Citation for published version (APA):
Trentelman, H. L. (1986). Families of linear-quadratic problems : continuity properties. (Memorandum COSOR; Vol. 8603). Technische Universiteit Eindhoven.
Document status and date: Published: 01/01/1986
Document Version:
Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers)
Please check the document version of this publication:
• A submitted manuscript is the version of the article upon submission and before peer-review. There can be important differences between the submitted version and the official published version of record. People interested in the research are advised to contact the author for the final version of the publication, or visit the DOI to the publisher's website.
• The final author version and the galley proof are versions of the publication after peer review.
• The final published version features the final layout of the paper including the volume, issue and page numbers.
Link to publication
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal.
If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please follow below link for the End User Agreement:
www.tue.nl/taverne
Take down policy
If you believe that this document breaches copyright please contact us at: openaccess@tue.nl
providing details and we will investigate your claim.
Department of Mathematics and Computing Science
Memorandum COSOR 86 - 03
FAMILIES OF LINEAR-QUADRATIC PROBLEMS
CONTINUITY PROPERTIES
by
HARRY L. TRENTELMAN
Eindhoven, The Netherlands February 1986
by
HARRY L. TRENTELMAN
Abstract. In this paper we investigate for g1ven one-parameter families of linear time-invariant finite-dimensional systems the parameter depen-dence of the linear-quadratic optimal cost, optimal control inputs, optimal state-trajectories and optimal outputs. It 1S shown that results that have been obtained in the past in the context of the problem of
'cheap control' can in fact be generalized to a much broader class of parameter dependent cost-functionals, including cost-functionals in which for every parameter value the weighting matrix of the control inputs is singular. Essentially, only two assumptions on the parameter dependence of the cost-functionals are required in order to have continuity of the
optimal cost and optimal control inputs with respect to the underlying parameter. One assumption is concerned with the continuity of the weighting matrices with respect to this parameter, the other with the monotonicy of
the weighting matrices with respect to the parameter. Instrumental in our development is a characterization of the linear-quadratic optimal cost in terms of the so-called dissipation inequality. The results obtained are applied to the problem of 'cheap control' and to a problem of 'priority control'. The latter provides an example of a family of quadratic cost-functionals with a polynomial parameter dependence.
1. INTRODUCTION
It is a well-known fact that optimal control inputs and optimal state trajectories for the infinite horizon linear-quadratic time-invariant optimal control problem are in general distributions. The corresponding optimal cost may be obtained by calculating the maximal solution to the so-called dissipation inequality. These issues have been the subject of detailed studies as for example [1], [2], [3] and [4].
We will consider the finite-dimensional linear time-invariant system
Ax(t) + Bu(t)
xeD)
(1 • 1 a)y (t) Cx(t) + Du(t) (1.1b)
together with the quadratic cost-functional
'"
f
lIy(t)U2 dt • (1. 2)D
Here, it will be assumed that u(t) E 1Rm, x(t) E 1Rn and yet) E
m
P . 1t will be a standing assumption in this paper that the paIr (A,B)lSstabilizable with respect to
C-
:= {sEel
Re s <a}.
The linear-quadratic problem associated with the system (A,B,C,D) 1S
the problem of finding the infimal value of the cost-functional (1.2) with respect to an appropriate class of control inputs u and to calculate, if it exists, an optimal one. The problem is called regular if the mapping D appearing in (1.1b) is injective and singular if this is not the case. In the latter case in general the optimal controls will no longer be smooth functions (as in the regular case) but will be distributions.
Traditionally, one method of trying to capture the singular case has been to regularize the cost-functional (1.2). Instead of minimizing (1.2) one then minimizes its 'regularization'
00
J
"yet)" 2 2 2 + E "u(t)" dt. (1 .3)Of course, minimizing (1.3) is the same as solving the linear-quadratic problem for the system
(A,B,(g)'(~I»
which, for c > 0, is a regular problem. For E=
0 the original problem is recovered. In the past, it has been the subject of a considerable amount of papers to characterize the behavior of the optimal cost, the optimal control inputs and the optimalstate-trajectories as the parameter E tends to zero ([5] to [17]).
In the present paper we shall consider the more general situation that the mappings C and D depend more or less arbitrarily on a real parameter E
lying in some closed interval, say [0,1], and study the continuity proper-ties of the optimal cost, the optimal controls and the optimal state-trajectories as functions of this real parameter. More precisely, we shall consider the system (1.1a) together with the output equations
y (t) = C(E)X(t) + D(s)u(t) , E
with associated cost-functionals
£: E [0,1] ( 1 .4)
E E [0,11 (1. 5)
and study the behavior of the optimal cost, optimal controls and optimal state-trajectories as functions of E. In (1.4) for each £ E [0,1], C(E)
and D(d are assumed to be linear mappings from lRn to lRP and from lRm to lRP respectively.
We stress that 1n this set-up £ no longer needs to have an
interpreta-tion in terms of 'small control weighting'. In fact, no (uwumpf/"ions uyill be made on the injeetivity of the mappingD D(c). Thus, our context will. also capture for example the situation that for all E E [0,1] the linear-quadratic problem for (A,B,C(c),D(c» is singular.
