Академический Документы
Профессиональный Документы
Культура Документы
Concepts from graph theory and from the geometric theory of linear systems are
used to derive explicit conditions for determining when the closed-loop spectrum of
a multi-channel linear system can be freely assigned or stabilized with decentralized
control.
Snmmmry~This paper studies the effect of decentralized part of a changing or random environment. Much
feedback on the closed-loop properties of jointly controlla- work dealing with decentralization in economic
ble, jointly observable k-channel linear systems. Channel
interactions within such systems are described by means of systems centers around linear programming and
suitably defined directed graphs. The concept of a complete the decomposition principle of[4]. There is a
system is introduced. Complete systems prove to be precisely growing literature concerned with applications of
those systems which can be made both controllable and
observable through a single channel by applying nondynamic decentralized control to frequency and voltage
decentralized feedback to all channels. Explicit conditions regulation problems associated with power shar-
are derived for determining when the closed-loop spectrum ing in multi-area electrical power pools [5, 6]. A
of a k-channel linear system can be freely assigned or
stabilized with decentralized control. conventional multiloop industrial controller not
involving 'cross-couplings' can be viewed as a
INTRODUCTION form of decentralized control, although the term
IT zs sometimes desirable or required to construct decentralization is rarely used in this case.
a system's feedback control or decision strategy Recent research concerned with decentral-
based on a constrained measurement or informa- ization has led to a number of interesting and
tion pattern. If the information pattern consists useful results. Gersho and Karafin[7] and later
of disjoint subsets of 'local' measurement vari- Sandberg[8] proposed a form of decentralized
ables, one subset for each local input variable, and control to synchronize data transmissions in a
ff the information fed back to each local input network of remotely located transmitting and
depends only on the input's corresponding subset receiving stations. Decentralization has been
of measurement variables, then the information successfully applied to a single-commodity net-
pattern is decentralized and the overall feedback work flow problem by Lau, Persiano and
strategy is called a decentralized control. Varaiya[9]. Aoki and Li[10] have studied the
Decentralization arises in many contexts. extent to which the state of a linear system can
Examples can be found of team decision and be recovered from information available to local
adjustment processes [2], [3] where the objective control agents. Chu[l 1] has considered a decen-
is for an organization of n members to optimize tralized version of the linear regulator problem
the expected value of a cost or payoff functional for a system consisting of a string of identical
while each member is allowed to observe only subsystems.
McFadden[12] and later the present
authors[13] studied the problem of stabilizing a
*Received 7 August 1975; revised 25 February 1976.The specially structured linear system using a simple
original version of this paper, based on the principal results decentralized decision strategy appropriate to
of[l], was presented at the 6th IFAC Congress which was
held in Boston/Cambridge, Mass. during August 1975. The
economic systems; similar problems have re-
pubfisbed Proceedings of this IFAC Meeting may be ordered cently been discussed in[14]. Another study of
from: Instrument Society of America, 400 Stanwix Street, the stabilization question for more general linear
Pittsburgh PA 15222. This paper was recommended for
publication in revised form by associate editor E. J. Davison. systems was undertaken by Aoki[15] who
TDepartment of Engineering and Applied Science Yale developed a number of interesting but partial
University, New Haven, CT 06520, U.S.A. results. A solution to the stabilization problem
~The work of the authors was supported by the U.S. Air
Force Office of Scientific Research (Grant No. 72-2211) and for a large class of linear systems was recently
by the National Research Council of Canada, (Grant No. obtained by Wang and Davison[16].
28140). In the present paper we study the effects of
479
480 J . P . CORFMATand A. S. MORSE
decentralized feedback on the closed-loop prop- (A, B) is the characteristic polynomial of A mod
erties of k-channel, jointly controllable, jointly (A[~).
observable, linear systems (I). Channel interac- If ~ is A -invariant, A I~ is the restriction of A
tions within such systems are usefully character- to ~ . If ~ ~ ~ and ~ is A-invariant, A [l(~/~)
ized in qualitative terms by means of suitably denotes the map induced by A in ~ / ~ .
defined directed graphs (2). The concept of a The root set of a real polynomial a (,I) is the
complete system is introduced in 3. Complete symmetric set of complex roots of a (A) repeated
systems prove to be precisely those systems according to multiplicity. The spectrum of (C, A,
which can be made both controllable and B) (i.e. the spectrum of A) is the root set of the
observable through a single channel by applica- characteristic polynomial (c.p.) of A. A set of
tion of nondynamic decentralized feedback to all complex numbers is stable, if all elements in the
channels (3). set are points in the open left half complex
It is well-known[17] that the closed-loop plane; a (A) is a stable polynomial if its root set is
spectrum of a linear system can be freely stable.
adjusted or assigned with centralized control The (unique) monic least common multiple of
provided the system is both controllable and two polynomials a(A) and ~5(A) is written as
observable. However with the added constraint a (A)v/3 (A); the monic greatest common divisor
of decentralization, controllability and observa- of a(~,) and/3(A) is denoted by a(A)^~(A). The
bility are no longer sufficient as pointed out complement of a subset s in a set k is written as
in[15]. In the sequel (4) it is shown that if the k-s.
graph of a k-channel system ~ is strongly The complement of a proper variety in a
connected, then the closed-loop spectrum of vector space is called a robust set. Properties of
can be freely assigned just in case ~ is complete. robust sets are discussed in the appendix.
