Almost disturbance decoupling with bounded peaking
Citation for published version (APA):Trentelman, H. L. (1986). Almost disturbance decoupling with bounded peaking. SIAM Journal on Control and Optimization, 24(6), 1150-1176. https://doi.org/10.1137/0324070
DOI:
10.1137/0324070
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.
ALMOST DISTURBANCE DECOUPLING WITH BOUNDED
PEAKING*
HARRY L. TRENTELMAN’
Abstract. Thispaperisconcerned with a generalization of the almost disturbance decouplingproblem by state feedback. Apart from approximate decoupling from the external disturbancesto a first to-be-controlledoutput, we require asecond output tobeuniformly bounded with respect to the accuracyof
decoupling. The problemis studiedusingthegeometric approachtolinear systems. Weintroducesome newalmost controlledinvariant subspacesand studytheirgeometricstructure. Necessary and sufficient conditionsforthesolvabilityof the aboveproblemare formulatedintermsofthesecontrolledinvariant
subspaces.Aconceptual algorithmis introducedtocalculatethe feedbacklaws needed to achieve thedesign purpose.
Key words, almost disturbance decoupling, almost invariant subspaces, linear systems, geometric approach, high gain feedback,output stabilization
AMS(MOS)subject classifications. G3-B28,G3-BS0, G3-C05, G3-C15, G3-C35, G3-C45,G3-C60
1. Introduction. In this paper we are concerned with the problem of almost disturbancedecoupling bystatefeedback asintroducedbyWillems
[20].
Thisproblem dealswiththesituation in which we cannot achieve exactdecoupling from the external disturbances to an exogenous output channel as, for example, in[22],
but only approximate decouplingwithanydesireddegreeof accuracy.In general,thefeedback gain necessaryto achieve this will increase asthedesireddegree of accuracyincreases. Itmay then happenhoweverthatsome of thestate variablestendto peakexcessively.Itisofconsiderablepracticalinterest toknow whenit ispossibleto achieve disturbance decoupling within any desired degree of accuracy, while this peaking phenomenon will not occur.
The systemthatwe will be considering in this paperis given by the equations Yc Ax
+
Bu+
Gd,(1.1)
z
Hx,
z
H2x,where the control
u(t),
thestatex(t),
the disturbanced(t)
and the outputsz(t)
andz2(t)
are real vectors offinite dimensions. Wewill assume that the vectorz2(t)
is an enlargement ofz(t),
i.e., thereis amatrix M such thatH
MH2.
Iffor any positive real number e afeedbackmatrixF,
canbe chosen such thatintheclosed loop system with zeroinitial condition, forall disturbancesd(.)
in the unitballofLp[O, )
wehave(1.2)
IIz ll
<-thenwesay that for the system under consideration the
Lp-almost
disturbance decoup-ling problem from d toz
is solvable. After choosingF
to achievethis approximate decoupling, the outputz2(t)
ofcoursedependsoneand,forcertain disturbancesd(.),
itmay then happen that
z=ll
,.o-
ase 0,i.e.,asthe accuracy of decouplingincreases.* Receivedbythe editors April28, 1983,and in revisedformJune1, 1985.
"
Departmentof Mathematics andComputing Science,EindhovenUniversityofTechnology,5600MB Eindhoven, the Netherlands.Thisresearch wassupported bytheNetherlands Organization forthe Advance-mentofPureScienceResearch(Z.W.O.).As
an example, considerthe system(1.1)
withA= 0 B= G=
0
H1-"
(1
00),
HE
10 Define afeedbackmatrix
F
byItcanthen be verified that the impulse response from thedisturbance d to zl isgiven
by
+3
9W
e(t):’-"
nle(’--JF)tG
e-3t/e(1-t+--
t2
k e
282
/and that
Wl,
e.Hence,
for any 1-< p-<o,
theabove feedbackmatrixF
achievesLp-amost
disturbancedecouplingfrom d toZl.Onthe otherhand, however,the impulse responsefrom d to z2 iscalculatedto be1+3
9t2
t e+2e
2W2
(t):=
HEeA+a)tG
e-3/2._7.
t2
2e 27 81t2
---
+
--andit canbeverifiedthat
o(1 / e)
->3ase->0, i.e.,wehaveobtained almost disturbancedecoupling from dtoz
atthecostof highlyundesiredpeaking behaviour of the outputz(t).
The question which we ask in this paper is this: When is itpossible to choose
F
such that simultaneously(1.2)
holds and there exists a constant C (independentof
e)
such that
for
all disturbances d(.)
inthe unit ballof
Lp[0,
)
wehave1.3)
z=
<--
c
for
all e? That is, the outputz2(t)
is bounded uniformly as e tends to zero. Ifthis behaviour isachieved,wesay thatwehaveLp-bounded
peaking from dto z2. Problemsofthis kindhavebeenconsidered before.Francisand Glover
[3]
considered abounded peaking problemin the context ofcheap control. More recently, Kimura[9]
found conditions that guarantee bounded peaking in the context of perfect regulation. We willstudy the above problem using the bynowwell known concepts of almost controlled invariantand almost controllabilitysubspace[19],
[20].
Wewillalsousetheapproach offrequency domain description of geometric concepts asinitiated in Hautus[5].
The outline ofthis paperis as follows. In 2we will introduce somenotational conventions usedin this paper andstate somepreliminary results andbackground.
Section 3 contains a description of the mainproblemwe willbe concernedwith in thispaper. In 4 we will introducethedisturbancedecouplingproblemwithoutput stability. Thisproblem is anextensionofthe
(exact)
disturbance decoupling problemas treated in
[22].
Its solution willbe needed to solve our mainproblem, but isalso importantin its ownright.In 5we willderive anecessaryconditionforthe solvability of(ADDPBP)p.
This condition willbe in the form ofa subspace inclusioninvolving an almost controlled invariant subspace. Section 6 contains an investigation ofthe geometric structure of the almost controlled invariant subspace that was introduced in 5. In 7 these structural results will be used to prove that for certain classes of systems the subspace inclusion derived in 5 in fact constitutes a necessary and sufficient condition for solvability of(ADDPBP)p.
The sequences of state feedback maps thatachievethe design purposewillbeconstructed explicitly.Section 8 contains somecorollaries ofour main resultand some extensions. In 9 a numerical example is worked out to illustrate the computational feasibility of our theory. Finally, the paper closeswith someconcluding remarksin 10.Several technicaldetails ofproofs in this paperare deferred to AppendicesA,
B andC.2. Preliminaries andbackground. Inthis section we will introduce some notation used in this paper and review somerelevant facts oncontrolled invariant and almost controlledinvariantsubspaces. Alsosomebasicfacts onthe convergence of subspaces willbe given.
2.1. In this paper the following notation willbe used: If is a normed vector space, we will write
I1"
for the norm on.
If l"[0, c), is a measurable function, thenwe will denote[[/(t)ll
pdt ?-|esssup k. t>o lip if 1=<
p<
c,
ifp=.
If
II/11
<
o,
we will say that lLp[O, oo).
If M is a square matrix thentr(M)
will denoteitsspectrum. IfA1
andA:
aresetsof complexnumbers thenA1
u)A:
willdenotetheir disjointunion. Foranypositive integer n, we will denote _n :=
{1,
2,,
n}.
Considerthe system
(1.1).
Letu(t)
:=",
x(t)
:=
",
d(t)
:=q,
z(t)
Lrl
:=Rp, andz(t) Lr
2:=Rp2. Let
A, B, G,
H1
andH2
be real matrices of appropriatedimensions. We will write ’[i := ker
Hi
(i 1,2),
:=imB andAF
:=A+ BF. The reachable subspace will be denoted by(A[
):=
d+A
+...+A"-.
