair (~, x~ as described in Section II. A pair of points XO and x I are indistinguishable (denoted xOlx l ) if (~,x~ and (~,Xl) realize the same input-output map, i.e., for every admissible input (u(t),[tO,t l])
}:xo(u(t), [to,tJ])=~XI(U(t), [to,/I]). ) ] =[
a:
o
B.ixo) _OA (Xo)B.ixo)
1
[(~.j(XO))'(~.k(XO»)]=(~).
Indistinguishability J is an equivalence relation on M. ! is said to be observable at xO if I (x~ = {x"] and I is observable if I (x) = {x} for every x E M. Notice that the observability of ~ does not imply that every input distinguishes points of M. If, however, the output is the sum of a function of the initial state and a
447
ax
function of the input, as it is for linear systems, then it is The gradient d
h2
h2
1h"
hl
2
From this it follows that § is closed under Lie differentiaU-indistinguishability is not, in general, an equivalence relation on U for it fails to be transitive. This is related to the fact that I restricted to U is not necessarily complete. (See Sussmann [27] for a fuller discussion of this point.) However, we can still define ~ to be locally observable at XO if f~r every open neighborhood U of xO, I u( x 0.> = { x°), and ~ IS locally observable if it is so at every x E M. On the other hand one can weaken the concept of observability; in practice it may suffice to be able to distinguish xO from its neighbors. Therefore we define ~ to be weakly observable at XO if there exists a neighborhood U of XO such that I (x~ n U = {XO} and ~ is weakly observable if it is so at every x EM. Notice once again that it may be necessary to travel considerably far from U to distinguish points of U, so we make a last definition, ~ is locally weakly observable at XO if there exists an open neighborhood U of XO such that for every open neighborhood V of XO contained in U, Iv(x~ == {x"] and is locally weakly observable if it is so at every x EM. Intuitively, ~ is locally weakly observable if one can instantaneously distinguish each point from its neighbors. It can easily be seen that the relationships between the various forms of observability parallel that of controllability, i.e., ~
locally observable Jj, ~ locally weakly observable
=>
~
observable
~
~
weakly observable.
JJ
tion by elements of Cj" also. Let ex *(M) denote the real linear space of one forms on M, i.e., all finite C~(M) linear combinations of gradients of elements of COO(M). These are represented by row m-vector valued functions of x. The pairing between one forms and vector fields denoted by (w,h>E COO(M) is just multiplication of 1 X m and m X 1 matrix valued functions of x. We define a subset of ~X,*(M) by d§o={dq>:
Lh(w)(x) =(
a;:* (S)h(x»)*+w(x) ~~(x)
where wE 'X *(M), hE 'X(M) and * denotes transpose. The three kinds of Lie differentiation are related by the following Liebnitz-type formula
Lh l (W,h2 ) = (Lh rw,h 2 )
+ (w, [ h"h 2 ]>.
If w = dq;, then L h and d commute i, ( dcp) = d ( i; (cP)). From this it follows that d§ is the smallest linear space of one forms containing dgG which is closed with respect to Lie differentiation by elements of ~ 0 (or 'J). Elements of d§ are finite linear combinations of one forms of the form
In general there are no other implications but for autonomous linear systems it can be shown that all four are wherejj(x)=f(x,uJ ) for some constant uJEn. As before equivalent. The advantage of local weak observability we denote by d§ (x) the space of vectors obtained by over the other concepts is that it lends itself to a simple evaluat!ng the elements of d~ at x. The space d§ (x") algebraic test. To describe it we need some additional determines the local weak observability of L at xo. L is tools. said to satisfy the obseroability rank condition at XO if the Let CCX)(M) denote the infinite dimensional real vector dimension of' d§ (x") is m, L satisfies the obseroability space of all Coo real valued functions on M. Elements of rank condition. if this is true for every x E M. CX(M) act as linear operators on CCX)(M) by Lie difTheorem 3.1: If ~ satisfies the observability rank conferentiation. If hECX(M) and cpECOO(M) then L,,(tp)E dition at XO then ~ is locally weakly observable at xc. C~(M) is given by The proof depends on the following. Lemma 3.2: Let V be any open subset of M. If xu, x 1 E V and xOlvx 1 then q;(X~=
Proof" If xOIVX I then for any k >0, any constant controls u 1,. •• ,Uk EQ, small Sl'·· ,Sk ~o and gi' i= I,. · . ~n, we have f·
1
-1
Here y;(x) denotes the flow offi(x)=f(x,u i ) . Differentiating with respect to Sk" • • ,Sh at 0 yields
-1
Fig. 1. Dynamics and observers of Example 3.4.
g is spanned by functions of this form so the lemma
0
follows.
maximal rank. A necessary and sufficient condition for
Proofof Theorem 3.1: If the dimension of dg (x") = m the quotient topology on M / R to be Hausdorff is that R then there exists m functions {J)I'·· ,qJmE§ such that be a closed equivalence relation, that is, the graph of R be a dfP(x~,· · • ,dCPm(x~ are linearly independent. Define a closed subset of M X M. An equivalence relation R which map admits a Coo structure on M / R compatible with the projection is called regular. A necessary and sufficient condition for regularity is that the graph of R be a The Jacobian of
It is not hard to see using the continuity of solutions of I v(x~ = {xo} so ~ is locally weakly observable at xo. 0 differential equations with respect to initial conditions From. this we see that if ~ satisfies the observability that for any Coo system L the relation I is closed. Howrank condition then ~ is locally weakly observable. The converse is almost true as we shall see later on in Theo- ever, the following example due to Sussmann shows that it need not be regular even when the controllability and rems 3.11 and 3.12. observability rank conditions are satisfied. Suppose (~, x~ is a realization of an input-output map Example 3.4.· Let uEQ={(Ut,U2):Uj >O) xEM=R, Y which is not observable in any of the above senses. We 2 ER and tum our attention to finding such a realization (~',z~ of x= UtI} (x) + u2f2 (x) the same input-output map. To understand the difficulties involved consider the following. YI=gl(X), Y2=g2(X). Example 3.3: Let uEfl=R, xEM=R,yER 2, x(O)= We choose ./;,gj :R~R to be C'" functions with the xO=O, and following graphs. See Fig. 1. x=u Y.=cosx Y2=SlnX. Since 11 and 12 have no common zeros the system satisfies the controllability rank condition and is locally Clearly the system satisfies both controllability and ob- weakly controllable. Since dg, and d82 have no common servability rank conditions. Therefore, it is locally weakly . zeros -the system satisfies the observability rank condition controllable and locally weakly observable. It is not ob- and is locally weakly observable. k=xo+2k." servable because XO and x are indistinguishFrom the graphs of 81 and 82 we see the only possible able for any xO and any integer k. . pairs of indistinguishable points are x and - x where To obtain an observable system with the. sa~e 10- Ixl> 1. Since J;(x) = I if x > I and J;(x) = -I if x <; -1 it put-output behavior as the original we must identify XO can be seen that these pairs are indistinguishable. and x", that is, define a system ~' on M' = the unit If we quotient the state space M = R by the equivalence circle by 0 (0)"" 0° = 0 and relation of indistinguishability, the result is clearly not a manifold for it looks like a circle with a ray attached. At 9=u Yl=COSO Y2=sin(J. this point one can ask what additional assumptions on ~ Note that (~,x~ and (~',8<>:> realize the same input-out- are needed in order to insure that I be a closed and put map. regular equivalence relation. Sussmann has proved the This example seems to imply that one can obtain an following. observable realization from one that it is not by "factorTheorem 3.5 [27}: If (~,xc) is a weakly controllable ing" by the relation I of indistinguishability, the new realization and either system lives on the state space M' = M / J. However, given a) ~ is analytic or an equivalence relation R on M it is not always true that b) }; is symmetric (i.e.,VuED 3vED such thatf(x,u) M / R with the quotient topology is Hausdorff and admits = - f(x, o) \:Ix EM), then I is a closed and regular equiva Coo structure in such a way that the canonical projec- alence relation. Moreover, there exists a system ~' on tion 'IT: M -+ M / R is a submersion, i.e., aC 00 map of M' = M / I such that (~', I (x~) is an observable and
S"
449
weakly controllable realization of the same input-output map. If ~ is analytic, then ~' is also locally observable. In some sense Theorem 2.3b is the dual of Theorem 3.5a and Theorem 2.4 is the dual of Theorem 3.5b. What we would like to discuss now is the dual of Theorem 2.3a. Proceeding analogously we might expect that I is a regular equivalence relation if the dimension of d§ (x) is constant over M but Example 3.4 shows this not to be the case. Just as we have used the relation WA in addition to A when studying nonlinear controllability so must we introduce another relation for nonlinear observability. We call this new relation strong indistinguishability, XO and x 1 are strongly indistinguishable (denoted by xOS/x J) if there exists a continuous curve a: [0, l]~M such that a (0) = xo, a(l)= Xl and xOla(s) for all s E[O, 1]. Clearly S/ is an equivalence relation, xOSlx 1 implies xOlx1, and ~ weakly observable at X O implies Sl(x~={xo}. As we shall demonstrate in a moment, if the dimension of d~ (x) is constant over M then S/ is a regular equivalence relation and the quotient M', possibly non-Hausdorff, inherits a locally weakly observable system ~' which realizes the same input-output map as I. Before proving this we must introduce some more machinery. Let X(x) denote the space of all tangent vectors at x which annihilate every element of d§ (x)
for i= 1,··· .k, Moreover if fE~T and dcpEd§ then Lf(dcp) E d§ so there exists functions J.Lir(t) such that k:
Lj, (dq;;)( Yt(xO, to)) = ~ llir(t)d
fori=l,.··,k. Combining these three equations we obtain
for i = 1,· .. ,k. This is a linear homogeneous differential equation so there exist invertible linear transformations A(t): Rk~Rk such that
(~j( t») = A(t)(A;j( to)). The lemma follows since d§ (x~ is spanned by m
dcp;(xO)= ~ A;j(tO)dxj )=1
for ;=}, .. ·,k and for small It-tOI, d§(Yt(xO,t~)a / aXYt(xo, to) is spanned by
X(x)= {TE TxM:
lemma.
dfl'i('Y,(XO, to»)
~'Y' (xO, to) = L ~\i t)dx x
j
j
fori=l,··,k. 0 Remark: The above depends heavily on the fact that the dimension of dg (x) is constant so that d§ is "locally finitely generated." For similar results see Hermann [10]. A curve a: [0, 1]--:)oM is piecewise Coo if it is COO at all but finitely many points of [0, 1] and left (right) limits of a and all its derivatives exist at every point of (0, I] ([0, I». We define another equivalence relation on M,xoHx l if there exists a continuous and piecewise C:lO curve a : [0, I] ~M such that a(O)=xo, a(I)=x 1 and
Lemma 3.6: Suppose the dimension of d§ (x) is k for every xEM. LetJ;(x) be a time-dependent vector field on M such thatJ,(·)E~ for every t [for example, if u(t) is an admissible control and ft(x) = f(x,u(t»). Let Yt(XO,t~ be the flow of it, i.e.,
~ 'Y,(XO,tO)=J;(y,(XO,tO)) 'to(xO, to) = xo. Then
~a(s)EX(a(s».
D if. S·InceYt+s ( x 0 ,I0\J=Ys «Yt X 0,10\j, t)·t rroot: ,1 SUIfiIces t0 consider only small It - tOt. Choose dcp.,··· ,dCPk E d§ which are linearly independent on a neighborhood U of xo. A straightforward calculation yields
0) ax
0)
d( . ( 0 aYt 0 dt (/(Pi 'Y, (X ,t) (X .t )
= LJ,(dq>;)('Y,(XO, to»
Corollary 3.7: Assuming the dimension of d{J (x) is constant, if xOHx· then xOSlx 1• Prooi: '). Let (u(t), [to, 11]) be any admissible control with flow Yt(x, t~. The ith component of the output at time t when the system is started at a(s) at time to is
Yi( t) = gi( Yt( a (s), {O)).
The derivative with respect to s is given by
~~ (XO,tO).
On the other hand we can choose functions A;j(t) such that
which is identically zero by the last lemma. This next result shows that 'J is a collection of symme450
try vector fields (in the sense of Sussmann (26]) for the relation H. Corollary 3.8: Assume that the dimension of d§ (x) is constant. Let jElff with flow Yt(x). If xOHx l then 'Y,(xo-,HYt(x l ) for all tER. Proof: Let a: [0, l]--+M be a continuous and piecewise C«J curve such that a(O) = xo, a(l) = xl and d/dfa(s)EX(a(s». Define a curve, f3:[O, l]~M by fJ(s) ==y,(a(s». Clearly fJ is continuous, piecewise C«J, and P(O)=r/(x~P(l)='Yt(xl).Moreover for cpE§
(tkp(P(s». ~ P(s»=
=
~fP(fJ(s»= ~ fP(Yt(O:(s»)
=(dqJ ( Yt(O:(S» ) =0. So djdsp(s)ex'(p(s».
aYt
Z/(W)==Wk
d
ax (0: (s». d<: o:(s»
u)
where 'lTk: Rm~Rk is the projection on the first k factors. Therefore z': V-+R k is a homeomorphism into and (V,z') a coordinate chart. In the coordinates (U,z) and (V,z') on M and M', the projection 'IT is just "'k and clearly a
0
Theorem 3.9: Suppose the dimension of d§ (x) is k, then SI is a regular equivalence relation and there exists a loca.lly weakly observable system ~' on the k-dimensional non-Hausdorff manifold M' = M / Sf which has the same input-output properties as I. More precisely if 'IT: M~M' is the canonical projection then (I,x~ and (I',fT(x~) realize the same input-output map for every XO E M. If ~ is (locally)(weakly) controllable then so is I'. If ~ satisfies the controllability rank condition then so does ~' and moreover M' is Hausdorff. Proof: An outline goes like this. From Coronary 3.7 we see that H equivalence implies SI equivalence. We first show that H is a regular equivalence relation and that we can define a system ~' on M' - M / H with the same input-output properties as ~. We then note that ~' satisfies the observability rank condition, hence I' is locally weakly observable from
submersion. M is covered by a countable number of charts (U,z) and so M' is covered by the corresponding (V,x'). Verifying that changes of coordinates on M' are C IX) reduces to checking that every rp':M'-+R is c» on (V,z'). But if fPe§ then on (U,z) d
coo.
so cp(Z)=CP(Zl'·· ,Zk). Clearly cp'(z')=cp(zi,··· ,Zk) is This shows H is a regular equivalence relation on M. Next we define ~' on M' locally. In the coordinate system (U,z) on M, the dynamics of ~ are given by
i=
;~ (x(z»j(x(z),u).
In particular for i= 1,.. ·· ,k
which it follows that H:= SI. Remark: By Corollary 3.7 xOHx· implies xOSlx1 which by definition implies xOlxl which by Lemma 3.2 implies
acp.
s, = a; (x(z»j(x(z).u)" Lf(·.I1)(fPj)(x(z».
cp(X~_cp(XI) for every q>E~.
Given any xOeM there exists tApt,··· ,drpkEdfJ which The right side is an element of § and hence pulls down to are linearly independent at xo. After reordering the x a functionf:(z',u) on V'. We define the dynamics of I' on coordinates we can suppose that d({JI'···' dCfJk' dxlc + t' · · · ,dx". are linearly independent at xo. Define ZI(X)='P;(x)-q>;(x~, i== 1,.·· .k, Zi(X)== x;- x?, i= k+
(V,z') by
I,.·· .m and
Clearly if z(t) is a curve in U generated by the dynamics of ~ under the control u(t) satisfying z(to.>==zo and if z'(t) is the curve in V generated by the dynamics of I' under the control U(/) satisfying z'(t~==w(zc; then for
i'=f'(z',u).
u== {x :lz;(x)1 < t:}.
If E is chosen sufficiently small, then (U,x) is a coordinate chart around xo. Claim If X 1,x 2 E H ( U ) then x 1Hx2 iff cp;(x 1) =
small It -
tOI
Z'(t) = w(z{t». From this it follows that if ~ is (locally) (weakly) controllable so is I'. Moreover each gi E § and hence pulls down to a funcon M' which we use to define the output of I' on tion (V,Z') as
451
g;
y =g'(z').
The outputs of ~ from
ZO under u(t) and I' from 1T(Z~ under u( t) are the same because g = g' o 'IT and so
g(z( t» == g' 0 w(z(t»
= g'(z'( t).
Notice M' is of dimension k.
On the other hand suppose cp(X~=rp(XI) for every cp E §. Using the controllability rank condition we choose 11'· .. ,1m E 'J which are linearly independent at xo. We also choose dep),· · . ,dCPk Ed § which are linearly independent at xO. Then the k x m matrix with i-jth element
We tum now to the relationship between ~ and § of ~
and 'fJ' and ~ of ~'. Note that I
j'(7T(X),U)=
~: (x)j(x,u).
It foUows from this and the definition of ~ and 'F' that exists an f E ~ such that
I' E (j' iff there
j'(7T(X»=
~: (x)j(x).
(fJ.' ··
In particular
.
c:f'(7T(X»=
a:
a
is of rank k at xo. Without Joss of generality we can assume the first k columns of this matrix are linearly independent at xo. The elements of this matrix are all in § so the first k columns are also linearly independent at Xl and therefore dcpJ' ... ,dcpk are linearly independent at Xl. As before we can construct a coordinate cube (Uo,z) around Xo using fIJI' . · . ,CPk'Xk + I'· .. , Xm • In a similar fashion using the same e and ,qJk but perhaps different Xk+I'··· 'Xm we construct a cub.e (U',z) around x'. Given 8 > 0 define m - k slices Sao and Sal in UO and U I by
(x)c:f(x).
sJ == {x E Vi: z;(x) =0,
and since 'IT is a submersion this shows that if }: satisfies the controllability rank condition so does }:'.
1,··· ,k and
i=k+I,··,m}.
Iz;(x)f<8,
Let Y$i(X) denote the flow of Ji(X) for i= 1,· .. .k and cl denote the cube of side 28 around 0 in R k • Define maps
Furthermore g'ow=g
Pj : C; x Si~M
and so Lfg;(x) =
j=
by
4' (g;)('lT(x»).
Q.
Pi
Repeated applications of this formula show that cp' E § '.iff there exists on q> E § such that
where x E
(sl' ... , sn' x) = Ysick
0
•••
I ,.of I
'V
(x)
si and Is;1 < 8 for j = 1,· .. ,k. Of course si is a
m - k dimensional cube so we can view these maps as
cp/( w(x») = q>(x).
z;
In particular the coordinate functions = cp; are in § so I' satisfies the observability rank condition and therefore is locally weakly observable. This also implies that H and Sf are the same relation. We have already seen that xOHx l implies xOSlx l • Suppose the converse does not hold; there exists xo,x 1 EM such that xOSlx· but xOHx l • Let a(s) be the arc of ~-indis tinguishable points joining XO,x 1, since ~ and ~' have the same input-output properties, '1T(a(s» is an arc of L'·indistinguishable points joining '1T(x~ and ?T(X 1). Moreover this arc is not constant because '1T(X~*'1T(XI). This contradicts the local weak observability of ~'. If I satisfies the controllability rank condition then M' can be seen to be Hausdorff. Given XO~Xl EM first suppose there exists a q>E§ such that qJ(X~*
pJ: C8m~M
I
VO= {x E M :lqJ(x) - cp(xO)1 < Iq;(x I) - q;(xO)I/2}
where C;' is an m-cube of side (Sl~· · . ,sm)E C;' then
2~.
More precisely if
pi (s),' . · ,sm) = f3} (s},· · . ,Sk'X) where xESj with z;(x)=O for i=l,··,k and Zj(x)=s; for i==k+ I,··· .m, For sufficiently small 8 these maps are diffeomorphisms onto open neighborhoods Wi of xi contained in UJ. In fact (Wj,SI'· · . ,sm) are coordinate charts at x/, If 8 is chosen sufficiently small then we can assume that for every xE Wi and for every (Sl'·· ,sn)ECl we have
Now suppose there exists pOE Wo, r' E Wi such that pOHp I. In particular suppose pO = Ys~ 0 • • • 0 y},( q~ where qO E $8°. Let .
U 1== { x E M :I«p(x) - q;(xl)1 > l«p(x I)-
ql = 'V~ 1 s.
0
•••
a
'V~ I
(pI).,
5",.
then by assumption ql E Vi and by Corollary 3.8, qOHql.
Since fP is constant on H equivalence classes
7T- 1(7T(Ui »)= u' and so '1T(U~ and '1T(V I ) are disjoint open neighborhoods of '1T(x~ and w(x').
Because qOES80, xOHqo so XOHql. This implies that epi(X~. =cp;(ql) for i= 1,··· .k and so by assumption (J);(ql)= cp;(x l ) for j= J,. .. .k, By our previous claim, q 1Hx 1 and so xOHxl •
452
We have just shown that if '1T(W~n'1T(WI)*0 then '1T(X~=='1T(XI). It follows that if 11'(X~*'1T(Xl) then these are disjoint open neighborhoods of 1T(X~ and etx'), D The following simple example shows that if ~ fails to satisfy the controllability rank condition then M' need not be Hausdorff. Example 3.10: Let M=R 2\ {( X1,O): X1
x=O
y=x 1•
The dimension of d~ (x) is one for all x so we can fonn the quotient ,M' =: M / Sf. However this is a non-Hausdorff manifold because '77'(0, 1) and '1T(0, -1) are distinct points without a pair of disjoint neighborhoods. Notice that factoring M by the closure of the relation SI is not a way around this problem for the closure of Sf is not a regular relation. Theorem 3.5a and b and Theorem 3.9 exhibit three situations where realizations can be made observable in some sense. If I is analytic then I' is locally observable, if ~ is symmetric then I' is observable and if dg (x) of ~ is of constant dimension then }:' satisfies the observability rank condition and hence is locally weakly observable. The converse of this last remark is "almost" true. Theorem 3.11: If I is locally weakly' observable then the observability rank condition is satisfied generically. Proof: For any system the observability rank condition is satisfied an open subset of M, possibly empty. Suppose there exists an open subset U of M where the dimension of d§ (x) < m. Without loss of generality we can assume dimd~(x)=k<m for xEU. Choose xOEU, an open set V such that xO E V ~ U and a COO function cp: M-+R such that q>(x) = 1 for all x E V, cp(x)=#=O for all xE U and cp(x)=O for all xEM\U. Consider thesystem I' defined by
on
x=cp(x}J(x,u) y=g(x)
in a similar fashion to the proof of that theorem to show the dimension of d§ (x) is constant. Let x I = Ys (xO:> where y denotes the flow of f E§'. The adjoint of (a/ax)'Ys(x~ carries one forms at Xl to one forms at XO according to the rule
W(XI)rH,'(X I )
a 'Ys(XO), ax
for wE~*(M). This map also has a Campbell-Baker-Hausdorff expansion
a
00
k
x
k=O
.
w(x1)-a 'Ys(XO)= ~ {Lf)*w{xO)sk'· In particular, if wEH then the adjoint of (d/dX)'Y.r(Xo-, carries dg(x 1) into d§(x~ so the dimension of dg(x l ) is less than or equal to the dimension d~ (x~. A similar argument using the adjoint of (a/ax»)' -sex)~ shows the reverse inequality. Therefore, if xOWAx1 then the dimensions of dg (x~ and d § (x 1) are the same. Since I is weakly controllable, this implies that the dimension of d§ (x) is constant. Suppose dimd~(x)=k then we apply Theorem 3.9 and identify strongly indistinguishable points. But ~ weakly observable implies there are strongly indistinguishable 0 points so k must equal m. To relate the results of this section to the well-known linear theory, we consider the following. Example 3.11: Consider the linear system
no
x=Ax+Bu y=Cx. In Example 2.5 we saw that §" was spanned by { Ax , A'B*·jr=O ' . , · .. , m -1 , J.= 1" · .. I} where B •.'} denotes thejth column of B. Let C;* denote the ith row of C then for r ;> 0, p:> 0
LAxC;*yA rx = C;.A
x(O) = xO.
LA'B
The state space of ~' is U and ~' is complete. It is easy to see ~/(X) =~(x) for all x E U. From this it follows that dl3'(x)=dl3(x) for all sE U~ and so dimd§(x)==k for all xE U. We can apply Theorem 3.9 to ~' on U so ~' is not locally weakly observable. Since ~ and ~' agree on V, ~ is not locally weakly observable either. 0 Recall that for analytic systems, weak controllability, local weak controllability, and the controllability rank condition are equivalent. With regard to observability an analogous result holds for analytic systems which are weakly controllable. Theorem 3.12: If ~ is a weakly controllable analytic system then ~ is weakly observable iff it is locally weakly observable iff the observability rank condition is satisfied. Proof: It suffices to show that weak observability implies the observability rank condition. By Theorem 2.6 ~ satisfies the controllability rank condition, we proceed 453
.j
r+l x
Ci.APx = Ci.A r+PB.j
LAxC;.APB.j = LA'B.,C;.APB.j =0. )
Therefore by the Cayley-Hamilton Theorem § is spanned by
r=O,.·· ,m-l} and d § (x) is spanned by
{Ci.A r : i= 1,··' .m, r=O,'" ,m-l}. Clearly d~ (x) is of constant dimension. The observability rank condition reduces to the well-known linear observability condition that
IV.
C CA
=m.
rank
If the observability rank condition fails then we define X(x) to be the .constant linear space of column vectors orthogonal to d§ (x), Two points x~ and Xl are H equivalent (or Sf equivalent) if x l - xOe X(x) because
a(s)= XO+s(xl - Xo) is a curve tangent to X(x) from x O to x '. Factoring M by X(x) results in a locally weakly observable system which because of linearity is also locally observable (Sl =1 for
linear systems). Example 3.12: Suppose the linear system is nonautonomous as in Example 2.6. As in formula 2.1 we add time as a state variable which we can observe directly. Then direct calculation shows that the observability rank condition is equivalent to
MINIMALITY
A linear system is said to be minimal if it is controllable and observable. As is well known, two minimal linear systems initialized at 0 which realize the same input-output map differ only by a linear diffeomorphism of the state spaces, [I], [4]. A nonlinear system which is observable, weakly controllable and either analytic or symmetric is called minimal by Sussmann [29]. He has shown that two minimal nonlinear systems which realize the same input-output map from their respective initial states differ only by a diffeomorphism of the state spaces. A nonlinear system ~ is locally weakly minimal if it is locally weakly controllable and locally weakly observable. Two locally weakly minimal realizations of a given input-output map need not be diffeomorphic as is seen by Example 3.3, but the following theorem shows they must be of the same state dimension which is minimal over all possible realizations. Let L, ~' be two nonlinear systems with control set O=O'~RI states spaces M and M' of dimension m and m' and output with values in R" given by ~:i=f(x,u)
C(t)
y=g(x) '2:' : i
=J'(z, u)
y=g'(z).
=m
rank
for every t E R where
AoC(t)==
~ C(t)+ C(t)A (r),
A~ C (t) == ~ A~-IC (1)+ A~-IC (t)A (t). See [3] or [5]. Example 3. J3: Consider the bilinear system of Exam..
Theorem 4.1: Suppose (~,x~ and (L',Zo.> realize the same input-output map. If ~' is locally weakly minimal then m » m', Proof: Without loss of generality we can assume that ~' satisfies the controllability and observability rank condition at zoo (If not, by Theorems 2.5 and 3.11 we can find a control (u(t),[tO,t l]) and a corresponding L' trajectory (z(t),[tO,t l]) with z(t~=zo, z(tl)=ZI such that these conditions are satisfied at z '. Moreover if .\"1 is the endpoint of the corresponding ~ trajectory then (L, x I) and (~', z I) realize the same input-output map.) Since ~' satisfies the observability rank condition at ZO we can find a neighborhood V of ZO and functions qJ~,. • • ,cp;", E §, such that the map
pie 2.8
ep'=col(fPi,.·· ,cp;"): V~Rm' I
x=Ax+ ~ u;B'x. i=J m
If M=R (or a subgroup of G/(h,R», let C be a nXm (n x h) matrix and
is a diffeomorphism into. Because I' satisfies the controllability rank condition, as in Theorems 2.1 and 2.2 we can find controls u 1, • • • ,U m and corresponding vector fieldsf'i(z)=f'(z,u i ) with flows y'~(z) such that the map
y=Cx,
,r,'(' T S · · · S ) =v rm I'
then
g = {C;xDx: i= 1,··· .n; DxE'J} d§ ={ CixD:i= 1,··· ,n; DXE'F}. See [3] for further details.
rrm
I S'"
0
I ( 0)
I '." I 5, ,-
• • • "1
,
is a diffeomorphism from some open subset S of the positive orthant of Rm' into M'. Let q>j, 4>, I'. v'. 'I' be the corresponding objects of L. Since (}:,x~ and (~', z~ realize the same input-output map <J> 0 'l' =4>' 0 '1" and so 454
'I'
(20) - " A generalization of Chow's theorem and the bang-bana theorem to nonlinear control problems," SIAM J. Contr., vol.
S~Rm'~M~Rm'
12, pp. 43-52, 1974.
is a diffeomorphism. But clearly this is only possible if [21] C. Lobry, "ControUabilite des systemes non Iineaires," SIAM J. COftt~,vo1.8,pp. S7~S, 1970. m)m'. [22J - ,~lques aspects Qualitatifs de la thme de la commaacle," These de doctorat d'etat, Universite de Grenoble, Grenoble, Corollary 4.2: If (~,x~ and (I',z~ are locally weakly France, 1972. , minimal realizations of the same input-output map then (23) T. Nagano, "Linear differential systems with singularities aDd an application to transitive Lie algebras," J. Math. Soc. Japan, vol. 18, their state spaces are of the same dimension. pp.398-404, 1966.