If the graph of ~ is not strongly connected, then
spectrum assignment is possible if and only if (1) 1. FORMULATION
the dimension of ~ equals the sum of the The model to be considered is a k-channel
dimensions of a set of strongly connected linear system ~ ,- {C,, A, BI ; k} described by the
subsystems of ~ and (2) each subsystem is equations
complete (5).
k
If for a particular system, free spectrum
assignment is not possible, then the spectrum of ~(t) = Ax(t) + ,-i
y~ B,u,(t) 1
the closed-loop system which results from y,(t)=C,x(t), i E k - { l , 2 . . . . . k} (I)
application of any specific linear decentralized
control, contains a uniquely determined subset where u,(t) and y,(t) are vectors of control input
which cannot be shifted no matter what control and measured output variables associated with
is applied. By characterizing this subset directly channel i; the time functions u,(.), x(.) and y,(.)
in terms of the system's parameters (4, 5), a take values in ~ , ( = ~ =,), ~ ( = ~ ' ) and ~,(= ~ ' , )
complete solution is obtained to the problem of respectively, and A, BI . . . . . B,, C, . . . . . C~ are
stabilizing a k-channel linear system with de- constant matrices or linear maps. We deal
centralized control. These results significantly primarily with systems for which k I> 2, for if
extend the arthors' earlier work[18], and encom- k = 1 decentralization is not an issue.
pas the contributions of[16]. To rule out trivialities, only those systems for
which
Notation
In the sequel all vector spaces, maps and linear
transformations are defined over the reals; ~ " is
B, # ' i E k (2)
the real n-space; d(~) is the dimension of a
vector space ~. Both linear maps and matrices are are considered. We ustially assume that ~ is a
denoted by capital Roman letters A, B . . . . ; prime jointly controllable, jointly observable system;
k k
(') means transpose. i.e. (A[]~ ~,) -- &oand A [C, IA] -- 0. Note how-
An m-input, p-output, n-dimensional linear I--I |--I
where z,(.) takes values in ~,( = ~ , ) . Note that matrices, AF for A + X B,F~C,, and Xn~
i--I
the nondynamic local feedback law u , ( t ) = {C, An, B~; k} for the k-channel system rep-
F,y,(t)+ v~(t) is a specialization of (3) obtained resented by (5).
by setting n, = 0 and G, equal to the p,xp~ Closely related to the spectrum assignment
identity. problem is the problem of determining when Y
If the local controllers (3) are applied to (1), can be stabilized with decentralized control. The
there results a closed-loop system (1), (3) of stabilization problem has been studied [16] where
k
dimension n + X n, with spectrum (i.e., eigen- it is shown that ~ can be stabilized with
|--I decentralized control just in case the polynomial
value set) depending on the properties of the
local controllers X,. We say that the closed-loop ax(A) =- ^ det [ A I - An], (6)
FE~
spectrum of X is freely assignable with decentral-
ized control if for some fixed set of positive called the fixed polynomial of X, has all its roots
integers (q,, q2. . . . . q,} and for each symmetric in the open left-half complex plane. The neces-
set of complex numbers A, compatible with {q,, sity of this condition is a direct consequence of
q, . . . . . q,},? there exist local controllers X, the following proposition which can be deduced
which, when applied to X, result in a closed-loop directly from the results of[16].
system (1), (3) with spectrum A. The main
objective of this paper is to develop explicit Proposition 1. Let A ( ~ , k ) denote the spectrum
conditions in terms of the matrices which of the closed-loop system described by (1) and
characterize X, for determining when the closed- (3); and write Ax for the root set o]~,x(A ). Then
loop spectrum of X is freely assignable with
decentralized control. This is the spectrum Ax C A(~,, k)
assignment problem for decentralized feedback
for all linear controls ]~,, i E k.
systems.
One possible approach to the spectrum assign- In the sequel, Proposition 1 will be used to
ment problem is to first apply to Y, local arrive at an alternative solution to the stabiliza-
nondynamical controls of the form tion problem with solvability conditions expres-
sed directly in terms of the problem data. Our
u , ( t ) = F~y,(t)+ v,(t), i E k, (4) approach will in addition, enable us to illustrate
when the stabilization problem is well-posed, and
the objective being to select the F, so that the when it is not.
resulting closed-loop system
2. CONTROLLABLE, SINGLE-CHANNEL
SUBSYSTEMS
J--I
y,(t) = c~(t),
'
1
i ~ k.
t
j (5)
The principal problem to be considered here
and in 3 is as follows: For given X and fixed
j E k, find conditions for the existence of an
is both controllable and observable through a F ~ 3g such that (Cj, Ap, Bi) is a controllable,
single channel, say j. For if this can be observable system. If an F with the required
accomplished, then standard techniques[17] can property exists for some j, then it turns out that
be used to construct a second local control such an F exists for each and every j E k; and
system when this is so we say that ]~ can be made
controllable and observable through a single
tA set of complex numbers A is compat/bk with {q,, channel with nondynamic decentralized control.
q2. . . . . q,} if A can be expressed as a disjoint union of
symmetricsubsets, the number of elementsin the lth subset We begin by studying the simpler problem of
equalling q~. determining when an F E P can be found so
482 J.P. CORFMAT and A. S. MORSE
that (A~, B~) is a controllable pair. This problem and is therefore best treated in graph-theoretic
has in effect been solved in[19] for the case terms.