A
collection of subspaces,
2," ",k
will be called a chainin if=
:
=.=
k.
If0#b we will denote :=span b.
If Vc is AF-invariant, the restriction of
AF
to Vwill be denoted byAF[
Wewill write
AF
I./V
or/F
for the quotient mapinducedbyAF
onthefactor space/V (see [22]).
IfV
and are bothAF-invariant
andV,
thenAF[
V/
willdenote the map inducedby
AF[V
onthe factorspace
V/ogr. Wedefinethe canonical projectionP"/V
byPx:=x+
V:If B :=PB,
then(AF, B)
will be called the system induced in/V.
If H"Lr
is a linear map and VckerH,
then H"/V Lr
is defined byHP H. Adistributionf
9’/ (i.e., thespace offinite-dimensionalvalued distributions with support on[0,
oo))
will be called a Bohl distribution ifthere existN0f/t(i)
+f-1
Heref-l(t):=
KeLtM,
vectors
f
and matricesK, L,
M such thatf=
Y,i-(o) denotes Dirac’s delta and (i) its ith distributional derivative,
f
will be called regulariff
0 (i 0,.,
N)
and impulsive iff_ 0.2.2. Wewill now reviewsomebasicfacts from geometriccontroltheory. If cg
is a subspace, then
V*(’[)
will denote the largest(A, B)-
or controlled invariant subspace in Y{and*(Y0
will denote the largest controllability subspacein Yf[22].
IfCg
is asymmetric subset ofthe complex plane C(i.e.,hCg:>
Cg
andCg
contains atleast one point of the real axis),theng*(t
)
willdenote the largest stabilizabilitysubspace in Y(
([5]
or 11]).
Asubspace //’ac willbe calledalmost controlled invariant ifforallXo l/’a and for all e>0 there is a state trajectory
x(.
such thatx(0)=
Xo and d(,, x,(t))<-e for all t. Asubspacea
c willbecalled an almost controllabilitysubspaceifforallXo,
x
Y
there is a T>
0 suchthat forall e>
0there is astatetrajectoryx
(.)
such thatx(O)=xo, x(T)=Xl
andd(Ya,x(t))<-e
for all t. Basic facts on these classes ofsubspacescanbefoundin[19]
or[20]
(see
also[17]). A
subspacea
c isalmost controlled invariant if and only if+
,,
where is controlled invariant anda
is an almost controllability subspace. Asubspacea
is an almost controllabilityk
in such that subspace ifand onlyifthere is a map F"
T
07/ and a chain{
i}i=l=I+AF:+’"
"+AkF-k.
We will say that is a singly generated almost controllabilitysubspaceifthereis amap F"%
a vectorb andaninteger k>
0 such that,
d03AFd"
"03AkF-L
Again, for Y’c
,
T’*(YQ
will denote the largest almost controlled invariant subspace inY"
and*(Y{)
the largest almost controllability subspace in’’.
We will denoteRb*(Y’):=
+
A*(Y’)
andT’b*(Y():= T’*(Yc)
+
b*(Y{).
The subspace//’b*(Y’)
plays an important role inthe problem of almost disturbancedecoupling. In fact, in [20] the following resultwas obtained:
PROPOSITION2.1. Considerthesystem Ax
+
Bu,
z Hx. Thenfor
alle>
0 thereexists a map F’gT-ll such that
IIHexp[t(A+BF,)]GII,<-_e if
and onlyif
imGcT’b*(ker
H).
LetY{:=kerH.The space
Vb*(Y{)
willbe called the space of distributionallyweaklyunobservable states with respect to the output z.
b*’(Y’)
will be called the space of strongly controllablestates with respect tothe output z. Forthisterminology see[6].
Aproof of the following result canbe foundin [1, Lemma 1]. LEMMA2.2. LetY{
.
Then the following equalities hold"(i)
b*(Y’)
f’l Y’-"*(Y{),
(ii)
*(:]’)
f’)V*(Y’)
*(Y0,
(iii)
*(Y()
f)*(Y{)
*(Y0.
[3This paperwill sometimesdeal with a newsystem
(A, BW),
obtainedbydel.eting
the part of the inputmatrixBlyingin*(Y{).
Thissystemisobtainedb.y
taking=
suchthat0]
q (
f3 o//.,(y/))= and by letting Wbeamap suchthat =imBW(see
also
[1]).
The supremal almost controllability subspace contained in Y{with respect tothisnewsystem(A,
BW)
willbedenoted by*(Y{).
Wewillcorrespondingly denoteo
+
A*,(Y[)
by*(Y().
The following result follows from[1, Lemma2]"
LMMA2.3.
Assume
nowthatT"
is(A,
B)-invariant. Let Fbe such that(A
+
BF)
Let
(AF, B)
be the systeminduced in/T"
and P"-
/V
thecanonicalprojection.Wethen havethe followingresult:
LEMMA2.4.
If
t
isan almostcontrollabilitysubspacewith respectto(A, B),
thenPt
isan almost controllability subspace with respect to(AF, B).
Proof
LetPxo
andPx
be inPta,
with Xo,Xla.
There is a T>0 and, foralle>0, atrajectory
x(.)
such thatx(0)=
Xo,x(T)=x
andd(a,X(t))<=e
for all t.It can be seen immediately that
z
(t):= Px,(t)
is a trajectory of the system(AF, B).
Moreover,
z(0)=Pxo,
z(T)=Px
andd(Pa,z,(t))=infRllPr--Px(t)ll
Wewill also needthe following proposition, which is provenin
[17,
Thm.2.39]
(see
also[15]
or[16]).
PROPOSITION 2.5. Considerthe system
Ax
+
Bu.Leta
beanalmost controllabil-ity subspace.Suppose
A is a symmetric setof
dim(A
g)-dima
complex numbers. Then thereis an(A,
B)-invariant subspaceV
anda map F"-
such that VO)(A[)
andr(AF]
)
A.To concludethis section, we shall recall some factson left-invertibility oflinear systems. Again consider the system
Ax
+
Bu,
z Hx.Assume
that the map B is injective.We
willsay that the system(A,
B, H)
is left-invertible ifthetransfermatrixH(Is-A)-IB
is an injective rational matrix. The following resultwas provenin[22,
Ex.
4.4] (see
also [6, Thm.3.26]).
LEMMA2.6.
(A,
B, H)
is aleft-invertible
systemif
andonlyif
*(ker
H)=
0. [32.3. In the following, we will review some basic facts on the frequency domain approachtothe geometric concepts ofthis paper. Wewilldenote
[s]
(respectively,(s), +(s))
for the set ofall n-vectorswhose components are polynomials(respec-tively, rational functions, strictly proper rational functions) with coefficients in R. If
c
=R",
then5’’[s]
(respectively,5’[(s), r+(s))
willdenote thesetofall(s) [s]
(respectively,
(s), +(s))
withthe property thatsO(s)
5’[for alls.Slightly generalizing adefinitionbyHautus[5],
ifforagiven x thereare rationalfunctions:(s)
(s)
and
,
to(s) (s)
such thatx=(Is-A)(s)+Bto(s)
forall s, we willsay that x has a to)-representation.For a description of
(almost)
controlled invariant subspaces in terms of(,
to)-representations,wereferto
[5], [12], [13]
and[17].
Weshall need the following fact:LEMMA 2.7. Let
.
Then we have" x*b(7[) if
and onlyif
x has a(,
to)-representation with
(s)
’’[s]
andto(s)
a//[s].
U
2.4. Finally, we will recall some facts on the convergence of subspaces. In this paper we will use the common notion ofconvergence ofsubspaces in the sense of Grassmanian topology. Let {%; e
> 0)
be a sequence of subspaces of of fixed dimension.Itcanbe proven that%
-
(e
0)
ifand onlyifthereis abasis{v,
,
va}
for Vand there are bases
{vl(e),’’
’,Vq(e)}
of%
such that, for all i, vi(e)vi as e-
0(convergence
in).
