Linear feedback : an algebraic approach
Citation for published version (APA):Hautus, M. L. J., & Heymann, M. (1978). Linear feedback : an algebraic approach. SIAM Journal on Control and Optimization, 16(1), 83-105. https://doi.org/10.1137/0316007
DOI:
10.1137/0316007
Document status and date: Published: 01/01/1978
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.
LINEAR
FEEDBACK--AN ALGEBRAIC
APPROACH*
M. L. J. HAUTUS" AND MICHAELHEYMANNAbstract. The algebraic theoryof linearinput-output mapsis reexamined withthe objectiveof
accomodatingtheconcept of(state)feedbackinthistheory. The conceptsof extended and restricted lineari/o maps (andlineari/smaps) areintroduced andinvestigated. It isshownhow "fraction
representations"oftransfer matrices arisenaturallyin this new theoretical framework.
Conditions aregivenforwhen the changecaused to a linearinput-outputmap byan(open loop)
"cascadecompensator"canalsobeaccomplishedbyutilization of(closedloop) statefeedback.In
particular,it is shownthat thechangecaused toalinearinput-output mapbycascading (composing)it with an input space isomorphism, can also be effected by feedback, provided the input space isomorphism in "bicausal", i.e. it does notchangethe causal structure of theinput-output map. Furtherdetailed characterizationsoffeedbackarealso given especiallyin connection withthenewly introducedconcepts ofdegreechainanddegreelist.
1. Introduction.Probablyoneof themostimportantcontributions to linear
systems theory since the introduction of the concepts of controllability and observability, has been thediscovery by
R.
E. Kalman(1965) (see
also Kalman(1968)
and Kalmanetal.(1969,
Chap.10))
thatthetheoryoflinearsystemscanbenaturally accommodated in classical module theory. This observation led to a
completely satisfactory theoryof realization,i.e. thetheorythat links
(external)
input-output descriptionswith(internal)statespace descriptionsofsystems,the
most recentcompletediscussion of which canbefoundinEilenberg
(1974,
Chap.16).
Yet,
despite the power of the module theoretic approach in attacking the realizationproblem,there seemedtobenoapparentcontactbetween thistheoryand even some of the most elementary control theoretic questions of linear
systemsespeciallyinsofar as theconceptof feedbackis concerned.
Two,
completelyunrelated, approacheswere usedto studyfeedbackprob-lems: One is the so called "geometric" approach, forwarded and
promoted
byWonham and Morse
(see
e.g. Wonham(1974)),
which has been successfully applied tosolve suchproblemsasdecoupling, regulator design, designofmodelfollowing systems, and investigating feedback invariant structures,
(see.
e.g.Wonham andMorse
(1970), (1972), Morse
andWonham(1970), (1971),
Morse(1973), (1975),
Wonham andPearson(1974)
and Wonham(1973)).
Thesecondapproach, which was widely used and was developed mainly by Rosenbrock
(1970)
and by Wolovich(see,
e.g., Wolovich(1974)),
used polynomial matrixtechniques forthe study ofa varietyof control theoretic questions. This latter
approach,whoseprimary powerderives from thesurprising usefulness of fraction
*Receivedbythe editorsMay21, 1976,and in revised formJanuary25, 1977. This work was
supportedinpartbytheU.S.ArmyunderResearchGrantDAHCO4-76-G-0011through theCenter for MathematicalSystemTheory, UniversityofFlorida, Gainesville,Florida.
tDepartment of Mathematics, Technological University,Eindhoven, The Netherlands.This work wascompletedwhilethe author wasonleave at theCenterfor MathematicalSystemTheory, UniversityofFlorida, Gainesville,Florida.
$Departmentof ElectricalEngineering,Technionmlsrael Institute ofTechnology, Haifa, Israel.