2. A CLASS OF ADMISSIBLE INPUTS
In this section we shall spec ify the class of control inputs with respect to which we will perform the minimization of the cost-functionals
(1.5). In the sequel, let V denote the testfunction space of all smooth
functions in JR with compact support. Let V' be the distribution space of
all continuous linear functionals on V (for details on distributions, see
[18]). Let
V:
denote the subspace ofV'
of all distributions with supportin JR+ • Let £2 1 (JR +) be the space of locally square-integrable functions
, oc
on JR+. This space may be identified with a subspace of V~ by defining the
+
functional value of 1J! E £2 1 (lli) at q:> E V to be • oc
<1J!.q:»:=
J
1jJ(t)q:>(t)dt . JR+Distributions in £2 1 (JR+) will be called regular distributions. Linear
• oc
combinations of the Dirac distribution 0 and its higher order derivatives
will be called impulsive distributions. I f m E :IN then V,m and £m
2 1 <JR+)
+ , oc
will denote the spaces of m-vectors with components in
V'
and £2 1 (JR+)+ , oc
respectively.
A
distributional control input uE
V,m will be called+
admissible if it is in the space U
d1st " of impulsive-regular distributions
defined by
== {u E V~m
I
u = ul + U
z
with ul impulsive and u2 regular} .Given (1.1a), let
and d + (t)
[0,00». If u E V,m then we define the corresponding state-+
trajectory of (1.1a) by +
:= d
Xo
+ K*
u , (2.1)where
'*'
stands for convolution of distributions in V~. We note thatx(xO'u) is in V~n. In fact, i f u E U
dist then x{xO'u) is impUlsive-regular,
i.e. x{xO,u) == Xl +x
2 with Xl and x2 n-vectors of impulse and regular
X(XO,u)(oo) := lim x2(t) •
t-?o:>
(2.2)
Given an initial condition Xo E JRn the subclass of U
dist consisting of all admissible inputs that yield stable state-trajectories is defined by
u
s dlst :ab (x ) 0 := { u E U distI (
x xo,u) ()
O}00
=
.
Of course, if Xo E JRn and u E U
d1st . then the corresponding outputs yc;(xO'U) := C(E:)x(xO'u) + D(E:)u will also be impulsive-regular distribu-tions. Thus, in general the integrals (1.5) are not well-defined. Let
r~(JR+) denote the space of all p-vectors whose components are
square-integrable over JR+. If u E U
d' happens to be such that y (xO,u) rt. r P 2(lR
2 )
1st E
then we formally define
Incorporating this formal convention, we shall now define the optimal cost for the infinite-horizon linear-quadratic problem with stability for the system (A,B,C(E),D(e:» by
J* ( ) e: Xo := In . f { ( Je: XO,u )
I
uE
Ust.ab(x )} dlSt 0n s t a b .
Given Xo
E
JR , a control u;E
Udist(xo) IS called optimat for th problem if3. CONTINUITY OF THE OPTIMAL COST
In this section we shall show that under fairly mild conditions on the functions e: 1-+ C(E;) and E 1-+ D(E;) the function E: 1-+ J:CxO) is continuous
in c = 0 for every Xo E JRn• For the time being, take a fixed E: E [0,1]. Recall that we assumed (A,B) stabilizable. This implies that for all Xo Ust.ab(x ) . dlst 0 IS non-empty and J e Xo *( ) < 00.
Using an argument similar to the one in [3, theorem 6.12] it can be shown that in fact
J*. (X
o)
= inf {J (xO,u)I
u EU~:~bt(xO)
n r
m2 1 OR+n,
c u l S , oc
i.e. the optimal cost can be obtained as the infimum over all regular . . Ustab ( ) 1 . . 11 k ( [ 1 ]) h h . 1 1nputs 1n d' x
O ' A so, 1t 1S we - nown see t at t e opt1ma 1St
cost depends quadratically on x
O' i.e. there exists a nonnegative semi-definite symmetric matrix p+ E llinxn such that
E
(3.1)
Combining these facts it becomes possible to cha.:acterize the matrix p+ c in terms of the so-called dissipation inequality ([1]). Given
(A,B,C(E),D(E», define a map Fc: llinxn -~ lli(n+m)x(n+m) by
F (p) := [ATP T + PA ; CT(C)C(E) B P + D (E)C(E) PB ; C T (c) D ( c)
1
D (c)D(E) (3.2) EA matrix P is said to satisfy the dissipation inequality for the system +
(A,B,C(E),D(c» if Fc(P) ;>, O. I t was proven in [I] that Pc can be characterized as the maximal element in the set of solutions of the dissipation inequality (see also [4] for a quick proof):
LEMMA 3.1. F (p ) + ~ O. Moreover, if P is a nonnegative semi-definite c E
+
x n matrix such that F (P) ;:;
o
then P ::; P •E E
We shall now impose the following assumptions concerning the depen-dency of the mappings C(E) and D(E) on the parameter E. In our first assumption C(E) and D(E) are interpreted as functions from [0,1] to lliPxn
pXm .
and lli respect1vely:
(A. 1) E ~C(E) and E ~D(E) are continuous 1n 0.