Wewillneedthe followinglemma,which canbe provenbystandard means:
LEMMA
2.8.Suppose
Vl,,
Vq areindependentvectorsandv,(e)
-
vfor
alli. Thenfor
esufficientlysmall,vl(e),"
.,
Vqe)
arelinearlyindependent.Consequently,ifV
-
Vand
t’
tV,
whereV
f’)/4/"{0},
thenfor
e sufficiently small%
f’)/V
{0}
and%
O)3. Mathematicalproblemformulation. Considerthe system
(1.1).
Wewill assume thatz2(t)
is an enlargement ofz(t),
that is, thereis a matrix M suchthatH1
MH2
or, equivalently,
(3.1)
kerH2
=:X2
c’’
:=kerH.
Fromnow on,
(3.1)
willbeastandingassumption.Throughoutthispaperwewill also assume thatB is injective.Consider the following synthesis problem. Fix 1-<_p<_-. We will say that the
Lp-almost
disturbancedecouplingproblemwithbounded peaking(ADDPBP)p
issolvable ifthereis aconstant C andfor all e>
0thereis afeedback mapF"
suchthat,with the feedback lawu
Fx
inthe closedloopsystem forx(0)
0for all dLp[O, ),
the following inequalities hold:
(3.3)
IIz=ll
CIIdll,
Notethatif wetake
H
H,
we obtainthe originalL-almost
disturbancedecoupling problem,(ADDP)p,
without the requirement of bounded peaking(see [20]
or[17]).
Anotherinteresting specialcaseis totake
H
=/,whichcorrespondstothe requirement of bounded peakingof
theentirestate vector.Inthe present paper, necessaryand sufficientgeometricconditionsfor the solvabil-ity of the aboveproblemwillbe derived for the casesp 1, p 2 and p oo. Wewill firstshow how the solvability of
(ADDPBP)
canbe
expressedin terms of the closed loop impulse response matrices from the disturbance d to the outputsz
and z,respectively. If
F"
q/isa statefeedback map, then denote the closed loop transition matrixby(3.4)
T( t)
:=e(’+)’and let
(3.5)
’(s)
:=(Is-A-
BF)-’
denote its Laplace transform.Wethen have the following:
LEMMA3.1. Fixp
{1, oO}.
Then(ADDPBP)p
is solvableif
andonlyif
there is aconstant Cand
for
all e>
0 afeedback
mapF"
all such thatIIn
LOlls,-<-e
andIIH=
TOII,
<=
C.(ADDPBP)
is solvable and onlyif
there is a constant C andfor
all e>0 afeedback
mapF
such thatH T
(s)O
andH
’
(s)O
areasymptotically stable and such that supIIH,
L(i,o)Oll
<--
and supIIU=L(i)OII-<-
C.Proof.
Theprooffollows immediately from the fact that for p 1 and for p oo theL-included
norm of the closed loop convolution operator from d toz
equals exactly theL-norm
ofits kernel,i.e.,IIHTGIIL,.
The secondstatementfollows from the fact that the L.-inducednormof theconvolutionoperator from d toz
equals theH-norm
supaIlU,(i,o)Oll
(see,
for example,[23).
[34. Disturbancefleeouplingwithstabilityconstraints. Prior toconsiderations involv-ing the peakinvolv-ing behaviour of the enlarged output z2, we should make sure that the
outputz2is in
Lv[O,
oo)
atall.Hence,
animportant part of thesolutionof theproblemposedin 3is to constructthe required feedback maps
F
insuch awaythat,for any dL[0,
),
inthe closedloopsystemwithx(0)=
0 wehavez2Lv[0,
co).
Therefore,in this sectionthe followingvariation onthe well known
(exact)
disturbancedecoupling problem[22]
will beconsidered. Again, consider the system given by(1.1).
The usual disturbance decoupling problem is concerned with the determination ofa feedback map F" 0// such that in the closed loop system the external disturbance d does not influence a specified output Zl. We will consider the more general situation inwhich simultaneously we demand stability of thesecond, larger, output z2.
Inthis section, Cg, the stability set, willbe agiven subset of the complex plane C which is symmetric. Asymptotic stability is thus obtained by taking C g=
{h
cC"Reh<
0}. A
rational matrix or rational vector is called stable if all its poles areinCg.
Wewillconsider the followingproblem"(DDPOS)
the disturbance decouplingproblem with outputstabilization is said to be solvable ifthere is a feedback map F
such that
H(Is-AF)-G=O
andH2(Is-A)-IG
is stable.Inordertobe abletoformulateconditionsfor the solvability of the above problem, introduce the following subspace:
DEFINITION 4.1.
*Vg(ff’l,
’2)
will denote the subspace of all points xc1
for whichthereis a(:,
to)-representationwith(s)c
l,+(s),
to(s)c
q/+(s) andH2:(s)
is stable.Thus,interpretedinthetimedomain,
Vg(Y(,
’{2)
isthesubspace consisting ofall points in which a regular Bohl state trajectory starts that lies entirely in Y{. The components ofthistrajectory modulo’2
are stable. It follows immediately from the definition thatVg(Y{1, ’{2)
is contained inV*(Y{).
By the assumption(3.1),
if atrajectoryliesin ’{2,the same is truefor’{1.Consequently we alsohavethe inclusion
ff’*({2
(22c/’g(’{l {2)
We notethat Definition 4.1 is ageneralizationof a definitionbyHautus
[5].
His spaceS
(see
[5, p.706])
coincides withVg(Y{1,
’{2)
if ’r is taken to be.
The following theorem canbe provento be completely analogousto [5, Thm.4.3]:
THEOREM 4.2.
(,,
)=
*(c,)+
*().
D
Notethatitfollowsfrom the above theoremthat
Vg
({1,
ff2)iscontrolledinvariant. The nexttheorem provides the key step in thesolution ofDDPOS. The result states thatwhat can be done in Definition 4.1 by open loop control can in fact be done by state feedback"THEOREM 4.3.
ere
existsa mapF" such that(4.1)
AFg(Y, Y2)
Cg(Y,
(4.2)
AF*(Y2)
*(Y2),
(4.3)
(AF
g(Y[,
Y2)/*(Y2))
cCg.
Proof
Duringthisproof, denoteg
:=g(Y,
Y2).
Since*(Y2)
g
andsince both spaces are controlled invariant, they are compatible(see
[22, Ex.9.1]). Hence,
there is a map
Fo’
such thatAo*(Y2)
*(Y2)
andAFog
cg.
Let @:=@
g
and let Vbe anymatrix such that @ imBEConsider the controllability subspace
(AFo[).
Bythe
facts thatg
andAFog
g,
this controllability subspaceis contained inY.
Since any controllability subspaceisalsoastabilizability subspace,it mustbecontained inthe largest stabilizabil-ity subspace(Y)
inY1.
Itthen follows that(Y),
so(4.4)
(Y{)=
.
Next,
we claim that(E{1)
isAFo-invariant.
First, since it is(A,
B)-invariant, we haveAo(C,) (C,)+.
On the other hand,Ao(C,)Ao,
Hence,
again using(K1)g,
we obtainAFo(E{,)((E{1)+)g
(E{1)
+
(
g)=
(E{1).
The last equality follows from(4.4).
Using
(4.4)
and [5, Prop. 2.16], we deduce that the pair(AFol
(E{1),
BV)
is stabilizable.Let
PI"
g
g/*(E{2)
be.the canonicalprojection. Let(AFo,
BV)
be thesystem induced ing/*(E{2).