This work was completedwhile the author was on leave attheCenterfor Mathematical System
Theory, UniversityofFlorida, Gainesville,Florida. 83
representationsoftransferfunction matrices, seemedtobeespeciallysuccessful in
providingconvenient andquitepowerful computational algorithmswitha capa-bilityofyieldingvarious abstractresults.Also,fractionrepresentationsoftransfer matrices provide a convenient vehicle for studying such problems as minimal
realization
(see
e.g.Heymann
(1972),
Forney (1975)
and Fuhrmann(1976))
and feedback invariants asinHeymann (1972).
Probably the most strikingparadox in this state of affairs is the fact that, historically,themoduletheoreticapproachtolinearsystemsseemedtosupport
the prevailing viewpointthat transferfunction matrices in the form
H(z)/(z)
(if(z)
apolynomial)are thenatural(and
theoreticallysound)
concreterepresenta-tions of linear input-output maps
(see
Kalman et al.(1969,
Chap.10)).
Therepresentationoftransfer function matricesasmatrix fractions seemed therefore
to be nothing more than a useful technical trick. This discrepancy has been
recognizednotably by Eckberg
(1974)
andbyFuhrmann(1976)
whoattemptedtoreconcile the two representationswithinthe module theoreticframework.
Yet,
both of these attempts provide a rather ad-hoc accommodation. Specifically,
whilebothEckberg and
Fuhrmann.
makeaverysound case forviewingtransfer matrices as matrix fractions, no successful contact is made with the theory ofinput-output maps.
More
importantly, there is no satisfactory contact withfeedbacktheory.WhileinFuhrmann
(1976)
thereis noattemptinthisdirection,inEckberg
(1974)
thetreatmentoffeedback isnotverysuccessfulinthatit fails inexhibiting module theory as a powerful or even a convenient framework for
dealingwiththefeedback concept altogether.
Themainpurposeofthepresentpaperis to reexaminethe module theoretic
settingof linearinput-output mapswiththeexplicit objectiveofaccommodating
theconceptof
"state
feedback"withinthisframework.In
thetheoryof realization, theconcept"canonical"(equivalentlyreachable andobservable)
playsaverycentral andfundamentalroleinthatitdefines whatisessentiallyauniquestatespace.
Yet,
thepropertyofbeingacanonical realizationis not invariantunder feedback (i.e. acanonicalstatespacecan be modifiedby
state feedbacktobecomenoncanonical andvice
versa).
Sincethe input-outputmapdefinesuniquely
(or
essentiallyuniquely) onlyacanonicalstatespace, it isclear whythe conceptof statefeedback somehow seems incompatible with the "classical"module theoretic setting oflinearinput-output maps. Itiseasilyseen
however that reachability is invariant under feedback. The importance of the
matrix fraction representation for theinput-output maps derives from the fact that the representation essentiallyfixesareachablestatespace. Specifically, the
representation determines uniquelya reachable realization.
As
aconsequence, the conceptoffeedback(and
especiallyitseffects)
canbe studiedatthe level ofinput-outputmapswithoutgoingthroughthe
process
ofconstructingstatespace descriptionsfirst.Hence,
inthestudyoffeedbackfromaninput-output pointof view, theprocessof(concrete)
realization canessentiallybebypassed.It
hasbeentechnicallywell known forquitesome time that the modification caused to the input-output map of a linear system by state feedback can beaccomplished by cascadingthe system withalinear device (sometimes called a
compensator).
Conversely,the effecton alinearinput-outputmapbycascadingitfeedbackimplementation.
Yet
thisconverseproblemismuch less wellunderstoodeven on apurely"technical"
(rather
thantheoretical) level.In
this paper we give necessary and sufficient conditions (in a module theoreticframework)
foracascade"compensator"
tobe realizablebyafeedbackimplementation. To formulate these conditions certain revisions in the way an
input-outputmapis viewed are
necessary
inordertoaccommodate feedback.We
shall adopt here a point of view that was
already
taken previously in theunpublishednotesof
Wyman
(1972) (see
alsoSontag
(1976)).
The
paper
is organized as follows: In 2 the concepts of extended and restricted lineari/o
maps are introduced and theirrelation is investigated(the
latter concept coinciding withthe standard linear input-outputmapdefined in
Kalman et al.