Our second assumption deals with the monotonicy in E of the nonnega-tive semi-definite symmetric matrices
o
T
Q(E) := (C(E) D(E) (C(E) D(E) (3.3)
•
o
J
(xT(t)UT(t»Q(E)(~~~~)
dtWe shall assume that in a neighbourhood of 0 Q(s) 1.8 a monotonically nondecreasing function:
(A.2) There exists a 0 > 0 such that for all 0 ~ £1
Q(E:
1) ;£ Q(E:2)·
It turns out that under the above two assumpt ions the optimal cost J:(xO) is continuous in 0 for all xO:
THEOREM 3.2. Assume that (A.l) and (A.2) hold. Let
Xo
E JR • n ThenPROOF. By (3.1),
J~(xO)
=
x6 P;Xo
for all £ E [0,1]. We claim that lim p+ =: P exists. Indeed, as a consequence of the assumption (A.2) the£+0 £ +
function £ ~ P is monotonically nondecreasing on
[0,8].
Since also p+0
E £
for all e, this proves our claim. Next, we note that by the assumption (A.l) the function
(p, E:) ~ F (P)
E:
1S continuous in (p,O) for every P E JRnxn and thus in particular in (P,O).
We may therefore conclude that
Since F (P+) ~ 0 for all E, this yields FO(P} ~ 0, i.e. P satisfies the E: E:
dissipation inequality for the system (A,B,C(O),D(O». By lemma 3.1 we
h f h P < p+ t ere ore ave = O. f · .1.nd Po + ;£ PE: + f or all E:
On the other hand, again by the assumption (A.2) we
+
E [0,8]
and thus also Po ~ P. We conclude that- +
P ;: PO'
REMARK 3.3. An argument similar to the above was used in [4] to obtain the corresponding result for the special case that C(E:)
=
(e)o
andD(E)
=
(~I)
(see also [5] and [9]). Note that for the latter case the assumptions (A.l) and (A.2) are trivially satisfied.REMARK
3.4.
Instead of considering the linear-qudratic problem with stability, one could also consider the version of this problem in which we do not require stability of the optimal state trajectories (see also [2]). Again consider the system (1.1a) with cost-functionals (1.5). Instead ofstab
minimizing over Udist(xO)' now minimize over the (larger) class Udist and define the associated cost by
I t can be shown that under the assumptions (A.1) and (A.2) the function
E ~~(xO) does not need to be continuous in 0 for all x
o'
i.e. theanalogue of theorem 3.2 does not hold for the linear-quadratic problem without stabilUy. In order to obtain a class of counterexamples we shall briefly recall the notion of invariant zero ([20J, [21], [22]). Given a system (A,B,C,D), let
V*
denote the associated output-nulling subspace and let R* denote the controllable output-nulling subspace. It is well-known that there exists a mapping F: lRn ..;. lRm such that (A+BF)V*
cV*
and(C+DF)V*
= {OJ. Moreover, any such F automatically satisfies (A+BF)R* c R* and it turns out that the spectrum a(A+BFI V*/R*)
is independent of these Fls. This fixed spectrum is denoted by a*(A,B,C,D) and is called the set of invariant zeros of (A,B,C,D). Now, let J*(xO') denote the optimal cost for the linear-quadratic problem for (A,B,C,D) with stability and let J*(xO) denote the optimal cost for the problem without stability.
Obvious-ly, in general
J*(x
O) ~ J*(xO)' However, it follows from the results in[2] that if and on ly if J*(x )
o
a*(A,B,C,D) cE-n for all
Xo
E lR {sE
tiRe s ' O}NOW, take any system (A,B,C,D) with D
=
0, (A,B) stabilizable and (C,A) detectable, such that at least one invariant zero lies inc+ := {s E tiRe s >
a}.
Define C(E) :=(g)
and D(c) :=(~I)
(with I the m x m identity matrix. Obviously, for c > 0 the system (A,B,C(s),D(c))defines a regular linear-quadratic problem. This problem is also standard (i.e. CTCc)D(c) = 0) and consequently (see [23]), since (A,B) is
stabilizable and (A,C(e» is detectable, for c > 0 we have J:(x
o)
J~(xO) for all xaE
mn. However, caused by the presence of an invariant zero 1n(:+, there exists an
Xo
E mn such that the strict inequality '17;(xa) < Ja(x
o)
holds. For this Xo we must have J~(xO) 7 Ja(xa) (C
+
0) (by theorem 3.2) and consequently J:(xO)
fr
~(xO) (c+ 0).
4. CONTINUITY OF OPTIMAL CONTROLS AND STATE-TRAJECTORIES
This section 1S cOllcerned with tli(' continuity proper! ies of the
optimal control inputs and optimal state-trajectories. We shall prove that if these exist then considered as functions of E they are continuous in c
=
0, provided that (A.1) and (A.2) hold and provided that the system (A,B,C(O),D(O» is left-invertible.Let us first consider the linear-quadratic problem with stability for
a giv('11 system (A,H,C,D) and recall some facts on the existence imd
ulllque-ness of optimal controls for this problem. The existence of optimal con-trols is known to be closely connected to the position of the set of
invariant zeros o*(A,B,C.D) (see remark 3.4) with respect to the imaginary axis to := {s
Eel
Re s=
a}, while the uniqueness depends on the inver-tibility properties of the system. The system (A,B,C,D) is calledleft--1
invertible (right-invertible) if its transfer matrix C(Is-A) B + D is injective (surjective) as a matrix over the field of real rational func-tions. (A,B,C,D) is called invertible if it is both left- and right-invertible.