REFERENCES [1] B. D. O. Anderson and 1. B. Moore, Linear Optimal Control. EDalewood Oiffs, NJ: Prentiu-Hall, 1971. (2] W. M. Boothby, An IntrotktiOll to Differential Manifolds and Rientl1l1lian GetJnwtty. New York: Academic, 1975. (3] R. W. Brockett, "System theory on group manifolds and coset spaces," SIAM J. Cont,., vol. 10, pp. 265-284, 1972. [4 -Flllite Dimensional LiMQT Systems. New York: Wiley, 1970. [s W. L Chow, "Uber Systeme von Linearen Partiellen 00ferentialgl.eichungen erster Ordnung," Math. AM., vol. 117, pp.
l
98-105, 1939.
(6] H. D'Angelo, Linear Time-Vary;ng Systems.
[7] (8) [9]
(10]
[IIJ
[12] [13] (14] [IS]
[16] (17]
(18] [19]
Boston, MA: Allyn and Bacon, 1970. E. W. Griffith and K. S. P. Kumar, "On the observability of nonlinear systems, I," J. Math. Anal. Appl., vol. 35, pp. 135-147, 1971. G. W. Haynes and H. Hermes, "Non-linear controllability via Lie theory," SIAM J. Contr., vol. 8, J?p. 450-460, 1970. R. Hermann, "On the accessibility problem in control theory," in l"t. Sy"". 011 Nonlinear DifjeremioJ Equations and Nonlinear M«lIimics. New York: Academic Press, 1963, pp. 325-332. -D~rent;aI Geometry and the Calculus of Variations. New York: Academic, 1968. -CtThe differential geometry of foliations, II," J. oj Math. and Meek, vol. II, pp. 303-316, 1962. -"Some differential geometric aspects of the Lagrange variational.problem," J. Math., vol. Ill, no. 6, pp.634-673, 1962. -"EJDsteDce in the large of parallelism homomorphisms," TraM. Amer. Math. Soc., vol. lOB, pp. 170-183, 1963. -.-"Cartan connections and the equivalence problemfor geometnc structures," Contribution to Differential Equations, vol. 3, 1964, pp. 199-248. H. Hermes and J. P. Lasalle, Functional Analysis and Time Optimal Control New York: Academic, 1969. Y. M.-L. Kostyukovskii, "Observability of nonlinear contIoUed systems," Alllomal. Remole Comr., vol. 9, pp. 1384-1396, 1968. -~imple conditions of observability of nonlinear controlled g'Stems," AlllomtJI. Remote Contr., vol. 10, pp. IS75-1S84, 1968. S. R. Kou, D. L FJliot, and T. J. Tarn, "Observability of nonlinear system.s," Inform. COIIIr., vol. 22, P. 89-99, 1973. A J. Krener, "A BeneralizatioD 0 . the PontJyagin maximal princi-
f
ple and the bang-bang principle," Ph.D. dissertation, University of California, Berkeley, CA, 1971.
455
(24] J. P. Serre, Lie Groups and lie Algebras. New York: BenjamiD, 1965. [25] H. 1. S~ andY. J. Jur~jevic, "ControUability of nonlinear systems, J. Differentitll EquDtioru., vol. 12, pp. 95-116, 1972. [26] H. J. SUSSIDIJm, "Minimal realizatioDl of Donlinear systems," in Geometric M«hodr ill Systemr TIwry, MaYDe and BrOckett, Ed., Dordreeht, The Netherlands: D. Riedel, 1973. ,(27] - " A generalization of the closed subp-oup theorem to quotients of arbitrary manifolds," J. DiffemuiiIJ Chonwtry, vol. 10, pp.
ISI-I66, 1975. (28] -"Observable realizations of finite dimensional nonlinear autonomous systems," Math. Systems '1'It«Ny, to be published. . [29J -"Existence and uniqueness of miDimaI realizl.tioDS of nonlinear systems I: Initialized systems," Math. Systems T'heory. .
Analysis of Recursive Stochastic Algorithms LENNART LJUNG
INTEREST in identification algorithmsand their use in adaptive control commenced very early, the former having its roots in the literatureon time series analysisand the latter in dynamic programming which was immediately perceived, on its introduction in the 1950s,to be the natural tool for decision making under uncertainty. Adaptive control required recursive identification,stimulatinga substantialactivityin this topic. An obvious starting point for analyzing these recursive algorithms was the olderliteratureon stochasticapproximation [5], [8].However, in recursiveidentification, unlike the models consideredin the traditional stochastic approximation literature, the regression (observation) vector 4J(t) is allowedto dependon all previousvalues of theparameterestimate,preventing use of the oldertechniques. It was obviously important to answer questions about the convergence of such identification algorithms: for example,
ingon the currentparameterestimateanda so-calledobservation vector, 4J(t) which is drawn from the "true" system,but may be noisy;entriesof 4J(.) may be noise-contaminated inputsand outputs of the true system so that it is not necessarily a sequence of independent, identically distributed random variables commonly demanded by stochasticapproximation. Lastly, yeo) is a sequenceof positivescalars.Since there is interest in parameter convergence, and since in general Q varies randomly, it is clear that parameter convergence can only occur if yet) --+ 0
as
t --+
00
(2)
Thisis the typeof assumptionthatbecameveryfamiliarin theliterature on stochastic approximation. It carries with it the great advantage that analysis of a fundamentally stochastic system can be replaced by analysisof a deterministic system,and there is some degree of comfort in a theorem which says that parameters converge to a deterministic limit. On the other hand, • under what circumstances does the algorithm converge? there is also a great disadvantage. With y going to zero, less and • what is its rate of convergence? • what meaning can be attached to the point to which conver- less attention is paid to the "noisy" observations. This means that shouldthe plant whoseparametersare being estimated,and gence occurred? from whom observations are derived, not have constant paramThese questionshad startedto present themselvesin analyses eter values, but have parametervalues that change in time (perconducted several years previously, such as references 1-5 of haps slowly) the ability to track those varying parameters vanthis paper by Ljung. Among other things, these references had ishes as y goes to zero. In effect, only strictly time-invariant shownthatconvergence neednot necessarilyoccur,and alsothat plants can be identified. the convergence property could be associated with the stability It is important to understand at a broad level the basis of the properties of a deterministic ordinary differential equation de- argumentwhich allowsthe analysisto be done by consideration spite the fact that discrete-time identification algorithms were of an ordinarydifferential equation.There is a general appeal to being considered. averaging theory ideas. More precisely, suppose that in EquaThepaperby Ljungstartswitha verygeneralformof recursive tion (1) yet) is very small over a long interval. Then over this algorithm: interval, x(·) does not change much, and presumably 4J takes a large number of values with some statistical regularity properx(t) = x(t - 1) + y(t)Q(t; x(t - 1), 4J(t» (1) ties. These properties may well be x dependent, especially if a In this equation, x(·) is a sequence of vectors, of dimension n, controlleris being implementedto generateplant inputs,and the constituting parameter estimates. It is hypothesised that in the parametersin that controllerare computedusing the estimate x. background there rests a system described by a "true" parame- In this circumstance, the sequences 4J(t) and x(t) will certainly ter, and one would hope that the parameter estimates converge not be independent. to the true parameter. In Equation (1), Q is another n-vector, Considering x(t) as approximately a constant over a long inmost commonly not dependent explicitly on time, and depend- terval, and assuming that f/J(t) takes a range of values during 457
this interval, one agrees to replace say, s, successive updates of Equation (1) by an update involving a single gain t+s
Ly(k)
(3)
t+l
and to replace Q by its average value conditioned on x(t - 1). The result is still then a difference equation. However, one can regard it as an Euler approximation of an ordinary differential equation, and the ordinary differential equation comes out to be something like Equation (4) below, where f(x) is the expected value of Q conditioned on x.
(4) The above paragraph sets out a heuristic justification for the assertion that the behavior of the recursive algorithm can be linked to the behavior of the deterministic ordinary differential equation. Of course, the paper sets out a series of assumptions, under which formal theorems can be stated. Broadly speaking, these theorems are of the type: if the differential equation has an invariant set with an associated domain of attraction, and if the recursive equation gives a parameter estimate which from time to time lies in the domain of attraction, then the recursive equation will have a solution converging to the invariant set. If the invariant set is a point, then the unknown parameter vector can be considered as being learnt (identified). However, if it is not a point, the unknown parameter cannot be identified, although some of its properties can be learnt. The second group of results focuses on the question of learning the whole parameter vector. A third important result of the paper relates the trajectories of the deterministic ordinary differential equation and the stochastic recursive equation, putting a bound on the probability that these can never be far apart. Thus the trajectories of the differential equation illustrate how the parameter values will change on a transient basis-at least for a typical run. The value of the paper for applications is that a whole collection of important situations are captured by the general theorems of the paper. Important applications include equation error identification and self-tuning regulators. Stochastic approximation turns out to be a special case of the theory. Readers familiar with equation identification will probably be aware that in the presence of noise, bias in the parameter estimates is to be expected. It is a measure of the power of the theory of the paper that in this case, where parameters may
converge but not to the correct values, one nevertheless has a hard result. Where did the paper lead to? As might be expected, a series of incremental advances and a widening of the example classes were able to be found. Secondly, the paper flagged the great importance of averaging ideas for recursive identification and for adaptive control. The averaging derives from the fact that the time scale associated with the time-constants of all but the identification process, and the time constant of the identification process itself, are very different. This separation of time scales found application in more than one book dealing with adaptive control; for example, [1] and [6]. Systematic use was made of the the techniques of the paper to analyze and classify recursive identification algorithms (including the extended Kalman filter) in [4]. A simpler approach to averaging analysis that preserves its full power is presented in [7]. The same family of problems, using different machinery, is addressed in [3] and [2]. The paper also gave results underpinning later work addressing the very difficult issue of the behavior of algorithms in which the adaptive gain sequence does not go to zero. There had been earlier analysis of such algorithms, essentially for FIR systems (this was particularly driven by the school of Widrow, and many results can be found in [9]), but these earlier results fell completely short of addressing situations where there was closedloop control and the systems were not finite impulse response. Such situations are being addressed in isolated papers, but there seems as yet no textbook distillation embodying the key ideas. REFERENCES [1] B.D.D. ANDERSON et al., Stabilityof AdaptiveSystems, MIT Press (Cambridge, MA), 1986.
[2] A. BENVENISTE, M. METNIER, AND P.PRIOURET, AdaptiveAlgorithmsand Stochastic Approximation, Springer-Verlag (New York), 1990. [3] H.J. KUSHNER AND D.S. CLARK, StochasticApproximation Methodsfor Contrained and Unconstrained Systems, Springer-Verlag (New York), 1978. [4] L. LYUNG AND T. SODERSTROM, Theory and Practice ofRecursive Identifiction, MIT Press (Cambridge, MA), 1983. [5] H. ROBBINS AND S. MONRO, "A stochastic approximation method," Ann Math. Stat., 22:400'407,1951. [6] S. SASTRY AND M. BODsoN,Adaptive Control, Stability, Convergence, and Robustness, Prentice Hall (EnglewoodCliffs, NJ), 1989. [7] V. SOLO AND X. KONG, AdaptiveSignal Processing Algorithms:Stability and Performance, Prentice Hall (Englewood Cliffs, NJ), Information and System Science Series, 1995. [8] M.T.W. WASAN, Stochastic Approximation, Cambridge University Press (New York), 1969. [9] B. WIDROW AND S. STEARNS, Adaptive Signal Processing, Prentice Hall (EnglewoodCliffs, NJ), 1985.
B.D.O.A.
458
Analysis of Recursive Stochastic Algorithms LENNART UUNG, MEMBER, IEEE
A1JIIIw:f-
raa.a-
ft
......... '
rn.ew.n.
. . are rllMe Is tIIat tile
~
All IDaportIIIt
"......" or the· ......., e.a., Itodtastlc
I. INTRODUCTION
ECURSIVE algorithms, where stochastic R tions enterare common in fields. In the control ........................................, ......... and estimation literature such algorithms are widely dis. . . . . . .., depead oa
C8MIdnII
e.traI It II
equtIoa
CIa.~
.,
bIe~
_'..IIIc
oIl11e
.................................. nearrelltia ' \
observa-
many
.,.....
willi
• well daD 1M
The ...
tile .". CGIdnIL
"die
cussed, e.g., in connection with adaptive control, (adaptive) filtering and on-line identification. The convergence analysis of the algorithms is not seldom difficulL As a ...... ..... be
tbe ,• to JII'ObI-Ia 1tIIII1IIca.
rule, special techniques for analysis are used for each type of applicati9n and often the convergence properties have to be studied only by simulation.
In this paper a general approach to the analysis of the asymptotic behavior of recursive algorithms is described. In effect, the convergence analysis is reduced to stability analysis of a deterministic, ordinary differential equation. This technique is believed to be a fairly general tool and
to have a wide applicability. Applications to various probCopyrlaht CO J9T1 byThe Institute of Electrical and Electronics Engineers. Inc. Printed in U.S.A. Annals No. 708ACOOS Reprinted from IEEE Transactions on Automatic Control, Vol.AC-22,August 1977, pp. 551-575. 459
lems have been published in [lH4] and some theory was presented in (5]. The objective of the present paper is to give a comprehensive presentation of formal results and useful techniques for the convergence analysis, as well as to illustrate with several examples how the techniques can be applied. In Section II a general recursive algorithm is described and discussed. A heuristic treatment of the convergence problem is given in Section III and this leads to the basic ideas of the present approach. Section IV contains a discussion of the conditions which are imposed on the algorithm in order to prove the formal results. These theorems are given in Sections V and VI. The theorems suggest certain techniques for the convergence analysis, and these aspects are treated in Section VII. Several examples of how the theorems may be used, some of them reviewing previous applications are given in Section VIII.
observation q>(t) as the (lower dimensional) output of a dynamical system like (2), ep(t)= C(x(t -l»cp(t). However, this case is naturally subsumed in the present one, since
and the proofs for this case are given in [7]. Throughout this paper it is assumed that the estimates are desired to converge to some "true" or "optimal" value(s). Since Q(/,X(I - l),cp(/» is a random variable, with, in general, nonzero variance, convergence can take place only if the noise is rejected by paying less and less attention to the noisy observations, i.e., by letting
y( t)--+O II.
THE ALGORlmM
A general recursive algorithm can be written
x(t)=x(t-I)+ y(t)Q(t;x(t-I),fP(t»),
(I)
(3)
as
1--+00.
(4)
In tracking problems, when a time-varying parameter is to be tracked using algorithm (I), this condition is not feasible.. Then y(/) usually tends to some small, nonzero value, the size of which depends on what is known about the variability of the tracked parameters and about the noise characteristics. This case is not treated here, but it is reasonable to assume that analysis under the condition (4) also will have some relevance for the case of tracking slowly varying parameters. Suppose we have a linear, stochastic, discrete-time system, governed by a linear output feedback law, which at time t is determined by x(t - 1). Then the behavior of this overall system can be described by (2), with ,,(t) consisting of lagged inputs and outputs. Therefore the algorithms (I) and (2) can be understood as archetypical for adaptive control of a linear system. Indeed this setup is useful for analysis of certain adaptive controllers, as will be exemplified below, but the basic algorithm (I), (2) also covers several other cases of interest.
where x(·) is a sequence of n-dimensional column vectors, which are the objects of our interest. We shall refer to x(·) as "the estimates," and they could, e.g., be the current estimates of some unknown parameter vector. They could, however, also be parameters that determine a feedback law of an adaptive controller, etc., and we shall be precise about the character of x(·) only in the examples below. The sequence y(.) is assumed throughout the paper to be a sequence of positive scalars. The m-dimensional vector cp(t) is an observation obtained at time I, and these are the objects that cause x(t - 1) to be updated to take new information into account. (The notion "observation" does not have to be taken literally. The variable q> may very well be the result of certain treatment of actual measurements.) The observations are in" general functions of the previous estimates x(·) and of a sequence of random vectors e( .). This means that the observation is a random variable, which may be affected by previous estimates. This is the case, e.g., for adaptive systems, when the input signal is determined on the basis of previous estimates. If the experiment designer has some test signal at his disposal, this may be included in e(· ). The function Q(.; ., .) from R x R" x R m into R" is a deterministic function with some regularity conditions to be specified below. This function, together with the choice of the scalar "gain" sequence y(.) determine entirely the algorithm. We shall not work with completely general dependence of ,(I) on x(·), but the following structure for the generation of cp(.) will be used:
The algorithm (1), (2) is fairly complex to analyze, being a time-variant, stochastic, nonlinear difference equation. Notice also that the correction x(t)- x(t-l) depends in general via fP(/) implicitly on all old x(s). Therefore, while (I) certainly is recursive from the user's point of view, it is not so for analysis purposes. In this section we shall illustrate heuristically how a differential equation can be associated with (1), and bow it seems reasonable that asymptotic properties of (1) may be studied in terms of this differential equation. The formal analysis and results follow in the next two sections. Consider
fP(t)==A(x(t-l»q>(t-l)+B(x(t-l»e(t).
X(/)=x(t-I)+y(t)Q(x(t-I),.,,(t»),
III.
(2)
HEURISTIC ANALYSIS
(5)
where for simplicity we let Q be time independent.. As Here A (. ) and B (.) are mlm and mlr matrix functions. Remark: It is perhaps more natural to think of an remarked before, cp(l) depends on all previous estimates: 460
I
,(/)- I
j-I
(
) IIt A(x(i-I» B(x(j-I»)e(j).
(6)
(5) by I
i-j+}
T,==
I
y(k).
(12)
Ie-I
Now, if (2) is exponentially stable, then the first terms in We therefore have some reason to believe that the (6) will be very small and, for some M, sequence of estimates x(·) asymptotically should follow the trajectories x D ( •) of (11). I A(x(i-l»)B(x(j-l»e (We could also have related (9) to the difference equation Moreover, it follows from (5) and (4) that the difference (13) X(/)- x(l-l) becomes smaller as I increases. Therefore, fqr sufficiently large I, we have x(k)~x(/); I ~ k ~ 1- but the differential equation is easier to handle since it is time-invariant.) , 2M. Hence, It now seems reasonable that asymptotic properties of Ie k-«! algorithm (I), (2) may be studied in terms of the differencp(k)~ I [A(x(t»] JB(x(t»e(j) tial equation (11). This heuristic discussion is perhaps not j-le-M very convincing, but along a similar path, though with far Ie • (7) more technical labor, formal results to this effect can be JB (x(t»e(j) £ ~(k;x(t» ~ I [A(x(t» proven. These are given below. j-I
4p(t)~j.~M C.¥.
u).
t-
for I ;> k ;> t - M. Furthermore,
IV.
AssUMPTIONS ON THE
ALoolllTHM
Q (x( k ~ 1),cp(k»~Q (x( t),q;(k; x( I») == /(x(t» + w(k)
In order to prove the formal results, certain regularity (8) conditions on the functions Q, A, and B and on the driving "noise" term e, have to be introduced. Some of where these are fairly technical, but it is believed that none is very restrictive. Several sets of assumptions are possible, f(x)- EQ(x,q;(k;x») and we shall give a few. In particular, there is a possibility to treat the sequence e(·) either in a stochastic or in a and hence w(k) is a random variable with zero mean. deterministic framework. Using (8), we can approximately evaluate . We shall start by giving a formal definition of ~ used in the previous section. Let 1+$ x(t+s)-x(t)+ I y(k)Q(x(k-l),cp(k» Ds == {xlA (x) has all eigenvalues strictly i-,+1 1+3
I
~x(t)+f(x(t»)
I+.s
y(k)+
Ic-I+ 1
~x(t) + f(x(t»)
I
inside the unit circle}.
y(k)w(k)
k-t+ I
Then for each xEDs' there exists a A=A(x), such that
t+.r
I
A(X)< I.
(9)
y(k),
(14)
t+ I
Take xeDs and define the random variables ~(t,i) and where the last step should follow since the last term is a v(t,A,c), A< 1, by zero mean random variable which is dominated by the q;(t,i):aA(i)~(t-l,i)+ B(i)e(t); ~(O,i)==O second term. Expression (9) suggests that the sequence of estimates more or less follows the difference equation (15) x D ( T+~T)= x D ( T) +~"f( x D
(,,»
(10)
V(t,A,C)=-AV(t-I,A,c)+cle(t)l;
V(O,A, c) ==0.
(16)
Let DR be an open, connected subset of Ds . The regularity conditions will be assumed to be valid in DR. Now, the 1+$ first set of assumptions is the following. I y(k). A.I: e(·) is a sequence of independent random vari1+1 ables (not necessarily stationary or with zero means). It is useful to interpret (10) as a way of solving the A.2: )e(t)1 < C with probability one (w.p.l) all t. differential equation (4'T small), A.3: The function Q(t,x,cp) is continuously differentia.. hie w.r.t x and f for x E DR. The derivatives are, for fixed (II) x and cp, bounded in t. A.4: The matrix functions A ( .) and B ( .) are Lipschitz where the (fictitious) time" relates to the original time t in continuous in DR. where 4,. corresponds to
461
A.5: lim,.... ooEQ(/,X,q>(I,X» exists for xE DR and is denoted by f(X). The expectation is over e( · ). A.6: Ify(/)= 00 A.7: Iiy(t)P < 00 for some p. A.B: y(.) is a decreasing sequence.
x and
·(I + v( I,A,e») - k o { 1- 1,X,A,C) ]; (17)
kC)(O,x,A, c) ==0. (ISb)
Notice that for the common choice y(/)= I/t, (18a) imConditions A.S and A.. 9 are motivated by technical arguplies that ments in the proofs, but they have so far not appeared to be restrictive.. For example, it is easy to see that the z(t,x)-! Q(k;x,qi(k,x») sequence y(t)= Ct:" satisfies A.6-A.9 for 0< a c; 1. t k-l If we would like to alleviate A.2, further regularity conditions on Q are required. This gives us our second set and analogously for k,,(t,x,A,C). The assumptions then are of assumptions. (~(i,p) denotes a p-neighborhood of i, as follows. i.e., ~ (x,p) {xfli- x]
±
=
462
remaining in DR and boundedness of D can be avoided. Corollary J: Consider the algorithm (I), (2) subject to assumptions A, B or C. Let i5 be a closed subset of DR (possibly is= DR = R") such that (20) holds. Assume that there exists a twice differentiable function ~i (x) in DR" such that
convergence of (18) are given. In fact, conditions B imply that C.3 and C.4 hold w.p.l. It may in this context be remarked that there is actually a tradeoff between condi.. tion A.7 == B.9 and conditions B.2 and B.7. The largest p for which B.2 and B.7 need to hold is twice the p for which A.7 holds. Therefore, if y ( · ) is subject to (17), then only the fourth moments of e, Q, and ~i have to be bounded. This is discussed further in [5] and [8], and we shall not pursue it here. V. MAIN
V'(x).f(x) ~ 0
(22)
Then either x(l)~Dc={xlxEDR and
THEOREMS
V'(x)!(x)=O} \V .
The functionJ(·) defined in A.S, B.6, or C.3 is the basic object of interest. As the heuristic discussion in Section III or indicated the differential" equation
{x(t)} ( 19)
p ( x ( t) --+ ~'B (x *, p)) > 0
inf Ix( t) - xl-+O.
The phrase "w.p.l" naturally does not apply in the case with assumptions C. In order to verify the stability condition (21 ) analytically, usually the Lyapunov theory has to be applied. Theorem 1 can be given a formulation, which does not refer to any differential equation, but directly relates to a Lyapunov function associated with f(x). In that way also the assumption about the trajectories of (19)
0
for all p > o.
(23)1
Furthermore, suppose that Q(I,X·,~(t,x*» has a covariance matrix bounded from below by a strictly positive definite matrix,
x(t} ED and \cp(t)1 < C infinitely often (i.o.jw.p.l (20)
xEDr
has a cluster point on the boundary of DR'
Our second theorem concerns the set of possible convergence points. It can be used to prove failure of conver.. gence by showing that the "desired" or "true" parameter value does not belong to this set. Theorem 2.' Consider algorithm (1), (2) subject to assumptions A or B. Suppose that x* E DR has the property that
will be relevant for the asymptotic behavior of the algorithm (1), (2). The exact relationships between (19) and ( I), (2) are given in three theorems. The first one concerns convergence of (I). Theorem J: Consider the ~gorithm (I), (2) subject to assumptions A, B or C. Let D be a compact s.!!.bset of DR such that the trajectories of (19) that start in D remain in a closed subset DR of DR for T > O. Assume that 1) there is a random variable C such that
2) the differential equation (1.2) has an invariant set D, with domain of attraction D A :> D. (21) Then x(t)~Dc with probability one as 1--+00. 0 Remarks.' By (20) is meant that there exists with probability one a subsequence tk , p..2ssibly depending on the realization w, such that x(tk ) E D and Icp(lk)1 < C (w), k = I, 2" · · . This condition, which we may call the "bounded.. ness condition," is further discussed in Section VI. An invariant set of a differential equation is a set such that the trajectories remain in there for - 00 < T < 00. The domain of attraction of an invariant set D, consists ofthose points from which the trajectories converge into D; as T tends to infinity. It is obviously an open set. See, e.g., [9]. An interesting special case is when the invariant set D, is just a stationary point of (19) say x", with f(x*) = O. Then the theorem proves convergence of x(t) to x*. By x(t)~Dc is meant that
p. I as l-;~ "X;
(24)
and that EQ (I,X, <j)(t, x) is continuously differentiable with respect to x in a neighborhood of x* and the derivatives converge uniformly in this neighborhood as I tends to infinity.
Then (25a)
f(x*)=O
and H (x*) =
fx f(X)lx~x. has all eigenvalues in the LHP (Rez
< 0). 0
(25b)
The matrix H (x*) defines, of course, the linear differential equation obtained from (19) by linearization around x*. Therefore this theorem essentially states that the algorithm can converge only to stable stationary points of the differential equation (19). If j(x)= -(d/dx)V(x), which might be the case if the algorithm is based on, criterion-minimization, then J/ (x) can be chosen as a Lyapunov function for the differential equation (19). Since id] dT)V(X(T»= -If(x(T))1 2, \ve see
463
'P(A)-The probability of the event A.
x
that the stationary points of (19), together with the point {co}, form an invariant set with global domain of attraction. Moreo.ver, if the stationary points are isolated, it follows from Theorem 2 that .only stable ones, i.e., local minima, are possible .convergence points. It also follows from Theorem I that the estimates cannot oscillate between different minima. Collecting all this we obtain a corollary to Theorems 1 and 2. Coro/lary 2: Suppose that DR:: R", that f(x) = (dj dx) V(x) and that V(x) has isolated stationary points. Assume that 'cp(t)1 < C La. w.p.l. Then, w.p.I, X(/) tends either to a local minimum of Vex) (i.e., V"(x) positive semidefinite) or to infinity as t tends to infinity. Finally, our third theorem relates the trajectories of the differential equation (19) to the paths of the algorithm (I), (2). The result is formulated as follows. Let X(/), 1= to,' · ., be generated by (I), (2). The values can be plotted with the sample numbers I as the abscissa. It is also possible to introduce as before a fictitious time T by I
T,=
~ 1'(k).
(26)
k-l
t'
Suppose that the estimates X(/) are plotted against this time T. See Fig. l(a). Let XD(T,Tto,X(/ o» be the solution of (19) with initial value X(/o) at time Tto. Also plot this solution in the same diagram, as in Fig. l(b). Let I be a set of integers. The probability that all points X(/), I E I, simultaneously are within a certain distance E from the trajectory is estimated in the following theorem. Theorem 3: Consider algorithm (1), (2) under assumptions A or B. Assume thatf(x) is continuously differentiable, and that (20) holds. Assume that the solutions to (19) with initial conditions in D are exponentially stable, and let I be a set of integers, such that inf I'T; - 'Tjl == 60 > 0 where i=l::j and i,jE/. Then for any p~l there exist constants K. f(} and To that depend on p, D, and 60, such that for E < EO and to> To. P {SUPIX(t>-xD(Tt;Tto,X(to»)I>f} <; lEI.
t> 10
~ Jrf
(
.
xfli)
(a)
X
Fig. 1. (a) and (b) To illustrate Theorem 3.
preting (27) is that the gain sequence y(.) can be scaled so that x(·) stays arbitrarily close to x" (.), with an arbitrary high degree of probability. A result that is related to Theorem 3 in the case A (x)=O is given in [35]. Interesting connections between stochastic-approximation type algorithms and a corresponding differential equation have also recently been made by Kushner [36], [37], using weak convergence theory. The proof of Theorems I, 2, and 3 are long and technical. They are given in Appendices I, II, and III, respectively. The idea of the proofs of Theorems I and 3 follow the discussion in Section III. However, a considerable amount of technicalities are required to rigorously justify the "approximatively equal" signs. In Appendix V extensions of the theorems, e.g., to continuous time algorithms and to cases when the limit in A.5 or B.6 does not exist are also commented upon. VI.