(Chap.
10)).
In 3 the concepts of abstract realization and semirealizationarediscussed and theconceptoflineari/s
mapassociatedwith a(reachable)
semirealization is introducedandinvestigated.Theorems3.5 and 3.9givecharacterizations of linear
i/s
maps.In
4 the results of 3 arespecializedtothe casewhere theinputandoutput spacesare finitedimensionaland inparticular
forthe case inwhich the realizationorsemirealization (i.e. the associated state
space)
is finite dimensional. Specifically, it is shown that with every reachable realizationthere is associatedaconcreterepresentationofthe extended lineari/o
mapin terms ofa (matrix)fraction. This establishes the naturalness of fraction
representationsandexplaitheirusefulness in thestudyof feedback.
In
asimilarmanner it is shown how linear
i/s
maps are concretely represented as matrixfractions in association with reachablesemirealizations oflinear
i/o
maps.In
5the concept of feedbackis abstractly introduced and itsrelation withlinear
i/s
mapsisinvestigated.Thenotionofabicausalisomorphism (inthe extendedinput
space)
is defined andinvestigated. Itis seen thateveryfeedback transformationcanbeimplementedin
"open
loop"throughabicausalisomorphismof theinput space. Conversely, conditionsare
found for the implementability ofa
bicausalisomorphismof theinputspaceas afeedback transformation
(Theorems
5.7and5.10).
The sectionisconcludedwithTheorem 5.13,whichstatesessentially thatifthelinear
i/o
mapis rational("rational"beingappropriately defined)theneverybicausalsomorphismof the inputspacecanbeimplementedasfeedbackinsome
finite dimensional reachable
(although
not necessarilyobservable)
state space.Thisresult has the intuitiveinterpretationthat thechangecausedto a linear
i/o
map by
(externally)
modifying its inputstructure can also be accomplished by (internally) implementing feedbackprovided the externalinputchangedoes notalter the causal structure of the linear
i/o
map and is reversible. The paper is concluded in 6,where thestudyoffeedbackisfurtherexpanded.In
particular,itisnoted thatfeedbackcanbeinvestigatedby studyingthestructureof the kernel of a (restricted) linear
i/s
map. Since this kernel is a submodule of the input module, the study of feedback is generalized by studying the structure of anessentially generalsubmodule.Inthisconnection, theconceptofthedegreechain ofasubmoduleis introducedandinvestigated.The main result of 6isTheorem
6.10, which gives a complete characterization of "feedback equivalent" sub-modules. This reestablishes from a new viewpoint the central role of certain feedback invariants which have been introduced previously.
In
addition, it is shown how certainpreviouslyknown(but
notwellunderstood)
factsfind anaturalaccommodation within this new theoretical framework (in particular
Forney’s
conceptof minimal basis and Wolovich’sconceptof column
properness).
Throughoutthispaperitwill beassumed thatthereaderisfamiliar withthe
now classical module theory oflinear systems as can be found for example in Kalman
(1968),
and Kalman et al.(1969,
Chap.10).
2. Extended and restricted linear i/omaps.
We
shall begin by introducingsome notation. Throughout the paper
K
will denote a fieldandU
and Ywill denote K-linearspaces.ThespaceU
willbe referredtoasthe input valuespaceand
Y
asthe outputvalue space(of
anunderlyingdynamicalsystemE). We
shall make finitedimensionality assumptionsonU
andY onlywhenexplicitlystated.We
let7/denotethesetof integers. IfS is a K-linearspace (in particularU
and
Y),
we consider the set of all sequences s(st)t
("
",s-a,So,sl,")
possessingthefollowing properties" (i)st Sfor all 7/,and(ii)thereexiststo67/ such that st 0 for all
< to.
These sequences will be identified with(formal)
S-Laurentseries inz-1 i.e series oftheform
(2.1)
ss,z-’.
t=to
We
shalldenotethesetofS-LaurentseriesbyS((z-))
oralternativelybyAS.