The following results were proven 1n [2]:
PROPOSITION 4.1. Consider the linear-quadratic problem with stability for the system (A,B,C,D). For X
o
E: mn , let J*(xO) denote the optimal east.IJ'hen the foUo1.JJing holds:
(i) For aU Xo E mn there exists an optimal control u~ if and only
(ii) For aU
Xo
E:m
n there exists exactly one optirmZ control ~o 1:f and only if a*(A,B,C,D)n
r
O= 0
and (A,B,C,D) Zeft-invertible.( iii) J*(xO) = 0 for aU
Xo
E:m
n if and only if a*(A,B,C,D) c ~-. oThus, returning to our original problem, in order to be able to speak about optimal controls u* for the linear-quadratic problem for
E
(A,B,C(E),D(e» we shall assume that for e sufficiently small these sys-tems do not have invariant zeros on the iITk,ginary axis:
(A.3) There exists
n
> 0 such that for all 0 ~ e ~n
we have a*(A,B,C(E),D(E»n
r
O= 0.
Next we shall label the following invertibility condition on the 'boundary system' (A,B,C(O),D(O»:
(A.4) The system (A,H,C(O),D(O» IS left-invertible.
By proposition 4.1, (A.3) and (A.4) will guarantee that the linear-quadratic problem for (A,B,C(O),D(O» has a un1que solution for every
xo.
A fortiori however, (A.l) and (A.4) imply that for E >a
sufficientlysmall (A,B,C(E),D(E» is also left-invertible. Consequently, we can make the following observation:
LEMMA 4.2. Assume that (A.1) hoZds. Then the following statements are equivalent:
(i) 3 n>O such that for an 0 S
E
;'i n foY' aUXo
E:m
n thelineaY'-quadmUc problem for (A,B,C(d ,D(d) has a unique optirml contI'ot input u*
E,XO
( H) (A . .)) and (A.4) hold. o
Before stating our ma1n result, a few words on the type of convergence we shall consider are at order. A sequence of distributions ~n
E
V~ is said to converge to ~ E V~ in distributional sense if <Wn'~> ~ <W,~> (n ~ 00) for all ~E
V.
In the obvious way, a sequence uE
V,m
is said ton +
converge to u E
V,m
in distributional sense if this convergence holds+
componentwise. Finally, a sequence Y
n E £~(JR+) :is said to converge to
y E £~(JR+) strongly if convergence holds componentwise 1n the topology of
+ £2(JR ).
Now consider the system (l.la) together with the cost-functionals (1.5). For
Xo
E "IRn let:= x(xO,u* )
s,xO and
be the optimal state-trajectory and optimal output for (A,B,C(s),D(s». Our main result is the following:
THEOREM 4.3. Assume that (A. 1), (A.2), (A.3) and (A.4) hold. Then for all n
Xo
E JR we hmJe sense. Moreover, u* s,xo y* s,xO -+ u* O,xO -+ y* O,xO and x* -+ x*s,xo O,xO (£
+
0) in distributional(£
+
0) strongly.In order to establish a proof of this theorem we shall proceed as
+
follows. Let Po be the nonnegative semi-definite symmetric matrix defining the optimal cost for (A,B,C(O),D(O» via (3.1). Again, let FOCP) be given by (3.2). By lemma 3.1, FO(P;) ~ O. Since FO(P;) is also symmetric it admits a factorization
(4.1) (see also [3]). Assume that in (4.1) the matrix (C+ D+) has linearly independent rows. This can always be achieved by taking the number of rows equal to the rank of FO(P;). Thus (4.1) assigns to (A,B,C(O),D(O» a
uniquely defined system (A,B,C+,D+). Consider now the linear-quadratic problem for (A,B,C+,D+). Given Xo E JRn and u E U
dist let 00
J+(xO'u) :=
J
Uc+ x(xO,u) + D+ uU2 dto
if C+ x(xO,u) + D+ u is square-intregrable and J+(xO'u) := +"" otherwise. Let the corresponding optimal cost be given by
The following important lemma relates the optimal costs for the linear-quadratic problems associated with the systems (A,B,C(O),D(O» and
LEMMA 4.4. Let Xo E 1Rn and u E Udist' Then we have J+(xO'u) < <X> if' and
onLy if Jo(xo'u) <: 00. For every u E
U~~:~(xO)
we
have(4.2)
PROOF. A proof of this can be found in [3, lemma 6.21]. o
As an immediate consequence of the above we see that for every
Xo
Emn the linear-quadratic problem for (A,B,C(O),n(O» has a unique solution u* if and only if for everyXo
E 1Rn the problem for (A,B,C+,D+) has aO,XO
unique solution u*
+,xO and that 1n this case uO
*
,xO = u: ,xO for every xo. This fact will be instrumental l.n our further development. In the sequel, a system (A,B,C,D) will be called minimum-phase if its invariant zeros are contained in the open left-halfplane, i.e. if o*(A,B,C,D) cC-
(seealso-[24]) •
LEMMA 4.5. (A,B,C(O),D(O» has no invariant zeros on the imaginary axis if' and onLy if' (A,B,C+,D+) is minimum-phase.
PROOF. Taking infima on both sides in (4.2) we find that J;(x
O)
=
0 for all xo. It thus follows from proposition 4.1 (iii) that O*(A,B,C+,D+) cE-.
Consequently, it suffices to show that o*(A,B,C(O),D(O»
n
CO=
0
if and only if o*(A,B,C ,D ) + +n
to=
0.