It can easilybe seen, for example, by using arank test(see
[4]
or[5, Thm.2.13]),
that the latter systemis stabilizable.Hence,
there is a mapF1
onthefactor space such that(AFo +
BVF1)
Cg.
Now,
letF1
beany mapong
such thatF
FP
anddefineF1
arbitrary on a complement ofg.
Define F :=Fo
+
VF1.
Since
FI
*(C:)=
Fol
*(C=) ("
I"
denotes"restrictionto"),
wethen haveAF*(E{=)
*(E{)
and itcan beverifiedthat Fig. 1 commutes.Weare now in apositionto prove themain resultofthis section. THEOREM 4.4. DDPOS issolvable
iff
imGg(E{1, E{2).
Proo
()
Choose Fasineorem
4.3. Then(AFlim
G) E{1,whichyields the decouplingfrom d toz.
FIG.
Let
/2
be as in the Fig. 1 and lettF’.’-AFlt/’g(ff{1,{2)/ff’*(ff{2).
ThenHE(IS-AF)-IG
I2(Is--,F)-IP1G,
which is stablesincetr(AF)
cCg.
(==>)
IfF
is such thatHI(IS-AF)-IG-O
andH2(Is-AF)-IG
is stable then ford let
:(s)
:=(Is-
AF)-iGd
andto(s)
:=F(s).
Then clearly Gd(Is-A)(s)+Bto(s)
and H_(s)is stable. [3Remark 4.5. Ifin the above problem we take
HI
HE
H,
DDPOS reduces to the ordinary disturbance decoupling problem DDP(see [22]).
Inthis case we have,denoting if{:=ker
H,
Vg(7{1,
’{2)
Vg*(20
+
V*(20
V*(2{)
If we takeHI
=0 andHE--H,
we arrive at OSDP as studied in Hautus[5].
The solvability ofthis problem requires the existence of a state feedback F such thatH(Is-AF)-IG
is stable.Necessary and sufficient conditions can be found by noting that
Vg(2{,ff’2)=
Vg*()+
V*(20.
As
also noted in [5], if we takeHI
=0,HE
H and im G=,
the above theorem provides necessaryand sufficient conditions forthe solvability of the output stabilizationproblem, OSP.5. Anecessarygeometric condition for thesolvabilityof
(ADDPBP)p.
Inthis section we shall establish a necessary condition for the solvability of(ADDPBP)p.
This condition will be inthe form ofa subspace inclusion. The proofis rathertechnical and some ofthe details are deferred to Appendix A. In the rest ofthis paper, the stability setwillbe takento be Cg{A
CIRe
h< 0}.
Consider the system
=
Ax+Bu,
z
HlX, z2H2x
and assume that(3.1)
issatisfied. The following subspace willplay an important role inthe sequel:
DEFINITION 5.1.
//’b(ff{1, ’{2)
will denote the subspace of all x X that have a(s
c,
to)-representationwithso(to) ’{(s), to(s) (s)
andH2(s)
isproper and stable. Interpreted inthetime domain,Vb(ff{,
ffQ) consists exactly of those points in that can serve as an initial conditionforsome Bohl distributional trajectory thatlies
entirelyin :7{1,while thevectorof components of the trajectory modulo
2
isthesum ofa stable regular Bohl functionand a Diracdelta.It follows immediately from the definitionand [12,Thm.
4.1]
that//’b(’{, ’tr2)
is contained inVb*(ff{),
thesubspace of distributionally weakly unobservablestates with respecttotheoutputZl.Itisalso immediatethatVg
(ff{,:7{2)
is contained inVb
(’{,
TQ). Weare now in apositionto state the mainresultof this section:THEOREM 5.2. Fixp
{1,
2,oo}.
Then thefollowing holds:{(ADDPBP)p
issolvable}:=>{im Oc T’b(’{1,’{2)}.
Inthe remainder ofthis section we will establish aproof of the abovetheorem. Again,considerthe system Ax
+
Bu,
Zl HlX,z2H2x.
Assume
that fore>
0,u(t)
is aregularBohl input. LetXoX. Let
zl.(t)
andz2.(t)
be the outputs corresponding totheabove input andinitial conditionx(0)
Xo. Denotei.(s)
forthe corresponding Laplace transforms ofzi,(t).
Wethen have the followinglemma:LEMMA5.3. Suppose thateither
of
the followingconditions issatisfied:
(ii)
l,e(S)
andJ2,e(s)
arestablefor
all e, sup,oaII ,.(fo)ll- 0
-0 adthereexists a constantCsuch that supo,en
z,e
(ito)]]
<-Cfor
alle.Then Xoe
Vb
(Y{1, Y{2).
71A
detailed proof of Lemma 5.3 can be found in Appendix A. The idea of the proofisthe following. First we notethat the initial condition Xo abovehas for each e>0 a(,
to)-rcpresentationXo=(Is-A)e(s)+Btoe(s).
Heretoe(s)
isthe (rational)Laplace transform of
ue(t).
Using the asymptotic behaviour as described by the condition (i) or (ii) above, we then analyse the limiting behaviour for e->0 of the sequences ofrational vectorse(s)
andtoe(s).
This willleadto a(:,
to)-represcntationfor Xo with the properties described in Definition 5.1. To conclude this section we apply Lemma 5.3 to obtainthe following:
Proof of
Theorem 5.2.Assume
that(ADDPBP)p
issolvable.Let
Xoeim G. LetFe
be as in Lemma 3.1 and define
ue(t):= FeTe(t)Xo.
Then, depending onp, one of the conditions (i) or(ii) in Lemma 5.3 is satisfied. Itfollows that Xoe Vb(X1,Y{2).
D
6. Thegeometric structure ofSra(f’l,;/’2). Inthe sequel,it will turn outthat under certain assumptionsonthe system
(1.1)
the subspaceinclusioninTheorem 5.2isalso asufficientconditionfor the solvability of(ADDPBP)p.
Inordertoprovethis andto be abletoconstruct the required feedback maps,we need more detailed information on the geometric structure ofthe subspace ]’/’b(Y{1,Y{2)
as introduced in the previous section. In the present section, we will first show that the subspace //’b(Y{,Y{2)
can always be written as the sum of the subspaceg(Y{,
X2) (see
4)
together with an almost controllability subspace depending onY{
and 3’{2. Using this result, we will show thatif either*(Y{1)-
{0}
orY{2{0},
then T’b(Y{1,X2)
admits a decomposition into the direct sum ofWg(X1,
Y{2)
togetherwith anumberof singly generatedalmost controllability subspaces,with aparticular position with respectto the subspaces Y{andY{2. The main resultofthis section willbe the following theorem:
THEOREM 6.1.
Assume
that]*(Y{1)
{0}
orthat3’{2{0}.
Then thereisan integerm’
,
thereare integersrl,,
r,,,,M
andvectorsbl,,
b,,, andthereis a mapF"
->
all such that(6.1)
//’b(Y{1,
Y{E)
o//.g(y{,
Y{2)@
()
’i,i=1 where with j=l
(6.2)
and --2(6.3)
@
AJF-1/i
cY{2. j=lIf in thestatementof the above theoremoneof the integers riissuch thatri-1
<
1 orri-2<
1, then the correspondingsumsin(6.2)
or(6.3)
are understoodto be equal to{0}.
Itwill turn out in theproofof Theorem6.1 that in the case that*(Y{1)= {0}
the integer
m’
may be chosentobe equalto m (=dim).