k = 2. Central to the problem's solution is the We define the graph of 2, written Gz, to be a
concept of a remnant polynomial. The remnant pair (k,F) consisting of k nodes labeled 1,
polynomial of a (singie-channel) linear system 2 . . . . k, with node i representing the ith channel
(C, A, B ) is defined as of 2, and a function F mapping k into its power
set according to the rule F ( j ) = { i : H , ~ 0 } ;
ifr it
P(C' A' B ) = I , ~ , ~,(A) if r < '
(7)
H, ~- C , ( A I - A)-'Bi is the transfer matrix from
channel input uj of X to channel output y,. The
graph may be represented by a diagram consist-
ing of a set of k nodes, labeled 1, 2 . . . . . k, with a
directed arc drawn from node j to node i just in
where r is the rational function rank of the case uj influences y, in the sense that H , ~ 0. The
transfer matrix C ( A I - A ) - ' B and ,(A), graph of ~ thus represents the open-loop pattern
=(A) . . . . .
, (A) are the transmission polynomials of signal flow between the channels of E.
of (C, A, B)[20], ordered so that ,(A) divides It is perhaps natural to expect that if (A~, Ba) is
,_,(A), i~ {2,3 . . . . t}; (C, A, B ) is called a controllable for some F, then each node of Gz
complete (single-channel) system if its transfer must be reachable from node j along a directed
matrix is nonzero and if p(C, A, B ) = 1. patht in Gx. This is, in fact, the case.
The main result off19] is as follows.
Proposition 3. If for some j E k and F E $~,
Proposition 2. Let (C, A, B) and D: V ~ be (AF, Bj) is controllable, then for each i ~ k, with
fixed with (A [~ + ~ ) = and C(A I - A)-'B # 0. i # j, there exists a path in Gx from node j to node
For all K : ~ ~ V, p(C, A, B) divides the uncon- i.
troUable polynomial of (A + DKC, B), the word
"divides" being replaceable by 'equals' just in While this p r o p e r t y of Gx is necessary for
case K is selected so that d((A + DKC[~)) is as (AF, Bj) to be controllable for some F, it is not
large as possible. sufficient for as we have already noted, when
k = 2 completeness of a certain subsystem of
By making use of the following lemma, it is is also required. To describe analogous subsys-
possible to restate Proposition 2 in a more tams for k-channel systems, we introduce the
convenient way. following notation: If s is a nonempty subset of
k , with elements i., i2. . . . . is ordered so that
Lemma l.t For the two-channel system il < i2 < . . . < is, then we define B. and C. so that
- { C , , A, B,; 2} and any decentralized feed-
back matrix F = d i a g . [F,, F2], (AFI~,) =
(A+B2F2C21~,) and u.p. (AF, B,) = u.p.
J
C,= B . = [B,,, B~ . . . . . B~ ]
(A + B=F,C=, B,).
and L e m m a 2 follows that for each i ~ k, with Theorem 2. Let ~ and j satisfy the hypotheses
i # j, there exists a path in G~ from node j to of Theorem 1, and write ap(7,) for the uncontrol-
node i if and only if
lable polynomial of (AF, Bj).
i. For all F E ~,
Ck-s(A/" - A )-'B. # O, s ~ k(j)
v p(C._.,A,B.) divides aF(A) (10)
semi)
Since {(C._s,A,B.):s ~ k(j)} is the set of all
complementary subsystems of `2 containing
input channel ], we can state ii. If, in addition, for each i E k with i # j there
is a path in Gx from node j to node i, then a F(A )
L e m m a 3. The transfer matrices of all c o m - is independent of F E ~_j and
plementary subsystems of `2, which contain input
aF(A)divides H p(Ck_.,A, BJ (11)
channel j, are nonzero if and only if for each .Gk(D
i ~ k, with i # j, there exists a path in the graph
of `2 from node j to node i. Theorem 2 implies that the uncontrollable
spectrum of (AF, B~), namely the root set of
aF(A), contains the root set of v O(Ck-,, A, B.)
Proof of Proposition 3. Let (Ap, Bj) be o~kU)
controllable. Fix s E kfj); then as a fixed subset for all F E 4. The theorem also
implies that if node j of Gz is connected by a
CA + `2 B,F,C,I~.~CA + "2 B,F,C, If#j).
t Ek--s I Ek directed path to each other node of Gz, then the
Since (A + `2 B,F,C,, Bj) is controllable, it uncontrollable spectrum of (Ap, Bj) is contained
follows that CA + `2 B,FaC,[~.) = ~. in another fixed subset, namely the root set of
i ~ k---a H p(Ck-.,A, B.), whenever F is selected so
i E I~/)
If C . - . ( M - A )-'B. ffi O, then C . - . A 'B. ffi O, i >~
0; thus ( A i ~ . ) C k e r n e l C.--. and (AI~I.)= that d((A~l~j)) is as large as possible. From the
CA + `2 B,F~C,I~.). Therefore ~ E k e r n e l C.-. obvious relationship between the two fixed
| Ek--s subsets bounding the root set of at(A), it follows
or equivalently, C~_, = 0. This clearly contradicts that if node j of Gz is connected to each other
(2); therefore C k ~ A I - A ) - * B . # O . Since s E node of Gx, then F can be selected so that (Ap, Bj)
k(j) is arbitrary, it must be true that the transfer is stabilizable just in case each polynomial p(C~_,,
matrices of all complementary subsystems of `2 A, B J , s E k ( ] ) , is stable.
containing input channel j, are n o n z e r o . It In the sequel, use will be made of the following
follows from L e m m a 3 that Proposition 3 is true. remarks, which pertain to a n y linear system ((7, A,
[] B).