(27) where N~sup i; iE I, which may be 00. Remark: In the proof of Theorem 3 it is assumed that the exponential stability of the solution x D ('r, Tl o ; x( 10» is ensured by a quadratic Lyapunov function for the (linear and time-varying) variational equation around this solution, cf., e.g., [10]. Although the proof of Theorem 3 provides an estimate of K from given constants, we do not intend to use (27) to obtain numerical bounds for the probability. The point of the theorem is that a connection between the differential equation (19) and the algorithm (1), (2) is established. In particular, we notice that, due to A.7 there is a p such that the RHS of (27) becomes arbitrarily small when 10 increases, and E is fixed. This means that the estimates stay close to the corresponding trajectory with higher and higher probability as 10 increases. Another way of inter-
x~ (7J 1f.(J .X(~J)
(b)
y(j)p !«
.. ...
THE BOUNDEDNESS CONDITION
In this section we shall discuss condition (20). The reason why it is required is twofold. First, obviously x( I) must be inside DR (with fP(t) not t09 large to prevent an immediate jump) for the differential equation to be valid at all, and also inside DA to get "caught" by a trajectory converging to Dc. Second, and perhaps less obviously, even if DR == DA = R" it may happen that x(t) tends to infinity. The reason for divergence is that if Q (/,X,
464
the measures to keep x(t) in a bounded area may not be completely arbitrary to obtain convergence. A feature that can be used when DR == DA = R", is to introduce a saturation in Q(.; ., .) so that IQ(r; x,
x(t)- [x(t-l)+ y(t)Q(t,x(t-l),cp(t)) ]0 1,02
if x(t - 1) ED) a value in a given compact subset of R m if x (t - 1) e D I
Z
(29)
It
2
if zED
={ some valuet in D2
if z r= D I •
It should be clear that D., D 2 cannot be chosen arbitrarily. Loosely speaking, the trajectories of (19) that start in D2 must not leave the-area'Dl~ Otherwise there may be an undesired cluster point on the boundary of D r- This may be formalized as follows. . Theorem 4: Consider the algorithm (28), (29) subject to assumptions A, B or C. Let D( C DR be an open bounded set containing the compact set D 2• Let fj == D I \ D 2 (D. "minus" D 2). Assume that D2 C DA , with DA defined as in Theorem I. Suppose that there exists a twice differentiable function If(x) > 0, defined in a neighborhood of D with properties sup U'(x)f(x) <0,
(30)
xeD
U(x) > C 1
for x f£ D 1
U(x)
forxED 2•
(31)
Then Theorem I holds without assumption (20). The proof of Theorem 4 is given in Appendix IV. Assumption (30) clearly makes U ( .) a, Lyapunov function in iJ, while (31) formalizes the- intuitive notion of trajectories from D2 never leaving Die We may remark that (30), (31) hold, e.g., if the trajectories of (19) do not intersect the boundary of D 1 "outwards" and D2 is sufficiently close to DrVII. How TO
USE THE THEOREMS
The intuitive content of the theorems of Section V is that the algorithm (1), (2),
x(t)=x(t-l)+y(t)Q(t,x(t-I),cp(t»
can be studied and analyzed in terms of the differential equation (33)
where f(x)== ,--.CX) lim EQ(t,x,~(/,X»).
(34)
(3~).
where D1 ::::> D2, and [ Z ]D D
(32b)
The precise statements about the relations between (32) and" (33) of Theorems 1-3 may be summarized in a somewhat looser language as follows. a) x(t) can converge only to stable stationary points of (28)
A(x(t-I»)cp(t-l)+ B(x(t-l»)e(t) cp(t) =
(32a)
b) If x(·) belongs to the domain of attraction of a stable stationary point x· of (33) i.o. w.p.l , then x (I) converges w.p.1 to x· as t tends to infinity. c) The trajectories of (33) are "the asymptotic paths" of the estimates x(')t generated by (32). These statements are fairly attractive intuitively, and they suggest certain unified techniques to analyze recursive algorithms. We shall illustrate this below, but let us here point out some aspects. By the result a) the possible convergence points of (32) may be determined and studied. That a possible convergence point must be a zero of (34) is fairly obvious and it may be derived without reference to any differential equation. However, the observation that among these stationary points only stable ones are candidates for being limit points of (32) is a most important complement and it is probably less obvious without the present interpretation in terms of the differential equation. Perhaps the main use of result a) is to prove failure of convergence. It may be remarked that usually an algorithm is constructed so that the desired limit indeed is a stationary point. Consequently the possible lack of convergence is then due to the unstable character of the stationary point, so it is the complement (25b) that is the key result for proving divergence. Result b) is the result by which convergence can be proved. In many cases it is not easy to find a proper Lyapunov function to prove global stability of (33), and sometimes the RHS of (33) is quite complex. For certain algorithms, though, in particular those arising from criteria-minimization, it is possible to do this analytically, and some examples will be given below. While analytic treatment of (33) may be difficult, it is always possible to solve it numerically when the dimension of x is not too large. In that way insight can be gained into the global stability properties of the differential equation, the stationary points and their character. In view of result c) the trajectories thus obtained are also relevant for the asymptotic behavior of the algorithm. Therefore, numerical solution of (33) is a valuable complement to simulation of (32). Due to the time scaling (26) in the differential equation, this reveals more rapidly the asymptotic properties and the stationary points of the algorithm. Since the estimates change more and more
465
slowly, due to (4), it is seldom not difficult to decide from simulations only whether the estimates have settled around a limit value or are just converging slowly. In addition, it might be difficult to tell from a simulation if a certain effect is an inherent feature of the algorithm or just depends on random influence. Numerical solution of (33) may resolve such questions. The regularity conditions A, B, or C are usually not very difficult to verify as demonstrated below. Notice in particular that the observation cp may have a "dummy" character, since it may have to be extended to fit into the assumptions (2) and A.I. For example, if the sequence ( e( ·)} in (2) is stationary process with rational spectral power density, then it can be modeled as the output of a stable, linear filter with independent random variables as input. Hence, by extending the vector cp and adjoining this filter to (2), assumption A.I will hold. Even if this may lead to a vector tp with a large dimension, the complexity of the algorithm and of the analysis is not affected. What matters is the calculation of
frequently used in Tsypkin's work, see, e.g., [15], [16], clearly can be understood as a stability condition for (33) with V(x)= lIx - x*lt 2 as the Lyapunov function. Tsypkin's condition
is then a variant of the "boundedness condition." The convergence results thus obtained are, however, essentially restricted to the case cp(.) being independent random variables (A (: )==0) and y(.) satisfying (17) which is quite restrictive for control and estimation application. These conditions are inherently tied to the use of martingale theory in the proofs and cannot easily be dispensed with. OUf Theorem 1 when applied to (36) is thus more general in that (J>(.) may be dependent (gerierated as white noise through a linear filter) and y(.) has only to satisfy A.7. This is satisfied, e.g., for y(/) = Ct :" 0< a < I, while (17) admits only 1/2< a ~ 1. Notice that slowly decreasing gain sequences may be of interest in practice to achieve fast convergence of the sequence of estimates. We must, however, admit that we in return EQ (/,x,qi( t,x») require more regularity of Q and of e(·). On the other and this can often be done without explicit expressions for hand, nonsmoothness of the involved functions is seldom a problem in applications, and we believe that our version A( ·),B(·), 9'( .), and even for Q(" ., ·). In the next section we shall apply the method to a few of the convergence theorem is more widely applicable. In addition, Theorems 2 and 3 are important results for examples and illustrate how the techniques of the items convergence analysis, and we are not aware of similar above may be used. previous results for the Robbins-Monro scheme. In many applications it is of interest to minimize a VItI. EXAMPLES function ErpJ(x,q»=P(x) with respect to x. If the derivative of J with respect to x can be calculated, the stationary Example l-i-Stochasttc Approximation Algorithms points of P(x) can be found as solutions of
a
Consider the problem of solving (35) for x. Here "E." denotes the expectation with respect to qJ, while the vector x is considered as a fixed parameter. Quantities Q(x,cp(t», t == 1,2,· · " are available for any x, where the distribution of the random vector q>( •) does not depend on ·x. Robbins-Monro (11] proved that under
certain assumptions the scheme ("the Robbins-Monro scheme")
x (t) =x (t - 1) + 'Y ( t)Q (x ( I - 1),9>( I) )
(36)
gives a sequence of estimates that converges to the (a) solution of ,(35) in the mean square sense. Convergence w.p.l of (36) has then 'been studied in several papers, e.g., [12H14], and the theorems of these studies do not differ very much conceptually ftom Theorem I. The functions used in the convergence theorems of, e.g., [12] or [13] can be interpreted as Lyapunov functions for the differential equation (33) and condition (20) of Theorem I is ensured by further conditions on this function, rather than by the more practically oriented 'Theorem 4, see, e.g., [12, condition A); or [13, condition B, p. 184]. It can also be
remarked that the conditton (37)
This is a problem that can be solved using the Robbins-Monro scheme and then Corollary 2 of Theorems 1 and 2 is quite useful. If the derivative of J cannot be calculated it seems natural to replace it with some difference approximation. This was suggested by Kiefer and Wolfowitz [17] and their procedure has also been used for, various control and estimation problems. Kushner has in several recent papers discussed interesting variants of this procedure, see e.g., [18], [19]. Our theorems are not directly applicable to the Kiefer-Wolfowitz scheme as they stand, since condition A.3 (or 8.3) is not valid. The reason is that the function Q in this case increases to infinity with t. For the case of additive noise to the function to be minimized, however, it can readily be shown that Theorems 1-4 hold anyway. Details are given in [5] and [7J. , Stochastic approximation algorithms have been applied to a broad variety of problems in control theory, see e.g., Tsypkin [15], [16], Fu [20], and Saridis el aI.. [21]. The approach is known as "learning systems," and in this framework estimation and identification problems, adaptive control, supervised and unsupervised pattern recognition, etc. can be treated. An approach that is related to stochastic approximation
466
is suggested by Aizerman et al. [13]. Their "Potential Function Method" can be applied to various problems in machine learning. Therefore the Robbins-Monro scheme appears in various disguises in many control and estimation algorithms, and consequently the described techniques can be applied 0 to these. A particular example is given below.
-J
c(t)=(x A (/)+x B (t»)/ 2 where
x A (t-l)+ y(t)[ qJ(t)- x A (t-l)], if qJ ( t) is classified as A x A ( t - 1), XB(/)
(39)
otherwise.
is defined analogously. Clearly, x A (I) is the mean
value of the outcomes classified as A. This scheme is discussed by Tsypkin [22] and Braverman [23]. Let cp(/) have the distribution shown in Fig. 2 consisting of two triangular distributions. The probability of outcomes in the left triangle is A. We assume that cp(.) is a sequence of independent random variables. Clearly, it is desirable that the classification rule, the number c(t), should converge to some value between - I and + 1. Introduce
x(t)==
XA(t) ] . [ xB(t)
Then (39) can be written
x(t)=x(t-l)+y(t)Q(x(t-I),cp(t»)
(40)
where
-I
0
Fig. 2. Probability density function of the random variable to be classified by the automatic classifier.
Example 2-An Automatic Classifier A classifer receives scalar valued signals tp(t) which may belong to either of two a priori unknown classes A and B. The classifier must find a classification rule, i.e., a number c(t) such that qJ(t) is classified as A if cp(t) <; c(t) and B otherwise. The number c(/) can, e.g., be determined as follows:
-t
Clearly, the algorithm (40) together with the observation equation
(41 )
is a simple case of (32). Since cp(.) = e( .) is bounded we may use assumptions A. Obviously A.I to A.5 are satis . . fied, and let us assume that y(.) is such that A.6-A.9 hold. (Here A.3 holds in virtue of our somewhat artificial modification of QA; but this example will illustrate that a heuristic use of the present convergence results will reveal important features of the algorithm.) E"Q(x,fP)=!(x) is readily computed as follows. For a given x the corresponding classification point is c(x) = (x A + x B)/2. fA (x) is then the mean value of the distribution left of the point c(x), minus x A • fB(x) is found correspondingly. The algebraic expression for f(x) as a fune .. tion of x and "'A is lengthy and is omitted. We first note that by construction, the estimates are confined to the area D: 3> x B :> x A > - 3. Therefore condition (20) of Theorem 1 is trivially satisfied. Analytical treatment of the differential equation x = f(x) is not easy, but its trajectories can easily be determined by numerical solution and they are shown in Fig. 3 for two choices of A. For the case A= 0.5, [Fig. 3(a)] there is convincing evidence that the point x* = ( - 2,2) is a stable stationary point with global domain of attraction. Therefore, for A=O.5 it follows from Theorem 1 that x(t)~x* w.p.l as t --+00, which gives a correct classification rule c· = O. The case A=O.99 [Fig. 3(b)] corresponds to a common situation where errors that occur rather seldom (I percent), "outliers," shall be detected. In this case there are two stable stationary points of the differential equation. x* = (-2,2) and x*·=(-2.3, -1.4). There is obviously a non.. zero probability that x(t) belongs to the domain of attraction of x·* i.o. Therefore Theorem 1 shows that for "'A = 0.99, and for any starting value x(O) there is a nonzero probability (that depends on x(O)) that x(t)~x·* as t~ 00. This gives an asymptotic classification rule = - 1.8, that classifies 39 percent of the "correct values" as out.. liers. For this case simulations of the classifier are shown in Fig. 4. In fact, the simulation leading to the undesired value c** appeared only after several (257) attempts and from simulations only it might have been tempting to conclude general convergence to c·. In this example it is cumbersome to find a suitable Lyapunov function for the stability problem. However, as seen in Fig. 3 numerical solution of the differential equation yields sufficient insight into the stability properties. Such detailed information can naturally be obtained only if the dimensionality of the problem is small.
c··
and A(
Q
A X
,x
B
)_{cp-x
,qJ -
0,
A
,
if
if
i(x x
i<x
A A where the values for + B) - 8 c cp ~ + X B) + 8 are such that QA is a continuously differentiable function
of qJ and x. Here 8 is some small positive number.
467
.....---------n
(a)
Fig. 4. Simplations of the classifier (40) for the case A-O.99.
the unit circle, i.e., D (q - I) is an exponentially stable filter. Even if an exact description of the system is impossible, a fJ can be determined that gives a model (44) which describes the recorded data as well as possible. Often 8 is determined by minimizing a criterion based on the equation error
,/ ;'
3 .,
(47) (b)
Fig. 3. Trajectorie, for the ODE that is aasociated with self-learning classifier (40). (a> A-O.S. (b) A-O.99.
Example 3-Equation Error Identification Methods A common way of modeling dynamic systems is as a vector difference equation (VDE),
Several algorithms based on the idea of somehow minimizing (47) have been suggested in the literature, see e.g., [24] and also [25] for a comprehensive treatment. The probably best known method of this type is the least squares algorithms, see, e.g., [24]. Then the sum N
~ lJy(t)- 8T~(t)1I2
(48)
1-1
y( t)+ A 1y (t - I)+ ' " +AIV"(t-n) · mimnuze . . . d w.r.t 9 to 0 btai . 9(N)based on IS tam the estimate = 8.u(/-1)+··· + B"u(t- n), (42) measurement up to time N. An important and well known feature of this method is that the sequence of estimates where y(t) and u(t) are column vectors and A; and B; are can be obtained recursively as matrices of appropriate dimensions. Introduce
8=(A." .A"B I " I
.
9 (t) ... 9 (t - 1)+ y(t)%(t) [y( t) - 9 (t -1) T\fi( t)
B,,)T
"'(/)= ( -Y(/-I) T... -y(t-n) Tu(t-I)···u(t-n) T)T. (43)
=iJ T1IJ( t).
(44)
(49a)
%( r) = R - 1 (I - I)",( t )
/[ 1+ y(t)(\fi(t)TR -I (t-l)\fi(t) - I)] == R - I (t)l/; ( t)
Then (42) can be written y (t)
r
R(t)=R(t-l)+y(t)[ \fi(t)\fi(t)T - R(t-l)]
(49b)
(49c)
We may remark that (44) also covers several other inter- (usually (49c) is written in terms of R -1(/), which makes esting estimation problems, not necessarily related to sys- it of "Riccati type"). For the minimization of (48) y(/) has tem identification. to be taken as 1/ t. Other sequences 'Y ( .) correspond to Usually, the true system cannot be described exactly in criteria where old measurements are discounted, which
the form (44). Suppose that it can be described as
Y (I) =8[1/I(t) + v( t)
often is relevant in practice. Let us assume that the input to the process is de(45) termined as
where v( .) is a disturbance that can be modeled as
V(t)=D(q-l)e.(t). (46) Here D (q - I) is a matrix with rational functions of the backward shift operator q -I as entries and e l( .) is a stationary sequence of independent random vectors with finite moments. It is assumed that the denominator polynomials in D(z) (z replacing q-l) have all roots outside 468
where F(q-I) and H(q-I,9) are matrices with rational functions of the backward shift operator q - I as entries. Let el( · ) be a stationary sequence of random vectors with finite moments, that are mutually independent and also independent of et(·). Moreover, H(q-I,8) is a causal operator that allows output feedback terms in the input.
This feedback law may depend on ,the current parameter estimate as is further discussed in Example 5. It is clear that the rational filters in (46) and (50) can be represented in a state space form,
Since e;(·) are stationary, q5(t,X) w~l approach stationarity exponentially, for all x, such that 8 makes the closed-loop system stable. Therefore the limits
f( i) = lim E~ (t,x)[ j( I,;) - 8T~ (I,X)] T (59a)
v(t) = (J O· · · O)zo(t)
I~~
(51) ZII (t
G(i)= lim E~(t,x)f{t,X)T
+ 1) == All( 8 (t) )Z., (I) + z., + B.e2(t + I) + BuyY (t + I); U(/)=(I O· . · O)ZII(t),
(52)
are well defined where y(/,X) and \fI(t,X) are the corresponding parts of ~(t,i), and
where zo(·) and z..(·) are the corresponding state vectors of appropriate dimensions. We may now form the "observation vector,"
lim EQ(t;x,~(I,X»= 1-+00
cp(t)== [ Y(I) T ,1/1(1) T ,Zo(l)T ,ZII(t)T]T
(59b)
1-+00
i -(1(O)_ _) 1 [colG(8)-R
(60)
(53)
so 8.6 is satisfied. Moreover, from (57) and (58) it follows that B.7 holds, since all moments of ~(t,x) and V(/,A,C) exist. Conditions B.8-B.ll about the sequence y(.) are e1(t») cp(t)=A",(8(/-l»cp(t-I)+ B", ( e ( t ) (54) assumed to be satisfied. 2 The conclusion therefore is that the differential equation where the matrix Atp( •) is formed from (45), (43), (51), (52) in an obvious manner. Its eigenvalues are the poles of the ; 8('1")= R -1 ('I")j(8('I") (61a) filters D(q-I), F(q-I) and of the closed loop system which is obtained for (42) with a constant feedback (50) d using 8 (t - I). There are also a number of eigenvalues in (6Ib) d'T R(T)==G(8(T»)- R(T) the origin, arising from the shifting of the vector 1/1(/). Notice that Arp(9) depends on 9 only since the feedback can be associated with the algorithm (49). In the remainfilter H(q-l;8) does. Let us take ing part of this example, we shall assume that the feedwhich obeys
back matrix H does not depend on 8 (i.e., there is no (55) adaptive feedback), that the matrix F(z) has full rank a.e. z and that e;(·) are full rank processes. (Adaptive feedback is further discussed in Example 5.) This means that (56) the matrix ACJ>(·) does not depend on fJ, cp(/,X)==cp(t), and y(t,X)= y(/), so the values in (59) are directly defined in
X(/)=( 8(t)T colTR (t)( Then (49) takes the form
x(t)==x(t-l)+y(t)Q(t;x(t-l);cp(t»
with an obvious definition of Q(I;X,cp) from (49). There- terms of input-output covariances. In particular, the fore the algorithm (49) together with (54) is of the general matrix G is independent of 8; G(fJ)= G. form (32). Let us check if assumptions 8 of Section IV are Introduce satisfied. Conditions B.I and B.2 are satisfied due to our assumptions. By straightforward calculations it is readily r = Ecp( t)v( t) T (62) shown that B.3 is satisfied in the open area DR == {xiR > and we have, using (45), OJ, [cf., (55)] e.g., with
f(9)= G·«(Jo-8)+r.
:JC,(x,cp,p, v) == (191 + p)(l + Icpl + V)2 /( 1- pi R -11)2 (57) for p=p(x)<
l/IR -II. Then
Hence, (61) can be rewritten as
B.4 will be satisfied with
~(x,cp,p,v,w)-(191+p)(lcpl+2w+ v)/(I-pIR
2
; 8 ( r] = R - I ('1") G[(80 + G - 1,) - 8 ( r) ]
-II) . (58)
d .
Condition B.5 is satisfied if the matrix H(q-I;(J) is With Lipschitz continuous in 8. For condition B.6 we define
~(/,x)::z A",( i )~(t -
(63)
l,x) + B", (e l (I») e2 ( / )
and
(i = (iT colTR)T). 469
d"R(T)==G-R(T).
(64a) (64b)
we have
{A(t)}. Take in this example the observation to be
~ V(O (T),R (T»)= -20 TG9+OT (G- R(T»O == -
iT (1")( G + R (1"»)9 (1") <; 0
11 ( /) ) cp( t) = ( colA( t)
(69)
(65) and in (32b) A (x):=O, B(x):=/. The estimate is as before given by (55). The condition C.I has already been verified and XI can this time be taken as
so that V is a Lyapunov function for (64) or (61) that assures that the stationary point
'X,(x,cp.P, v) = I + Ixt + p+ IAI + c.
(66)
(70)
has a domain of attraction equal to DR' Therefore condi.. tiIon.(21) 0 f Th eorem l 1S i sa t'ISC·red WI with D D T0 Ch ek cCondition C.2 IS. trivially satisfied, The vanable z(t,X) A == R. , . condition (20) we note first that the assumption on full defined by (18a) IS given by rank and finite moments of "e;( .) implies that CI > G > 81 for some 8 > 0, C < eo. Therefore also (8/2)1 < R (t) < !R -I [1J(s) - A(s)8] 2CI and IqJ(t)1 <2C i.o, w.p.l. We also note that (49a) can t $=1 for large 9(t-l) be written z(t,i)= II
•
•
0
0
0
0
•
•
±
1
(J ( t) ~ (I -
col
R - I (t) \fI( t)'"(t) T )9 (t - 1)
R- 1
that
==
w.p.l as
t
~ [A(s)- i] s==1
which shows that 9(t) can, w.p.l, not tend to infinity. Hence x(t) belongs to a compact subset of DR i.o, w.p.l and condition (20) is satisfied. Theorem I now implies 9(t)~9·
t
1-+00.
col
{[! ± t
1J(S)]-[+
±A(S)]e}
s-1
s=1
(71)
[+ ±A(S)] -colR s-1
In particular, we see that the least squares estimate is
\.
consistent only if r =0, which essentially is the same as where requiring that v(·) is a sequence of uncorrelated random variables, and that the current u(t) is uncorrelated with future v(s), s ~ t. colR Other variants of equation error methods are treated analogously. These facts are, of course, well known, [24], (25], at least Therefore, if the sequences (1J(s)} and for the case y( t) = 1/ t, but one reason for this example is that that the analysis extends into less trivial problems. 0 1 t - ~ 1J(s)-+.q 1 £-1 Example 4-Equation Error Methods-Assumptions C
x==( e ).
The model (44) and the criterion (48) appear in several contexts, like curve fitting, etc., where a stochastic framework often is not imposed. Let y( I) == 1/t and denote
"'(t)y(t)T ==1)(t)
(67a)
1/J( I)1/;(t) T == A( t).
(67b)
{~\(s)}
are such
(72a)
and
(72b) then C.3 holds with
(73)
Then the algorithm (49) can be written
8(/)-8(1-1)+!R -I (/)[1J(/)-A(/)8(/-l)]
(68a) Since A and B do not depend on x (see C.?) condition C.4 is satisfied if
t
R(t)-R(/-l)++[A(/)-R(/-l)].
(68b)
1 I - ~ IA(s)1 t 3-1
converges
(74)
We shall in this example illustrate how assumptions C can be applied to the algorithm (68) to infer convergence which is implied by (72b) if A(s) == ~(s)1/;(S)T. The directly from the properties of the sequences (.,,(t)} and boundedness condition of Theorem I has been verified in 470
Example 3. The differential equation
methods, like Landau's scheme [28] etc. Since a companion paper [29] is devoted to the last two schemes we shall d' - not go into any detail here. Further analysis of some dr 8(T)= R -I (T)[ 71- A8(T) ] (75a) recursive identification schemes using the present approach can also be found in [I], (30), and [31 [, ;R(T)="A-R(T) (75b) It was remarked above that the structure (32) can be understood to be patterned after adaptive control of linear is as shown in Example 3 globally asymptotically stable systems. We shall therefore conclude with such an example. with stationary point (76) Example 5-Selj-Tuning Regulators
if A is positive definite. Hence Theorem 1 implies that 8(,) given by (68) converges to e* defined in (76) for any sequences {A(s)}_and {ll(s)} such that (72) holds with a positive definite A. This is by no means surprising since (68) in fact only is a way of recursively writing -I
9(/)=
L~I A(S)] S~I 7I(s)
using the matrix inversion lemma, However, it illustrates that our method of analysis does not bring in "unnecessary" assumptions, even though it is concerned with a more general algorithm. Moreover, for a stochastic approximation version of (68), 8(/)=8(/-l)+! [71(/)- A(t)8(1 -1)] I
(77)
We shall in this example discuss an application to the self-tuning regulator, described in [32]; see also [2] and [4]. This regulator is based on least squares identification, (49), and the output feedback law is determined from the current parameter estimates. Usually the feedback law is chosen to be a minimum variance regulator, [32], but here it could be a general linear regulator as in (50), where perhaps in most cases F is zero. In this case the matrix ACJi«(J) in (54) does depend on () since the feedback term does. However, the point now is that, in contrast to conventional analysis of the least squares algorithm, most of what was said in Example 3 still holds. Up to (61) the development was quite general. This differential equation is valid also in the case of adaptive feedback. although G and r now are functions of 8. If v(') is a sequence of independent random variables, then r=O and (64) and (65) hold. (Equation (65) does not hold if r~O depends on fJ.) We therefore still find that the points defined by
we have exactly the same convergence result under the condition (12), since the associated differential equation then is form an invariant set with global domain of attraction. (78) Clearly ·BoE D~ . and whether D~ contains more points depends on the choice of feedback law and model order. In this case no explicit expression for 8(/) is available to There is a further complication before Theorem I can be infer this result directly. In fact, the algorithm (77) has applied. In this case the area Ds is unknown" i.e., the area been studied in [26], [27] using fairly elaborate methods, of such (J that inserted in a constant feedback law (50) d d" e(,,)=~-Ae(T).
since it was found that the usual stochastic approximation
makes the closed-loop system stable (Acp(9) in (54) has all
convergence results could not be applied due to the correII lation in the sequences {A(s)} and (1I(s)}. These four examples have all been for the case where A(·) in (32b) actually does not depend on x, The convergence part in this case can, at least under further assumptions, not seldom be treated by more conventional statistical methods. When A ( .) does depend on x, conventional approaches become much more difficult, and in fact, also in the proof of Theorem I, a major burden is to keep control over the coupled stability questions in (32a) and (32b). The inclusion of x-dependent A -matrices becomes necessary for more elaborate recursive identification methods in which the observed data are processed through filters that are formed from current parameter estimates. This is the case, e.g., for the extended Kalman
eigenvalues inside the unit circle). Therefore we cannot guarantee stability by projecting 9 into Ds . Hence condition (20) of Theorem I has to be verified by other consid.. erations, e.g... by showing that the overall system has a certain stability property as in [33]. But when this is shown, Theorem 1 proves convergence of 9(/) into D, w.p.l. Let us repeat that this holds for the case of arbitrary feedback law, but under the assumption that r(·) is white noise. For general noise c( ·) the convergence analysis is more cumbersome. but it can be performed in certain special cases" [2], [4]. [29]. We refer also to these papers for more details on how Theorems 1. 2. and 3 can be used in the analysis of self-tuning regulators. Numerical solution of the associated differential equation has turned out to be a valuable tool here, and it has been used in [34] as well as in the references above.
filter, the extended least squares method, output error
471
IX.
CONCLU.SIONS
W." The abbreviation "i.o." denotes as before "infinitely often." U
Recursive algorithms like (I) have been analyzed in various contexts. However, we would like to stress again that when cp in (I) is generated as in (2), the analysis becomes more difficult. The reason is that (1) no longer is recursive in x for analysis purposes: the whole history of x(·) enters in each step of (I). Moreover, the coupled stability problems between (1) and (2) are intricate. But the structure (1), (2) is nonetheless common in estimation and control problems; a typical example is adaptive control of linear stochastic systems. The analysis of this case is also known to be usually very difficult. • With the present approach we are able to give a general treatment of (1), (2) under assumptions that do not appear to be restrictive. The examples indicate that the theorems may be applied to rather diverse problems, and perhaps the technique also may serve as a basis for a unified approach to the analysis of adaptive controllers. In addition, an extension is obtained for the conventional convergence results in the simple case where A ( .) in (2) is independent of x. We may remark that the analysis is restricted to the asymptotic behavior, convergence, possible convergence points etc. of the algorithm. Two related algorithms which are associated with the same differential equation may differ noticably in transient behavior and convergence rate. In the described theory, we would like to stress the intuitive content of the theorems and the methodology of analysis as outlined in Section VII. I t is no doubt important to appreciate" the exact formulations of the theorems and to know the exact conditions under which they are valid. But it is perhaps equally rewarding to use the properly defined differential equation as a general instrument for analysis in a more heuristic fashion. This may be exemplified in Theorem 3, which has a fairly technical formulation. and' is probably more valuable as a "moral
ApPENDIX
I
PROOF OF THEOREM
I
Outline
In order to prove Theorem 1, we shall first show that the estimates provided by the algorithm locally and asymptotically follow the trajectories of the associated differential equation. This will be done in Lemma 1 under assumptions C and the proof of this follows the intuitive outline of Section III. After that the local behavior of the algorithm is thus established, this is used to prove that all cluster points of the sequence {x(t)} must belong to Dc. This is done by means of a Lyapunov function, the existence of which is inferred from the stability condition (21). A possible clusterpoint outside the set D, would yield a decreasing value of the Lyapunov function along the corresponding trajectory of the differential equation. Since Lemma 1 proves that we follow this trajectory asymptotically, a contradiction is obtained. The value of the Lyapunov function is decreasing outside D, and it is not possible to return a given point outside D, infinitely often. This is the intuitive path of proof. The many technicalities tend to obscure the simple idea, and the proof will be structured as much as possible to enhance the basic ideas. Finally it is proved in Lemma 2 that assumptions A or B imply assumptions C for the case y(n) = 1/ n. An outline of the general proof will also be given, but the details of this are omitted, and the reader is referred to [7] for them. Lemma 1: Consider the algorithm ( l ), (2) under assumptions C. Let .~ E DR and define the number m(n,a'T) such that
support" for studying the trajectories of the differential
m(n.~'T)
equation, than in its literal sense.