Inthe case that Y{2{0}
it will appear thatm’
may be chosentobe equalto m dimV*(Y{1)
and alsothatinthis casethe integers rimay be takentobe either1 or 2. Since
]/’g(ffl,
{0})--(see
Theorem4.2)
thetheorem thus states thatT’b(Y[I,
{0})
is equaltothe direct sum ofT’*(Y[I)
together withanumber of singlygeneratedalmostcontrollabilitysubspaces which are equal to eitherspan{hi} (with0#bi
)
orspan{hi,AFb},
with{b,Av-b}
linearly independentand
b
/’1
QThe resultof Theorem6.1 willbeinstrumental inthenextsection, wherewe will considerthe sufficiency of thesubspace inclusion imGc
b(Y(l, Y{2)
for solvability of(ADDPBP)p
andpropose a"scheme" for calculationof the required feedback maps.Intheremainderof the present section we willestablish aproofof Theorem 6.1.
We
introducethe following subspace:DEFINITION 6.2.
?b(Yfl,
if{2) will denote the subspace of all x that have a(:,
to)-representation with(S)fffl[S], (.O(S) 0[S]
andH2(S)
is constant (i.e., if,(s)=Y’,oX,S’
thenHExi=O
for i->1).Interpretedinthe timedomain, b(Y’l,
Y’2)
consistsexactly of those points in that canbe driven to 0 along apurely distributional Bohl trajectory that liesentirely inY(,
while thevector of components ofthis trajectory moduloX2
is a Dirac delta.It follows immediately from the definition and Lemma 2.7 that every point in
b(X, Y2)
is strongly controllable withrespectto the outputz.
Moreover,
it is also immediate that every point x thatis strongly controllable with respectto the outputZ2, is an element of
b(fffl,
if{E).Hence,
theinclusion /b*(fff2)c
b(fffl, Yf)C
b*(Y()
holds. In fact, we havethe following nice result: THEOREM 6.3.
(i)
b(ff/,, Y[2)
+
A(b*(Y/2)
f-)Yfl),
(ii)
T’b([1,
if/2)T’g([,
Y[2)+b(Y[, Y/2).
Proof.
(i) Suppose thatx=(Is-A)s(s)+ Bto(s),
with(s)
[l[S],w(s) ll[s]
N
and
H2:(s)
is constant. Lets(s)=Y.,=oX,
S andw(s)=Y
N+--o uis. Obviously,:(s)=
Xo+S(s)
and w(s)=uo+stoi(s), wheresl(s)X2[s]
andWl(S)
q/Is].Hence,
x=Buo-Axo+
SXo+
s2(s)-As(s)+
BSWl(S)
and by equating powers itfollows that(6.4)
x-Axo
+
Buo,
(6.5)
-Xo(Is-A)l(S)+ Btol(S).
Therefore,
Xob*(X2)
(see
Lemma2.7).
Since also XoY{, we obtain x3+A(b*(’{2)f’iY/’).
Conversely, ifx=Buo-Axo
with Xob*(ff/’2)fqY{1, there iss(s)X2[s]
andtol(S)e ?/Is]
suchthat-Xo=(Is-A)s(s)+Bto(s).
Defining then:(s)
:=Xo+
s(s)
andto(s):=
Uo+
sto(s),
we obtaina(:,
to)-representation of x withs(s)
Y{[s],
to(s)
//Is] andH2(s)
H2xo
isconstant.(ii) Assumethatx T’b(Y{,
Y{2).
Thereis a(:,
to)-representationforx withs(s)
Y((s),
to(s) //(s)
andH2:(s)
properandstable.Decompose s(s)
:l(s)+ :2(s)
andto(s)
tol(s)
+
to2(s),
where:l(s)
andto(s)
are polynomialvectorsands2(s)
andto2(s)
arestrictlyproper. Obviously,
s(s)
fffl[s],s2(s)
ff/’.+(s),
toi(s)a//[s]
andto2(s)
+(s). Moreover, H2s(s)
is constant andH2:2(s)
is strictly proper and stable.Now,
sincetheleft-handsideof this equationisproper and the right-hand sideis a poly-nomial vector, both sides must, in fact, be constant. Thus, there is a vectorx
such that
x=(Is-A)l(S)+Bto(s)=x-(Is-A)2(s)-Bto2(s).
It follows thatXG
?b(Ytrl, ’f2)
and x xl7/’g(Yfl,
r2).
Since xxa +
(x
x),
we obtain that xFg(Yf,
Yf2)+b(Yf, Yt’2).
The converse inclusion follows immediately from the definitions.The importance of the abovetheoremisthatitshows,togetherwithTheorem4.2,
that
Thus,the space
Vb(’(a, ’(2)
can,in principle, be calculated usingexisting algorithms. The stabilizabilitysubspace and the controlled invariantsubspace appearing in(6.6)
can be calculated using the invariant subspace algorithm ISA [22, p.
91]
and a construction as in [22, p. 114]. The almost controllability subspaceb*(Y’2)
can be calculated using the almost controllability subspace algorithm(ACSA)’ [20].
Forany fixed subspace ’’c,
this algorithmis defined by(6.7)
’+1(’/)
+ A(’
(’)
’/);ff(’r)
{0}.
Itis well known, see [20], that
(6.7)
defines a nondecreasing sequence of subspaces which reaches a limit after a finite numberofiterations.Moreover,
this limitequalsff"(ff’)
b*(ff{).
Inthe sequel, denote(6.8)
i(ffCl,
ffLr2):-’-
ffi(ffLr2)
f’)ffLrl
Using the properties of the sequence
-i(ff.)
statedabove, togetherwithTheorem6.3,
the following resultisimmediate:.LEMMA 6.4.
i(’’1,
’’2)is a nondecreasingsequence which reaches a limitafter
afinite
numberof
iterations. This limit equals"(Y{, Y{2)=
Rb*(Y{2)
Y{1. Consequently,(6.9)
b(27{, Yff2)
?""
A"(?7{, YLr2).
[Other properties of the sequence
ff(3’{,
’2)
areproveninLemmaB.1, AppendixB. Using these properties, we obtainthe following lemma:
LEMMA 6.5. Assumethat
*(71)=
{0}.
Then there is a chain {i}’= in anda map F allsuch that(6.10)
b(’{, 72)
AF’"
"AF,,
(6.11)
-
AF-’iC
Xl, i=l(6.12)
)
i-AF
,
i=2(6.13)
dimi-dim
A-
dim(7[, 7’)/-1(’c,
3’c2)
].Proof.
See Appendix B.We are now in a position to establish a proof of Theorem 6.1 for the case
*(C,)
{0}"
Proof of
Theorem 6.1(Case
1"*(’/’)=
{0}).
Duringthis proofwe will denote(7’, [:)
by,
V
(J’, J2)
byV
andVg
(7’1, ’’)
byVg.
According to Theorem 6.5 we have thatV
Vg +
.
We claimthatthe latter sum is a direct sum. Indeed, this follows immediately from the facts thatV
cV*(’[)
and c*(J),
whileV*(’’)
*(’)=
*(’[1)= {0) (see
Lemma2.2).
By Lemma 6.5 there is a chain{@)’=1
in @ andamap Fsuchthat(6.10)
to(6.13)
hold. Let@
be the firstsubspace in thechain which is not zero, i.e.,@
{0)
and@
{0}
forj-14-1,-..,n. Choose a basis for @ as follows. First choose a basis {b,...,b,}
for@.
Extend this to a basis{b,,
b,, b,
,
b,_)
for@_(here,
d
:=dim @i).Continue thisprocedure until wehave abasis for @.AFJi,
V,
the followingvectorsformabasisforb
By
the fact thatdim dimAFbl,
Arba,
AF
-1bd_,
AtF
-
b,AF
-1bd,
AF
-
bdrl,
AFbl,’",
AFbd,
AFbdr+l,’’’,
AFbd,_,’’’,’’’,AFbal,
It may immediately be verified that the above list of vectors can be rearranged to
At’-1
hi},with obtain m subspacesi
:=span{b, Afl,r,-2b,}
cspan
{b,
AFbi,AF
and
Ar-3h
span
{b, A,.