4.84 J . P . CORFMAT and A. S. MORSE
and write aF.~(A) ="c.p. (A~,+ B,K)I(AF + B,) = u.p. (A + B2FTC2, BO. From this and
B,KIB, ). Since u.p. (AF+ B,K, B J is clearly Lemma 4 applied to the system (Cz, A, B J with
independent of K, we can write c.p. (AF + D = Bz, it follows that u.p. (AF., B 0 divides (u.p.
B,K)=aF.K(A)(u.p. (A~, BO). But c.p. (A~+ (A, B3))(p(C2, A, B,)). Hence the lemma is true
B,K) can also be written in factored form as for k = 2.
~(A)~SF.x(A) where ~(A) == ^ c.p. (A~+B,K); Let p ~>2 be fixed and set q = p + 1. Suppose
~.K that the lemma is true for k = p. Assume that the
thus
lemma's hypothesis holds for k = q; i.e. by
a~.x (),)(u.P. (AF, B 0) =/z (X)/SFX(A), Lemma 3,
(F, K) ~ 3 ~ ( ~ (21)
Cq_.(AI-A)-'B.O s E q(1) (24)
Since aF.X(X)=c.p. (A~ + B , K ) restricted to
Write ~ " (resp. ~ , ) for the linear space of all
the controllable space of (AF, B,), it follows that
matrices F" - diag. [F,, F, . . . . . Fp ] ( resp., F,).
for each fixed F ~ ~, a K E g can be found for
Fix s* E p(l); then s* E q(l). Since
which aF.~(A) and ~(A) are coprime. This and
(21) clearly imply that
C,_,. [ C , ,.],
= LC~ J (25)
/t(A) divides u.p. (AF, B,), F ~ 3; (22)
it follows from (24) that
In addition, since by hypothesis u.p. (A~,BO
divides 8 (X) for F in some robust set ~ C :~, ~ (A)
must divide 8 (A). [C~-"] ( k I - A)-'B.. # 0 (26)
By Lemma A.1, the set ~ of pairs (F, K ) E
~ ) ~ such that ~,,.~ (X) and 8(X ) are coprime, is Since s* U {q} E q(l), (24) holds for s ffi s* U {q}.
robust in ~(]~g. Since ~* is robust in ~;, so is
Using the fact that q - (s* U {q}) = p - s*, we can
_~ ~ _~*. Therefore _~ n ((~ n ~ * ) x ~ ) is robust
therefore write
in ~ . This means that there exists a pair (Fo,
Ko) such that F o E _~ n _~* and ~FoX,(X) and
C)-,*(AI - A )-'[B.., Bq ] ~ 0 (27)
8(A) are coprime; but u.p. (AF, B J divides/~(A)
for F ~ _ , so fiFo~o(X) and u.p. (A~o, B 0 are
From (26), (27) and Lemma 5, it follows that
coprime. It follows from (21) that u.p. (A~., B,)
there exists a robust set _~(s*) C ~o such that
divides ~z(A); hence by (22), u.p. (A~o, B0 =
with Fq E _~(s*),
/z(A). But Fo ~ _~* and from (20), degree (u.p.
(AF, B,)) is independent of F ~ _~*. Therefore
C , ~ ( A I - A - B,FqCq )-'Be ~ 0 (28)
degree (u.p. (AF, BJ)=degree tt(X), F ~ _~*.
But this and (22) can be true only if u.p.
and
(AF, B J = ~z(A)forF ~ _~*. Fq.
p (C,-.., A + BQF~C,,B,.) divides
Lemma 7. Let 2 "1 {C,, A, Bi; k} be ~xed and
suppose that for each i E k, with i ~ 1, there
p([CT],A,B,.)p(C,-..,A,[B..,B,]) (29)
exists a path in Gx from node 1 to node i. Then
u.p. (Av, B,) divides (u.p. (A, b~)) Since s* is arbitrary, the preceding statement is
valid for all s* E p(1). Thus (28) and (29) hold for
(.y~, p(C,_., A, B.))
all s* E IR1) and all Fo in the robust set
n
whenever F E is selected so that d((Ad~,)) is a* ~ re(I)
is no less than n for all real and complex values standard Euclidean metric, it follows that E t is
of A. Since rank (hi - A) = n for all A not in the open in Ex, with respect to the restriction of this
spectrum of A, it is easy to see that the rank metric to Ex. Thu~ if ~ can be made single
condition on M(A) is equivalent to the require- channel controllable and observable with decen-
ment that no prime factor of c.p. A be a factor tralized feedback then so can any system
of every n th order minor of M(A ), the minors of E Ex, provided ~ is sufficiently close to 2. In
M(A) being viewed as polynomials in ~. Since other words, the problem of finding an F E 3~
the coefficients of each of these polynomials for which 2p is single channel controllable and
depend algebraically on the parameters of (C, A, observable, is well-posed at any point in E~.