~
y(t)~aT
asn-+oo.
(1.1)
n
Assume that
ApPENDICES
The proofs of Theorems 1--4 are given in Appendices I-IV. They are slightly compressed versions of the proofs in [7], in that some calculations of technical nature are omitted. Proofs for a simple special case of the algorithm (1), (2) are also given in [6]. These are naturally less technical than the present ones, and reveal perhaps more clearly the underlying ideas. The basic path of the proofs of Theorems 1 and 3 follows the heuristic outline of Section III. Some notational conventions in the proofs should be noted. "C" will denote any constant, that need not be the same in different parts of the proof. Important dependencies of the constants will be given as arguments, while indexed constants are "global" throughout the proof. "~5 (x,p)" denotes as before an open p-neighborhood of x. Realizations in the sample space n will be denoted by
s. p)
(I.2a)
is sufficiently small
(I.2b)
x ( n) E ~~ (
where p = p( x)
and that (1.3)
Then, there exists a value aTo=~To(;'p) and a number No= NoC'i,p) such that for ~T
(1.4) where m=m(n.. ~'T) and
472
ql (n,m,x)~O
as n-+oo
(1.5)
and
p = p( x)
(1.12)
where A(X) is defined by (14). The idea obviously is that since the matrix A (X) is where C1 and C2 depend on x and p, but not on n> No or exponentially stable, exponential stability of the timevarying difference equation (2) given by A(x(k» will be aT. Proof of Lemma 1: The proof is structured into 4 guaranteed if x(k) varies in a sufficiently small neighborhood of x. The formal proof consists of straightforward steps. In Step 1 an explicit expression for xC)~ where calculations given in [7], but omitted here. The "sufficiently small" p mentioned-in (1.2b) refers to the fact (1.7) that p should be so small that (1.12) implies (1.11). From (2) it follows that is derived. This expression shows that if certain terms are small, then xC}) is close to what would be obtained if Q(t,x(t ~ 1), cp(t» is replaced by Q(t,x,~(t,X) [with [ IT.A(X(k»]B()e() ~(t,X) defined by (15)]. As a first step to show that these cp(t)= [ k-n j-" ks«] terms indeed are small an expression for cp(t)- q;(t, X) is derived in Step 2. This expression is used in Step 3 to (1.13) prove that assumptions C.3 and C.4 imply that the terms are small. A complication so far has been that in order to which together with (1.11) and (1.3) implies that prove these things it must be assumed that x(s) remains in t for n <; s <; l- In Step 4 it is a small neighborhood of Icp( t)1 < C·C.,X I-n + ~ eX I-JIB (j)lle(j)1 (1.14) proved that x(s) actually will remain in this neighborhood j-n up to s = m(n, 41") if 41" is chosen sufficiently small (and dependent only on X) and n sufficiently large. This will if (1.12) holds. Since the variable conclude the proof. ep (t) ~ cp ( t) - ~ ( t; x) Introducej= j(,,) such that
IT A(X(~»lcp(n)+ ~
x
(1.8) obeys and
ep (t + 1) == A ( i)ep (t) + [ A ( x ( t)) x(k) E
k = n,n + 1,. · · ,j - l.
~ (x,2p),
A (x) ] ~ ( t)
(1.9)
+ [B(x(t»)- B(x) ]e(t+ 1) (1.15)
Step I-An Expression for x(j) Directly from the algorithm (I) the following expression we have that is obtained: j
~
x()=x(n)+
y(s)Q(s;x(s-I),cp(s»
1-1
+ ~ A(x)'-S[{A(x(j»-A(x>}cp(j)
s-,,+1
j-n
j
==x(n)+ ~ y(s)Q(s;x,iP(s;i»)
l
n+l
+ { B (x(j») - B (x)} e(j + 1)
(1.16)
j
+ ~ y(s) [ Q(s, x(s -1 ),«p(s») - Q(s; x,~(s; x») J
If KA and KB denote the Lipschitz constants of A(.) and (1.10) B (.), respectively (see assumption C.2), then from (1.16) and (14) it follows that
n+1
where q>(s; i) is defined by (IS). End Step 1. The first goal is to show that the last term in (1.10) is "small," and in order to do this we first consider qJ(s) -
lep(t)1 <;
f.A(X)/-n {1
~(s;i).
+ c ~ A(X)'-i[ KAlx(j) - ~Hq>(j)1
Step 2-An Estimate for Iq>(t)- q;(t,X)1 'We claim that I
II A (x(k»)";
k-"
c[
1+A( -) 1-"
X]
2
j-n
+ KBJx(j) - ille(j + 1)1].
~ CX(x),-n
(1.11) Clearly, with v(n,A,c) defined by (16),
c{I«p(n)1 + ,,( n,x)l} c v(n,A( x),c)
if x(k) E ~ (i,2p);
n ( k <; t
(1.17)
for sufficiently small
for some constant c. 473
(1.18)
Now, inserting (1.14) and (1.18) in (1.17) gives
We thus find that
lcP (/)1 c A(i)'- nv( n,A(x),C) + c- max (KA , Ks -)
j
x(j) = x(n)+ ~ y(s)f(x)+ R. (n,j)+ R2 (n,j)
(1.23)
s-n
where I-I
+
j
j
R 1 (n,j) = ~ y(s)[ Q(s,i,lj)(s;i")- f(i)]
~ ~ A(i)/-JXJ-k)e(k)1
$.
j:snk-n
(1.24)
n
and
Sum first over j in the double sum and introduce c' and A where such that
A< 1.
·{I + V(S,A,C)}
Then 19'(t)-q;(t,i)1 (A(X),-nV(n,A(x),C)
+ max
n<j
~(j) -
(1.26)
and
x1 ·V(t,A,C).
(1.20) Q4(n,}) ==
ma~ n<s<j-t
{'X1(i,cP(s; x),2p,V(S,A,C») j
.v(s,A, c)'"'(s) }. ~ heX) s- n.
In particular,
(1.27)
,s-n
h(x),-nv(n,~(i),c)'-; C(t,A(X),C)
Step 3a-maxn<J<m(n,~1')IR.(n,j)I-+O as n-+oo Let the maximum be attained forj=j*(n). Solving (18a) we obtain
and therefore also
Iq;(t)- ~(t,i)1 (V(t,h,C)
(1.21)
j*(n)
~ y(k) where A can be taken as (3+A(X»/4 and c is a constant k=n+l that is obtained from c.' c in (14), C in (1.11), c~ in (1.19).. KA , KB , and maxIB(x)l(xE~(x,p». However, since ·[Q (k,x,cP(k.x») - z(k - Lx)]. assumption C.4 shall hold for any finite c, there is no point in tracing the expression for c through the calcula- Now, let n tend to infinity. By assumption C.3 tions. lim z(n,x)= lim z(j*(n),x)=f(i) We summarize this step as follows: Choose 2p in (1.9) n--.oo n--.oe so small that (1.12) implies (1.11) for this p. Then (1.20) and hence and (1.21) hold for n <: t
z(j*(n) ..x)=z(n.i)+
n--oc
IQ(s,x(s - 1),
End Step 3a. Step 3b: There exists an N 1 such that
· [lx(s-l)-xl+lcp(s)-~(s;x)l]
(
max
n
Ix(k) -
forn>N 1,
il' 9(1(x,~(s,i),2p,V(S,A,C»)
·{I + V(X,h,C)} + %1(x,q;(s,x),2p,v(S,A,C») 'A(x)s-lI v(n,A,c)
wherethe last inequality follows from (1.20).
where kf) is the limit of kv(t,x,A,c) defined in C.4. This step is completely analogous to Step 3a, using the defini(1.22) tion (ISb) and assumption C.4. End Step 3b. Step 3c-q4(n,)(n))-+O as n-+oo 474
It now follows from (1.33)-(1.35) that
From (18b) y( s )·v(s,A,c)' %.( x,q>(s,x), 2p, v(s,A,c»
Ix()) -
il < 2p
== kv(s) - kv(s - 1)+ y(s)ko(s - I). and hence that (1.30) and (1.31) hold. End Step 4. The RHS tends to zero according to C.4 and C.6 and the From (1.33)-(1.35) it also follows that claim is proved. End Step 3c. IX(j) - xl ~ If(x)I~T + qs( n,m(n,~,,» Expressions (1.23)-(1.27) together with Step. 3 would complete the proof of Lemma I, if }(n) were equal to +4jii1T'k:, + tx(n)- xl. (1.36) m(n,4T), that is if x(k) remains in ~(i,2p) for all k up to m(n,~7'). This is proven in the final step. Since the RHS does not depend on), this inequality holds Step 4 for all n « ] <; min, ~T) for sufficiently small aT and large There exists a ~TO== a"o(x,p) and a No=: No(x,p) such n. that for ~T < LlTo and n> No, the expression (1.9) will hold Therefore, from (1.23)--(1.27) and (1.36) for j == m(n, ~T). By the definition of j(n) x( m(n,4T») = x(n) + aTf(x) + q. (n,m,x) + q2(m,n,x) x() - I) E ~~ (x,2p) (1.28) (1.37)
and
where
max
n<.k<'j-I
Ix(k) -
xl < 2p.
We shall show that (1.28) and (1.29) imply that also x(j)E~~(x,2p)
(1.39)
aT.
By induction it then
for sufficiently large n and small follows that x(k) E
~ (i,2p)
and
Iq2(n,m,x)1 c 2k:,.11"[ I/(x)l- a1" +4pad~:' + Ix( n) - xl]
k==n,···,m(n,d'T).
(1.31)
(1.39)
We have
for aT < ~TO and n > No. The expression (1.37) with (1.38) and (1.39) is the assertion of the lemma since q.(n,m(n,dT)X} lends to zero j according to Steps 3a-3c and (1.1) [remember that qs is
xl ( Ix(j) -
x(n)1 + Ix(n) -
xl
tR.
and C.4 hold and such that condition (20) of Theorem I
+4p·aT·~ + Iq4(n,j)1 +p
holds. It follows from the converse stability theorems (see e.g.,
strictly less than V(x(nk » if x+x*. The formal proof is somewhat lengthy and involves several elaborate choices of constants.. The result is, however, intuitively clear. The proof of Step 1 extends to (1.45). Step 2: From the above result it is quite clear that x* must be a cluster point to x(n), since V(x(n» has a tendency to decrease everywhere in DA except for x==x*. Step 3: If there is another cluster point to {x(n)} than x·, say X, the sequence must move from x* to x infinitely many times. But then V(x(n) is increasing, which contradicts the result of Step 1. Hence, only one cluster point exists and convergence follows. Step 1: Any cluster point of (x(t)} outside D; yields strictly decreasing values of V. From assumption (20) of Theorem 1 there exists a subsequence, such that
Then (1.44) implies that by a proper choice of e, aT, and K (see [7] for a detailed account), we have V(x(m(nk,aT») < V(x)-aT6/64
k>K.
(1.45)
End Step 1. This means that if x is a cluster point different from x* the sequence x(n) will i.o.. be interior to D={xlxE~ and V(x) < V(X)-~T8/64}. The conclusion that xED R follows from our assumption that trajectories that start in DR remain in there. This is r~quired in o!der to apply Lemma I to new points in D. The set D is compact. Consequently, another cluster point must exist that yields a smaller value of V. Moreover, since x(t)E~(x,2p) nk < t <; m(nk,d'T) we have from (1.14) k"
1
w
Xk - k'le(k + l )1
k=-=k'
(1.46) Since D is compact, there exists at least one cluster point and hence 1
as k--+oo..
(1.41)
Consequently, for arbitrarily small E: > 0,
Step 2: Suppose (1.. 45) holds for any subsequence {x(nk ) } that converges to a point different from x* . Then lim infV(x(t»)=O.
(1.47)
1'-+00
(1..42)
Consider inf V(x) taken over all cluster points in DA • Let this value be U. Since the set of cluster points in DAis compact, there exists a cluster point X, such that V(x)= U. If now U > 0, V'(x)f(i) will be strictly negative (= where m is defined as in Lemma I. Denote nk= k~ and 8) and from (1.45) V (x (k» takes a value less than Um(nk,AT)==k", and use the mean value theorem. This 5~'T/64 i.o.. , which contradicts U being the infinum. gives Hence, U=O, which means that x· is a cluster point. End Consider now
0
S~p2
V[x(k")]- V[x(k')] = V'(f)[x(k")-x(k')]
Step 3.* From (1.45) and (1.47) it follows that
t~~supV(x(t»=O.
=V'(i)[x(k")-x(k')]+[€-it
(1.48)
. IIy Let p* = p(x*) be the region for w hich (2)·IS exponentia stable for x(k) E ~ (x*, p*), as in (1.11), (1.12). where ~ and f' belong to ~(i,E+41"). Suppose that Now take e P(X), and we can in view of (1.42) apply Lemma 1 to x(nk ) , which gives (1.49) lim sup V(x(n»= W>O. .V"(t.')[x(k")-x(k')] ~
(1.43)
n~oo
x(k") - x(k') :=A".f(x) + ql (k',k",x)+ q2(k',k",x)
Take W < W such that
where q; are subject to (1.5), (1.6). Insert this into (1.43)
{xIV(x)( W}cB(x·,p*)
and consider the interval 1=[ W /3,2 W /3].. (We may of course choose any subinterval; there is no particular rea(1.44) son to divide it up in thirds.) Since x· is a cluster point and since V (x(n» is where supposed to have a subsequence tending to W, this interR3(~T,nk,i) =(f - x)TV"(f')(x(k") - x(k'») val I is crossed "upwards" and "downwards" infinitely many times by V(x(n). + V'(x){ ql +q2}· shall now proceed to show that (1.49) would imply that there must be a subsequence of V(x(n» that belongs Now suppose that the cluster point x is different from the to I, by proving that in ~(x*,p·) the "step size" x(n+ I) desired convergence point x·. Then V'(X)f(i)== - 8, 8 > - x(n) tends to zero. o,
V[ x(k")] - V[ x(k')] :aA'TV'(x)f(x) + R) (~T,nkx)
We
476
We shall now proceed to prove that assumptions A or B imply assumptions C.3 and C.4. This will be done only for the case Y(t) == 1/ t, which anyway should be the most common one. We may also remark that since the correction term Q(·, ., .) may be time-varying, this case also includes cases where
First, let x(iik ) be a subsequence tending to x", such that Jcp(nk)1 < C. (The existence of such a sequence follows from Step 2 and the stability argument (1.46), using the fact that p =p(i) is bounded from below by a positive constant in D.) For t > nk' but such that x(t) "remains" in B(x·,p·) we have, (1.46)
c(t) y( I) = -t- ; c( I)--+C
Icp(t)1 < eX 1-"*fcp(nk)1
(1.56)
be redefining the factors of the product
t
I
+ CIBt
as 1-+00
Xt-kle(k+ 1)1
y(/)Q(t,x,
ie-nit
Lemma 2. Suppose the algorithm (I), (2) is subject to assumptions A B. Assume that y(/)= lit. Then C.3 holds w.p.1 as well as Steps 3b and 3c of the proof of Lemma I. If, in addition
with v(t,A,c) defined by (16). Hence,
or
ly(t)Q(t,x(t),cp(t»1 c y(t)1Q(t,x·,O)1
+ y(t)X.(x·,O,p·, v(t,X,c) )(Ix(t) - x·1 + Icp(/)1)
E ~(x, ep( I,X), p(x), v( I,A, c)( I + v( I, A, c»)
< Y(/)IQ (t,x·,O)1 + y(t)X1(x· ,O,p*, v( t,~,c») ·(p*+v(t,A,c»
converges at t~(X), then also C.4 holds w.p.l. Proof: For y(t)= I/t we have from (18a)
(1.50)
±
where the first inequality follows from assumption C.l. (1.57) z(t,x)= 1. Q(k;x,lj)(k,x». It follows from Step 3a. of the proof of Lemma 1 that t k-I the first term of the RHS of (1.50) tends to zero and from Step 3c that the second one tends to zero. Consequently, We shall apply the following ergodicity result of Cramer inside B (x·,p·) the step size tends" to zero, and hence and Leadbetter [38, pp. 94-96) to (1.57). there will be a subsequence of V(x(n» entirely in the Let f(k) be a sequence of random variables with zero interval I. Consider now a special, convergent sequence of means and covariances "upcrossings," that is a subsequence of this. kP+s P Let the sequence n:C be defined such that O<2p
».>
V(x(nk+sk»>2 W I3 where SIt
I
I
t
k == I
-I
(1.53a)
f(k)-+O
w.p.l as t~oo
II
(The proof in [38] is given for the continuous time case, is the first s for which V (x(n" + s) ~ J (I.53b) but it goes through without changes for the discrete time x (nk)-+.i
as k~oo.
(I 54) case.)
· In order to calculate the covariances, introduce for From (1.51) and (1.52) it-is clear that V (x) == W /3 and let k > s the variable V'(x)j(i) =- 8. From (1.45) we now have that
V(x(m(nk'~1'»)< W/3-8~T/64.
~(k,x)
(1.55) by
This means that V(x(n k + si» ~ I where Sk = m(ni,~T) -
n".Suppose now that there is a Sk<·S" such that V(x(n,,+
Ij)sO(k + I,x)" A (x)lj)sO(k,x) + B (x)e(k + I);
Ij)sO(s,x) =0.
This variable is independent of ep(s,i) according to
sk»>2W/3. However, V is continuous and x(n,,+s), assumption A.l==B.l, and (s<m(nk,~f')-nk)
will belong to an arbitrarily small neighborhood of x(nk) for sufficiently small A". according I~( k,x)- cp~(k,x)t (C-A(X)Ic-.r,q;(s,x)1 to Lemma I. Therefore (1.55) implies a contradiction to .( <; V(k,h(X),C». the existence of a subsequence ni with properties (I.SIHI.S4). Hence, no interval I may exist, W must be zero and (1.48) follows, End Step 3. Hence, using assumption A.3 or B.3 Equation (1.48) implies according to the properties of IQ(k,i,q>(k,x»- Q(k,i,~(k,x»l V(·) that x,,-+x*. This concludes the proof of Theorem 1. c %l(x,q)(k,x),O,v(k,A(X),C»)CA(i)k-.rI~(s,i)t. End proof of Theorem 1. 477
Moreover, from A.2, A.3 (according to which %1 and are bounded) or B.7,
We shall need the estimate (1.61) in Appendix III, and let us therefore prove (1.61) for p = I. The case with general p is proven analogously, though with more techni.. cal labor. The details are given in [7]. Denote for short
~(s,i)
EIQ(k,x,q>(k,x)- Q (k,x,~(k,x»12 <; C·A(X)"-J. (I.59a)
QIc = Q(k,x,ip(k,i»)
Now I COy
and assume that Qk is a scalar and that EQk =0. Then
{Q( k,i,,(k,i», Q(s,i,q;(s,x»)} 1
2
= ICOY{ Q(k,x,~( k,x»
- Q(k,x,~(k,x»
EC~n Y(k)Qk) = k~n .r~n y(k)y(s)EQkQ.r
+ [ Q(k,x,~ (k,x»
],Q (s,x,~(s,x»}1
m
"J
(n)2 C ~ ~ A1k -JI c Cy(n)2(m- n) (1.62) k -n s'==n
-IO+cov{ [Q(k,x,~(k,x"»-Q(k,x,~(k.x»],
where the first inequality follows from assumption A.8 Q(s,x,~(s,x»}1c C.).(X)k-$, (I.59b) and (I.59b). It remains now only to prove that where the second equality follows since ~(k,X) and ~(s,X) are independent, and the last inequality follows from Schwarz' inequality, (1.59a), and assumption B.7.
(m(n~~T)-n)(c/y(n).
Since y(.) is decreasing
Hence, (1.58) holds and
! t
m
2~T ~ ~ y(k) )(m - n)y(m)
±[Q(k,x,~(k,i»- EQ(k,x,~(k.i» ]_0
n
or
k-l
w.p.l as 1-+00
which according to assumption A.5 = B.6 implies that
!
t
±Q(k,x,~(k,i»-f(x)
(1.63)
w.p.l as
(m-n)(c/y(m).
(1.64)
Moreover, from assumption A.9 __I y(n+ I)
I~OO.
k-I
I_
(1.60) we have
Notice that (1.60) holds for any given x w.p.l. For a given y ( n + 1) > y ( n)( 1- cy( n + I)., realization '" outside a null set it does not immediately follow that it holds for all x E DR. To conclude that, we or, upon repetition first note that (1.60) will hold w.p.l simultaneously over a m dense denumerable, subset of DR. (The union of a de- y ( m) > y ( n) II (1 - cy (j )) numerable number of null sets is a nullset.) Hence, for a n+1 given realization outside a null set (1.60) holds for all x in a dense subset of DR- But since Q(., ., -) and f( .) are -..,y(n)exp (- cYU)).--c'y(n). continuous in x, we may extend this set to DR itself. n+1 The proof of the second part of Lemma 2 is analogous, (Under assumptions A it is trivial.) This, together with (1.64) implies (1.63) and the proof of End proof of Lemma 2. (1.61) is complete for p = I. The proof for general sequences {yen)}, subject to A.S, Proof of Corollary 10 Theorem 1: The function V(x) A.9 is achieved by first proving that with the property (22), whose existence is assumed in the corollary plays the role of the Lyapunov function m(",~.,.) throughout" the proof of Theorem I. Since. in the formulaE ~ y(k)[ Q(k,x,q;(k,i») tion of the corollary there is no guarantee that x(m(n, ~'T» k-" will belong to DR even if x(n) does, the possibility of a 2p cluster point on the boundary of DR cannot be excluded. - EQ(k,x,ci(k,i»)] < C·y(n)p (1.61) Moreover, if D is unbounded the existence of a cluster point in D is not guaranteed. But if no cluster point exists, then the sequence (x(n)} will tend to infinity, which in and then proving convergence by the Borel-Cantelli that case will be a boundary point of DR" End proof of corollary. Lemma and assumption A.7.
:2
478
that T(m) will, with probability one, not tend to zero as m tends to infinity. Hence, x(m) will not converge to 0 (= x·) with nonzero probability. To prove (25a) we first assume thatf(x*)*O. The general case is proven by linearization around x· We first note that according to Lemma 2, assumptions and the additional terms are taken care of by appropriate A or B imply that (1.4)-(1.6) hold w.p.l. approximations. Like in the proof of Lemma 1, this leads Let .!IT( x*) - .!l",* be "the sufficiently small" ~". as de- to several technicalities, but the basic idea remains the fined in Lemma I. same as above. Take To keep the Appendices within reasonable size, these technicalities will not be given here. The full proof can be (11.1 ) p. < ~"'*lf(x·)1/4 found in [7]. Proofs for related algorithms are given in [6] ApPENDIX
II
PROOF OF THEOREM
2
and let Sl·"-{~Ix(t)-+~(x·,p·)} with P(O*)=P*>O. If x(t) converges to ~ (x·,p·), it is in particular inside ~1 (x·, 2p*) infinitely often. Therefore, there is a cluster point x inside ~(x·,2p·). Also, inside ~(x·,2p~) the observation equation (2) will be stable for sufficiently small o", according to Step 2 of the proof of Lemma I. Hence, we can select a subsequence {x (nk)} tending to X, such that Icp(nk)1 < C, i.e., (1.3) holds. Then Lemma 1 implies that
and [1]. ApPENDIX
III
PROOF OF THEOREM
3
The idea of this proof is to apply Lemma I to obtain local estimates of how much x(t) differs from the corresponding trajectory, and then linking such estimates together making use of the stability property. Heuristic Outline of How the Estimates are linked Together
asymptotically as k -+ 00 for '" EO· except on a null set. The idea of how the local estimates are extended to This means according to (11.1) that x(m(nk'~"'·» is outside ~ (x·, 2p·), and the assumed convergence is con- global ones can be geometrically expressed as follows, cf. Fig. 5. tradicted. Assume that the estimate at time Tic is in the interval A. To illustrate the basic idea of the proof of (25b), conThe trajectories that start in A belong at time Tk + .!l1"/( to sider the special case the interval B which is smaller than A since the trajectory Q(t,x,(JJ(t + 1») == Ax + e(t + I) is stable. Now, the estimates obtained by the algorithm differ from the trajectories with a small quantity accordwhere A is an ntn matrix and e(·) is a sequence of ing to Lemma 1. Denote this distance by C. During the independent random variables with zeto mean values. time interval .1'T , the estimates have not diverged from k Suppose that A has an eigenvalue A with ReA> 0, and let the nominal trajectory if A <; B+2C. L be a corresponding left eigenvector. Let T(n)- Lx(n) To achieve this, A and ark must be chosen with care. and E(n) = Le(n). The condition on cov Q implies that The interval ~7'k must be large enough to let the trajecto{E(n)} is not identically zero. Then the algorithm (I) can ries converge sufficiently, "and small enough to limit secbe written ond-order effects and the noise influence. The formal proof will be developed in 5 steps. In Step I T(n+ 1)=T(n)+ y(n+ l)[AT(n)+E(n+ 1)] the details of the application of Lemma 1 are given. Step 2 deals with the implications of the stability assumption. In and Step 3 an interval for t1Tk is selected, corresponding to ~Tk not being too small nor too large as described in the heuristic outline. In Step 4 it is shown that the estimates T(m)=f(n,m)· {T(n)+ n+1 stay with the given e-region as depicted in Fig. 5 if certain stochastic variables are less than a given value. Step 5 where calculates the probability that they are less than this value. Step 1-App/ication of Lemma 1: f(n,m)= (1 +;\;y(k»)-exp {A Y(k)} (11.2) Order the set of indices J =-{ n;} such that
f
Pk~(k)}
f
IT
n~1
n+l
n I < n2 < . · .
and
Denote ~7'k = Tn I - T". Then by taking i Lemma I we obtain
k
~k=y(k)
< nk < nk + I < ·..
I. ...
II (l+Ay(j»-I. n+1
Since T(n) and the sum of random variables are independent and r(n,m) tends to infinity as m increases, it follows
~
= x(n,.) "
In
x( nk+.) - x( nk) + aTJ[ x( nk)]
479
(111.1)
Step 2-Stability of the Trajectories
According to the assumptions of the theorem there exists a function V(Ax,1') that is quadratic in Ax and such that
A
c so
(111.9)
Pia- S. To illustrate the idea of the proof of Theorem 3.
along solutions of the variational equation of (19). Since we assume V(aX,'T) to be quadratic in ~x it is no additionalloss of generality to assume that
where Q2(nk,nk+ I,X) is given by (1.6) and (1.39). Since
V (AX,T) == IAxl 2
since it is always possible to make a (time-dependent) (111.2) change of metric. Moreover ql(nk,nk+l'X) is given by (1.38). From Lemma 1 we also know that (111.1), (111.2) are valid only if A'1k is sufficientty small and nk so large that
Iq. ( nk' nk+ .,x(nk»)1 < p12.
Then (111.9) implies that Ix D ('T + 41"; 1",x + 4x ) - x D (T+ d1" ;1",x )I « I - Ad'T )IAx l (111.10)
(111.3)
for some I >A>0. We first note that the c01!!tant C2 in (111.2) can be taken End Step 2. to be globally valid in D, according to the e~ression Step 3-Selection of an Interval for A'Tk (1.39) and since 1/(x)1 and kf) are bounded in D. Mor~ To obtain the upper and lower bounds for ~Tk disover, there is a common lower bound to ~10(X) for xED, cussed under the heuristic outline, let which can be realized as follows. . (460M ~T oM4 4"\ rz: ) The radius 2p, which is so small that (1.12) implies (1.11) (0 = mm -A- , 3;\ 'X v pM · (II1.11 ) depends on x and is a measure of how fast A (x) changes in a nei&.hborh~d of x. Since A (x) is Lipschitz continuChoose « (0. Then automatically ous in D and D is compact, the radius P(X) will have a positive lower bound as x varies over D. Denote this by p. all k (111.12) Let [cf., (1.34)] since 4'Tk > 80 according to the assumptions of the theoa'To= inf (111.4) rem. An upper bound for &1"k can be obtained by possibly xED 11(x)I' 16~ extending the set I, cf., beginning of Step 1. Do this so that whi.£h is strictly positive, since I!(X)I and k., are bounded in D. (111.13) Then (111.1), (111.2) is valid for tJ.'Tk < ~TO if (111.3) holds. By solving the differential equation (19) from 'T to The resulting set I may depend on E. According to T + dT with x as the initial condition we have (111.11), automatically !:!Tk < ~TO. End Step 3. Ix D (1" + £\T;T,X) - ( x + 41" f (x»)1< L4T 2 (111.5) Step 4-The Estimates Remain in an e-Neighborhood of the where L can be taken globally in D. Introduce the follow- Corresponding Trajectory if q.(nk + .,nk,x(n/<) is Sufficiently ing abbreviated notation Small, all nk E I Assume that D X;D (j) = x (1",; 1"~,x(n;»). (111.6) (111.14a) Iq1(n k + r- nk, x (nk )) I< r ( f) Equations (111.1), (111.2), and (111.5) now imply in the notation of (111.6) where
(--L _1_)
e
Ix( nk + I) -
D Xk
(k + 1)1 « c2 + L)aT2 + q. (nk + I,nk,x( nk »)
A2( 2· 3
(111.7)
(III.14b)
r(E)<8/2.