,--,
This completes the proofofTheorem6.1 for the case that
*(ff{1)=
{0}.
Inthe remainder ofthissection, we will set up aproofof Theorem 6.1, the case that if{2
.
{0}.
Inthe following, let beasubspaceof such that[
*(ff{1)]
Let W be a map such that =imBW and let
(Y{1):
+A({),
where(Y{)
denotes the supremal almost controllability subspace contained in ff{ with respectto the system(A,
BW) (see
also Lemma2.3).
Define(6.14)
(Y{)
:=+
A(
if{l).We will show that if {2
{0},
then b(Y{,Y{2)
has a decomposition into the direct sumofg(Y{1,
Y{2)
(which,inthatcase, isequalto(Y{))
andthesubspace(Y{)"
LA 6.6. Let { bea subspace
of
.
en
(6.15)
V(C,, {0})=
(C,
{0)
(C).
Proo
Inthis proof, denoteg
:=g(Y{,
{0}).
Also, let :=*({).
Since({0})
,
itfollows from Theorem 6.3 that(c, (0)
+
+
A[
++A[()
C]
++A[+(
C)].
Now,
notethat1
*(Y{)
(see
[22,Thm.5.5]).
Consequently,A
*({)+
g
+
.
Hencewe.find(c, {0)=
+
+
A(
g
+ 1
+
+
A(
{).
Again, by the fact that
*(Y{1)
cg,
we have(c,, {0})=
+
+
A(
C).
Finally,since
g
*(Y{)
and(Y{1) ({),
itfollows from Lemma2.3 that the sumin(6.15)
is direct.Using the abovelemma we may now obtainthe followingproof of Theorem 6.1, the case that Y{2
{0};
oofofeorem
6.1(Case
2:{2{0}).
Weclaimthat(Y{)=
A(
{).
To
rove
this,assumethat thereis a vector 0 x such that xA,
with avector in Y{. Define := span{}.
SinceA
+
,
is controlled invariant. Since also {,we findthat*({).
Itfollows that*({)
{0}
and hence that x=0. This yields a contradiction.Next,
we claim that dimdim
A(
{1). Assume
the contrary. Then wemayfind a vector 0 xY{
suchthat Ax=0. It follows that span
{x}
is a controlled invariant subspace contained in { and hence that x*(ff{)
{0}.
Again, this is a contradiction.Now,
choose a basis for(ff{1)
as follows" first choose abasis b,...,b
of {1. Extendthis to a basis {b,..., b, b+,...,bm,}
of.
By
the above, the vectors {b,..., b, Abe,..., Abe, b+,..., b,}formabasisfor(Y{).
Thesevectors can berearrangedintoone- and two-dimensional singlygeneratedalmost controllability subspaces with the properties
(6.2)
and(6.3).
This completesthe proof of Theorem6.1. 17. The main result. In the present section we will combine our results of the previous sections to show thatif the system
(1.1)
is such thatit satisfies at least one of the followingtwoproperties:(7.1)
the system(A,
B, HI)
isleft-invertible,(7.2)
the mappingH2
isinjective,then the subspace inclusion imGc
b(’, X2)
is both a necessary and sufficient condition for solvability of theLp-almost
disturbance decoupling problem with bounded peaking(ADDPBP)p
for the values p 1, p 2 andpRecallfrom 5 that forthesevaluesof p thelattersubspaceinclusion wasalready
showntobeanecessary conditionwithouttheextraassumptions
(7.1), (7.2).
Herewe shall, in fact, prove that if either(7.1)
or(7.2)
holds then imGc//’b(/’, ffLr2)
is a sufficientconditionfor solvability of(ADDPBP)p
for
all 1The following resultisthemainresult ofthispaper:
THEOREM7.1. Assumethatatleastone
of
thetwoconditions(7.1), (7.2)
issatisfied.
Let p{1, 2,
oo}.
Then(ADDPBP)p
issolvableif
andonlyif
imIn orderto obtain a proofof the latter statement, we will prove the following"
LEMMA
7.2.Assume
thatat leastoneof
the two conditions(7.1), (7.2)
issatisfied.
Let
T(t)
and"(s)
bedefined
by(3.4)
and(3.5).
Then the following statements are equivalent"(i) Thereexigtsa constantCandasequence
{F;
e> 0}
such thatIIH
TII,-
0(e
O)
andH2
T
G L,--<
CV
e.(ii) ThereexistsaconstantCand asequence
{F;
e>0}
suchthat,for
all e,H
G
and are stable andsupllnZ(i,o)ll-O
(-0)
and supII/-/=L(i,o)ll-<-
C,
.
(iii) imG //’b(ff,
2).
Note that the implications (i)(iii) and
(ii)(iii)
follow immediately fromLemma5.3. Alsonotethat oncewehave proven the abovelemma,aproof ofour main result Theorem 7.1 may be obtainedby combining Theorem 5.2 and Lemma3.1. We stress that the implications (iii)=>(i) in the above, in fact, yields sufficiency of the subspaceinclusionim G
b(’’l, ffLr2)
for solvability of(ADDPBP)pfor
all 1<-_p
<__oo. The idea of the proofof the implication (iii):=>
(i) ofLemma
7.2 is as follows. First we notethat left-invertibility of the system(A, B, HI)
isequivalenttothecondition*(:)
{0} (Lemma 2.6),
while injectivity of the mapHE
is equivalentto’2
{0}.
Thus,under the assumptions ofLemma 7.2, //’b
(’’1, 2)
may bedecomposed according to(6.1),
(6.2)
and(6.3).
Each of the singlygeneratedalmost controllabilitysubspacesLi
appearing in this decomposition will then be approximated by sequences of con-trolled invariant subspaces{;
e>0}.
If we then define:=
//’g0))’l L,
the sequence{o//.;
e>0}
will converge to//’b(’l, ’2).
In this sense, imG is "almost contained" inthecontrolledinvariantsubspace.
Thesubspaceo//.
in turn is"almost contained"in ffr(where
thelatter "almost" should be interpretedintheLl-sense,see also[20]),
while its distancefrom’/’2
is uniformly bounded with respectto e. Using the structure of the above, we will construct a particular sequence of feedback maps{F;
e> 0}
suchthat(A
/BF)
F
F.
Finally itwillbe shown that this sequence has the properties required by (i) and (ii) in Lemma7.2. To startwith, we will show howasinglygeneratedalmostcontrollability subspacecanbe approximatedby control-led invariant subspaces. Let be and let :=g0)’."0)AkF-g.
For i_k and e>0,definevectorsin by
(7.3)
xl(e):=(I+eav)-lb,
x,(e):=(I+eav)-Iavx,_(e).
Notethat thematrixinversions in the above expressionsare definedfor e sufficiently small.
Moreover,
it canbe seenimmediately thatx,(e)
A
-
b(e
0).
Thus itfollows from Lemma 2.8 thatfor
e sufficiently small, the vectors {xi(e), _k} are linearly independent. Foreach e, define a subspacee
by(7.4)
e
:=span{x(e),
.,
Xk(e)}.
Assume
u q/ is suchthat bBu
and define amapFe
"*e
Uby(7.5)
Fex,(e)
:=e-’
u, (i_k).The main properties of the sequences
{e;
e> 0}
and{F
e;e> 0}
are summarized in the following lemma:LEMMA7.3. For _kwe have
x(
e -->Ai-
as e-->O. Consequently,e
->.
Eache
is controlledinvariantand, withFe
defined
by(7.5), (AI
+
BFe)&e
e.