B), it must be that the resultant of each prime We would like to be able to assert that 'almost
factor of c.p. A together with each n th order every system' in Ez is in E t , i.e. that E~ is
minor of M(A ), is a polynomial in the parameters dense in Ez, but to prove this seems to be a
of (C, A, B). From this it directly follows that the difficult matter. Note that the usual approach, i.e.
class of systems in ~ which are not complete, is a show that E~ is the complement of a proper
variety in 7. Since ~ always contains at least one variety in Ex, will not work here since in general
complete system (e.g. select ( C, A, B) E ~ with Ex is a proper variety in g rather than a linear
(A, B) controllable and C(AI - A )-~B # 0), the subspace. Thus the problem of deciding whether
variety is proper. The complement ~ * of this or not joint controllability, joint observability
variety, which is the set of complete systems in ~, and completeness are generic properties of
is therefore robust. This implies that .~* is open systems in Ex, remains unsolved.
and dense in ~f (with respect to the standard
Euclidean metric), and thus that completeness is Remark. The topological constraints imposed
a generic property of single-channel systems in on the graph of 2 can lead to some unusual
~f. consequences. For example, if - {C,, A, B,; 3}
Suppose we now write g for the linear vector is a two-dimensional, three-channel, jointly
space of all k-channel systems {CY~'~, A ~'~'~, controllable, jointly observable, strongly con-
B~'~"o; k}. Using essentially the same arguments nected system possessing a transfer matrix with
as above, it is quite straightforward to show that zero diagonal blocks (i.e. the transfer matrix
if ~g* is the set of jointly controllable, jointly from channel input i to channel output i is zero,
observable, complete systems in ~g, then ~g* is i E 3), then the two eigenvalues of A must be
robust in ~. Thus one can also state that equal.
joint-controllability, joint-observability and com-
pleteness are generic properties of systems in ~g. 4. S T R O N G L Y CONNECTED SYSTEMS
From a practical point-of-view, such an
By making use of the preceding results it is
assertion can be very misleading in that it is
now possible to obtain complete and construc-
based on hypotheses which fail to take into
tive solutions to the spectrum assignment and
account possible dependencies between model
stabilization problems for strongly connected
parameters. For example, if there really is no
systems, W e begin by reviewing some basic
internal physical connection between the ith properties of linear systems,
channel input and the jth channel output of the
It is well known that the closed=loop spectrum
process modelled by ~, then the transfer matrix
of a linear system (C, A, B), resulting from the
between channel input i and channel output j of
application of (centralized) linear control, con-
must be exactly, not approximately, zero, and
sists of two subsets: one is controller dependent
the class of admissible model parameter pertur-
and freely assignable with dynamic controllers of
bations should preserve this property, if conclu-
sufficiently high order[17]; the other, called the
sions based on perturbation analysis are to make
unassignable spectrum of (C, A, B), is controller
sense.
independent and is therefore a fixed subset of the
With this in mind, suppose we call two systems
closed-loop system's spectrum, no matter what
in ~, topologically equivalent if both systems
the controller design. It is also known that if
have the same graph. For the topogical equival-
a(A) is the characteristic polynomial of A rood
ence class Ex of ~ ~ g to contain at least one
( A I ~ ) and if/3(X) is the characteristic polyno-
complete system, it is necessary (cf. Lemma 8)
mial of A restricted to ( A [ ~ ) f 3 [C[A], then the
that be strongly connected. Assume, therefore,
root set of a(A)~(A) coincides with the unas-
that ~ is strongly connected; and write E !
signable spectrum of (C, A, B).t It is natural to
for the set of all jointly controllable, jointly
observable, complete systems in E~. Since E |
tNote that the roots of a(A) mKl ~(A) are the 'uncontrolla-
coincides with the intersection of ~* and Ez, and ble' and 'controlhble but unobservM~le' eilBmVMues of (C, A,
since _g* is open in E, with respect to the B ), respectively.
call a(A)jS(A) the unassignable polynomial of Theorem 5. Let E - {Ci, A, Bi; k} be a jointly
(C,A, B). controllable, jointly observable, strongly con-
For fixed ~ - {C~, A, Bi ; k}, it is possible to nected system. Then with decentralized control
characterize the unassignable polynomial of (Cj, i. the spectrum of E can be freely assigned if
At, Bj) as a function of j E k and F E ~. For and only if ]~ is complete.
this, let k* denote the set of all proper subsets of ft. ~ can be stabilized if and only if for each
k, and write _~l for the set of F E for which s E k*, p(Ck-,, A, B.) is a stable polynomial.
the rank of the n th order Hankel matrix
Remark. In view of the discussion in 3, it is
c,B, C,A,B, ... evident that the spectrum assignment problem is
well-posed at any solution point within the
n,(j)-| i .,
-
topological equivalence class of E, provided ~ is
strongly connected; using similar arguments it
is as large as possible. Clearly each .~j, j E k, is can easily be shown that this is also the case for
robust in as is the intersection ~ * t f=~ ~ .
the stabilization problem. However, for systems
which are not strongly connected, quite the
opposite can be true as will be illustrated by
Theorem 4. Let E I {Ct, A, B~; k} be a jointly example at the end of the next section.