(III.I4c)
According to (111.11),
if (111.3) holds. Introduce M=C2+L.
r(E)= M.16 ·
(111.8)
The estimate (111.7) with the expression (1.38) for q. is the basis for the rest of the proof. End Step 1.
Assume also that
480
txf(k)-x(nk)l
c E.
(111.15)
Since ~'T'k now is assumed to be less than ~10 and since (111.14) is assumed to hold, Lemma 1 can be applied, giving according to (III.7) and (1II.J4) [notation as in (111.6)]
fx(nk+1)-xf(k+ 1)1 <; M·a'Tf+r(E).
according to assumptions B.7 (or A.3) and 8.10. Analogously to the first term of (111.19) its 2p-absolute moment around the mean is bounded by C·y(n)p. (Much better estimates are possible, but uninteresting.) Denote the two first terms of (III.19) by q~ (m,n,X) and the last one by (111.16) qf(m,n,X). Then
Now
q, (m,n,i)= q~ (m,n,i)+ qf (m,n,i)
Ixf(k + 1)- x(n/c+ .), (Ixf(k + 1)- xf(k + 1)1
where
+IXkD(k+ 1)-x(nk+I)1
(111.21 )
«I-A~'Tk)txf(k)- x(k)1
and
+ M·i1'Tf+ r(E)
Iq~ (nk+ .,·nk,x(nk»)1 < r( E)/2,
~2E2·3
~(l-.;~~hk)E+M·~TI+ M.16 =
E
+ M[ (~Tk -
4A~ )( ~Tk - ~~ ) ] ~ E
(111.20)
(111.22)
Chebyshev's inequality gives, using (111.21) (111.17)
P(lq~ (nk+l,nk,x ( nk»)1> r(f)/2) 2p
where the second inequality follows from (111.10) and ~( 'C-y( nk)p· (III.23) (111.16), the third from (111.15) and (III.14b), and the last one from (111.13). The probability that (111.14) does not hold for some The conclusion is that if x(nlc) lies in an e-neighborhood »;» N I , nJc E I is thus bounded by of the trajectory, so will x(nlc+) if (111.14) holds. If (111.14) holds for all nk E I, then
rtE»)
sup Ix(nk) - xf (k#)1 < e. nk E /
End Step 4. It remains now only to estimate the probability that (111.14) holds for all nJc E I. Step 5-The Probability That (111.14) Holds The expression for q.(n"+I,nk,x(nk» may according to
(1.38) and (1.22) be decomposed as follows
Iql(m,~,i)I<1
f
y(k)
Combining the conclusion of Step 4 with this result gives the assertion of the theorem. The number To mentioned in the theorem equals N 1 defined by (111.22). Remark: The inequality in (111.24) is somewhat wasteful. For example, if y(t)= l/t, then with
k-n
· [Q(k,x,q;(k,x»- EQ(k,X,q;(k,X»]1 implies that nk+l~nk(l +a'Tk ) for small A'Tk and taking
III
+
I
Ie-"
p = 1 in (111.24) an upper bound for the LHS of (111.24) is
y(k)XI(i,~(k,x),2p,v(k,A,C»)
given by
·A "-lIv(n,~,c)
+ I~Tf(X)-
f
(
y(k)EQ(k,X,q;(k,X»I· (111.19)
k-n
_2 r(E)
)2.C'
f
I
nk-NINI(l+~.,o)k 2
<: _1
(--1..-) .C. 4M <. C/(5. r(E) hE
NI The last term in this expression is deterministic and tends to zero according to assumption A.S =B.6 .and according End proof of Theorem 3. to (1.1) (m - m(n,~1"». Let N I be such that this last term is less than r(E)/2 for nk> Nt. ApPENDIX IV For the first term of (III.I~) the estimate of its 2p-ab-
solute moment, (1.61) holds. The second term of (111.19) has a mean that is bounded by
PROOF OF THEOREM
(111.25)
4
In virtue of the projection we know that X(l) belongs to a compact area i.o. that is part of DR. We could therefore apply Theorem 1 directly, apart from the fact that the
III
C' ~Y(k)Ak-"~ l~A y(n) 481
projection algorithm (28), (29) differs from the algorithm (1), (2) treated in Theorem 1. It therefore suffices to show that the "projection" takes place at most a finite number of times w.p.l. After the last time x(t) is forced into D 2 the projection algorithm coincides with the basic algorithm (1), (2) and the proof of Theorem 1 is valid. If indeed, the estimate x(t) were outside D. infinitely often, then it would have to pass from D 2 to outside D. i.o., i.e., to a higher value of U (x( t» (see (31» in spite of the force trying to decrease U according to (30). In Step 3 Theorem 1 it was proved that this is of the proof impossible, and hence the projection facility in (28) is used only a finite number of times. Also, the estimates cannot remain in D from a certain time on, since condition (30) shows that (using Step 2 of the proof of Theorem 1) they will be forced into D2•
0'
ApPENDIX
ON SOME
V
e«
m(n.~1')
y(k)f(x,k)-+a'Tf(x)
Equation (V.4) if of course just a special case of the general, nonlinear dynamics (3). To treat this case let i(X) denote the limit of the recursion
and define ~(t,X) through ~(t,x")
I,x) + B (i (x) )e(t).
(V.6)
Iz(t)-z(t,i)I< C· max li-x(k)1 r « k<1
if z(r)=z(r,x)
(V.?)
li( x) - z( l,x)1 <; CJL I. Ii (x) - z(O,x)I; #L < 1 (V.8)
as n--+oo
Jz(1)1 + Icp( t)1 < c i.o.
(V.9)
To prove this we note that assumption A.4, (Y.7) and (V.8) imply that
'A(z(j»)-A(i)l
+ cJLJ-nli (x) -
z(n)l.
If we use this estimate in (1.16) together with (V.9) we arrive at the same expression as in (1.17) and the proof continues without further changes.
(V. I) ACKNOWLEDGMENT
I would like to express my sincere gratitude to Profs.
K. J. Astrom, H. J. Kushner and Ya, Z. Tsypkin for
±
(V.2)
f(x,k)-f(x)
important discussions on the subject of this paper. I have
also benefitted from the referees' competent suggestions on how to improve the readability of the Appendices.
k-l
Notice that (V. I) or (V.2) very well may hold even if A.5 =B.6 does not hold. It is consequently unnecessarily restrictive to assume convergence of f(x,t) as t~oo. Moreover, if (V.I) does not hold, it is sufficient to require the existence of a twice differentiable function V(s) [which plays the role of Lyapunov function for the in this case undefined differential equation] such that m(n,A.,.)
V'(i)
=A (i (.~))~( t -
The theorems will then hold in the same formulations as before, if the following three conditions on (V.4b) hold:
which, for the case y(k) = 1/ k is equivalent to
t
(V.5)
z(t.x)=h(z(t-I,x),x),
k-n
1
(V.4b)
and
EXTENSIONS
Continuous Time Algorithms: It should be clear that the theorems and proofs given here are valid also for continuous time algorithms, with proper and straightforward modifications. The Case when Assumption A.5 = 8.6 does not hold: The basic assumption, defining the RHS of the differential equation is A.5 =8.6. X) be denoted by f(i, I). In the proof Let EQ (t, X, it is only used that ~
:(/)=h(z(t-l)~x(t».
[
I
y(k)f(i,k)
1
< -6(i)·~T
(V.3)
k-n
for all sufficiently large n, where 8 (X) >0 for xfi. Dc. More Complex Generation of the Observations: In some cases, like for the extended Kalman filter, which is presently being analyzed using the methods of this paper, a more complex mechanism replaces (2): cp(t)==A(z(t--l»cp(t-l)+B{z(t-l»e(t) (V.4a) 482
R~FERENCES
(I) L LjUD& T. SOderstrom, and I. Gustavsson, ·'Counterexamples to generaJ convergence of a commonly used recursive identification method," IEEE Trans. Aulomat. Contr., vol. AC·20, pp. 643-652, Oct. I?,S. (2] K. J. Astrom, U. Borisson, L. Ljung, and B. Wittenmark, "Theory and a~lications of adaptive regulators based on recursive identification, in Proc. 6th IFAC Congress, Boston, MA, 1975. An extended version of this paper will appear in Automauca, Sept. 1977. . (3] L Ljunl and S. Lindahl, "Convergenceproperties of a method for stale estimation in power systems," Int. J. Contr.; vol. 22, pp. 113-118, July 1915. (4] L. LjUDI and B. Wittenmark, "Analysis of a class of adaptive regulatOrs," in Proc. 1FA C Symp. on StocNutic Control, Budapest, Sept. 1974. [5] L. Lj\IDI, "Convergence of recursive, stochastic algorithms. n in Proc. IFAC Sy"". on StocJuutic Control, Budapest, Sept. 1974. (6) ··Stroq coDveraence 01 a stoehutic approximation algorithm," Annals of Sttltiltics, to be published; alSO available as Rep. LiTH-ISY-I-0126. LiDkopinl University, LiDkOping, Sweden. [7] wrheorems for the asymptotic analysisof recursive, stochastic alloritlulls," Dep. Automat. Contr., Lund Institute of Technology, Lund, Sweden, kep. 7522.
(8] "Convergence of recursive stoeltastic a110rithms," Division of [27J D. C. Farden, "Stochastic approximation with correlated data," Automatic COntrol, Lund Institute of Technology, Lund, Sweet,.., Dep. Elect. EDI., Colorado State University. Fort Collins, CO, Rep. 7403, Feb. 1974. aNR Tech. Rep. 11. [9J W. Hahn, Stability of Motion. Berlin, Germany: Springer·Verla& [28] I. D, Landau, 6&Unbiased recursive identification USinl model 1967. reference adaptive techniques," IEEE TrtUU. AU'Omtl'. Con" .• vol. (IOJ R. W. Brockett, Finite DimerLr;OMI LinetU Systems. New York: AC·21, pp. 194-202, Apr. 1976. Wiley, 1970. (29) L. LjUDa, "On positivereal transfer functions and the convergence [11) H. R.obbins and S. Monro, .tA stochastic approximation method," of some recursive schemes, this issue, pp. 539-551. AM. MflllI. StilI., vol. 22, pp. 400-407, 1951. (30] T. SOderstrom, L. LjUD& and I. GUitaVSlOn, "A comparative study (12] J. Blum, "MultidimensiotW stochutic approximation methods," of recunive identification methods," Automatic Control, Lund AM. "'lItla. Slat., vol. 25, pp. 737-744, 1954. Institute of TechnolOI)', Lund, Sweden, Rep. 7427, 1974. (13) M. A. Aizermann, E. M. Braverman, and L. I. Rozonoer, "Metod {31] L. LjUDI! ''Convergence of an adaptive filter algorithm." Int. J. Potentsialnykh fUDktsij v teorii obuchenija mashin" (The method Comr.; ) 977, to be published. 01 ~teIltial functioDi in the theory of machine learning). I zd. [32J K. J. AstrOm and B. Wittenmark, "On self-tuning regulators," Ntlillca, Moscow, 1970 (in R.ussian). Automtl,;ctl. vol. 9, pp. 185-199, 1973. (14) M. T. Wuan, Stoclttu'ic Approximtltion. Cambridge University [33] L. LjUDI and B. Wittenmark, "OIl a stabilizina pr~ty of adap~~I~. . tive reaUlators," in p,.ri"U 4th IFA C s,,,,,. IdMIijicat;on, Part 3, (IS) Va. Z. Tsypkin, Adtlption lind ua,n;", in Automatic Systems. Tbilisi, pSSR, Paper 19.1, pp. 380-389, 1976. New York: Academic, 1971. [34] K. J. Astrom and B. Wittenmal'k, "Analysis of a self-tuning [161 - , Fou1lda';01l$ of 1M TMory of Learning Systems. New York: reaulator for nonminimum phase systems," in Proc. IFAC Synp. Academic, 1973. on S,ocluu,;c Control, Budapest, 1974. [17] J. Kieferand J. Wolfowitz, "Stochastic estimation of the maximum (35) D. P. Derevitskii and A. L. Fradkov, "Two models for analyzing of a regression function," Ann. Malh. Slat., vol. 23, pp. 462-466, the dynamics of adaptation alaoritbms,·t AutOmtl'. and Remole 19S2. Comr., vol. 35, pp. S9~7, 1974. [18J H. J. Kushner, "Stoehutic approximation algorithms for the local (36] H. J. Kushner, "General CODveracnce results for stochastic apoptimization of functions With non-unique stationary points," proximations via weak CODVerpace theory," Brown University, IEEE TI'tIIII. Auto,,",t. COIIIr., vol. AC·17, pp. ~SS, 1972. Providence, RI, LeDS Tech. Rep. 76-1, 1976. (19) H. J. Kushner and T. Gavin, 6IStochutic approximation type (37] - , "Convergence of recursive adaptiveand identification procemethods for constrained systems: Algorithms and numerical redures via weakconvergence theory," IEEE TralU. Automat. COllI'., suIts," IEEE TrarLr. Automat, Cont,., vol. AC·19, pp. 349-3S8, to be published. 1974. (38J H. Cramer and M. R. Leadbetter,KStati01ltUY and &Iated (20) K. S" .Fu, "Learnilll system theory:' in System 1Mory, L. A. Zadeh Stoc1l4flic Processes. New York: Wiley, 1967. and E. Polak, EdI. New York: McOraw·Hill, 1969. pp. 425-466. (39] N. N. Krasovskij, Stability of Motion. Stanford, CA: Stanford (21] G. N. Saridia, Z. J. Nikolic, and K. S. Pu, "Stochastic approximaUniversity Press, 1963. tioD a110ritbms for sxstems identification, estimation and decom- (40) W. Hahn, StiUJi/ity of Motion. Berlin, Germany: Sprinpr-VerIa& position 01 mixtures, IEEE Trans. S,st, Sci. Cybem; vol. sse-s, 1967. pp. 8-1S, 1969. (41) V. I. Zubov, Methods of A. M. Llt1J11lN1V and tlteir Applications. (22] Va. Z. Tsypkin, "Self..learninl-:-What is it?," IEEE Trans. AutoGroningen, Holland: Noordhoff, 1964. nult. Co,,'r., vol. AC·13, pp. 608-612, 1968. (23] E. M. Braverman, "The method of potential functions in the problem of training machines to recognize patterns without a teacher." Automtlt. R.1IIOt. COItl,., vol. 27, pp. 1748-1770. 1966. (24] K. J. Aatrom and P. Eykhoff, "System identification-A survey," AlltOlfUlt;ca, vol. 7, pp. 123-164, 1971. (2S] J. M. Mendel, DUcrele. T«ltniqws of ParalMte, utiffUJlion. New York: Marcel Dekker, 1973. [26J L. L. Scharf and D. C. Parden, "Optimum and adaptive array processing in frequcnce-wavenumber space," in Proc. IEEE Conf. ....... LjIIDa (S'74-M'75), for a photograph and biography see page 5SI of this issue. on Decuion and Control, Phoenix, AZ, 1974, pp. 604-609. tt
483
Discrete-Time Multivariable Adaptive Control GRAHAM C. GOODWIN, PETER J. RAMADGE, AND PETER E. CAINES
ADAPTIVE control, so great is its appeal, has been studied for almost forty years. Its early history is one of diverse, but interesting, heuristic endeavors. One of the first problems studied, and one that has formed the focus of much subsequent research, is that of adaptivemodelreference control in which the controller employs an estimate of the unknown plant to cause the output y to track y* = H r, where r is the reference, H a prespecified reference model, and y* the desired output. Another consistent theme is certainty equivalence, regarding (and using) the estimated plant as if it were the true plant; the motivation is obvious ... if the estimated parameters converge to the true parameters, the certainty equivalent controller converges to a satisfactory (optimal) controller for the true plant. However, adaptive controllers can perform satisfactorily even if, as is often the case, the estimated parameters do not converge to their true values; this was dramatically shown in the paper by Astrom and Wittenmark [1], also included in this volume. Early use of Lyapunov theory to establish stability was restricted to minimum phase plants with relative degree one. Plants with a relative degree higher than one posed a major problem [17] that was not satisfactorily resolved until the late 1970s; indeed, the paper by Goodwin, Ramadge, and Caines was a major contribution to the resolution of this formidable problem and a substantial stimulus to subsequent research. Suppose the system to be controlled is described by
where YE := (1/ E)y, UE := (1/ E)u, () is a vector of the coefficients of the polynomials Rand S, and the regressor vector ¢ has components qi y E, i = 0, ... , n - 1, and qi UE, i = 0, ... , n. The estimation equation y = ¢T e may be used in many different ways to provide an estimate of the unknown parameter e. Suppose the desired closed -loop transfer function is H = (N / E) so that the desired output is y* = H r, and that b.; = Sn = 1 is known. To determine the control, the system equation may be written in the form
e
E2Y = RYE l
Ey = Ry
where 1/J:=E2 ¢ (the vector 1/1 has components qi YE l , i = 0, ... , qiUEl' i = 0, ... , n). The degree of E 1 is n so that 1/JT emay be written as 1JT e + u. If eis known (and the system is minimum phase), the control u may be chosen (as u = E2Y* 1JT ()) to cause the output to satisfy E2Y
= 1JT e + U = E2Y*
so that the tracking error y - y* decays to zero (and all signals remain bounded). If () is unknown, a tempting strategy is to employ the certainty equivalent control u = E2Y* - 1JT with the result that E 2 y* = 1/JT and the output is now given by
where E = E 1E2 is monic and Hurwitz, and £1 and E2 have degrees nand m, respectively. Hence (ignoring exponentially decaying terms)
e
e,
E2Y = E 2y* + 't/JTe The tracking error er := Y - y* satisfies et
+ Su
= 1/JT e
n - 1, and
Ay =Bu where A and B are polynomials of degree nand m, respectively, in the shift operator q, and A is monic; for continuous-time systems q is replaced by the derivative operator d / d t. The system may be re-parameterized (as shown in [1] and [15], the latter in 'state-space language') to obtain
+ SUEl
y
= e + er
e
e
where := ¢T is the estimated output, e := Y - Y = ¢T is the prediction error (Monopoli's augmented error), := () is the parameter error, and er := (y - y*) is an estimate of the tracking error eT. Since
eT
e
e
= [(1/ E2)l/!T]e - (1/£2)[ l/IT OJ
e e
it is also known as the 'swapping error'; eT is zero if is constant, but otherwise depends on the rate of change of (in the continuous-time case it is proportional to (d / dt)e). Monopoli [15] proposed use of the certainty equivalent controllaw in conjuction with a simple gradient estimator (d/ d t)O = ¢e; however, as was later pointed out, convergence of the
485
tracking error to zero was not established except for the case of unity relative degree. The first proof of convergence (for continuous-time systems) was achieved in [4] by modifying the controller; nonlinear damping was added to ensure convergence of eT to zero; this enabled convergence of er (and boundedness of all signals) to be proven. Because of the complexity of the controller, and of the accompanying analysis, the paper [4] did not have the impact it deserved. The second and much simpler proof (for discrete-time, systems), presented in this paper by Goodwin, Ramadge, and Caines, achieved convergence by modifying the estimator; the rate of change of 8 was reduced (thereby reducing eT ) by the introduction of error normalization (replacing e in the estimator algorithm bye := e/[I, + 1c/J1 2]1/2). Error normalization was independently proposed in [3]. The relevant result is Lemma 3.1 of the paper by Goodwin, Ramadge, and Caines, which essentially states:
(a) Suppose the estimator is such that the normalized prediction error e = e/[1 + 1c/J1 2]1/2 lies in f 2 • (b) Suppose the controller is such that the regressorvector c/J satisfies the growth condition 1c/J(t)1 2 ~ c[1 + max{le(r)1 2 IrE [0, t]} for all t Then:
~
o.
(i) Ic/J(t) I is uniformly bounded, and (ii) e(t) ~ 0 as t ~ 00.
The result is simple to state and prove. It provided a simple means for establishing convergence (and boundedness of all signals) for a wide range of adaptive controllers and contributed to an explosive re-awakening of interest in adaptive control. An important feature of the result is its modularity: condition (a) on the estimator can be established independently of (b), i.e., independently of the control law. The result was presented in an electrifying seminar at Yale University (New Haven, Connecticut) in late 1978 and was rapidly extended to the deterministic continuoustime case in [16] and [21]. All three papers appeared in the same issue of the IEEE Transactionson Automatic Control; ironically, the paper by Goodwin, Ramadge, and Caines was the only one that did not appear as a regular paper. A parameter estimation perspective of the continuous-time results was given in [5], providing an analogous modular decomposition of the conditions for convergence. Stochastic versions of the paper by Goodwin, Ramadge, and Caines and of [5] appeared, respectively, in [7] and [6]. Research on convergence and stability has continued to this day. For example, the earlier nonlinear damping approach of [4] was extended in [11] to deal with 'true' output nonlinearities for which certainty equivalent adaptive controllers are inadequate, and backstepping [13] was developed as a systematic tool for designing adaptive controllers for linear and nonlinear systems with high relative degree. Furthermore, switching was introduced to enforce convergence when structural properties
(e.g., relative degree) are unknown [18], and modifications were made in the basic algorithm to ensure robustness [19], [9], [12]. Research results of this period were rapidly consolidated in texts such as [8], [20], [14], [2], [22], and [10]. The paper by Goodwin, Ramadge and Caines is an important milestone in the evolution of adaptive control. It contributed much to the richness of a subject that has progressed far and that now appears poised for further significant advances.
REFERENCES
[1] K.1. ASTROM AND B. WITTENMARK, "On self tuning regulators," Automatica, 9:185-199, 1973. [2] K. 1. ASTROM AND B. WITTENMARK, Adaptive Control, Addison-Wesley (Reading, MA), 1989.
[3] B. EGARDT, StabilityofAdaptiveControllers, Springer-Verlag (NewYork), 1979. [4] A. FEUER AND S. MORSE, "Adaptivecontrol of single input, single output linear systems," IEEE Trans. Automat.Contr., AC-23(4):557-569, 1978. [5] G. C. GOODWIN AND D. Q. MAYNE, "A parameter estimation perspective .of continuous time adaptivecontrol," Automatica, 23:57-70,1987. [6] G. C. GOODWIN AND D. Q. MAYNE, "Continuous-timestochasticmodelreference adaptivecontrol," IEEE Trans. Automat.Contr., AC-36(ll): 12541263, November 1991. [7] G. C. GOODWIN, P.J. RAMADGE, AND P. E. CAINES, "Discrete time stochastic adaptive control," SIAM J. Contr. Optimiz., 19(6):829-853,1981. [8] G. C. GOODWIN AND K. S. SIN,AdaptiveFiltering, Prediction and Control, Prentice Hall (EnglewoodCliffs, NJ), 1984. [9] P. A. IOANNOU AND P.KOKOTOVIC, AdaptiveSystemswithreducedmodels, Lecture Notes in Control and Information Sciences, Vol. 47, SpringerVerlag(New York), 1983. [10] P. A. IOANNOU AND 1. SUN, Robust Adaptive Control, Prentice Hall (EnglewoodCliffs, NJ), 1989. [11] I. KANELLAKOPOULOS, P. V. KOKOTOVIC, AND A. S. MORSE, "Adaptive output-feedbackcontrol of systems with output nonlinearities," IEEE Trans. Automat.Contr., AC-37(11):1666-1682, 1992. [12] G. KREISELLMEIER AND B. D. O. ANDERSON, "Robust model reference adaptive control," IEEE Trans. Automat. Contr., AC-31(2): 127-132, February, 1986. [13] M. KRSTIC, I. KANELLAKOPOULIS, AND P. V. KOKOTOVIC, Nonlinearand AdaptiveControlDesign, John Wiley, New York, 1995. [14] P. R. KUMAR AND P. P. VARAIYA, StochasticSystems: Estimation, Identification and AdaptiveControl, Plenum Press (New York), 1986. [15] R. V. MONOPOLI, "Model reference adaptive control with an augmented error signal," IEEE Trans. Automat.Contr., AC-19:474-482, 1974. [16] A. S. MORSE, "Global stability of parameter-adaptive control systems," IEEE Trans. Automat.Contr., AC-25(3):433-439, 1980. [17] A. S. MORSE, "Overcoming the obstacle of high relative degree," Journal of the Societyfor Instrumentand ControlEngineers, 34:629--636,1995. [18] A. S. MORSE, D. Q. MAYNE, AND G. C. GOODWIN, "Applicationsof Hysteresis Switching in Parameter Adaptive Control," IEEE Trans. Automat. Contr., AC-37(11):1343-1354, 1992. [19] S. M. NAIK, P. R. KUMAR, AND B. E. YDSTIE, "Robust Continuous-Time Adaptive Control by Parameter Projection," IEEE Trans. Automat.Contr., AC-37(2):182-197,1992. [20] K. S. NARENDRA, Adaptive and Learning Systems-Theory and Applications, Plenum Press (New York), 1986. [21] K. S. NARENDRA, Y.-H. LIN, AND L. S. VALAVANI, "Stable adaptive controller design, Part II: Proof of stability", IEEE Trans. Automat. Contr., AC-25(3):440-448, 1980. [22] S. SASTRY AND M. BODSON, AdaptiveControl: Stability, Convergence and Robustness, Prentice Hall (EnglewoodCliff, NJ), 1989.
D.Q.M. & L.L.
486
Short Papers
m.eaD-Iquue output is bouaded wheDevcr the sample mean-square of the HOWC\W, the questioD of stability remaiDI
Dllaet&-11IIIe M _ Adapdve Coatrol GRAHAMc. GOODWIN, ....... JIIIB, PETER. J. RAMADGI!, AND PETER. E. CAINES. MJ!I88I. IBI!B
.4....
nil
Ur
-..,--.It ...... .,.... ............................. . . . . . . . . . . . . . .'
6
II
•••• "rde
I.
tIIIt
uaauwerecllor atoebutic adaptive aJaorithmL The study of discrete-time determiIlistic aJaorithms is of indepeDdent intereat but also provides iDsiaht mto stability qucatioDl in the stoehutic cue [12~ (IS]. Recent work by 100000U ad Monopoli (13) bu been concerned with the exteDlioD of the results in (2) to the cliIcrete-time cue. AI for the continuous cue, the aupICIlted error method is used. In tbiI paper we preIeIlt 11ft' resu1tI nIatecI to clilCJ'ete-time ~ iatic adaptiw coatroL Our approKh diffen from previous work in several ~or respects altbouP ca1aiD upectI of our approach are
..
. - . ••• c-.. .. 'II'cl ........ wII-..
,
.-.u
noiIe is bouaded.
iaspirecl by the work of Feuer aDd MOlle(5]. The uaIysia preIeIltecl here does Dot rely upon the use of aupDeDted erron or awdlWy iD.putI. Moreover, the alpithms have a Vf!IIY simple structure and ue applicable to lDultipltHDput multiple-output systems
IN'raoOUCl10N
A_ problem ill control theory has been the questioD 01 the . . . . . . oIlimp1e.llot.lIy COD". . .t adaptivecontrol alpithmL By t1IiI we III8Ul aJaoritlulll which, for all iDitial I)'IteID aad aJaoritbm . . . C&1IIe the outputl of a liveD IiDear system to _ track a daind output IICIueDCet aDd achieve this with a bounded-iDput
with
rath. pnera1 UlUlDptioDI.
11le paper praenta a paeraI metbod of auIyIiI for dilcrete-time detenD.iDiltic adaptive control alpithmL The JDethod is iDUitrated by .tablilbiDa slobal converpDCe lor thJee simple alpithmL For clarity of prIMIltatioD, we shall first treat a simple IiDIIo-iDput siqIe-output alpitbm in detail. The reaults wiD then be exteaded to siqIo-input aiqIe-output aJaoritlulll iDdudiDa thole hued on recunive least squarea. Piully, die exteDIioD to DlaltipJe..iDput multiple-output systems will be prIMIlteeL SiDce the results in this paper were prIMIltecl a Dumber of other authon [16]-[18] have presented related resultl lor diacrete-time determiDiltic adaptive control alpitbms..