However, this follows immediately bycombining proposition 4.1 (i) and lemma 4.4 upon noting that the existence for all KO of an optimal control for (A,B,C(O),D(O» is equivalent to the existence for all Xo of an optimal control for (A,B,C+,D+). 0
REMARK 4.6. The fact that the system (A,B,C(O),D(O» has no invariant zeros on the imaginary axis if and only if the same holds for the trans-formed system (A,B,C+,D+) is in fact a manifestation of the fact that the invariant zeros on the imaginary axis are fixed under the transformation (A,B,C(O),D(O» ~ (A,B,C+,D+) defined by (4.1). This has also been noted
In [25J in the (dual) context of the singular filtering problem.
Our next result relates the invertibility properties of the original system and its transform:
LEMMA 4.7. Assume that (A,B,C(O) ,D(O» has no invar'iant zeros on the imaginary axis. .Then (A,B,C(O) ,D(O» is left-invertible if and only 1:f
(A,B,C+,D+) is invertible.
PROOF. It follows immediately from proposition 4.1 (iii) and the fact that J:(xO) = 0 for all Xo that (A,B,C+,D+) is right-invertible. Thus it suffices to show that (A,B,C(O),D(O» is left-invertible if and only if (A,B,C +,D+) 1S left-invertible. Again, this follows by combining proposi-tion 4.1 (ii), lemma 4.4 and lemma 4.5 upon noting that the existence for all
Xo
of a unique optimal control for (A,B,C(O),D(O» is equivalent tothe existence for all
Xo
of a unique optimal control for (A,B,C+,D+).0
REMARK 4.8. The above lemma is in fact also true without the premisse that (A,B,C(O),D(O» has no invariant zeros on the imaginary axis. A proof of this claim can be given using the geometric characterization of left-invertibility in terms of controlled invariant subspaces (cf. [22, ex.
4.4]. Since we do not use this stronger result here, the proof is omitted. Before proving our main theorem we still need one more preliminary result. In the following, a left-invertible system (A,B,C ,D) will be
called strongly detectable if for any
Xo
E: JRn and any u E: Udist such that Cx(xO,u) + Du = 0 we have x(xo'u) (00)=
0 (cf. (2.1) and 2.2».The latter definition generalizes the one in [3, def. 6.7], where a system is called strongly detectable if the above property holds for
regular inputs u. With our definition the following extension of [3, tho 6.8] holds.
LEMMA 4.9. Asswne that (A,B,C,D) is left-invertib . Then (A,B,C,D) 1"8 strongly detectable i f and only i f it is minimum-phase.
PROOF. Obviously, if (A,B,C,D) is strongly detectable in the sense of
our definition, it is also strongly detectable in the sense of [3, def. 6.7]. Thus, by applying [3, tho 6.8] we find that (A,B,C,D) 1S m1n1mum-phase.
Conversely, assume that (A,B,C,D) is minimum-phase and let u E:
U
d1St .n
and
Xo
E: lR be such that Cx(xO,u) + Du = O. The distribution x(xO,u) can be written uniquely as x(xO,u) = Xl + xWe want to show that lim x
2(t) = O. Now, S1nce u E Ud. it has a unique
t-7<X> ~st
representation u u
l + u2 with ul impulsive and u2 regular. Define x(O+)
:=
lim x2(t). It is well-known that x2
=
x{x(O+),u2) (see e.g. t+O[19, prop. 2.5]), i.e. x2 is the state-trajectory emanating from initial
condition x(O+) and regular control u2• Since obviously CX2 + DU2 = 0 it therefore follows from [3, tho 6.7] that x2(t) + 0 (t + 00). o
PROOF OF THEOREM 2 In this proof, take an arbitrary but fixed
Xo
and* * *
*
..
y by u , x and y • Let
°
> 0 be suff~c~ently~,xO ~ ~ ~
denote u* ,x* and
€,XO €,xQ
small such that Q{c
1) Q(£2) for alIOS £1 ~ £2 ~o (cf. (3.3» and such that for all € E [0,0] the linear-qudratic problem for (A,B.C(~),D(£» has a un~que solution u* (cf. lemma 4.2), Define
y
:= C x* + D+ u*. By lemma£ c + £ £
4.4 and the monotonic assumption (A.2), for all € E [0,0] we have
J (xO,u*) + .::
00
J
IIC(O)x: + D(O)u*U 2 dt - J~(xO) ~E: 0
"" ;;;;
f
i1C(s)x* + D(s)u*U2 dt - J~(xO)E: £
0
J:(x
o)
*=
- JO(xO).