controllable, jointly observable, strongly con-
nected system. L e t / ~ ( A ) denote the unassignable Example 1. The two channel system described
polynomial of (Cj, AF, Bl) and write az(A ) for the by the matrices
fixed polynomial of ~ as defined by (6). Then
c[0'00,00]
i. v p(Ch-., A, B,) divides ~tvJ(A)
sell* C2=[1 0 1 O]
jEk, FE~
[C~[Av])). From this, (41), (42) and the definition A, B,) are coprime, is robust in ~ . Since
of /z/(A), it follows that /z/(A) divides _~, N _~2 n .~* is robust in ~, it must be that
II p (C~_,, A, B,). This and (40) imply that _~ n ((_~, n _~2 n _~*) ~ ) is robust in ~ ( ~ .
~E( k ( / ) u E ~ j ) This means that there exists a pair (Fo, Ko) such
that F o E _~, n _~2 n _~* and ~p..~(A) and
/z~(A) divides H p(C~-,,A,B.), II p(C,-,, A, B.) are coprime; thus from (43), it
eEIx*
find an F which will make a single channel of Ep triangular structure of the form (48). Corres-
controllable and observable {cf. Theorem 3}. pondingly, the overall transfer matrix of
However it still may be possible to assign or admits a similar block triangular structure; the
stabilize the system's spectrum with decentral- ith diagonal block in the latter is the transfer
ized control. To determine under what condi- matrix of the k,-channel strongly connected
tions this can be done we use standard techni- system {C~, A, Bj; q,} described by the equations
ques from graph theory[21] to decompose Gx
into a number--say p - - o f strongly connected ~,(t)=Ax,(t)+ Z Bjus(t)
14squ
comPonents G~ - ( q , , r,); the subsets q, form a
(49)
partition of k (i.e. k can he written as the disjoint y,(t) = C,x,(t), j E q,
union of the qD; F, is a function mapping q~ into
its power set according to the rule F , ( / ) - { r : Note that in general this system is neither .jointly
r ~ (F(j)) t3 q,}, F having already been defined in controllable nor jointly observable. Nevertheless
2. The reduced graph of 2, written G t, is the realization theory insures the existence of a
pair (qX Fro) where qR t {q,, q2. . . . . qp} and F m is jointly controllable, jointly observable system
the mapping from qS into its power set, defined ~' - {Cj, A,, Bj ; q,} with the same overall transfer
in such a way that q, E F*(q~) if and only if there matrix as (49). We call E' the ith strongly
exist nodes ie ~ qj and ]o ~ q~ such that ie connected subsystem of E.
FOe). We say that a component G~ is isolated if We wish to relate the closed-loop spectrum of
I ~ (q,) is the empty set and q, ~ U I ~ (q~). Thus to the closed-loop spectrum of the k-channel
|q system determined by the ~', when identical
G~ is isolated if no arc enters or leaves the set q,. local controllers ~a are applied to both systems.
Observe that the reduced graph G a contains no For this write A(~, k) for the closed-loop
circuits (i.e. if q, ~ Fa(q~), then q~ ~ FS(qD; spectrum of ~, and with identical controls
thus we may assume that the set q - { q , , applied to the 2', let A,(Ej, q,) denote the
q2. . . . q~} is ordered in such a way that closed-loop spectrum of Y/.
Observe that the block-triangular structure of
r " (q,)C{q,, q~ . . . . . q,_,}, i =2,3 ..... p the transfer matrix of 2, with the transfer matrix
(47) of E' being the i th diagonal block, implies that the
parameter matrices describing 2 can be written
Associated with E, consider the k x k matrix as
Tx defined by , = [ ~ , , x . . . . . x], q ~ q,)
[!x
= [ 0 if i ~ F0){i.e. if C,(M-A)-'Bj =0} c,=[o, 6,,x ..... xl, q ~ q0
tu [ I otherwise
C,=[O ..... O,d,], (j ~'q,)
We call Tz the topology matrix of 2; clearly Tx is
the transpose of the incidence matrix[21] of Gz,
and it completely characterizes Gx. .~2 x
Let k, i E p E { l , 2 . . . . . p}, be the number of A= 0
elements in q,. Observe that is always possible to
relabel the channels of E in such a way that the X
set q, contains the k, first integers, q2 the next ks, 0 A,
etc.; to this relabeiling there corresponds a
permutation of the rows of Tx, together with the
same permutation of the columns, and when this Bj-- U E q,),
is carded out there results a matrix T[ which,
because of (47), has the block triangular form
T,, . . . T,p ]
B~ = (j E q=) . . . . .
0 T= . . . T~.
T~[ = (48)
0 . . . 7",,
where x-marks denote matrices of no impor- one of the strongly connected subsystems of X.
tance to the present analysis, and ~' =" {(~j, ~,,/~j; If the resulting decentralized control is then
q,} is a system with the same overall transfer applied to the original system, the spectrum of
matrix as X~ -{C'j, A,, B~; q,}. Since X' is jointly the resulting closed-loop system will coincide
controllable and jointly observable, the parame- with the disjoint union of Ao and the assigned or
ter matrices of ~ can be written as stabilized spectra of each of the p strongly
connected subsystems. A short calculation
~'~= [0, C'.x],(j E q,) shows that the total dimension of the decentral-
ized control system needed to achieve stability
or to assign a prescribed spectrum, never need
(j Eq,) exceed n, the dimension of X.