Moreover,
a matrix
of
(A
+
BEe)le
isgiven by(7.6)
Me:
Finally,
for
each e,e
c(A]),
the reachablesubspaceof
(A,
B).
Proof.
The claimx(e)
A-Ib
is immediate. SincethevectorsA-b
are a basis for,
itfollows from 2 and 4thate
.
Using(7.3)
and(7.5),
itmay be verified by straightforwardcalculationthat(Av
+
BFe)x(e)=-=1
eJ-’-xj(e).
Itfollows thate
is indeedAF
+
BFe-invariant
and thatamatrixof the maprestricted towe
isgiven by(7.6).
Finally, to prove thate
is contained in the reachable subspace, make a Taylor expansionto findthat(I
+
eAF)
-1m=0
(-e)mArfl
Itthenfollows immediately thatxl(e)E(AFI)
for all e. The samefollows forX2(e), X3(e)
etc.Wenotethataslightly ditterentconstructionleadingtoanapproximating sequence forasinglygenerated controllabilitysubspacewasdescribed in
[13].
Theconstruction describedby us howeverexhibits animportant propertywhich willbe formulated in the following lemma. The proof ofthis result is straightforward but rathertechnical andwill be deferred toAppendixC.LEMMA7.4. Let :=
)/k=
A-
it
be such that)
k=2
A
i--2F
rl
and)/k=
A-3/
772.
Letx(e)
andFe
beasdefined
above. Then the following holds" there is constant Csuch that
for
all _k:(7.7)
]]H
eA+")’x,(e)llL,O
aseO,
(7.8)
IIH_
e’+)’x,(e)ll,
<-Cfor
alle. [3Now,
in orderto completeaproof ofLemma 7.2,weneedonemorepreliminary result.Up
to nowwe have constructed a sequence of controlled invariant subspaces convergingtoasingly generated almost controllability subspace anddefined afeedback mapon eachof these controlled invariant subspaces.By
applying the decompositiontheorem, Theorem 6.1, and applying the aboveconstruction to each appearing in
(6.1),
we can find a sequence of controlled invariant subspacese
converging to)__m’
.
Intheobviouswaywe can define amapFe
one.
Nowthe question is,can we defineFe
appropriatelyon asubspace complementarytoe
?Thenextconstructiontheorem states that, indeed, we can. It is here that we will use the results on exact disturbance decoupling with stability constraints from 4. In the following,
Vb
:=Vb(:7/.1,
:7/.2),Vg
:=Vg(:7/.1, :7/.2)
andb
:=b(:7/.1, :7/’2)
are denoted:THEOREM 7.5. Considerthesystem
(1.1).
LetAbeasymmetricsetof
dim[((A
)+
Vg)/F
b]
complex numbers. Then there is a mapFI"
-->
71 and a subspace c suchthat the followingconditions are
satisfied"
(7.9)
(7.10)
(7.11)
(7.12)
(7.13)
(7.14)
(A+
BF1)
T’g
cT’g,
A+
BF1)
V*(:7/.2)
V*(:7/.2),
tr(A
+
BFI
/
*(’C2))
C,
(A+
BF)(T’g
))
tr(A
+
BF,
O )/
)
A.Proof.
LetFo"
->
q/ be amap that satisfies the conditions(4.1),
(4.2)
and(4.3)
of Theorem 4.3. Let P"
->
/Vg
be the canonical projection. Let(AFo,
B)
be the system inducedby(AFo,
B)
inthefactor space/Vg.
SinceVb
Vg
+
b
and kerPVg,
wehavePVb
Pb.
By
Lemma2.4andthe factthatb
is an almostcontrollability subspace,itfollows thatP’
b is analmostcontrollability subspacewith respecttothesystem
(/Fo,
/)"
By [22, Prop. 1.2],P(A[)=(.Folim
).
LetA be as above. It can easily be verifiedthat #A= dim[(,Folim
B)/tb].
Thus,we may apply Proposition 2.5 to find an(gFo,/)-invariant
subspace c/Vg
andamap/"
/Vg
suchthat(7.15)
P@
(olim
(7.16)
(AFo +
BF)
Z,
(7.17)
tr(fiFo+//
)
A.Now let c be any subspace such that po and
Lr
fqVg
{0}.
Define a mapFI"
->
0-// byF1
:=Fo
/FP. We contend that the subspace and the mapF1
satisfy theclaimsof thetheorem.To prove(7.9)
to(7.11),
notethatF11
Vg
Fol
Vg.
Theclaim(7.12)
can be proven as follows: From(7.15)
we haveP(Vb
+)=
P(AI3).
Hence,
since
Vg
cVb,t/’b
W--
t/’g
W(Al
tJ ). Assume
xeVb
c.
ThenPxePVbf’={0}.
Thus, x kerPf)
Vg
[-1{0}.
Itfollows thatVb
+
Vb
.
To prove
(7.13),
note by using(7.16)
thatP(A+BF)(Vg)=
P(AFo +
BFP)(
Vg
)Z)
(AFo +
BF1)
cLr
P(
Vg
09).
Finally,(7.14)
follows immediately from(7.17).
Weare now in a positionto complete the proof ofLemma7.2:
Proof of
Lemma 7.2. (i)=:>
(ii). This follows immediately from the fact that theL2-induced
norm of a convolution operatoris bounded from aboveby theLl-norm
ofits kernel
(see,
for example,[2]).
(iii)
=>
(i). Inthis partwe will construct asequence of feedback maps{F;
e> 0}
suchthat,foreachx Vb,Lx
0andH= Lx
-<-
c
forall e,forsome constantC. The construction is divided into five steps"
1. Decomposition.Apply Theorem6.1 tofind adecomposition
i--1
with
=)5__
AF
,
such that(6.2)
and(6.3)
hold.2. Approximation
of
singly generatedcontrollabilitysubspaces. Foreach,
apply the construction(7.3)
to(7.6).
Thus we find vectorsx(e)
(i),
subspaces := span{x(e);
}
and mapsF"
&
0// suchthatMoreover,
by applying Lemma 7.4, there are constants Cj such that(7.19)
[[H1
e(av+BFS)txJ)(e)iiL,’-’>O(e
--->0),
(7.20)
[IH2
e(Av+BFs’)txJ)(e)IIL,
<=
C foralle.3. Composition. Sincethe
Z#s
areindependent,itfollows from Lemma2.8that for e sufficiently small theZ#s
(J
_m’) are independent. Define :=Z#l""
"m’-It
follows that-
.
Define nowF
-
by definingFI
:=(F
+
Fs
(j
_m’).
4. Construction
offeedback
outsidet.
To define a mapon a complement of,
letAcCg
be asymmetric setofdim[((A[
)+
o//,,)/o//,]
complex numbers and apply the constructiontheorem Theorem 7.5to find asubspace c and amapFI"
such that(7.9)
to(7.14)
are satisfied. Inthe remainderofthis proof, denote by.
Wemay then prove the following:LEMMA7.6. Foralle sufficiently small the following holds"
(7.21)
T’g
b
Proof
By Lemma 6.6,T’b3=T’gY.
Since, for each e,(Lemma 7.3),
itfollows from(7.12)
that cT’gY.
Since-->
,
we obtain from Lemma 2.8thatVI (T’g
Y)=
{0}
for e sufficiently small. Theequality,(7.21)
now follows immediately by noting that for e sufficiently smalldim dim
.
5.Definition
of
the sequence{F;
e> 0}.
LetW
be an arbitrary subspace of such that=
T’gY
/4/’. In this (e-dependent) decomposition of defineF"
-->
byF]T’gY:=FIlT’gY,
FI=FI
andF
arbitrary onWecontendthat the sequence
{ F;
e> 0}
defined in thiswaysatisfiesthe condition(i) ofLemma7.2.Toprove this,firstlet x
T’g.