r +
Consequently, by theorem 3.2, y£ + 0 (£
+
0) strongly in £2(ffi ) (here r~s equal to the number of rows of (C+ D+». A fortiori this implies that y£ + 0 (£
+
0) in distributional sense. Now defineG (t) +
By definition of y£ the following convolution relation holds:
*
Since, by (A.3), (A.4) and lemma 4.6, the system (A,B,C+,O+) is invertible, -1 the convolution operator with kernel G+ has an inverse with kernel G+
(G-1 is an impulsive regular distribution: it can be characterized as the
+
-1 -1
inverse Laplace transform of [C+(Is-A) B + D+l ). It follows that
and therefore, SInce
Y
E -+ 0 as a distribution, u* -+
E (E '" 0) (4.3)
1n distributional sense. Note that, since
c:
1 is impulsive-regular, u is admissible, i.e. ~ E Ud1st ' • Also, it follows from (4.3) that ~ and Xo yield zero output for the system (A,B,C+,D+), i.e. C+ x(xo'~) + 0+ ~
=
0, or equivalently J+(xO'~)=
O.Now, we contend that in fact u is optimal for the Linear-quadratic problem for (A,B,C+,O+). In view of the foregoing, to show this it suffices to prove that
~ E Uds~abt(xo)'
Indeed, since (A,B,C ,0 ) is left-invertibleIS + +
and minimum-phase by lemma 4.9 it is strongly detectable. Thus. since u and Xo yield zero output. we have x(xo'~)(ro)
=
0 or, equivalently,~ E Ustab ( dist Xo . )
Obviously, by lemma 4.4, u is also optimal for (A,B,C(O),O(O» and must therefore be equal to the unique optimal control u
O'
This proves thatu;~ -+ u*
€ 0 (e
+
0) In distributional sense.To show that x~ -)0- x;, note that x~
verges to d+ Xo + K
*
u~
=x;.
Finally, we will show that y: -+ y~ consider the inner product space £~(JR+)
00
<Yl'Y2>:=
J
y;(t)Y2(t)dto
+
d Xo + K
*
u •*
E The latter
con-(e '" 0) strongly. To show this, with inner product
We will first show that y: -+ y~ weakly in this space, i.e. <y~,y> -+ <y~,y> for all y E £~(JR+). By theorem 3.2 we have
00 00
J
IIY~1I2
dt •o
Thus, there is a 0 > 0 such that the set {y* lEE [O,6]} is bounded in £:
.c~(JR+). Let (£n)nElN be a sequence in [0,6] such that En -+ 0 (n -+ 00).
The sequence (Y~ ) ElN is then bounded in .c~ (JR +) and therefore has a
sub-n sub-n +
sequence that converges weakly to, say,
y
E .c~(JR ).On the other hand, y*
=
C(E)X* + D(E)U* which converges indistri-£ £ £
butiional sense to C(O)x~ + D(O)u~
=
y~ by the foregoing. Thusy
=
y~,from which it 1S easy to see that y* -+ yO* (£
+
0) weakly. The proof is E:now completed by noting that
00 00 00
J
lIy: _y~1I2
dtJ
II *112 dt - **
J
II *112 dt= yr:; 2<ys'YO> + yo
.
0 0 0
The right-hand side of this equality converges to
o
as f:+
O.5. APPLICATIONS: CHEAP CONTROL AND PRIORITY CONTROL
In the present section we will apply the very general result of theorem 4.2 to some important special cases.
o
First we shall consider the special case that is commonly referred to as the problem of 'cheap control' ([5] to [17]). Consider the system (1.1) together with the cost-functionals (1.2) and (1.3). As before, assume that (A,B) 1S stabilizable with respect to
t-.
As noted in the introduction theabove falls within our context by taking C(e) :=
(g)
and D(£) :=(~l)'
Obviously in this case (A. I) and (A.2) are satisfied. COlldition (A.4)
requires that (A,B,C,D) is left-invertible. In the following, let n
<ker C
I
A> :=
n
i=l
ker CAi - 1
be the unobservable subspace of (C,A). The next lemma states that (A.3) is equivalent to the requirement that (A,B,C,D) has no invariant zeros on the imaginary axis and that (C,A) has no 'unobservable poles' on the imaginary axis:
LEMMA 5.1. (A.3) satisfied for
(A.B'(~)'(~I»
If and only i fc n )'<
PROOF. In this proof denote o*(A,B,(O)'(cr») by 0c' tt suffices to show
that a* = a(A
I
<ker CI
A» for all c > O. Let V* denote the output-nullingC
e
D Ssubspace of the system (A,B,(O)'(s!»' We claim that for s > 0
V:
= <ker eI
A>. Indeed, let Xo EV:.
There is a regular input u such that(g)x(xO'u) + (~I)u
=
O. This implies that u = 0 and that ex(xO'O)=
a,whence Xo E <ker e
I
A>. The converse is also immediate. Finally, sinceG* = G(A+BF
I
V*) for every F£ such that (A+BF )V* c V*, we conclude thatc £ c £ £
a* = G(A
I
<ker eI
A» (just take F = 0 upon noting that <ker C ! A> is£ £
A-invariant). o
Now, glven Xo • E lRnl , e t U
*
,,<*
b ' 1 ho
,~ ,xa ,~ and YO,XO*
e optima. for t e.
cost-funr·tional (1.2) and u* ,x* and y for the cost-functHmal
£,xO c,xo c,xO
(1.3). We have the following corollary:
COROLLARY 5.2. There exists 0 > 0 such that for all £ E [0,0] and Xo E mn u* exists and is unique if and only if (A,B,C,D) is left-invertible,
c,xo
a*(A,B,C,D) ncO
= 0 and
G(A! <ker CI
A» n CO0.
Tn that case IvC have u* -r u* and x1< -r xo* (£ -} 0) in
c,xa O,xa c,xO ,xa
diBtributional Bense and y:,XO -r y~,xo (c -} 0) stI!ongly.