L: o A,. The preceding observations can be summar-
ized in the following way.
Note that the set Ao= tJ (or(A,,) t3 or(A,2)) is the
Theorem 6. Let X = {C, A, Bt; k} be a jointly
complement of the union of the open-loop
controllable, jointly observable system. Then with
spectra of the ~' in the spectrum of X.
decentralized control, the closed-loop spectrum
For each j E k, let $~=H~zj+L~y~, us=
of X can be
Mjzj + F~yj denote a local control system Xs to be
i. freely assigned if and only if the sum of the
applied to the jth channel of X. Write A*~ for the
dimensions of X' s strongly connected subsystems
system matrix describing the overall closed-loop
is equal to the dimension of ~ and spectrum
system which results when controllers E~(j E
assignment is possible for each strongly con-
p,) are applied to X'. It is easy to see that the
nected subsystem.
system matrix A* associated with the closed-
ii. stabilized if and only if the complement of
loop system resulting from application of the Xj
the disjoint union of the spectra of ~'s strongly
to X, can be written as
connected subsystems in the spectrum of E is
stable and stabilization is possible for each
./~*] X X strongly connected subsystem.
0 x x
A*= 0 Theorem 6 together with its predecessor,
Theorem 5, thus provide explicit conditions for
". A* the solvability of both the spectrum assignment
and stabilization problems for k-channel linear
where systems.
So far in studying the spectrum assignment
and stabilization problems, we have implicitly
A*= IA'''
0
0 0 Atz
assumed the system data to be exact. However,
dil~culties may arise if the preceding results are
applied without qualification. To illustrate this,
Clearly ~ ( A * ) = AoO Or(A*). We are led to the consider the two-channel system with overall
following proposition which characterizes the transfer matrix
closed-loop spectrum of X in terms of the
closed-loop spectra of its strongly connected 1 1
subsystems. s-I s-l-e
0
S
Proposition 4. For aH linear controls Xj, j E k,
I o )- .J (50)
between channels. On the other hand, both the
stability and dynamic response of an entire
decentralized system are determined only after
with the a, positive. Then two possibilities arise: all local controllers have been applied, at least
i. The parameter a, exactly equals a2, the for strongly connected systems, and so if
equality remaining valid for all physically mean- channel couplings are 'strong', final system
ing perturbations of the process data. In this case tuning from just one channel would appear to
the process is truly of dimension two and it make sense.
makes sense to apply the theory developed Although some of the preceding solvability
above to stabilize the system: the closed-loop conditions undoubtedly hold generically, the
system will remain stable for small perturbations corresponding nongeneric cases cannot be ig-
of the process data. nored. For example, if a strongly connected
ii. The parameters at and a2 are equal to system is in some sense 'close' to being
within computer tolerances, but are not exactly incomplete, then it may be better to adopt an
equal. It then makes little sense to represent the incomplete model for purposes of design.
process by a model of dimension two and apply The subject of decentralized dynamic control
the theory to that model; for if a decentralized is quite new and it is evident from the preceding
control is designed on such a basis and then discussion, that there is much work to be done.
applied to the process, instability will inevitably
occur. This, of course, is because the process
really requires a model of dimension three, and REFERENCES
the theory should be applied to such a model to [1] J. P. CORFMAT: Decentralized Control of Linear
correctly predict that the pole at a, cannot be Multivariable Systems. Yale University Becton Center
Technical Report CT-67, October, 1974; also Yale
shifted and therefore that the process cannot be University Doctoral Dissertation, December, 1974.
stabilized with decentralized control. [2] T. A. M~OtSHS~: Computation in Organizations: The
Now suppose that in place of (50), the process Comparison of Price Mechanisms and Other Adjust-
is initially modeled in state space form; i.e. 2 is a ment Processes. Decision and Organization, 237-281.
North Halland, Amsterdam (1972).
minimal realization of (50). Then 2 will be of [3] J. M~HAK and R. RADNeR: The Economic Theory o[
dimension two if at = a2, and dimension three if Teams, Yale University Press, New Haven (1971).
[4] G. B. DA~rrzloand P. WotJ~: Decomposition principle
at ~ a,. Suppose now that all parameters of for linear programs. Operations Res. II(l), 1960.
are perturbed subject only to the constraint that [5] O. I. ELOmtD: Electric Energy Systems T~ory: An
the perturbed system has the same topology Introduction. McGraw-Hill, New York (1971).
[6] C. W. SANDERS,E. C. TACKEaand T. D. LINTON:A new
(graph) as ~. Then it is easy to check that by class of decentralized filters for interconnected systems.
perturbing the parameters of 2 ever so slightly, it IEEE Trans. Aut. Control AC-19, 259-262June (1974).
is possible to obtain a system for which [9] A. Omtsso and B. J. KAaA~N: Mutual synchronization
of geographically separated oscUlators. Bdl Systems
stabilization is possible, even f o r case ii when Tech. J. 1689-1701, December (1966).
a, ~ a,. Thus because of computer inaccuracy, [8] I. W. SAmmmto: On conditions under which it is
for example, one might well conclude that possible to synchronize digital transmission systems.
lk,U Systems Tech. J. 46(6), 1999-2020(1969).
stabilization is possible when in fact it is not. Of [9] R. LAU, R. C. M. P m t ~ s o and P. P. VAttA~A:
course the problem here is that the class of Decentralized information and control: a netwo~ flow
admissible perturbations is too large. example, IEEE Trans. Aut. Control AC-17, 446-473.