IeqUeDCL
11lere is a CODIidcrable amoUDt of literature OD continuous-time determiDiltic adaptive control alaoritbma [IJ. However, it is 0DIyrecendy that aJobal stability aacl conv-aeace of these aJaoritlulll bas been studied UDcler a-aeraI uaumptioDl. Much interest wa paerated by the iDDovative CODfipration pI'OpC)Ied. by Moaopoli (2) whereby the feedback piDI were directly .timated and aD aupleDted error sipal aDd awdlWy iilput aipaII were introduced to avoid the use of pure dif· fereDtiaton ill the aJaoritbm. UDfortunate1y, as pointed out in [3J the 8I'JUIIleD.tI livea ill (2) CODCeI"DiDa stability are incomplete. New proofs for related aJaorithma have recently appeared (4], (S]. In (4] Narendra and. VaiavaDi treat the cue where the difference in orders between the DUIUI'&tor aad deDomiDator 01 the system traDller function (relative etearee) is laa thaD 01' equal to two. In (5], Feuer and Morse propose a solution for paeralliDear systems without coDltrainta OD the relative dep'ee. The alpithma ill [5] use the auplented error concept and auxiliuy iDputi U in [2J. The Feuer and Morse result seems to be the moat paeral to date for siDale-input siqle-output continuous-time syatau. Howevw, these results are teclmicaJ1y involved and caDDot be directly applied to the dilcrete-time case. 11lere hal also been interelt in'dilcrete-time adaptive control for both the determiDiatic ADd Itocbutic cue. This ana baa particular relevance ill view of the iDcreuiDa ute of ctiaital tecJmolo&y in control applications
0""
II.
PaOBUDlI STATBMENT
In tbiI paper we shall be coacemed with the adaptive control of liDear time-iavariant fiDite-dimeDlioDal systemI haviq the lollowina state apace repreaentation: X(I+ 1)-Ax(I)+ B,,(I):
1(1)- Cx(t)
x(O)-xo
(2.1)
(2.2)
wha'e X(I), u(t), y(l) are the 11X 1 state vector, rX I input vector, and m x 1 output vector, respectively. A staDdard result is that the system (2.1), (2.2) can also be represented ill matrix IraetioD, or ARMA, form u
(6), (7). ldUDI [8], [9J bU proposed a pneral techDique for auaIyziDa conver· of diacrete-time Itocbutic adaptive alaoritbms.. However, in this auIyIiI a q..aioa which is yet to be reaoIved ccmcems the boUDdednesa of the .,... vuiabl-. For OM particular aJaorithm [10], it baa been arpecI iD [II] that the alpiduD poIIeII the property that the sample
aence
(2.3)
with appropriate initial conditioDS. In (2.3), A(q-I), BU(q-l) (;1,··· ,m;j-l,··· ,r) denote scalar polynomials in the unit delay opera-
q-'"
tor q-l aDd the 'acton represent pure time delays. Note that it is not UIUIIled tbat the system (2.1), (2.2) is completely controllable or completely observable, DOl' is it IIIUIDeCl that (2.3) is imclucible. The system will be required, however, to satisfy the conditioDa of Lemma 3.2. It is UIUIDed that the coefficientsin the matrix. A, B, C in (2.1), (2.2) are UDbown and that the state %(t) is Dot directly m.surable. A feedback coatrollaw is to be dcaiped to ltabilize the system and to
caue the outpu~ (y(I», to track a liven reference sequence {y·(I)}. Reprinted from IEEE Transactions on Automatic Control, Vol. AC-25 , 1980, pp. 449-456. 487
Specifically, we requireye'l and u(t) to be bounded uniformly in lim y~(I)-YI·{t)-O
i-I,-·· ,me
LentmtJ 3.2: For 1M syJtem (2.3) with r - m, and subject to
I, and
(2.4)
zdu-dIBII{Z)
' ....00
det
UI. KEY Tl!cHNlCAL LBMMAS Our aualysil of discrete-timemultivariable adaptive control a1p'ithma will be buecI OD the loIlowiDa technical results. UmmtI 3.1: 11 lim '-+00
: [
s(/)2 -0 bl(I)+~(t)cr(/)Tcr(')
(3.1)
for
(3.6)
Z~I-~B"'I(Z)
Izi < 1 WMn 14- I <j<", min tlu
i-I,··
·,m,
if max 1)',(t+4)1-~
(3.7)
O
w1Mre (b.(I», (~I», IIIttI (aI(/)} tire retIl XG1tIr ~ tIIItl (CJ(t)} i.r II 1WII Jl-tJ«IfJr .....,.eei . . Alb}«1 10
'11M ,.,. uin COfUItUttI m,. "4 whicll an ~ of T witA 0
1) fIIIi/DnII~ CtJIIdltlort
(3.2)
O
for aliI > 0 tIIIIl 2) . . . ~ COIIIlltitIII
lIer(t)H < C, + C2 max I.r(1')1
(3.3)
0<.,.<,
wIleN 0 < CI < 00, 0 < C2 < co, il follow8 tMt
(3.4)
lim 1(1)-0 ' ....00
tIItIl {UCJ(I)JI} i8 bouItdttd. Proof: II (I(/)} ill a bounded sequence, thea by (3.3) (IIcr(1)1I) is a bouDdecIleqUeDCe. Thea by (3.2) and (3.1) it follows that
hoof: The rcault ilataDdard ad simply follows from the fact that (3.6) eaaunI that the system hu • stable inverse. 0 III the remainder of the paper thae reaul. will be used to prove Blobal CODV...... of a Dumber of adaptive control aJaorithma. Sections IV-VII will be concemed with adaptive control of siqle-input ~ output ayatemI. SectiODI vm and IX will extend these results to the mul1iple-iDput multip~utput cue.. IV. SINGU-1NPtrr SlNGu-otrrPur SYSTBMS
It is well known that for the siD&le-input single-output (SISO)case, the system output of (2.1), (2.2) can be described by
lim 1(/)-0. 1-+00
Now IIIUJIle that (.t(t)} is unbounded. It follows that there exists a lubiequeDce (I,,) such that
lor(',.)1- 00 1,.-+00 lim
(4.1)
whero (u(t)}. (yet)} denote the input and output sequences, respeolively, mel A(q-I). B(q-I) are polynomial functiODS of the UDit delay operator q-l,
ad
I.r< t)1 < '8(,-)1
for I
< I".
..4(q-I)-I+G,9- 1 + ... +1I,.q-"
(3.S)
B(q-I)-bo+b1q-I+ ..• +b",q-"';
Now aloDa the aublequenee {t,,}
I
I
[ b.(..>+ ~1.)_( t,.)TG( '.) ]
18(1..)1
>
.r(t.)
1/'1
[K
d repJaeJltl the system time delay. The initial conditions of (2.1) arc replaced by iDitial values of y(t), 0> t > -11, and u( t). - d > t» - d - m.. The loIlowiq 'UI1IIDptioDl will be made about the system. A&ru1It'tiM Set 4: a) d is Imown. b) AD upper bound for " and m is known. c) B(z) hal all zeros strictly outside the closed unit disk. (This is neceaary to easure that the control objective can be achieved with a bounded-input sequence.) We note thalt by successive substitution, (4.1) can be rewritten u
uaiDa (3.2)
+ KII_(',.)112] 1/2
lJ(t,,)1
> x 1/ 2 + X I / 2 Ucr(t,.) fI :>
18(1,,)1
gl/2+ KI/, C. + C 1.r( t,,)1] 2
uaiDa (3.3) and (3.5).
Hence, lim ,--+ao
I[b.(I,,)+~(I,,)a(t,,):rcr(tll) ]1/2 I> .r(tJ
bo+O..
y(t+ d)-a(q- J)Y(/) + /l(q-I)u(t)
_1_ >0 K'/2('2
(4.2)
where
but this contradicta (3.1) IIld heDce the assumption that (8(/)} is unbounded it falae ad the remit follows. 0 ID order to 1110 this lemma ill proviDa pobal CODveqence of adaptive control alpithma it will be necessary to verify (3.1) (with .r(t) interpre.ted u the trackiDa error) aDd to check that IlllUlDPtioDa (3.2) and (3.3)
As previously stated, the control objective is to achieve
are satiafied. The nat lemma will be used to verify that the linear boundedDCSI conditioD (33) is .tiIfied by an important e.... of linear timo-invariant ayatemI. This c1ua comspoDda to tho. tiDear timo-invariaDt systemI for which the control objective (2.4) CUl be achieved with a bounded-input sequence aDd lor which the trackiDa error caD be reduced to zero if the
lim [Y(I)- y·(t)J-O
(4.3)
' ....00
where (y·(t)} is a reference sequence. It is assumed that (y·(t)} is kIlowD II priori and that
systaD parameten are known.
(4.4) 488
V.. SISO PROJECTION ALGoRITHM I Let '. be the vector of system parameters (dimensionp - n + m + d).
for all values of tp(1 - d) provided 0 <0(/) < 2.. This is satisfied by definition (5.8). Then, since 111(/)112 is a bounded nonincreuing function it converges. Setting
(S.I)
lben (4..2) can be written
(I). - ,,(1- d)T;(/-l) [and DOtin. that
(5.2)
y(t+d)_cp(t)T,O
tI(I)[ -2+11(1)
,,(I-d)T,(I-d)
]
[1 +f)(t-d)7'cp(t-d»)
where ,<1)7'-(y(t),·· . ..,(/-11+ 1),"(/),·· · ,11(/- m- d+ 1».
(5.3)
Now doIiDe the output trackiDI error as ~I + d)
'0 -
y.( 1 + d).
(5.4)
and hence
chOOliDa (II( t)} to satisfy
it is evideat that the tractiDI error is identically zero. However, since is UDbaowD, we replace (5.5) by the foUowinl adaptive aJaoritbm:
if1-
(5.14)
(5.5)
cp(/)T'O·y·(/+d)
i(t) -
is boUDded away from zero, with a(/) defmed u in (S.8») we conclude, from (5.12), that
- y(l+ d) - ,·(1 + d) • cp( I) T
By
(S.13)
'0
Now using (5.13)and (5.11) it follows that «(I). -4p(t-d) rz I(/-d)-
I) + a( t).(1- d)[ 1+ .p(t - d) Tq>( 1- d)]-1
d-I ~ ~ I-I
11(1- i)
A
· [y(t)- cp(t-d) T.1(/-1)
(S.6)
]
.
f(1-d{!
where i(t) is a }I-vector of reaIs depeadiDa on an initial vector i(O) and Then uaina (5.4) and (S.7) we have that ony(t'). 0<.,<1. u(1'), O<.,
a(/)-1
(5.8)
e(l)
(I)
[ 1+ cp(t - d) T«p( 1- d)] 111
This choice of pin constant prevents the computed coefficient of u(I) in (S.7) beiDa zwo. We also remark that the purpose 01 the coeffICient 1 in the term [I +f(t-d)Tf(/-d)]-1 of (S.6) is solely to avoid diviaioD by zero. ADy positive c:oDltaDt could be used in place of the 1. Apart from the llbove modificatioD, the al.orithm (5.6) is an orthogolUll projection of 1(/-1) onto the hypersurfacey(/)-tp(t-d)T,.O. In the auJyaiaof this alJorithm, we wiD rust show that the Euclidean Dorm of the vector i(,) '0 is a noniDcreainl fUDCtion alODI the trajectories 01 the a1lOrithm. This leada to a characterization of the limitin. behavior of the alaorithm which will allow us to use Lemma 3.1 to establish poba1 coDveraence.
i-I
[I +4fJ(t-d)T.,(t-d)]1/2 ,,(I-d-i)
( t - i)
•
. (5.16)
[I +cp(t-d- i)Trp(t- d- ;)]1/2 Now by the Cauchy-Schwarz inequality and the fact that la(t)1 < 2 0<
I
0(/- i)tp(t- d)T
fP(t-d- ;)
[I + '<,- tl)TCP(t_d)]1/2 [1+.(t- d-i)T.,(t-d- i)]1/2
(5.9)
f(1-1) · [1 +tp(/-d- i).ct-d- ;»)1/2
7',-
Proof:
-' d)T y\/-
[1 +cp(t-d-i)T'
lAmmJI 5.1: Along 1'- 801ution.r of (5.6), (5.7),
I--.CO
4-1
- ~ 0(/-;)
1<,)-
lim
(5.15)
Hence,
otherwise where y is a constant in the interval (e:.2-e:). y.".1 and 0<<<< I.
- y
f(l-i).
[I +.ct-d-i)TqJ(t-d-i)]
(S.7)
tp(1)Ti(t)-Y·(I+d)
tit) 1(/) -0. [1+9(t)T4p(t)]l/2
I
(5.10)
UIiDa the dermitiOD of i(t), (S.6) may be rewritten as
I
I
2(1- i)
< [I +cp(I-d- j)Tcp(I_d_ i»)1/2 • Then using (5.14) it follows that
i(t) - i(/-l) - a(t)4p{t- d)[l + q>(t- d)T.(1 - d)J-l
·,,(t-d)Ti(t-l). (S.lt)
Hence, ,,(t-d-;)
";(1)11
2-
Ui(t-l)tf2.tJ(t)[ -2+11(1)
,,(I-d)TCP(I-d)
[1+ .,( t - d) T, ( t - d)]
[1 +,,(t - d - i) Tq>( t - d _ ;)] 1/2
]
. i(1-I)Tf(I-d)f!I-d)Ti(,-I) <0 (5.12) [I +.p(/-d)T,,(t-d)]
.
489
(1-
i)
[I + ,,(1- d- ;)Ttp(t_ d- i)t / 2
1-
0
for ;-1,2, .. ' td-1.
(S.l7)
Hence, usiq (5.16), (S.I7), and (5.14)
where
-0.
lim
e(l)
1-+00
[I +e,(/-d)Tf'{/-d)]1/2
cp(t)T-( -y(t)_·· -y(t-n+ 1), - u(/-l)···
(S.18)
-u(t-m-d+ l),y·(t+d»
Thia eatabliahea (S.IO). 0 Note that we do Dot prove, or claim, that i) converps to However, the weaker condition (S.lO) will be sufficient to establilb CODV"eoce of the trackiDa error to zero and boundedneu of thesystem iDputiand output&. These are the prime properties of concern in adaptive control 77Ieorwm $.1: Subj«1 to AUIIIPf'liOlU 4a)-e); if tIIII iIlgoritlun (5.6), (5.1) II f/IIIIIW to lite 1Y.J/eIn (2.1). (2.2) (,. m-l), tMn (y(/» tIItd (u(t» tIIW boIwI64 twl
(ft!'-
'0-
",T ' P 'I'··· '~+d-I' R' o _ -". ••• ,«,,-It PoI ) .
It is evident that the traekiD& error can be made identicaUy zero by ehooaina (u(t)l such that (6.3) However, since '0 is UDknown, the control law will be reeursively estimated. The foUowina adaptive aJaoritbm will be considered:
(5.19)
lim [Y(/)- y·(t)]-O.
t-.ao
foralll
1
(6.4)
"(t) _.,(/)Ti(/)
(6.S)
Po
hotI: Lemma 5.1 CDIlII'eI that condition (3.1) of Lemma 3.1 is aatiafiecl, with 1(/) - e(/). the traekiq error, and 0(1)- cp(1- d) the vector defiD.ed by (5.3). Also 6.(/)-1, and 62( / ) -I. It foBows that the UDiform bouDdedaell CODdition (3.2) is satisfied. AasumptioD 4c) ad Lemma 3.2 ensure that Ju(k-tI)f<m,+m. max IY(1')1
i(t)-i(t-d)- ~.,(t-d)[l+cp(t- d)T'
where A, is a flXecJ CODStaq,t and i(t) is a p-vector of reals depending on d initial values 1(0),••• ,I(d-I) and on y(.,), 0<., cr, u(T). 0<1" < t - d - 1 via (6.4). Note that (6.4) is aetuaI1y d separate recunioDl interlaced. (It bas recently been pointed out (18] that it is also possible to analyze a _pe recursion without iaterlaciDa lIIiq a different technique but the same aeneral priacipals.) The aualysia of projection a1aoritbm n bas mudl in common with the analysis for projection alptbm I. We wiD therefore merely state the analop of Lemma S.l and Theorem S.I for the algorithm (6.4), (6.S). ummtJ 6.1: DejiM
TberefOl'et lIIiq (5.3)
lIt'
but
;(/)- i(/) -
Hence,
_
1<.,<1
1<.,<1
..
0
0
and it follows that the linear bounc1edness condition (3.3) is also satisfied. The reaultDOW foDows by Lemma3.1 and by noq that boundedness of {II,(/)II) eIIIURI boundedDess of {IY< '>I} and {IN(t)I}· 0 VI. SISO
(6.6)
TIJsJ IIlI(t+d)U1-1I1(/)lfz< 0 aJOIIg witla tlw lOhdioru of (6.4) and (6.5) tmd
lI,
'o.
PaOJBC1"lON ALoOIU1llM
IJo Ilo
0< -:- <2.
II
o
Lemma 6.1 is used to prove Theorem. 6.1 in the same lIl81U1er that
In this seetioo we present an algorithm differing from that of Section V in that the cootrollaw is estimated dirccdy. This approach is adopted in [5), and essentially involves the factorization of fJo from (4.2). A related procedure is used in the self-tuDing replator (10) where it is assumed that the value of IJo is mown. AD advaDtqe of the aJaoritbm is that the precautions required in Section V to avoiddivilioo by zero in the calculationof the input are DO lonler neceaary. However, a disadvantap is that additionalinformation is required; specificaDy, we need to know the sign of fJo and an upper bouDd lor ita mapitude. PactoriDl flo from (4.2) yields
Lemma S.I is used to establish Theorem S.1. We obtain the foUowina theorem in this way. '1'heomn 6.1: Subject to AUIIIPf'tioIu 4a)-c) and for 0
(6.6)
' ......00
We Dote that the condition 0
y(t+d)-JIo(crOy(t) + ... +c(_ly(t-n+ 1)+11(/)
VII.
+ lJ;u(t-I)··· + fJ:,,+d_I"(t-m-d+ 1». (6.1)
ADAPTIVE CoNn.OL USING
R.Ect1RsJvE LBAST SQUARES
The wide-spread use of recursive least squares in parameter estimation
Let
indicatesthat it may find applicationin the adaptivecontrol context We treat the unit delay cue 01 algorithm I with the projection (5.6) replaced
e(/+ 4)-y(t+d) - Y·(I+d)
by recursive least squares.
-Po(lI(t) + crQy(t)·•• ~_Iy(t-II + 1)+ 1I1 ( t - l)
The adaptive control algorithm then becomes
11
;(t)-i(t-l)+
~ ... + 1I.:.+4-1"(t- m-d+ 1)- ~y.(t+d») -1Jo( U(/)- .,(1)7'0)
a(I)p(t-2).(t-l)
[1 +a(t)cp(/-l)Tp(t-2)cp(t-l)] (6.2)
490
[ y( t) - cp( t - 1) Ti( t - 1) ]
(7.1)
P(t)-[l-
P(t-I),,(t)f(I)Ta(t+ 1) ]P(t-l) 1 + CP(t)TP(t-l)cp(t)a(t+ 1)
.(1)Ti(t>.y·(t+ I)
(7.2)
Now
(7.3)
1,(t)12 [I +a(t)cp(t-l)Tp(t-2)cp(t-I)]
where p(t) is . , Xp matrix and the recanion (1.2) is 8II1UIled to be iDitialized with p( -1) equal to any positive defiDite matrix. The ICI1ar G(t) in (7.1), (1.2) playa the same role u in SectionV and is required 0D1y to avoid the noapDeric pOllibility of divisioD by zero in (7.3) wh. evaluatiq II{t). Hence, G(I)-1 will almoIt always wort IDd for tl(1)-1 we observe that (1.1) aDd (1.2) are the .tuldard recursive leut IqlIU'eI aJaorithm. The sequence {tI{t)} may be choIeD u in (5.8). u.. 1.1: AItmg with 1M 8OIutItH&r of (1.1), (1.2), (1.3) , . /tIItctiolt Y(1)-i(t)TP(t-l)-li(t) U II bount/ed, IIOIIMIt'tiN, PJOIIinc1W&filfr jwIctioII tIIIIl lim
'-.00
(7.12)
Hence, from (1.11) and (7.12) -0.
(7.13)
''''00 [1 +2(Up(t-2»))U«p(t-l)U 2 ]
TbiI will be recopized as beiDa condition (3.1) with .f(1)- ce(t), b.(I)1, and bi') - 2(A..,J.p(t - 2)D. To establish the UDiform boundednesa condition (3.2) we proceed as foUows. From (7.2) and the matrix inveniOD lemJDa,
pet) -1- P(I-l)-I + a(t)cp(t)9(/)T.
(7.4)
Hence,
~ i(t)-i(/)-,o-
Proof:
2
1!<1)1
lim
r-
.(1-1) '(I-I) -0 [1 +a(t)cp(I-l)7'P(t-2)CP(I-I)]1/2
;>
le(t)12 [I +2U.(I-l)1I2
XTp(t)-l x >xTp(t-l)x
Prom (7.1),(S.2),
• ... a(I)P(I-2)tp(I-l)cp(t-l)T;(I-l) 1(1)-1(1-1).
[I +tI(t)CP(I-I)Tp(,-2).(t-I)]
>A.ua[p(/-l)-ll llx Il 2
(7.S)
for each xEA'.
(7.14)
Now choose x as the eigenvector corresponding to the minimum eigenvalue of [P( t) - I ~
Then uain& (1.2),
Then from (7.14)
A.m[P(I)-I] >A.m(P(t-l) -I].
;(I)_P(I-I)P(I-2)-I;(/-I).
So AaJp(t) -I] is a nondecreasing function bounded below by
Thus,
A-JP(-I)-I]-X- I >0. P(t-l)-I;(I)-P(t-2)-Ii(,-l).
Now defllliD& V(I)
Hence from (7.13), o
(7.6)
i(, -1)TP(t -1);(1 -I) we have
U
Y(t)- Y(t-l)_;(t)Tp(l-l)-I;(t)_;(t-I)Tp (t - 2) - Ii(t - I ).
VIII.
MULTIPU-1NPur
MULTD'LB-<>urPuT SYSTEMS
For the case m- r» 1, the system (2.1), (2.2) can be represented in the form
UIiq (1.6)
Y(t)- Y(t-I)- [i(I)-i(t-l)]TP(t-2)-1i(t-l)
_ -a(t) i(t-I('PC t- l7
(7.1)
where we have used (1.5). It is clear from (/.7) that V(I) is a bo~ DODDeptive, DODiDcreuina function and hence converp8. Thus. from fl.1), aDd since a(t) is bounded away froJDzero, •
•• u(t)
r-
T
lim 1(1-1) .(1-1),(1-1) '(1-1) -0.
,-.00 [I + tI(t)4p(t-l)Tp(t-2)cp(t-l)]
where Ak(q - I) and BIc/(q -I) 1
Hence, lim
e(/)
,.....ao
[1 +a(t)cp(t-l)Tp ( I - 2)cp(t - I»)' / 2
-0
(7.8)
where
and
4- 1<)<". min (dil}'
(7.9)
o
lim [.,,(1)-)'-(1»)-0.
I-.ao
(7.10)
Yl(t~dl) ] _ [al{~-I) [
,.(1+ tJ...)
~ ]y(t) a".(q-l)
0
~1I(q-l)
Proof.' From Lemma 7.1
e(l) -0. ''''00 [1 +a(t)cp(t-l)7'P(t-2)9(t-l»)1/1
i-I,···,nI,
(8.1) can be written
1'II«Jmn 1.1: Subj«t to Amlnpliolu 4tI)-c) if I. aI,orilltm (7.1), (1.2), (1.3) i8 ."Ii. to 1M IYstem (2.1~ (2.2) (r- m-l), IMn {Yet)}, {fI(I)} are botI1ItJ«J and
lim
(8.1)
q -tC-.B...... (q-l)
+
(7.11)
: [ I3lftl(q-l)
491
...
where ~(t) is !-P, (- ~ + m(m, +~» vector of reals depending upon an initial vector ',(0) andy,(.,.), 0<"
where
and ",(q-l)-Ii(q-I)BU(q-l)q44-~.
(8.3)
a
It can be seeD that (8.2) consists of a set of multiple-input siDIle-output (MISO) systems haviDa a common input vector. The foUowiq asaumptioDl will be made about the system.
11(0). 2) For I:> 1 the proceduJe of Lemma 9.1 below guarantees the 101vability01 the algorithm equatiou for II(I). Um1IIII 9.1: I" ortIttr tItat 1M IIttII1U t1/ cM/fidm/8 of u(1) in (9.5) i.r 1tOfLfbIIultv for Gill:> I U u III/fIdett for a(t) I" (9.4) 10 be cIIomt 4r jollowl:
A""".tltHt Set B: a) d1, - · · ,tl. are known. b) An upper bound for the order of each polynomial in (8.2) is knoWD. c) The system (8.1) satisfies condition ,lIIu-.t'Bu(z)
det[ z4a-4l(.B..,(z)
.. •
•.•
zd••-tllB.",(Z)] .... o '4.- ....B....(z)
rr:
f
for Izi < 1.
where
lim [y,(/)- yt(t»)-O
i-l,···,m
eipavalue of - R -1(~-l)Y(/) with (9.6)
Condition Ie) daervea commenL Firat, for ally output colDpOnentYI' 1
O<e< 1 and G(/)-l is not an
and
Y(/) ~
ro.,··· '011I]'
(9.7)
'. in (9.6) is the vector of coefficients of II{t) in ',{t - 1), that is, T;-
S,',(t-I)
(9.8)
when (9.9)
", in (9.6) is the vector of cbaDps in the coefficient of 11(/), that is,
where ""(/) is a reference sequence. It is assumed that each (yt(t)} is known a priori and that ly,·(/)f <ml < co for aU I, ;~ 1,- _. ~m.
0,- Sir CPI(t-~)(l + 9,(/- ~)TcpI(t-~»-I
-1»)].
·(,,(/)- ,;(t-~)T'I(1
(9.10)
IX. MIMO ADAPrIVB CoNTllOL Proof: The proof will be by induction and we first observe that This section wiD be concerned with the multivariable versions of the from (9.4) and (9.6)-(9.10) adaptive control algorithms introduced in Sections V and VI. The multivariable version of the allOritbmof SectionVII also foDows analoR( I) - R( 1- I) + a( I) V( I). (9.11)
gousty. A. MIMO Proj«tioft Algorithm I Let
'&
be
the
vector
of
parameters
in
(1,(9 -.)
and
Then: i) R(O) is nODlinpiar by the initial choice of ;~O), ; - I, ... t m. ii) Assume R(I-l) is nODSinplar. Then from (9.11), usinJ a(t)+Ot
fJll(q-I). • • fJ,.(q-I). Then (8.2) may be written ill the form
",(1+ ~)- cp,(t)T,&.
1<; <m
detR(t)-[detR(t-l)][det(l+a(t)R(t-I)-1 V(t»] -ldetR(/-l»)(Q(/»"'[ de~
(9.1)
where
-0
if and ODly if
~/»)I+ R(/-I)-I V(t)]
at/) is an eigenvalue of
-R(t-l)-I yet).
Defme ·1('+~)-YI(t+~)-,t(t+4)
_~(t)T'~_Y:(/+~).
(9.2)
It is evident that the trackiD& error may be made identically zero if it is possible to choose the vector u(/) to satisfy
1
(9.3)
Obviously (9.3) is a set of simultaneous equations in ..(t). Now the matrix multiplying 11(/) is nonsinplar since in (8.3) det(diaIJi(z»-t at z-O and Assumption Ie) holds. ~ence a unique solution "(1) of (9.3) exists at the true parameter value IJ.
But the definition of a(/) euures a(t)-I is not an ei&envalue of - R(t-l)-IV(t), hence R(/) is nonsingular.. However, by i) R(O) is nODSingular and it foDows by induction R(t), 1 ;> 0 is noDSiDplar. 0 We note that the above choice of a(t) hu been included for technical completeness and that a( I) - I will almost always work since it is a Dongeneric occurrence for t to be an eigenvalue of - R(/-l)-IY(/). Also since - R(1- 1)- I Y( t) bas only a fmite Dumber of eigenvalues it is always possible to find an a( t) to satisfy the lemma by computation of the eigenvalues of R( 1- 1)- I Y(I). 'I'Morem 9.1.· Subject to Ammption.r &I)-c) if tlte algorithm (9.4), (9.5) if applied 10 tlw system (2.1), (2.2) with r=m, then (y(/)} and (II(I)} an bou1ukdand
Consider the foUowilll" adaptive algorithm:
lim 1.1;( t) - y/.( 1)1-0;
i,(t)-~(t-l) +a(t).,(t- d,)[ 1+ CPi(t~t4)T,,(t- d,)]-I ·(YI(I)-CPi(t-~) T"';(t-I) )
cp,(t)T~(t)_,,"{/+ ~),
1 «t <m
1
' ....00
Proof: Usins Lemma (5.1) for each i, we have
(9.4) (9.5)
492
lim ' .... 00
.
[1 + 41>;(1 -
e.(t) f
~)Tcp;(1 -~) ]
1/2
-0
•
where d-max(d1t • • .Jt(J. i(t)T is aD ntXn' matrix of reala depeactiq II iDitial matrices I(f)T, 1 <;
The proof DOW followa that of Theorem. 5.1, except in the case that the vector )'(1) is unboundecL In this case there exists a subsequence (I,,) such that lim II,(/..)U- co
OD
I.. ~CIO
and IY/(I + ~)I < IYJ('-)('- + ~,-~I
for some I
gT + K - KTg i.r poritiw (9../5):
It thea follOWl by Lemma 3.2 that there exist coDltanti O
11.,('..)"
I
a) trace[ i(I+d)Ti(t+d)]
b)
lim t....