We note that in the above for £ > 0 the linear-quadratic problem for (A,B, (CO) , (£D
r
»
is regular and that therefore u* c,xO and x* c,xa are regularfor all xo. For c
=
0 the problem might become singular. Also note that asufficient condition for o(A ! <ker e
I
A» n to = 0 to hold is that (e,A) is detectable.Finally, we will discuss the following application. Again consider the system (1.1a). Instead of the output equation (l.'lb) however, consider the situation that we have N+1 output equations:
yO(t) = Co x(t) + DOu(t) y 1 (t)
,-
C1 x(t) + Dl u(t)
Here, we assume that yk(t) takes its values in, say, JRPk. For E > 0, define the following cost-functional:
00
1 2 2N N 2
Uy (t)U + ••• + E Hy (t)" dt. (5.2)
Keeping in mind that E 1S a small positive real number, the cost-functional
(5.2) could be interpreted as assigning to the various outputs specified in (5.3) weightings of different 'orders'. In practical control problems this kind of cost-functional could be used 1n systems with several to-be-controlled outputs to reflect the fact that there are different orders of priority in controlling these outputs ('priority control'). We would like to characterize the behavior of the optimal closed loop system as E
becomes small. First note that the above situation is captured in our context by taking
C{E:) := DCE:) (5.3)
Again, (A.l) and (A.2) are satisfied. Condition (A.4) is equivalent to left-invertibilityof (A,B,CO,D O)' Denoting Co DO ~ C 1 Dl C := D := C N DN
we have the following lemma:
LEMMA 5.3. Let C(E) and D(E) be given by (5.3). Assume that (A,B,CO,DO) is left-invertible. Then (A.3) is satisfied if and only if
*
a
*
~ ~ 0a (A,B,CO,DO)
n
t =0 and a
(A,B,C,D)n
t= 0.
*
-*
PROOF. Let V and V denote the output-nulling subspaces of
£
(A,B,C(E),D(E» and (A,B,C,D) respectively. First note that, since
(A,B,CO,DO) is left-invertible, (A,B,C,D) is left-invertible and there is 6>
°
such that for all E E [0,6] (A,B,C(c),D(c» is left-invertible.Consequently a*(A B C(£) D(E» = a(A+BF
I
V*) for any F such that, ' " £ £ £
(A+BF £)V; c: V~ and a*(A,B,e,D)
=
a(A+BFI
0*)
for any F such that (A+BF)O* c:0*.
Now, we contend that for all £ > 0 we have V~=
v*. Let E > O. ThenXo
E
V:
if and only if there1S
a regular u such that C(dx(xO'u) + D(E)u O. However, the latter equality is equivalent to ex(xO'u) + Du=
O. ThusXo
E
V:
if and only ifXo
E
0*.
From this it*
*
-follows that, for £ > 0, a (A,B,C(E),D(E» = 0 (A,B,C,D). This proves our
lemma. 0
Finally, for E E [0,1] and
Xo
E JRn, let J~(xO) denote the optimal cost associated with the cost-functional (5.2) and denote by u* x*E,XO' £,xO and y* the associated optimal control, state-trajectory and output.
E,XO
COROLLARY 5.4.
(ii)
(iii)
There exists ~ > 0 such that for all E E
[0,6]
andXo
E ~n u* E,XO exists and is unique if and only if (A,B,CO,DO) is left-invertible,
*
0 J ~ ~ 0o (A,B,CO,D
O) n t
=
I
and o·CA,B,C,D) n t=
0.
Asswning this to be the case, for aU
Xo
E JRn-we
have u* -+ u"'£,xa O,xO
and x* -+ x* (£
+
0) in distributional sense and£,XO O,xO
y:,xo -+ y~,xo (E
+
0) strongly. o6. CONCLUSIONS
In this paper we have investigated the parameter dependence of the optimal cost, optimal inputs, optimal state-trajectories and optimal outputs associated with a fixed linear system and a quadratic
cost-functional depending on a real parameter. The questions we have considered were mainly inspired by similar ones that have been studied before in the context of the problem of 'cheap control' or the 'nearly singular optimal control problem'.
It was shown that under two assumptions, namely continuity in c
=
0 of the output mappings C(c) and D(E) appearing in the cost-functional, and monotonicy in a neighbourhood of E = 0 of the quadratic forms defined bythese output mappings, the optimal cost is indeed continuous in E = O. This result generalizes and extends corresponding results obtained before in the 'cheap control' context. We stressed that our result is valid independent of any assumptions what so ever concerning regularity or singularity of the underlying linear-quadratic problems.
Under the same assumptions as above it was shown that as E tends to zero then the optimal control inputs, optimal state-trajectories and optimal outputs for the linear-qudratic problems with E strictly positive converge to those associated with the linear-quadratic problem for E
=
0, provided that the optimal controls, state-trajectories and outputs exist and are unique for c sufficiently small (including E=
0). Again, no assumptions were made on the regularity or singularity of the linear-quadratic problems. An important tool in our development was the dissipa-tion inequality as introduced in [1] and studied before in [3] and [4].Finally, we applied our very general results to derive the
correspond-~ng convergence results in the context of the 'cheap control' problem and in the context of an optimal control problem of 'priority control', wbere the cost-criterion was polynomially in powers of c2•
Many questions concerning the subject of this paper still need to be resolved. A very interesting open problem is the following. Again consider the systems (A,B,C(c),D(c». Assume that for c > 0 sufficiently small the optimal controls u* exist and are unique but that for c
=
0 the optimalc
controls exist but are not unique. This situation occurs if and only if there exists 6 > 0 such that o*(A,B,C(c),D(c»
n
~O 0
for c E [0,6],(A,B,C(c),D(c» is left-invertible for c E (0,6] and (A,B,C(O),D(O» is not left-invertible. We may then pose the following questions: do the optimal controls u* (c
+
0) converge as E tends to zero? If they do, is€
their limit an optimal control for the linear-qudratic problem associated with c
=
O? Is it possible to give a characterization of the limit in terms of the system (A,B,C(O),D(O»? The above questions are closely connected with the existence of a nonzero controllable output-nulling subspace of the system (A,B,C(O),n(O»). In [16] some preliminary results concerning these questions were obtained, again in the context of rcheapcontrol'. The general problem however seems to be quite difficult and its solution will undoubtedly involve extra assumptions on the £-dependence of the mappings C(£) and D(€).