Itis thenwell known that thesetAK
K((z-1))
of(scalar)
K-Laurentseriesisafieldwithconvolution asscalarmultiplicationand the obvious
(coefficientwise)
addi-tion.Also,withconvolution asscalarmultiplicationand the usual addition,
AS
is aAK-linear space. This,inparticular, impliesthatAS isalsoK-linear and also a
K[z
]-module
(where
K[z
]
istheringofpolynomialsinz).
ForanelementsAS
givenby
(2.1),
the order ofs is definedbymin
{t
ZIst
rs
0}
ifs#0,(2.2)
oFds!
ifs=0.Furthermore, for an elements
AS,
multiplicationbyz results inashiftofthesequence
(s,)
to theleft,that is(St),
_
Z(St)t
Z(St+l)t
We
now introduce the extended input spaceA U,
and the extended outputspace
A Y.
If a map[" A U- A Y
is K-linear, we say thatf
is time invariantprovideditsatisfies
f
(z
w)
zf
(w)
for all wA
U.
The following elementary but important result then follows
(see
alsoWyman
(1972))"
THZORM2.3.
A
K-linearmapf"
A
U-,A
Yis time invariantif
andonlyif
it isAK-linear.
A
K-linearmapf"
A U- A Y
iscalledcausalif ordf
(w)
->ord w andstrictly causal ifordf
(w)
>
ord w forallw6A U. We
nowintroducethefollowingDZlINITION2.4.
A
mapf"
A U
A Y
is calledanextended linear input-outputmap
(or
extended lineari/o map)
if it is K-linear, strictly causal, and time invariant.Theorem2.3 provides an algebraic characterization oftime invariance.
In
ordertocompletethe characterization ofextendedlinear
i/o
mapswe nowturn tothequestionofcausality.
Let
usdenotebyL
the K-linearspaceof all K-linearmaps
U
-
Y
and considerthespace
AL
of allL-Laurent
series. Thisspace
canbe identified with thespaceof AK-linearmapsA
U
A Y
asfollows"We
define the K-linearmapsiu" U AU:
uu (canonical injection),Pk"
A Y-> Y:
,
ytz-t--> Yk,and foreveryAK-linearmap
f"
A U-> A Y
andevery k 7:wedefine the K-linearmap
Ak"
U-> Y
by(2.5)
Ak
:= Ag(f):=/g
f
i.
The
L-Laurent
series associatedwiththemapf
is then givenby(2.6)
Zf(z
-1)
:=
Y At(f-)z
-tand iscalled theimpulse responseorthe
transfer
function
off-
Now,
ifZ
Y
Atz-t
isanyelement of
AL
wedefinethe action ofZ
onwY’.
utz-
A U
by(2.7)
Zw
:=Y,
(AkUt-k)Z-’.
k
If
f"
A U
A Y
is definedasthemapwhoseactionisgivenby(2.7),
it isthen easily verified thatz.
We
nowhave thefollowingimmediate characterization ofcausality: Themapf"
A U-+ A Y
iscausalif
andonlyif
Ak (f)
0for
k<
0 andisstrictly causalif
andonly
if
Ak
(f)=
0for
k<-O.
The followingisthenproved"THEORE2.8.
A
mapf"
A
U-+
A Y
isanextended lineari/o
mapif
andonlyif
itisAK-linearand
Ak
(f
0for
all k <- O.Foralinearspace S wedenoteby
S[z
oralternativelybyfS thesetof allN k
polynomials
Yk--O
SkZ with coefficientsSk inS. Obviously, fS isasubset ofAS
andit iseasilyseenthatit isalsoaK-linearsubspaceandeven a
K[z ]-submodule
of
AS. It
is not,however, aAK-linearsubspace.In
thedevelopment
ofthemoduletheoretic treatment of linearsystems,one of Kalman’sprimary objectiveswas toprovidearigorousframework forrelatinglinearinput-outputmapstostatespace descriptions, (i.e.theproblemof
realiza-tion).