SinceF
T’g
F1
T’g,
wehave by(7.9)
and the fact that
T’g
cY(1thatx(AvlT’g)=
Y(1 forall e.Thus, forall e,H1T(t)x=O
for all t. Let
Av,
andH2
be defined by the following commutative diagram (Fig.2),
in which
P
isthe canonicalprojection:AF
P
P
,/*(yc)
AF
’,/
(yc)FIG. 2
Wethen have
H2T(t)x=
fflzeAF,
tXfor all e. Itfollows from(7.11)
thatH2Tx
is in LI[0,oo)
with, obviously,Ll-norm
independent ofe. Tocomplete the proofit now suffices toshow that forallxe,,
N
rx
,-,
0andN.
rx
,
isuniformly bounded withrespectto e.Sinceb
isspanned by thevectorsA
-1b,itsuffices totgkex
A
--1b.
By
(7.21),
we haveb
c@N.
Thus, there are vectorsv(e)e
andcoecients
z0(e)
e such that(7.22)
A-lb,
v(e)+
Y,E
’o(e)xS)(e)
j=l i=1
(i, j)#
(s, l)
andthat’st(e)
1(e
0). Now,
for a 1, 2wehave7.23)
m’
j=li=1
By
(7.19),
note that for a 1 the last term converges to 0 as e0. Using(7.20),
it followsthat for a 2thelast term is bounded fromabove, independent of e.Finally, we will show that for both a 1, 2 the first term on the right in
(7.23)
tends to 0 as e 0. For this, let
A,
andH
be defined by the commutative diagram(Fig.
3)
(P
isthecanonical projection):AFt
FIG. 3
By (7.14),
note thato’(AF,)=
AcCg.
Moreover,
Since
v(e)
-->0,this^
expression tends to 0as(e
-->0).
Let
A,
andH:
be definedby thecommutativediagram(Fig.4)
(P3
isthecanonicalAFt
projection)"
FIG. 4
It can be verified that
tr(.F,)=tr(.F,)Wcr(AF,
lt/’g/*(?7[2)),
which, by(7.11)
and(7.14)
iscontained inCg.
Itfollows thatwhichagain converges to 0as e 0. This completes the proof ofLemma 7.2.
Remark 7.7. Itis worthwhile to pointout which freedominthe spectrum assign-mentwe havein
A
+
BF
whenweusetheconstructionofthe sequence{F;
e>0}
as in theproofofLemma 7.2.The lattice diagram (Fig.
5)
shows the hierarchy of the relevant subspaces in combinationwith the freedominthe spectrum ofA
+
BF.
Denote
o//.
:=g
(fill, fir2))NotebyLemma7.3that the spectrum of the map
A
+
BFI
consistsofaneigenvalue in -e-1 with multiplicityequalto dim[l/’b/t/’g].
fixed, independent ofe assignable, independent ofe tendsto -oo as eO stabilizable, independent ofe fixed, independent ofe FIG. 5
8. Somespecial cases and extensions. Inthis section we will consider somespecial cases of the maintheorem, Theorem 7.1,and spendafew words on someextensions ofthis result.Oneinteresting special caseof
(ADDPBP)p
isthesituationthatwetakeH1
H andH2
I. This corresponds to the almost disturbance decoupling problem withbounded peaking of theentire state vector. Denoter
:=ker/-/.Since,byTheorem4.2,
g(:][,
{0})=
o//g,([)
+
*({0})=
o//.g,(y[)
and since, by Theorem 6.3,b(Y[,
{0})=
3
+
A(
b*({0})
f3)
3+
A(
fqff),
wehavethe following corollary of Theorem7.1" COROLLARY8.1. Fixp{
1,2,}.
Then theLp-almost
disturbancedecouplingprob-lem with bounded peaking
of
the entire state vector is solvableif
and onlyif
imG*(
C+
+
A
n
c).
Next,
wewillspendsomewordsonpossibleextensionsof the results ofthispaper. First we wouldlike to pointoutthat,while(ADDPBP)p
is a nontrivial extension of(ADDP)p,
we might also consider an extension of theLp-Lq
almost disturbance decoupling problem(ADDP)’,
see[20]
or[17].
This would lead to the following synthesis problem"We will saythat the
Lp- Lq
almost disturbance decouplingproblem with bounded peaking(ADDPBP)’
is solvableifthere is a constant C and, for all e>
0, afeedback mapF
#/suchthatwiththefeedbacklaw uFx,
inthe closed loop system forx(0)
0there holds,for
all 1<=
p<-q<-oo,
forall dLq[0, )
Itmay beshownthat the solvabilityofthe above problemisequivalenttotheexistence ofasequence of feedback maps
{F,;
e>
0}
andaconstant C such that for both p 1 andp=oo
IIHTGII,-O(-O)
andIIH=LGII=,-<-
C for all e.Atheory analogoustothe oneabove may be developed aroundthis problem. It canbe shownthat,again under the assumption that either
(A, B, H)
isleft-invertible orthatH2
isinjective, anecessary and sufficient condition for the solvability ofthis problem isthatimGc
V(’tc,,
ff[2)
+
(9*(fftc2)
f"l’[1).
Formore details, the readerisreferred to
17].
Afinal extension of the results of the present paperisthesituation in which we requireinternal stability of the closed loop system. Thiswould lead to the following synthesis problem: We willsaythat the
Lp-almost
disturbancedecouplingproblem withbounded peaking andstrong stabilization
(ADDPBPSS)p
issolvable ifthe followingis true.There is a constant C and forall e>
0 and all real numbers S a feedback mapF,s:-
suchthat, with the feedback law uF,sX,
in the closed loop system forx(0)
=0for all dLp[0, c)
theinequalities(3.2)
and(3.3)
holdandsuchthatRetr(A+
BF,s)
<=
S.Thus, we require that the spectrum of the closed loop matrix can be located to the leftof any vertical line Res S in thecomplex plane. Itmaybeproventhat if at least one oftheconditions
(7.1),
(7.2)
hold,then for p {1,2,}
the latterproblemis solvable if andonlyif(A,
B)
is controllable and(8.1)
Wenote that if
(A,
B, H1)
is aleft-invertible system then the inclusion(8.1)
becomes imGc+
A[
b*(9’/2)
f3(see
Theorem6.3).
IfH2
is injective then(8.1)
becomes imGc*(3’g’,)+
+
A[CI
Again, for detailsthe readeris referredto
17].
9. A worked example. To illustrate the theory developed in this paper and to demonstrate its computational feasibility, in this section we will present a worked example. We will consider a linear system with two outputs and check whether
(ADDPBP)p
issolvable forthis system.Next,
we willactually computeasequence of feedback mappings thatachieves ourdesignpurpose. The system thatwillbeconsidered isgiven by(t)=Ax(t)+ Bu(t)+Gd(t), Zl(t)= H1x(t), z2(t)
H2x(t),
with0 -1 0 0 0 A= 0 0 1 0., B= 0 0
00
0 0 010
001
0 0 0 0 andH1
(0
0 0 1 0),H2
I5x5
0Thus,
5
andq/=2.
Denote9’ci
kerHi.
Theroute that we willtakeis asfollows. First,we will check whether the subspaceinclusion imGcVb(9’’l, 72)
holdsto see if(ADDPBP)p
issolvable. Itturns outthatthis is indeed true.After this,we willfollow closely the linesof the developmentin 7 and construct arequired sequence{Fn}.
As before,Cg={h
C[ReA
<0}
and the subspaces//’g*(ffrl)
andt/’g(Yffl,
ffff2)
are taken with respecttothis stabilityset. Letthe standard basis vectors in5
bedenoted byUsing the algorithm ISA