SiDce m is fiDifet there ail.. a further subsequence {t,,} of the sub{III} .uch that
MqueDC8
11.,(,.,)11
dI/iItit-, t"', tIlotw
1M tTtIjectorla of (9.14),
-trace[ i(t)Ti(t)] <0.
e,(t+~) -0, [I +,
I
Proof: a) We caD rewrite (9.13)
for at least one i, 1
UIiD& (9.1S) u
&lid (1,(,.,+4)} is ubouDded. The remaiDder of the proof then followl dlat 01 Tbeonm 5.1 where we note that
o
Ie,(I)I-IY,( t) - yt( 1)1·
Heace, from (9.14) B. NINO Pro}«tioIJ Al,orltllm II
z-o.
From (8.2). (8.3) the fact that 1'j(z)-1 for i-I,· .. ,lit uad by AuumptioD Ie) we can factor out the nonsiDgular matrix r o (-(IJ,(O)D
and
givinl
trace( i(1+tI)Ti(t+ d) -trace(i(I)Ti(t»
_-trace[(KT+K-KTK where
f(t)T,(t)
)
(I +.
.(i(t)T.,(t)[1 +.,(t)T.,(t)r l.ct)Ti(t» ]
<0 b)
aDd tim
1-+00
~
if gT + K- KTK is positive dcfiDite.
ill the proof of Lemma (7.1) it foDows that
traee(xr +K-XTK [I +.
Define
SiDce gT + K - KTX is positive defmite then
- : ]_[JlI(t;d ]_[JlW;d [ e",(t+ tI..) y",(t+tI..> y':(/+ t(..> l
l
) ]
)
or
yf(t+dt) -ro u(t)+C(q-l)y(t)+D(q-l)u(t-l)-r;1 : [ y':(t+d,..) [
- r 0< u( t) -
'OT.<
]1
tim
''''00
roll "." ][1 ·
d
1 +9
-..(t+t.(.J
0
Thia iJDpliea that (9.17) holds.
t»
(9.13)
UIiDa Lemma 9.2 and foDowiDa the proof of Theorem 9.1 we have the
t.
where 'oT ia an 111 X,,' matrix whose itb row coataiDa the parameten folJowiDl. '1'II«nm 9.2: Subject to AulurptioJu &I)-c), and K T + K - KTK po.r;from the ith fOWl of C(q-I). D(q-I), and E-r f(/) is an ,,'X 1 tiw definite, if Q/pritltm (9.14). (9.15) ;" appli«l to 1M .rystem (2.1), vectorcoataiDiDa the appropriate delayed veniou of y(t). u(t -I~ and (2.2) willi r- m, ,hDJ 1_ fJ«10I'.J y(I) awl II(t) tW bormd«IlIIId y·(t):
o'.
4p(1)T_( _ y(t)T, _ y(t- J)T,•.• , - fI(I_I)T,
lim IY,(t) - y;e(t)I-O,
-U(t-2)T, ... ,yr(t+d.),··· ,y,:(t+tJ..».
ADaIoaoUlly to Section VI we introduce the followina adaptive control
X.
alpithm, el(t+d.) ] i(t+d)T _i(i)T -p
:
[
lIe t) - i( If",,( t)
I «t <m.
1-+00
[l+'P(t)TfJ(t)]-I.(I)T (9.14)
_.(1+ d".) (9.1S)
o
NON'LINBAIl SYSTBMS
Altho. the aaalysia in the paperbaa been carried out for deterministic Jinear systems, it is clear that it could be readily exteDded to certaiD cIuIea of nODlinear systems of bowD form. The essential points are the form 01 (5.2) or (6.2), and the IiDear bound CODdiIiOD (3.3). The latter point would indicate that systems with cone bounded nODlinearitiea would satisfy the conditioD&.
493
xr,
CoNCLUSION
The paper hu analyzed a pneral class of discre.time adaptive control alaorithma and has abown that, under suitable conditions, they will bealobaI1y convcqent- The alaorithma have a very simple structure and are applicable to both sinl1e-input siD&10-0utput and multiple-iDput multiplo-output systems with arbitrary time delays provided only that a ,table coatrol law exists to achieve zero trackiD& error. The results resolve a loDa staDdiDg question in adaptive control reprdina the existence of simple. atobally converpnt adaptive algorithms.
(1) I. D. LaDdau. "A lurvey of model refa'OllCC adaptive tec1miqua-Tbeory aacl appticaltall." A........, vol pp. 353-319, 1974(1) R.. V. MoaopoIi. "Model ref..- adaptive COIltrol with all a teeI error Ii lEU TMu. AI....,. e-tr., vol. ACI', pp. 474-415, OcL 191.... (3) A. P , 8. R. BarmiIb, aDd A. S. Mone. NAD astable dyDUDic:al ayltem UIOCiatecl wi1b aodeI nllNIICe adaptive control," IEBB TPrIIIf. Coftt,•• vol AC23. pp. 499-5001' Jae 1971. (4J K. S. NaNDdra ucl L S. Valavui, -Slable adaptive COIltroUer cIeIipa-Direct coatrol.·,BU TNar. Alii....,. CfIIIIr.. vaL ACn. pp. S70-S83. A.... 1911. (5) A.'eur ad S. Mane, .. Adapti¥c QOIltrol of ...........t ~. tiMar .,......" IEEB n-t. A......,. C."..• wi. ACD, pp. 557-570. A.... 1971. (fi) G. A. DuIoDt ud P. L Bitaqer, --seu-hIIliq coatrol oI.litaaiUDl dioxictekiIa." IBEE n.u. Aw..... CMIr., 'VOl. A~Z3. pp. 532-531, A-. 1971. (7) K. J. Aatrim, U. IIoriIIoa, L liUDIt aDd B. WitteDmUk, -rbeory ud appticatiou of . . t replaton." A........, 19. pp. 457-476. 1977. (I) L li "AuIyIiI 01 recuniw .toebutic aJaoritJu.," IEEE TIwu. A...... C.",.., voL Aen. pp. '51-575. A-. 1m. (9) -,"OD poei,," nal traDII. hactio. aDd tIM COIlwrpllCe of IOID8 recursive . . . . .,.. 1BEE TIwu. C.."., wL ACn, pp. 539-551, A~ 1m. (10) K. J. A.aim ud B. Wkteamark. -oa leIf-tuaiAa replatan," A ~ vol. 9. lIP- 195-199, 1973. 1'1) L lsillDl ad B. WiUlllUDal'k. "OD • ltabiIiziDa property 01 adaptive replaton," ".,.,.II'AC rtIMtl/blltJlt. TbiIiIi. u.s.s.R., 1976. (11] 8. I!pftb. -A uaiIied approIdl to aaodeI rei..... adaptiYe.,..... aIlcllelf tuaiAa replatan," Dep. Automat. CoDtr., LuaclIDat. TedmoL Tech. Rep., Dec. 1m. [131 T.t......... L V. MoDopoIi, "DiIcrete IDOcIeI refereace adaptive CODtrol with aD ........ error 1ipaI," AIIItIfIIIIIktJ. vol 13. lIP- 507-517. Sept. 1m. (t4) J. L WiIkaI, 5MbIIUy,....". of D,,..,.aI S)..... New York: 1970. (IS) G. Co 000cIwiD. P. J. Ramadp. ucl P. Eo CaiDeI, MDiIcreCe time .todautic adaptive ooatrol."SIAM J. c.u,. OptiMiz•• to be pubIiIbecL (lfi) A. S. Mone, -cHoba1atability of parameter adaptive OODtrol.,...... Yale UDiv.• S A IS Rep. 1I02L Mar. 1979. (17] x.. S. Nueadra aDd V-H. LiB. -Stable ctilcrete adaptive ccmtrol," Yale UDiv., S • IS Rep. 79011' Mar. 1979. [II) B.1!pftb. -Stability of modcI refcreace adaptive and IeJf tUDiq rqulaton," Cep. Automat. CaatrOf Lad last. TecbAo1., TedL Rep., Dec. 1978.
to.
A.....,.
,A.,..,.
sy"". __
494
Feedback and Optimal Sensitivity: Model Reference Transformations, Multiplicative Seminorms, and Approximate Inverses G.ZAMES
T
HIS paper is credited with the introduction of the subject of H 00 control theory, a topic at the center of control research during the 1980s and 1990s. The symbol H'" refers to the Hardy space of bounded analytic functions. For continuous-time systems, this corresponds to the space of stable, proper transfer functions (analytic and bounded in the right half-plane). The H'" norm of a stable transfer function is its maximum modulus, that is, the peak value on the Bode magnitude plot. Such peak values for closed-loop transfer functions had always been an important measure of performance and robustness in control. In fact, it defines the induced L2-norm, and its importance in control comes from this connection. The paper was motivated mainly by three considerations: 1. The classical approach to feedback design for single inputsingle output plants, in the style of Bode and others, takes place in the frequency domain: The open-loop frequency response is shaped by the addition of a compensator, such as a lead or lag compensator. Zames wondered whether such controllers can be induced by some precise frequencydomain optimality criterion. If so, then multivariable extensions might follow readily, whereas classical methods had not extended satisfactorily. 2. In the H 2 (equivalently, the LQG, or Wiener-Hopt) design formulation, the exogenous inputs (e.g., disturbances or sensor noise) are modeled as stationary signals with known stochastic characteristics. Zames suggested the need to be able to handle a different set of exogenous signals. He reasoned that a natural choice was a weighted ball in the space of finite energy time functions. This leads to an Hoo-type minimization problem. One can think of the H 2 approach as relatively optimistic. White noise scatters its energy over all frequencies. Zames's scheme is instead a worst case approach. The disturbance energy is allowed to be concentrated at the worst possible frequency. 3. Often, plant uncertainty is modeled in the frequency domain. Zames argued that robustness with respect to state models is artificial in many situations, such as where high-
frequency dynamics are neglected in obtaining a lower order model. The paper has two main contributions. First, the formulation of the problem of minimizing the H'" norm of the weighted closed-loop sensitivity function for a nominal plant; and second, a study of the ability of feedback (as in the case of the feedback amplifier, discussed in the paper of Black elsewhere in this volume) to reduce uncertainty. Of the two, the first sparked the most research. The literature on H 00 control is vast (the reference list includes twenty-two books, a sampling of merely the books treating the subject in various settings and at various levels), so a survey of the entire landscape is not possible. Research in the 1980s was devoted mainly to developing a variety of solutions to the pure mathematical problem of H'" optimization. A beautiful collaboration arose between control theorists and operator theorists, and the H'" sessions at the conferences of the day were charged with excitement. Surprisingly, it turned out that the best way to solve ROO optimization problems was to convert them to state-space form. In fact, a breakthrough occurred in 1988 with Doyle and Glover's state-space solution (the formulas for the controller were announced in [11], the full treatment with proofs appearing later in [7]). Their solution involved two Riccati equations, just as for the LQG problem, but in addition a remarkable coupling condition. The 1990s saw a number of interesting developments. The two-Riccati solution highlights the fact that the ROO problem can be posed as a two-person zero-sum differential game: the control signal as the minimizing player versus the disturbance signal as the maximizing player. Basar and Bernhard [1] developed H'" theory in this way and extended it to time varying, finite horizon, and nonlinear problems. A more recent trend is the formulation of the H 00 criterion as a linear matrix inequality, e.g., [3], leading to a convex optimization problem solvable numerically by interior-point methods. Since one of Zames's goals began with classical control, it is fitting that H'" control has filtered down from research journals into undergraduate textbooks, e.g., [2]. In an undergraduate course, the treatment usually involves three problems. In the
495
nominalperformance problem, a suitable closed-loop transfer function, such as the weighted sensitivity function, is optimized in H'" norm for a nominal plant. The closed-loop shape is altered via the weighting function, and this supersedes the older approach of open-loop shaping. In stability robustness, a perturbation model of the plant is assumed and a controller is sought that maximizes the size of the perturbation while achieving closed-loop stability. Both of these problems are amenable to analytical solution. The third problem is performance robustness, where closed-loop performance and stability are guaranteed for all plant perturbations not greater in size than a prespecified bound. This problem has remained intractable, though it initiated a great deal of study, such as that by Doyle [6] on the structured singular value. The importance of Zames's paper is that it breaks a trend, a trend away from stochastic control as a way of dealing with uncertainty, and it initiates the trend to give robustness the place in control that it deserves. REFERENCES
[1] T. BA~AR AND P. BERNHARD, HOO-Optimal Control and Related Minimax Design Problems: A Dynamic Game Approach, 2nd ed., Birkhauser (Boston), 1995. [2] P. BELANGER, Control Engineering: A Modern Approach, Saunders College Publishing (Philadelphia,PA), 1995. [3] S. BOYD, L. ELGHAOUI, E. PERON, AND V.BALAKRISHNAN, LinearMatrix Inequalities in Systemand ControlTheory, SIAM (CherryHill, NJ), 1994. [4] T. CHEN AND B.A. FRANCIS, Optimal Sampled-Data Control Systems, Springer-Verlag (New York), 1995. [5] R.F. CURTAIN AND H. ZWART, An Introduction to Infinite-Dimensional Linear SystemsTheory, Springer-Verlag (New York), 1995. [6] J.C. DOYLE, "Analysis of feedback systems with structured uncertainty," lEE Proceedings, Part D, 133:45-56,1982. [7] J.C. DOYLE, K. GLOVER, P.P. KHARGONEKAR, AND B.A. FRANCIS, "Statespace solutions to standard H 2 and H'" control problems," IEEE Trans.
Auto. Contr., AC-34:831-847, 1989. [8] r.c. DOYLE, B.A. FRANCIS AND A. TANNENBAUM, Feedback ControlTheory, Macmillan (New York), 1992. [9] A. FEINTUCH, Robust Control Theory in Hilbert Space, Springer-Verlag (New York), 1998. [10] B.A. FRANCIS, A Course in H oo Control Theory, Springer-Verlag (New York),1987. [11] K. GLOVER AND J.C. DOYLE, "State-spaceformulaefor all stabilizingcontrollers that satisfy an H'" norm bound and relations to risk sensitivity," Syst. Contr. Lett., 11:167-172,1988. [12] M. GREEN AND DJ.N. LIMEBEER, Linear Robust Control, Prentice Hall (EnglewoodCliffs, NJ), 1995. [13] B. HASSIBI, A.H. SAYED AND T.KAILATH, Indefinite-Quadratic Estimation and Control: A UnifiedApproachto H 2 and H'" Theories, SIAM (Cherry Hill, NJ), 1999. [14] lW. HELTON AND O. MERINO, Classical Control Using H oo Methods, SIAM (Philadelphia,PA), 1998. [15] R.A. HYDE, H oo Aerospace ControlDesign:A VSTOLFlightApplication, Springer-Verlag (New York), 1995. [16] H. KIMURA, Chain-Scattering Approach to H'" Control, Birkhauser (Boston), 1997. [17] J. MACIEJOWSKI, Multivariable Feedback Design, Addison-Wesley (Reading, MA), 1989. [18] D.C. McFARLANE AND K. GLOVER, RobustController Design UsingNormalizedCoprimeFactorPlantDescriptions, Springer-Verlag (New York), 1990. [19] M. MORARI AND E. ZAFIRIOU, Robust Process Control, Prentice Hall (EnglewoodCliffs, NJ), 1989. [20] D. MUSTAFA AND K. GLOVER, MinimumEntropy H oo Control, SpringerVerlag(New York), 1990. [21] M.A. PETERS AND P.A. IGLESIAS, Minimum Entropy Controlfor TimeVarying Systems, Birkhauser (Boston), 1997. [22] S. SKOGESTAD AND I. POSTLETHWAITE, Multivariable Feedback Control: Analysisand Design, Wiley (New York), 1996. [23] A.A. STOORVOGEL, The H'" ControlProblem: A State-Space Approach, Prentice Hall (EnglewoodCliffs, NJ), 1992. [24] M. VIDYASAGAR, Control System Synthesis: A Factorization Approach, MIT Press (Cambridge,MA), 1985. [25] K. ZHOU WITH lC. DOYLE, AND K. GLOVER, Robustand OptimalControl, Prentice Hall (EnglewoodCliffs, NJ), 1996.
B.A.F.
496
Feedback and Optimal Sensitivity: Model Reference Transformations, Multiplicative Seminorms, and Approximate Inverses G. ZAMES, FELLOW, IEEE
A bsl,al:l-lo this paper, the problem of sensitivity reduction by feedback is fonnulated as an optimization problem and separated from the problem of stabilization. Stable feedback schemes obtainable from a given plant are parameterized._ Salient properties of sensitivity reducing schemes are derived, and it is shown that plant uncertainty reduces the ability of feedback to reduce sensitivity. The theory is developed for input-output systems in a general setting of Banach algebras, and then specialized to a class of multi variable, timeinvariant systems characterized by n X n matrices of II~ frequency response functions, either with or without zeros in the right half-plane. The approach is based on the use of a weighted .fem;norm on the algeb...a of operators to measure sensitivity, and on the concept of an approximate itwerse. Approximate invertibility of the plant is shown to be a necessary and sufficient condition for sensitivity reduction. An indicator of approximate invertibility, called a measur« of .dngularity, is introduced. The measure of singularity of a linear time-invariant plant is shown to be determined by the location of its right half-plane zeros. In the absence of plant uncertainty, the sensitivity to output disturbances can be reduced to an optimal value approaching the singularity measure. In particular, if there are no right half-plane zeros, sensitivity can be made arbitrarily small. The feedback schemes used in the optimization of sensitiviiy resemble the lead-lag networks of classical control design. Some of their properties, and methods of constructing them in special cases are presented.
d
FILTER Fig. I.
of sensiuvrty or plant uncertainty that are natural for optimization? how does plant uncertainty affect the possibility of designing a feedback scheme to reduce plant uncertainty? At a more practical level, the theory will be illustrated by simple examples involving single variable and multivariable frequency responses. The questions here are: can the classical "lead-lag" controllers be derived from an optimization problem? How do RHP (right half..plane) zeros restrict sensitivity? in multivariable systems without RHP zeros, can sensitivity be made arbitrarily small, and if so how? A.
I.
INTRODUCTION
I N THIS paper we shall be concerned with the effects of feedback on uncertainty, where uncertainty occurs either in the form of an additive disturbance d at the output of a linear plant P (Fig, I), or an additive perturbation in P representing "plant uncertainty." We shall approach this subject from the point of view of classical sensitivity theory, with the difference that feedbacks will not only reduce but actually optimize sensitivity in an appropriate sense. The theory will be developed at two levels of generality. At the higher level, a framework will be sought in which the essence of the classical ideas can be captured. To this end.. systems will be represented by mappings belonging to a normed algebra. The object here is to obtain general answers to such questions as: how does the usefulness of feedback depend on plant invertibility? are there measures Manuscript received October 8, 1979; revised December 4, 1980 and December 17, 1980. Paper recommended by A. Z. Manitius, Past Chairman of the Optimal Systems Committee. An earlier version of this paper [23J was presented the 17th Allerton Conference. October 1979. The author is with the Department of Electrical Engineering, McGill University. Montreal, P.Q., Canada.
v
u
Motivation
A few observations might serve to motivate this reexamination of feedback theory. One way of attenuating disturbances is to introduce a filter of the WHK (Wiener- Hopf- Kalman) type in the feedback path. Despite the unquestioned success of the WHK and state-space approaches" the classical methods.. which rely on lead-lag "compensators" to reduce sensitivity, have continued to dominate many areas of design. On and off, there have been attempts to develop analogous methods for rnultivariable systems. However. the classical techniques have been difficult to pin down in a mathernatical theory, partly because the purpose of compensation has not been clearly stated. One of our objectives is to formulate the compensation problem as the solution to a well defined optimization prohlem. Another motivating factor is the gradual realization that classical theory is not just an old-fashioned way of doing WHK, but is concerned with a different category of mathematical problems. In a typical WHK problem, the quadratic norm of the response to a disturbance c is minimized by a projection method (see Sections III'-A' and IV-C); in a deterministic version. the power spectrum
Reprinted from IEEE Transactions on Automatic Control, Vol. AC-26, April 1981, pp. 301-320.
497
Id(jw)1 is a single, known vector in, e.g., the space L 2( - 00,00); in stochastic versions, d belongs to a single random process of know'; covariance properties. However, there are many practical problems in which ld(jw)1 is unknown but belongs to a prescribed. set, or d belongs to a class of random processes whose covariances are uncertain but belong to a prescribed set. For example, in audio design, d is often one of a set of narrow-band signals in the 20-20K Hz interval, as opposed to a single, wide-band signal in the same interval. Problems involving such more general disturbance sets are not tractable by WHK or projection techniques. In a feedback context, they are now usually handled by empirical methods resembling those of classical sensitivity. One objective here is to find a systematic approach to problems involving such sets of disturbances. Another observation is that many problems of plant uncertainty can be stated easily in the classical theory, e.g., in terms of a tolerance-bind on a frequency response as in [2], but are difficult to express in a linear-quadratic-statespace framework. One reason for this is that frequencyresponse descriptions and, more generally, input-output descriptions preservethe operations of system addition and multiplication, whereas state-space descriptions do not. Another reason is that the quadratic norm is hard to estimate for system products (see Sections iII/-A' and IV-BI), whereas th~ induced norm (or "gain") that is implicit in the classical theory is easier to estimate. We· would like to exploit these advantages in the study of plant uncertainty. Finally, sensitivity theory is one of the few 'tools available for the study of organization structure: feedback versus open-loop, aggregated versus disaggregated, etc. For example, feedback reduces complexity of identification roughly for the same reason that it reduces sensitivity [12], [13]. However, it is hard to draw definitive conclusions about the effectsof organization without some notion of optimality, and such a notion is missing in the old theory.
that sensitivity reduction is possible if there
IS
such an
inverse.
C.
Background
Many of the ideas in this paper are foreshadowed in the classical theory [1], [2] of single-input single-output convolution systems, especially as presented by Horowitz [2], who derived various limits on sensitivity imposed by the plant, and stressed the need to consider plant uncertainty in design. The author posed the feedback problem in a normed algebra of operators on a Banach space, and introduced [4], (5] perturbation formulas of the type (l-P)-I-(J-Po)-I =(1- p)-I( r- Po)( j - Po)
·1
(1.1 )
which were used to show that high-gain feedback reduces the sensitivity of linear amplifiers to large nonlinear perturbations [3]-[5]. Desoer studied a related problem in [6], and recently [7] has obtained results for the case of P and Po both nonlinear (also see footnote 8). Perkins and Cruz [8] used perturbation formulas similar to (1.1) to calculate the sensitivity of linear multivariable systems. Porter [9] posed various sensitivity problems in Hilbert space, and in a paper with Desantis [10] obtained circle type conditions for sensitivity reduction. Willems (11] has stressed the Banach algebraic aspects of feedback theory. In [1]-[10], the disturbance is either a fixed vector, or lies in some band of frequencies, and sensitivity is measured in terms of an output norm. as opposed to an induced operator norm. The approach of using weighted operator norms, and relating optimal sensitivity to weighted invertibility via a fractional transformation was used in [12], but has since been reworked and expanded. D.
Two Problems
We shall be concerned with the system of Fig. 1. Here, P is a given plant with a single (possibly multivariable) input B. Weighted Seminorms and Approximate Inverses l' accessible to control, and an output y to which a disOne way of defining the optimal sensitivity of a feed- turbance d, not accessible to control, has been added. The back system, and of addressing some of the issues men- plant input v is generated by a filter whose only inputs tioned in SectionI-A, is in terms of an induced norm of the consist of observations on the plant output y and a refersensitivity operator. However, it will be shown in Section ence input u. Two types of problems will be considered. Problem J -Disturbance Attenuation: This problem will III'-B'that the primary norm of an operator in a normed algebra is useless for this purpose. Perhaps that is why be the subject of Sections V-VII. Suppose that u= O. The operator norm optimization has not been pursued exten- input-output behavior of the system between the nodes (2,3) can be modeled by the flowgraph of Fig. 2, which sively in the past. Instead, we shall introduce an auxiliary "weighted" consists of the plant P and a single additional operator F seminorm, which retains some of the multiplicative proper- in the feedback path. The disturbance d is uncertain in the ties of the induced norm, but is amenable to optimization. sense that it can be anyone of a set of disturbances. Plant uncertainty will be described in terms of belongingto Initially (through Section VI) P is assumed to be known exactly, but later (Section VII) to be uncertain .. We would a sphere in the weighted seminorm. like to characterize the feedback operators F which attenuate Approximate invertibility of the plant is one of. the features which distinguishes control from. say, communi- the response y to d in some appropriately optimal sense, and cation problems. We shall define the concept of an ap- examine the effects of uncertainly about P on disturbance attenuation. proximate inverse under a weighted seminorm, and show 498
d
v Fig. 4.
Fig. 2.
may be found in such texts as Naimark [19). Occasionally, it will be assumed that a normed algebra is a Banach algebra. i.e., has the property that every convergent sequence of elements of the algebra has a limit in the algebra. It will be assumed that all linear spaces and algebras are over the real field.
d
A. Algebras of Frequency Response Systems
y
The frequency response of a stable, causal, linear timeinvariant system is a function analytic in the right-half of the complex plane. An accepted setting for such functions involves the HI' Hardy spaces [15], which we shall employ with some modifications to accommodate unstable systems. The algebra consists of functions fi{· ) of a complex variable s=a+jw, each of which is analytic in some open half-plane Re(s»op possibly depending on p(.), and is bounded there, i.e., p(s)~const. for Re(s»ap • The functions in H': will be referred to as causal frequency responses. If jJ is in H':, then the domain of definition of p can be extended by analytic continuation to a unique, maximal, open half-plane of analyticity Re(s»opm' where ap m ~(Jp. In general, p need not be bounded on this maximal open RHP, but if it is, then it can further be extended to the boundary by the limit p(0pm +j", ~ lim CJ -+ CJ p{0+ jw), which exists for almost all fA) provided approaches the boundary nontangentially from the right. Assume that all functions in H': have been so extended. The algebra H«) (of stable causal frequency responses) consists of functions p of H~ for which 0, Ei; 0, i.e., the region of bounded analyticity includes the RHP. The norm IIpll=sup{lp(s)l: Re(s»O} is defined on H«J, making H«J a norrned and, indeed, Banach algebra. A strictly proper function in H:' satisfies the condition ft{s)--+O as Isl--+oo in Re(s)~ap. The symbols H:6 and H will denote the algebras of strictly proper frequency responses in H': and Boo, respectively. By a straightforward application of the maximum modulus principle, the nonned algebra H't of strictly proper stable frequency responses has the property that Ilftll=esssup{lft(jw)l: w real} for any Pin H't, i.e., the norm can be computed from jw-axis measurements. Spaces of Inputs and Outputs: For any integer 1~q< 00, the linear space Hi consists of functions Q( ·) of a complex variable, each u being analytic in some open half-plane Re(s»a., inwhich the restriction of u to any vertical line is in Lq and f~oola(a+jw)lqdwE;const. for all a>ou. Again, the domain of definition of each uin is extended by analytic continuation to a maximal open RHP of ana.. Iyticity and then, if u is Lq-bounded in this RHP, to its boundary by a nontangentiallimit. The space H" consists
Hr:
Fig. 3.
Problem 2-Plant Uncertainty Attenuation: (Problem 2 is the subject of Section VIII.) Suppose that d=O, and the plant P is uncertain to the extent that it can be anyone of a "ball" of possible plants centered around some nominal value PI. If the filter is linear, the behavior of the system between nodes (I, 2, 3) can be modeled by the flowgraph of Fig. 4. The filter can be characterized by a pair of operators (U, F). We would like to find operators (U, F) which shrink the ball of uncertainty but leave the nominal plant invariant; to find bounds on the optimal shrinkage and to look at its dependence on plant uncertainty. E.
Outline of the Paper See Synopsis following Appendixes. II.
SPACES AND ALGEBRAS OF SYSTEMS
The purpose of this section is to specify the meaning which will be attached to the terms "frequency-response" and "linear system," and to summarize their properties for later use. A feature of the input-output approach is that systems can be added, multiplied by other systems or by scalars, and the sums or products obtained are still systems, i.e., they form an algebra. Frequently, it will be assumed that the largest amplification produced by a system can be measured by a norm, typically the maximum frequency response amplitude over some region of analyticity; under this assumption the algebra of systems becomes a normed algebra. Normed algebras provide the natural setting for the study of system interconnections such as feedback. Their elementary properties will be used freely here, and
499
a+Jw
o
H:
inverse (I + P) - I exists in A. and for any F in A the products PF and F·P are in A .\,. The elements of A s will be called strictly causal operators. (The concept of an algebra of "realizable" systems was introduced by the author in [14]; the related notions of strong causality of Willems [11], and later strict causality of Porter, Saeks, and Desantis are compared in a paper of Feintuch [20].) An example of the space ~X of inputs and the algebra A of causal operators is provided by the space Hi of (trans.. forms of) inputs and the algebra H~ of causal frequency response operators. In this context the algebra H~) of strictly proper frequency response operators is an example of an algebra A.f of strictly causal operators. Henceforth, we shall refer to the strictly proper operators as strictly causal. In the case of stability, we shall need the fact that a stable input-output system produces a finite amplification of inputs that can be measured by a suitable norm, and that stable systems form an algebra (see, e.g. [5]). Accordingly, we postulate the following. ~~ is a Banach subspace of :X whose elements will be called bounded inputs or outputs [for example, H 2 or L 2(O . 8 is a normed subalgebra of A containing the identity I, whose elements will be called stable causal IIPIIHx=llp"H~' The stable operators in H 00 map H", which is a proper operators, under the following assumption: the norm of subspace of into H", and are in fact completely de- any P in B is the ~~ - induced norm, that is. termined by their behavior on H". They can therefore be IIPll ~ sup{IIPull/llull: u in ~f), u*O}~ the sup being represented by their restrictions of the form P: H" ~ H". finite.' In sections devoted entirely to stable systems, we shall If H ~ is taken as an example of A., then IHI ec IS an concentrate on operators of the form P: H" ~ H" without example of B. distinguishing them as restrictions of operators on Hi. B, is the subalgebra of B obtained by intersecting As
of functions u in Hi for which CJu ~O, and is a Banach space under the norm \I II = supo>o {f~:x Iu( 0 + jw)1q dw }l/q. Inputs and outputs will belong either to or one of the Hi spaces, 1~ q< 00. In the special case q= 2 it follows from the Paley- Wiener theory that every function of H 2 is a Laplace transform of a time function in L 2(0, (0) and vice versa. In the general case of H'I, some functions can be viewed as transforms of time functions, and the others as frequency functions that do not appear in physical applications and can be disregarded. Frequency Response Operators: Let q be any integer" 1 ~q~ 00.. which will be held fixed. For any causaJ frequency response jJ(.) in H~ an operator P: Hi -+ is defined by the multiplication h(s)=P(s)u(s). P will be caned a causal frequency response operator" and the algebra of all such operators win be denoted by the bordered capital H~. Similarly. for each of the algebras of frequency responses Roo. H:O~ and H~, an algebra of operators mapping Hi into itself is defined and denoted by the corresponding bordered capital. i.e., Hoo, ":0, or "0. In the case of normed frequency response algebras, the corresponding operator algebras are similarly normed, e.g.,
u
He:
H:
00».