REFERENCES
[1] J.C. Willems, "Least squares stationary optimal control and the
algebraic Riccati equation", IEEE Trans. AutOrrKlt. Control, vol. AC-16, pp. 621-634, 1971.
[2] J.C. Willems, A. Kitapyi & L.M. Silverman, tlSingular optimal control: a geometric approacb", SIAM J. ContI'. & Opt., vol. 23. no. 7. 1986. [3] M.L.J. Hautus & L.M. Silverman, "System structure and singular
control", Linear Algebra AppZ. 50, pp. 369-402, 1983.
[4] J.M. Schumacher, "The role of the dissipation matrix in singular optimal control", Systems & Control Lett., vol. 2, pp. 262-266, 1983. [5] B.A. Francis, "The optimal linear-quadratic time-invariant regulator
with cheap control". IEEE Trans. AutorrKlt. ControZ. vol. AC-24, pp. 616-621, 1979.
[6] B.A. Francis & K. Glover, "Bounded peaking in the optimal linear regulator with oheap control", IEEE Trans. Au tOrrKlt. Control, vol. AC-23, pp. 608-617, 1978.
[7) R.E. O'Malley Jr. & A. Jameson, "Singular perturbations and singular arcs - Part
1",
IEEE Trans. AutOrrKlt. Control, vol. AC-20, pp. 218-226,1975.
[8] R.E. O'Malley Jr. & A. Jameson, "Singular perturbations and singular arcs - Part 2", IEEE Trans. Automat. Control, vol. AC-22, pp. 328-337,
1977 •
{9] A. Jameson & R.E. O'Malley Jr., "Cheap control of the time-invariant regulator", J. f1ppZ. Math. Opt •• vol. 1. pp. 337-354. 1975.
[10] R.E. O'Malley Jr •• "A more direct solution of the nearly singular linear regulator problem", SIAMJ. ContI'. Opt., vol. 14, pp. 1063-1077, 1976.
[11] H. Kwakernaak & R. Sivan, "The maximally achievable accuracy of linear optimal regulators and linear optimal filters1l, IEEE Trans. Automat. ControZ, vol. AC-17, pp. 79-85, 1972.
[12] B. Friedland, ItLimiting forms of optimal stochastic linear regula-tors", Trans. ASME (J. Dyn. Syst. Meas. Contr.), sere G, vol. 93, pp. 134-141, 1971.
[13] S. Godbole. "Comments on: 'the maximally achievable accuracy of linear optimal regulators and linear optimal filters ttl, IEEE Trans.
Automat. Control, vol. AC-17, pp. 577, 1972.
[14] T. Fuji, "On the perfect regulation of optimal regulatorslt
, System
&
Control Lett., vol. 1, pp. 356-359, 1982.[15] J.C. Willems, "Almost invariant subspaces, an approach to high gam feedback design - part 1: almost controlled invariant subspaces", .IEEE TPans. Automat. Control, vol. AC-26, pp. 235-252, 1981.
[16] J. Grasman, "On a class of optimal control problems with an almost cost-free solution", IEEE TPans. Automat. Control, vol. AC-27, pp. 441-445, 1982.
[17] P. Sannuti, "Direct singular perturbation analysis of high-gain and cheap control problems", Automatiea, vol. 19, pp. 41-51, 1983. [18] L. Schwartz, Theorie des Distributions, Hermann, Paris, 1978.
[19] H.L. Trentelman, Akmost ln7~riant Subspaces and High Gain Feedback, Doctoral Dissertation, Rijksuniversiteit Groningen, 1985.
[20] H.H. Rosenbrock, "The zeros of a system", Int. J. Contr., vol. 18, pp. 297, 1973.
[211 H. Aling & J.M. Schumacher, "A nine-fold canonical decomposition for
linear systems", Int. J. Contr., vol. 39, pp. 779-805, 1984. [22] W.M. Won ham , Linear MUltivariable ControZ: a Geometrie Approach,
second edition, Springer Verlag, New York, 1979.
[23] V. Kucera, "A contribution to matrix quadratic equations", IEEE Trans. Automat. Control, vo1. AC-17, pp. 344-347, 1972.
[24] H.L. Trentelman. "A note on stabilization by dynamic high-gain output feedback". IEEE Trans. Automat. Cnntr'oZ, vol. AC-30, pp. 578-581, 1985.
[25] J.M. Schumacher, "A geometric approach to the singular filtering problem", IEEE Trans. Automat. Control, vol. AC-30. pp. 1075-1082,