Decentralized control of linear multivariable systems 495
[10] M. Aord and M. T. L~: Partial reconstruction of state complement y is open and dense in 3; with
vectors in decentralized dynamic systems. I E E E Trans. respect to the standard Euclidean metric; it is
Aut. Control. AC-Ig(3), 289-294 (1973).
[11] K. C. Cwo: Optimal decentralized regulation for a strong natural to call ~ a robust subset of 9;. Note that a
of coupled systems. I E E E Trans. Aut. Control AC-19(3), robust subset is never empty.
243-246 (1974). If ~ and V are robust in ~ , then so is their
[12] D. McFxDogN: On the controllability of decentralized
macroeconomic systems: the assignment problem.
intersection ~ N ~ . If ~ and ~_ are robust in
Mathematical Systems Theory and Economics 1, vector spaces ~ and q~ respectively, then the
221-240. Springer-Vering (1967). Cartesian product ~ x ~ is robust in the 'outer'
[13] J. P. CO~'~UATand A. S. MORSE: Stabilization with
decentralized feedback control. I E E E Trans. Aut. direct sum 3 ; ~ .
Control AC-I~6), 673.682 (1973). If av: 3~ -~ reals is a polynomial function in the
[14] M. Aora: On dynamic assignment problems and coordinates of F E 3~, relative to a fixed basis,
decentralized stabilization policies. UCLA, Dept. of
System Science Technical Report, 1974.
we sometimes say that aF is algebrically depen-
[15] M. Aord: On feedback stabilizability of decentralized dent on F. If My is a matrix with all entries
dynamic systems. Automatica 8(2), 163-173 (1972). algebrically dependent on F and if rank MF ~> r
[16] S. H. WANOand E. J. DAY,SON:On the Stabilization of
decentralized control systems. I E E E Trans. Aut. Con- for some F E 3~, then the set of F for which
trol AC-18(5), 473--478 (1973). rank MF ~> r is robust in ~. From this it follows
[17] F. M. BP.ASCH,JR. and J. B. PEARSON:Pole placement that if (Cp, AF, B~) is a linear system with
using dynamic compensators. I E E E Trans. Aut. Control
AC-~SO), 34-.43 (1970). coefficients algebraically dependent on F, then
[18] J. P. CmWMATand A. S. MoRSe: Spectrum assignment the set of F for which d((AFI~F)) is maximal, is
with decentralized feedback control. Proc. 1973Prince- robust in 9;. It also follows that if the transfer
ton Conference on Information Science and Systems,
228-231, March 0973). matrix of (C~, AF, Bp) is nonzero for some F,
[19] J. P. C O ~ ^ T and A. S. MORSE: Control of linear then the set of F with this property is robust in
systems through specified input channels. SIAM J.
ControlandOptimization 14(1),163-175,January(1976).
[20] A. S. MoRSe: Structural invariants of linear multivari- Suppose that av(A ) is a real monic polynomial
able systems. SIAM Z Control 11(3), 446-465. August with coefficients depending algebraically on
(1973). F ~ 3~. It is always possible to write a ~ ( A ) =
[21] R. G. BUSACK~aand T. L. SAA'rY: Finite Graphs and
Networks. McGraw-Hill, New York (1965). /z(A)/~v(A) where /z(A) = ^ av(A); ~F(A) is a
[22] W. M. WONH~ and A. S. MoRSI~:Decoupling and pole polynomial, with coefficients depending algebrai-
assignment in linear multivarible systems: A geometric
approach. S I A M J . Control $(1), 1-18, February (1970). cally on F, satisfying
^ /h~(~)= l (A.l)
FE~
APPENDIX
Properties of robust sets L e m m a A.I. Let y(A ) be any fixed polynomial.
In this appendix we discuss certain properties The set of F E ~ for which ~/(A ) and 13~(A) are
of robust subsets which are used throughout the coprime is robust in 3;.
paper. The assertions which follow can be easily
Proof. Write 3'(A)= H 3,~"(A) where the 3,~(A)
justified with very elementary arguments. With t-!
the exception of L e m m a A.I, proofs will are the prime factors of ~(A). For fixed i, there
therefore not be given. must exist F 3~ for which ~v(A) and ~!(A) are
For our purposes, a subset V of a real coprime; if this were false, y~(A) (being a prime)
n-dimensional vector space ~ , is a variety in ~ , would have to divide /3p(A) for all F E 3;,
if the coordinates of each element of V, relative contradicting (A. !). Therefore, the set ~, of F for
to a fixed basis for ~ , coincide with a zero of a which the resultant of y~(A) and ~j,(A) is
finite system of (real coefficient) polynomial nonzero, is nonempty and therefore robust in ~.
equations in n-indeterminates, and visa versa. If It follows that the set of F for which 3,(A) and
at least one polynomial in the system is not the
/3v(A ) are coprime, namely ~ ~!, is robust in 3;.
zero polynomial, then V ~ 4 ; in this case, V is a |-I