H:.
and B~ consisting of strictly-causal stable operators. It should be noted that B, is not a radical of 8" as P stable does not imply that ( 1+ P) - I is stable. We would like to take an axiomatic approach to the For the purpose of estimating the effects of small perproblem of sensitivity reduction by feedback" i.e., to single turbations, it will be assumed that Bs has the small-gain out the relevant properties of linear systems and postulate property; i.e., for any P in Os' if II P II < 1 then (J + P) - I is them as axioms. For example, the related properties of in B. If B is complete, i.e., a Banach space, this assumption causality.. realizability, and strong or strict causality have is redundant for then the series J- P + p2 -- ... converges definitions [see, e.g., [14] and [11]) reflecting the fact that to the inverse in B of (I + P). However, we have applicathe response to a sudden input to a physical system can not tions in mind in which completeness of B is replaced by anticipate the input, and cannot occur instantaneously. other assumptions. These properties of physical systems preclude the pathoA frequency response pE HOC which is not strictly proper logical phenomena associated with instantaneous response can not be realized exactly, but can be approximated by a around a feedback loop" and ensure that the feedback sequence of strictly proper responses of the form n( s +. operator (I + P) - I is well defined. However.. these details 11) - I p(s ), 11 == L2~ . . .. The sequence n( s + n) .. I is an are not relevant here. The only items of interest are that example of an "identity sequence." More generally, idencausal systems form an algebra of mappings, and that tity sequences will be used to construct strictly proper strictly causal systems form a subalgebra whose salient approximations to improper responses" and are defined as feature is the existence of the inverse (I + P) -) for all of its follows: an identity sequence {I" }c:= I for B is a sequence of members, i.e., a "radical." Accordingly we postulate the operators in B with the property that for any F in B the following. sequences II InF- F II and II FI" - F II approach 0 as n ~ 'YJ. ~ is a linear space whose elements will be called inputs It is assumed that 8", contains an identity sequence for B. or outputs. A is a linear algebra of linear mappings P: The following well..known (cf. Naimark {19. p. 162]) ~X,-+~, with identity I, whose elements will be called causal properties of a normed algebra will be crucial in many operators. A s is a radical of A, i.e.,I a proper nontrivial subalgebra of A with the property that for any P in A .~'l the
B.
More General Algebras of Systems
I The
properties of radicals are discussed in Naimark [19. p. 162].
2Whcnevcr the norm of .\' is not identified by a subscript. it should hi' taken to be the principal norm of the space to which .r belongs. 500
parts of the paper. For convenience they are proved in Appendix I. Let P and Q be in B. Proposition 2.1: a) If (I+PQ)-I IS In B, then (/+ QP)-I is in B~ and the formula P(I+QP)-I =(/+ PQ)-Jp is valid. b) If R is a radical in Band P is in R, then P has no inverse in B. (Strictly causal operators have no inverses in B.) c) If P and (1 + P)-1 are in B and tfPIJ
We shall be interested in situations in which P is at or near some nominal value PI' and the feedback F appears as an operator variable in an optimization problem whose object is to minimize response to d. Unstable operator variables are difficult to handle, and so our first step will be to show that F has an equivalent realization in terms of a stable operator. A.
The Model Reference Transformation
The flowgraph of Fig. 3 is described by the equations
FEEDBACK DECOMPOSITION: STABILIZING AND STABILIZED STAGES
We proceed to derive a decomposition principle to be employed in disturbance attenuation. Suppose that there is no plant uncertainty, and that the plant and feedback are constrained not to be simultaneously unstable. Under these hypotheses. any closed-loop stable feedback design can be decomposed into two stages: a first stage involving plant stabilization (which can be omitted for stable plants); and a second stage, involving a model reference scheme in which only stable elements are used, and which is automatically closed-loop stable. The choice of a stabilizing stage is independent of, and does not prejudice the choice of the second stage. Having established this fact, we shall be free to concentrate on the second stage of the disturbance attenuation problem under the condition that the plant is stable (or has been stabilized), without loss of generality. Consider the system of Fig. 2. The plant input v, output )', and disturbance d are all in ~. and satisfy the equations y=Pv+d (3.la) v= -F}'
(3.3a)
v= -Q(y-Ptv)
(3.3b)
in which .,Y, u, and d are in ~X, and P, PI' Q are in AJ' Equation (3.3) will be called a model reference scheme with comparator Q. as the output of the plant P with disturbance d added is compared to the output of a model PI of the plant without disturbance, and the difference actuates Q. The two sets of equations (3.1) and (3.3) are called equivalent iff every input- output triple i d, u, y) in 'X 3 satisfying (3.1) satisfies (3.3), and vice versa. Their equivalence will be established under the assumption that the equations Q=F{ 1+ P.F )--'
( 3.4a)
F=Q(I-PtQ)-1
(3.4b)
hold. If either equation in (3.4) is valid.. then so is the other. and
(3.lb)
in which P and F are operators in A s (see Remark 3.le)]. We shall refer to (3.1) as a feedback scheme with plant P and feedback F. Since P is strictly causal, the inverse (I+PF)-I exists in A. Therefore, for each d in ~, (3.1) have unique solutions for v and y in ~, given by the formulas y= (1 + PF) -1 d
y=Pv+d
(3.5)
To derive
(3.5)~
suppose (3.4a) is valid. Therefore,
!-P1Q=/-PJ F( 1+ P, F) ··1
=(I + PI F)( I + PI f~) - I -
PI F( I + PI F) --- I
=(J-tPIF)-l
(3.2a)
v= -F(/+PF)-'d. (3.2b) Let K 32 : :X~~"X, denote the "closed..loop" operator mapping d to v. K 32 is an operator in A given by K 32 =- F( J +
and (3.5) is true: here the expression 1-::: ( I + P 1 F )(I -+PI F) -I and the distributive law for multiplication on the right was used. Equations (3.4a) and (3.5) can now be used to give the identi ties
PF)-I.
The flowgraphs in this paper are simple, and will be approached informally in order to avoid lengthy definitions. Expressions for some of the subsidiary c.l. (closed- so (3.4b) is true as claimed. The converse proposition is loop) operators. which can be found by inspection, will be proved similarly. Assumption: For the present, and until the end of Section listed without derivation as needed. For a system to be physically realizable on an infinite VI, assume that P= Pit i.e., there is no plant uncertainty. TheorenI t. time interval, it is usual to postulate (14] that all c.1. input-output operators must be stable. though "open.. loop" a) Any ciosed.. loop stable feedback scheme (3.1) with operators such as the plant P and feedback F may be stable plant P E 18.~ and (not necessarily stable) feedback unstable. The set of c.l. operators for (3.1) consists of: FE A s is equivalent to a rnodel reference scheme ",'hose K 22 = ( J+ FP ) - I.. K 23=PK z2 • K 33=(I+PFr- l .. and K 32 branches are all stable . i.e.. QE8s and P, -= P where f' and specified above. Accordingly, the feedback scheme (3.1) Q are related hy (3.4). Conversely.. any model reference will be called c.l. stable if Kij is in B for i.. l> 2 or 3. scheme with stable branches is closed-loop stable.. and 4
501
guide to perturbation analysis. I t will also appear that model-reference schemes have a useful plant-invariance
equivalent to a closed-loop stable feedback scheme subject to (3.4).
property.
b) If (3.4) holds, and d and y satisfy either the feedback or model reference equations they satisfy the equation (3.6)
Proof' a) For any feedback scheme (3.1), if (d, u, y)ECX,3 satisfies (3.1) and F is given by (3.4b), then (d, v, y) satisfies (3.3), and conversely.Therefore, (3.1) is equivalent
to (3.3). If the feedback scheme is c.l. stable, then Q must be stable as it equals - K 32 • If, in addition, P is assumed stable, then all branches in (3.3) are stable, as claimed. Conversely, by a similar argument, any model reference scheme (3.1) is equivalent to a feedback scheme (3.3). Suppose that the branches of (3.3), namely, PI' Q, and p =PI are all stable. Then, all the cJ. operators of (3.3), namely, {K ij };, j : 2.3.4. S' must be stable because they can be expressed in terms of sums and products of the stable operators P, PI' Q, and I. The last assertion follows from the expressions for the diagonal c.l, operators K;; of (3.3), namely, K 22 =I-QP J , K 33 =/-P1Q, K 44 =K ss =/ and the fact, easily checked by inspection, that the remaining c.l. operators Kij' i=l=J, are products of the Kji by P, PI' Q, or I. It follows that (3.3) is c.l. stable. b) If P=P t and (d, y) satisfies (3.1) or (3.3), and (3.4) holds, then (3.6) is obtained by substitution of (3.5) into (3.2a). Q.E.D. The operator (J - PIQ) appearing in (3.6) will reappear as a factor in most expressions for sensitivity. It will be called the sensitivity operator and denoted by E. For equivalent schemes E=(I+P1F)-'. Remarks 3.1: a) The model-reference scheme has some remarkable features. Unlike most feedback arrangements, it is a realization which cannot be made unstable by any choice of Q, at least for stable plants in the absence of plant uncertainty. Under these assumptions.? any allowable feedback law can be realized in the form of an equivalent model reference scheme, with the guarantee that all branches will be stable, and the closed-loop system automatically stable. The design of Q, whether for small sensitivity or other purposes, canbe accomplished without concern for closed-loop stability. In engineering applications, model reference schemes are realizable in principle, but may have undesirable features. For example, they may have high sensitivity to errors in the realization of Q. Unstable inner loops, obtained whenever (J-p.Q)-1 is unstable, may present reliability problems. Even then, the fractional transformation remains advantageous from the viewpoint of theory, as potentially unstable feedbacks F are replaced by stable operators Q. In later sections on plant uncertainty, the flowgraph interpretation of the model reference scheme will provide a convenient 31( P is unstable. the parameterization or d. stableschemes by a single operator Q is obviously still possible. However, some of our other conclusions, concerning existence of a feedback realization with stable elements, structuralstability,or decomposition properties, may no longer
be valid.
b) Implicit in our notion of an allowable feedback is the view that each feedback realization involves a graph, and that although most of the internal details of the realization may be unimportant, closed-loop stability at all internal nodes is essential. c) Theorem 1 holds even if F and Q are in A but not strictly causal. However, strict causality is a prerequisite for physical realizability, and will therefore have to be assumed in subsequent theorems. B. Unstable Plants
The assumption in Theorem I that the plant P is stable will now be relaxed. Consider a plant Po EA s with disturbance d at the output, which is unstable but for which there exists a stabilizing feedback, i.e., an operator Fo E A which gives a c.l. stable feedback scheme on being fed back around Po. The stabilized system can be incorporated in a model reference scheme, by letting P be the stabilized c.l. operator Po( 1+ FoPo)- J and d be the stabilized disturbance (I+PoFo) - ld o, and Q (or F) can be selected as for a stable plant. At this point the question arises: "can Fo be selected independently of Q (or F), or could the prior choice of Fa prejudice the class of achievable systems?" In general, the choices are not independent, even for stable plants, because the application of two unstable feedbacks in succession may give a result different from the application of a single feedback equal to their sum. ConLet sider the follo"wing f~equency response example in p(s)=l a~dfl(s)=f2(s),,=s-I'4Theapplication of a single feedbackf(s) equal to!I(s)+h(s) gives a c.1. stable feedback scheme,with c.l. responses. 1, (s + 2) -I, and s(s + 2) -I. However, if the feedback is split into two branches, the c.l. response across either one of these branches is (s + 1)/s(s + 2), i.e., the system is not c.l. stable. Popular belief notwithstanding, c.l. stable systems do not form an additive group under feedback if the complete set of c.l. operators is considered. However, if feedbacks are constrained to be stable then choices are independent, as the following construction shows. Let Po E A s be an unstable plant which can be stabilized by either one of two feedbacks. Fu and Fh in A S1 and label the resulting feedback schemes (a) and (b), respectively. We would like to find an operator ~h EA s which on being fed back around scheme (a), as shown in Fig. 3(a)~ produces a two-stage feedback scheme equivalent to scheme (b). (Observe that the two-stage feedback scheme has extra nodes in the feedback branches to allow for the possibility of noise sources there.) Proposition 3.2: If Fa and Fh are stable, then the stable feedback Fo h ~ F; - Fu makes the two-stage feedback scheme c.l. stable and equivalent to scheme (b). Proof' The two-stage scheme is obviously equivalent to scheme (b). and is c.l. stable because its c.l, operators consist of: i) the c.l. operators K 22 , K 23 , K 32 , and K 33 of h () h sc erne Q or sc erne (b), which are stable by hypothesis, or ii) sums and products of the operators listed in i), and
502
Hr:.
inverses is hard to study. This is a serious limitation in feedback problems in which expressions such as (I + PF) - I play a major role. By contrast, the "Mm spec which is widely used in classicaldesign measures the maximum frequency response magnitude, and is essentially the induced operator norm II PI/ ~ sup {.II Px" L 2 / 11x II L 2 : x EL 2 } , which has the multiplicative property, and is therefore convenient to estimate in cascaded systems. By describing plant uncertainty in terms of a sphere of specified radius in a norm having such a multiplicative property, it is possible to obtain a genera) approach to problems involving disturbances/random processeswhich are unknown but belong to prescribed sets. as we shall see in Section IV. Minimization of an induced norm in effect amounts to a minimax solution. Minimax methods do not necessarily represent uncertainty with .greater fidelity than quadratic methods. However, the concern here is iess with fidelity than with the ability to handle product systems.
the operators Fa or Fb which are stable by hypotheses.
Q.E.D. The following decomposition principle can be obtained immediately from Theorem 1 and Proposition 3.2. Let Po E A s be any unstable plant stabilizable by a set of stable feedbacks in 8.1. An)' closed-loop stable feedback scheme employing a stable feedback around the plant Po is equivalent to a closed-loop stable scheme consisting 0/: i) a stabilizing feedback Fo ER.s which can be selected arbitrarily, followed by ii) a modelreference scheme with stable operators Q and
U
PI· It follows that under our hypotheses,' and in particular under the assumption that plant and feedback. are not
simultaneously unstable, the problem of, sensitivity reduction can be decomposed into two independent problems: stabilization followed by desensitization of a stable system. Henceforth, we shall confine ourselves to the second problem. III'.
ApPROACHES TO FEEDBACK-SENSITIVITY
B'.
MINIMIZATION
A'.
Quadratic versus Induced Norms
The main properties of feedback cannot be deduced without some notion of uncertainty. Suppose that the disturbance d is uncertain but belongs to some subset OJ) of possible disturbances in ~. From (3.6) it is clear that for disturbances to be attenuated, (I-P.Q) must be small on 6D, i.e...' Q must act as an approximate inverse of PIon "D. The various approaches to the disturbance 'attenuation problem are differentiated by the way in which uncertainty is described, and this approximate inversion is metricized and calculated. A typical WHK approach in a deterministic version could be viewed as follows: 6D consists of the set of disturbances d in L 2(0, 00) possessing a single, fixed, known power spectrum Id( jw >1 in L 2( - 00, 00), and the object of design is to find a filter Q that minimizes the quadratic distance If d- P.Qd II L'2 where in general Q depends on Id( j~ )1. In the stochastic analog of this problem 6D is a random process characterized by probability-covariance functions and metricized by a quadratic norm. This description of uncertainty has certain limitations that we would like to circumvent, namely, the following. I) The covariance properties of the random process must be known. In practice, they are often unknown elements of prescribed sets. This is merely a limitation on the class of random processes for which WHK is valid. More serious from the point of view of feedback theory is the following observation. 2) The quadratic norm on plants employed in the WHK method lacks the multiplicative property II PQ II L, E; II P II L, tI QII L,' and in general it may be difficult or impossible estimate the norm of a product PQ from the norms of P and Q. The product norm II PQII L, may be large even though II PilL, and II Q II L, are small. . Consequently, if plant ·uncertainty is metricized by the quadratic norm, its propagation through products and
to
Constraints on the Norm of a Sensitivity Opera/or
It is natural then to try to pose sensitivity reduction problems in terms of the minimization of norm of the sensitivity operator, and to employ a norm having multiplicative properties. The primary norm of a Banach algebra has such properties, but the following propositions show it to be useless for this purpose. . Proposition 3.3: If P and Q are in a Banach algebra B and 111- PQ II < I, then PQ has an inverse in B. Proof: Denote (I - PQ) by E. As II E II < 1, the power series [- E + E 2 - • •• converges to an operator which Q.E.D. inverts (1- E). As (I-E)=PQt (PQ)-I exists. Proposition 3.3 has occasionally been interpreted as showing that invertibility is necessary for sensitivity reduction. This interpretation is empty. In fact, since strongly causal operators never have inverses in B [see Proposition 2.1b)], we have the following. Corollary 3.4: If PQ is in B", then 111- PQ II ~ 1. It is impossible to make the sensitivityoperator less than this simply means that the 1 in the original II norm. in frequency response of PQ approaches 0 at infinite frequen-cies and (1- PQ) approaches 1. An obvious idea at this point is to make (/- P1Q) small in norm over some finite frequency band, i.e., over an invariant subspace. The next proposition shows that norms over invariant subspaces usually are not useful measures of sensitivity for optimization purposes. Let ~ I be a subspace of ~t II a projection operator onto ~ I' and suppose that ~I is invariant" under 8, i.e., RII=llRII for each R in B. Let a denote the norm of (1- p.Q) restricted to the subspace ~ 1 and optimized over all Q, i.e., a ~ infQEB,slip{II(I-PIQ)IIdll: de(j, and IIdll= I}. Proposition 3.5: For any P in Bs ' a= 1 or a=O. The proof is in Appendix I. If a =O. sensitivity can be made arbitrarily small over the subspace ~~ ,. In practice there are special cases involving "minimum phase" systems
Hr
4More generally, Proposition 3.5 holds if B is replaced by any of its
;0:~:Ting extensions. Note that Il is not necessarily in the algebra
503
2) Let oD be the set into which W maps flat disturbances. in which solutions that approach a=O may be useful. More i.e., WOi) I =uV. 6D is also the set typically, this resultis achieved at the expense of increasing the sensitivity without bound on complements of qB I; in such cases, the norm of a restriction of ([- P JQ) is not a candidate for minimization. Corollary 3.4 and Proposition 3.5 delineate some of the peculiarities of the sensitivity optimization problem. In one form or another these peculiarities were recognized in the classical theory, and are probably the reason why it stopped short of optimization. We shall try to circumvent them by introducing an auxiliary (semi) norm to which they do not apply-
oD can be described as the unit ball in the seminorm 11·11 r
defined on the range of W (which is a proper subspace of H by the equation IIdli r ~ ItW - Jdlt. There are many engineering problems in which the apriori information about disturbances is in the form of an upper bound to the magnitudes of their possible frequency responses. The seminorm description 2) is natural for such problems, and I) occurs in inverse problems. IV. MULTIPLICATIVE SEMINORMS AND The seminorm II· II, employs an up-weighting, and U·11 r ApPROXIMATE INVERSES employs a down-weighting. These two examples generalize into the notion of left and right seminorms, defined as Uncertainty in a disturbance (or plant) in a linear space folJows. can be specified in terms of belonging to a ball of disturbances (or plants) centered at some nominal value, and A. Seminomas lor Inputs and Outputs" of radius specified in some norm. Such a description of uncertainty may be cruder than a probabilistic description, Let "ll be a II·ll-normed linear space. Let II-II t be any but is usually more tractable in feedback problems. seminorm defined on all of ~, and 11·.11, a seminorm One of the axioms of a norm asserts that only the zero defined on some nontrivial subspace ~, of 09. The semielement has zero norm. This axiom is often not needed, norm 11·11 T is said to dominate 11·11 t iff II yilt ~ 11 y II r for all)' and with its elimination a norm is replaced by the slightly in ~,; this dominanceis denoted by 11·11, ~ 11·11 r: more general concept of a seminorm. Definition: A left seminorm is any seminorm defined on A ball in any seminorm can be shown to be a convex all of ~ with the property that U·II, E;; II· It. A right semiset.' Conversely, any convex setS in a linear space gener- norm is any seminorm defined on a nontrivial subspace '"Yr ates a seminorm (see Rudin (22, p. 24]) known as the of e;y with the property that n· It ~ 11·11 r: For example, if W is any H~ filter of unit norm, the Minkowski functional of that set. In linear spaces convex sets" of uncertainty can therefore always be described in expressions II y II f ~ II Wy II and II y II r ~ II W - Iy II define terms of seminorms. We shall employ seminorms to obtain left and right seminorms, on H OIJ and the range of W, a systematic approach to such sets of uncertainty (cf. the respectively. The range of W is a subspace of H«J, proper whenever w(s) has zeros in the right half-plane or at 00. objectives outlinedin SectionI-A). In the next section, we shall define classes of left and right seminorms. To motivate the definitions, let us find B. Weighted Seminorms for Plants seminorm descriptions for two disturbance sets which can be generated by the interaction of filters and certain "flat"
disturbance sets. Henceforth, W will denote a stable causal operator of unit norm, which will play the role of a weighting filter. For concreteness, W can be thought of as an operator in H W: BOO -.Hoc, with response w(s)=k(s+k)-I. U1) will denote a flat disturbance set (analogous to white noise) consisting of the unit ball in the space of inputs, in this
o,
case in H~:
o)
Definition: A weighted seminorm is any seminorm II· If ~.
on the B with the property that ". fI w ~ /I ·11. The terms "weighted" and "left" are synonymous. We shall use the term "weighted" to distinguish the left semi.. norms on B used as measures of plant sensitivity from the others. A weighted seminorm on B is induced by a pair (tI-It!, 11·11 ,), where fI-1f r is a left seminorm on the space ~ (of outputs), and 11·11, is a right seminorm on a subspace ~, C~ (of inputs), iff 11·11 w is defined for A E B by the equation
whose elements are frequency functions of unknown but IIAflw=sup{IfAullr/lfull,: uecj,and ffullr:;t=O}. bounded magnitude, Consider two situations. It follows that II I II w .so;; I. I) Let 6j) be the set mapped by W into flat disturbances, i.e., W6D=6J)I' Gj) can be described as the unit ball in the In control problems, weightings are often introduced by seminorm 11·11 I. defined on H" by the equation /I d /I f filters, which act on disturbances either before entering a ~ II Wd IJ. plant or after leaving it. For example, let We and w:. be linear mappingsin ~x ~, each of unit ~-induced norm. A 'Satisfying the following additional assumptions: I) if x is in the linear
space, then ax is in tbe set for somereal a; 2) if y is in the set then so is
rv-
beomsention: Wheneverx belongs to a space on which several norms are defined, the unsubscripted norm II x II denotes the principal norm. Wei&hted norms will be designated by subscripts.
504
~ II WrCVII for CEB, and II DII WI ~ II V -IDw:.U for DE B,. Then, II·II w has the multiplicative property claimed, as
left seminorm is defined on
tlCDIl w = II W,CD~II E; II WrCV 11·11 V -IDJt;.1I ~ II~-lull. The pair (1I·1I"It·II,) induces the weighted seminorm IIAII w = II W,A ~II on the space B. =IICllw,IIDll w l • Although weightings produced by filters will be emphasized in this paper. they can be produced by other A symmetric seminonn can be viewed as a special case of means. For example. a weighted seminorm on "0 is given a multiplicative seminorm on the space of products B· B,.. by the supremum over a shifted half-plane. in which Br =B and the multiplicativeinequality holds for each of two pairs of seminorms, namely, (11·11 w,II·II) and II PII w =sup {I fl{s)l: Re(s )~a}~ a>O. t
In gen.. eral, weighted seminorms lack the multiplicative property It CD II w ~ II C II w 11 D II W of algebra norms. In problems involving the attenuation of a single disturbance (or single random process) this need not matter, as multiplications can be avoided. However, in problems involving plant uncertainty, closed-loop perturbations have the product form ([- PQ)AP. We shall employ seminorms with weaker multiplicative properties suitable for such products. Definition: A symmetric (weighted) seminorm on the algebra B is a weighted seminorm 11·11 w on the space B which satisfies the multiplicative inequalities J) Multiplicative Seminorms-s-Symmetric Case:
UCDll w
(4.1)
Any operator W: 'X--+ 'X of unit 'i>-induced norm which commutes with all operators of 8 defines a symmetric (weighted) seminorm by the equation II A II w ~ II WA II
=
IfAWII.
Symmetric seminorms have the property that II J II w = I.
1) Multiplicative Seminorms-s-General Case: In multi-
variable systems the plant perturbation dP always appears on the right of the product ([- PQ) a P, and often does not commute with i l>: PQ). In such cases a more general class of multiplicative seminorms will be used. If ~p is strictly causal. then the product (1- PQ) a P lies in a proper subspace of B (which is a left ideal, although we have no immediate use for this fact). With such products in
(11·11,11·11 w)·
c. Approximate Inverses and Singularity Measures Many problems of feedback theory. both classical and modem, can be reduced to the construction of an approximate inverse. Let II ·II w be a fixed, weighted seminorm on the space B. . Definition: For any operator P in B, an approximate right inverse (in B,,)of P is any operator Q in 8" for which8 11/-PQ II w < 11/11 w; the right singularity measure (in Bs ) of P (under 11·11), denoted by J&{P), is
p.( P)=inf {1I/- PQ II w: Q in s.}. In general, p,(P) is a number in the interval OE;p,(P)< 1. The last inequality follows from the observation that Q 0 gives II/-PQI1 w =11111 w· In all of the following H co examples, II·II w will be the symmetric weighted norm defined by II A II w ~ II WA II for A EHQO, where WEH: is a fixed (strictly causal) operator of unit norm, For example, W can be the "low-pass" frequency response w(s)=k(s+k)-I, k>O. As 11·11 w is symmetric, it has the multiplicative properties defined in Section IV-B1). Example 4.1: Pa is a plant in H~ with frequency response Pa(s)=a(s+a)-I, a>O. The sequence of operators Qn in "0 with frequency responses tln(s)=a"-J(s+a) ·n 2(s + n) - 2, n = 1,2,. · ., satisfies the equation
=
mind, we make the following definition.
Ilw(I-Ptln)IIH =
Definition: Let H,. be any subspace of B, and B· B,
1C
denote the space of products {CD: C in 8 and D in B,.}. A multiplicative seminorm on the space of products B· B,. is any weighted seminorm on the space B · B,. with the following additional property: there is a left seminorm 11·11 wt defined on B, a seminorm 11·11 WI defined on B,., and the inequality IICDllw
2
sup Re(s);"O
W(S)( 1-
n
(.\'+n)
2).
(4.1)
The right-hand side (RHS) of (4.1) approaches 0 as n --. co. Therefore, the singularity measure in "0 of Pu under lI·1I K' is IJ( Pu'>=O. The operators Qn are approximate inverses.
and the sequence Qn is an example of what will be called an inverting sequence. Example 4.2: Pa is the operator of Example 4.1; P" is the "nonrninimum phase" operator in HOC with frequency responsep,,(s)=
'The definition can obviously be generalized for the case of perturba-
KRecall that 11111 w
tions appearing on the left.
505
__ I.
and 11111 ... = I if 1I·1i w is symmetric,
Section V-A and Corollary 6.l it will be established that in fact 1£( PI ) = Iw(P )1·
In these examples we have emphasized approximate inverses under a multiplicative seminorm, In passing, it may be worth mentioning that WHl( problems can be viewed as approximate inversion problems in which the weighted seminorm of (I - PQ) is obtained' by weighting by a fixed vector d E;:L2(0, (0), to obtain 11 1- PQ Uw = II(I-PQ)dU L 2. Here p,(P) is the irreducible error, However, U· II w lacks the multiplicative properties. No matter which seminorm is used p,(.) has the following property, Proposuio« 4.3: For any PII and Pb in B, p(PbPa ) -p(Pb ) ·
a) 111
Proof: If the contraryis assumed to be true, there is a Q in B, for which 11/- PbPQQIl w