In
his treatmentKalman considered the following "experimental" setup(see
e.g.Kalman(1968)
and Kalmanetal.(1969,
Chap.10))"
Allinputsterminate(i.e.becomeidentically
zero)
at t 0 and theoutputsareobservedonlyfort=>
1.Thisspecial set ofinputs
IU
willhenceforth be called the restricted inputspace,and elements w fU will be called restricted inputs. Since in thissettingoutputs
areobservedonlyfort
=>
1,itfollows thattwooutputsareindistinguishableif their difference isinlY.
Thus, ifweintroducethenotationfor
any
K-linearspaceS,
thenaccordingtothe foregoingwedefineF
Y
to be the restrictedoutput space.Let
A/Y
denotethesubset ofA Y
consistingof all elementst
oftheform
Y,__
ytzIt
isreadilyverifiedthatthere is abijectivecorrespondence
between
F
Y
and A/Y
with theproperty thatwithevery yF
Y
is associatedaunique / A/
Y
satisfyingzr7
y where7r"A
Y
F
Y
isthe canonicalprojection.Furthermore,
F
Y
inducesaunique modulestructureonA/Y
bytherequirementthat the above bijection become a
K[z]-module
homomorphism.We
shall henceforthidentifyelements ofF
Y
withthoseofA/Y
andrefertothelatter asrestricted outputs.
(It
ispreciselythe moduleA/Y
whichwasdefined inKalmanetal.
(1969,
Chap.10)
as theoutputmodule of a linearsystem.)
DEFINITION2.10.
A
map/:
fUFY
iscalled a restricted linear input-outputmap
(or
restricted lineari/o
map)
if it isaK[z]-module
homomorphism. Remark 2.11.In
the above definition of restricted lineari/o
maps
theproperties of causality and time invariance are automatically built in.
Wyman
(1972)
called the same aKalmanlineari/o
map and itis, infact, preciselytheinput-output mapderived inKalman’swork
(see
e.g.Kalman et al.(1969,
Chap.10)).
We
shall latermakeextensive usebothof extended and of restricted lineari/o
maps.In
additiontotheextended and restrictedi/o
maps,we introducethe conceptof linear
i/o
value map(2.12)
f:
fU-Y
asfollows:
Iff
isanextendedlineari/o
map,wedefinef
(associatedwithit) by(2.13)
f(w)
:=
p iT(w);
w fU.Alternatively,ifj isarestrictedlinear
i/o
map,weconstructthei/o
valuemap
f
by
(2.14)
f:=pf
where (with
F
Y
identifiedwithA/Y)
(2.15)
Pl"FY
Y:
Y
ytz-t--ya.
t=l
Conversely,if
f:
OU
Y
isanyK-linearmap,wecanregarditasalineari/o
map
by
associatingwith it themaps)v
and which areconstructedfromf
using theconditionsof time invariance andcausality.
In
particular,wehave(2.16)
](w)
Z
f(ztw)z
-‘-t_>o
and
(2.17)
f-(w)
Z
f((ztw))z
-t-1tZ
where thetruncationoperator
:
A
U-.
IIU
is definedby(2.18)
(Z
u,z-’)
:=Y
utz-’.
(Compare
also Kalman andHautus (1972,
formulas(2)
and(4)).)
It is easilyverifiedthatthe
maps/
and as defined in(2.16)
and(2.17)
are,respectively,arestrictedandanextended linear
i/o
mapand that the formulas(2.13)
and(2.14)
hold. The relation between
f,
f
andf
isindicated in thefollowingcommutativediagram"
AU-
AY.
U
-.
y
u
where isthe identity map,
]
isthe canonical injection, and 7r isthe canonical projection. The mapsf,
p and/1
are K-linear, the maps],
zr andf
areK[z
]-homomorphisms
andf
is aAK-linearmap.Themaps/r
andf
arecalled the restrictedandextendedi/o
mapsassociated withf.
3. Linear i/s
mas.
Let[:
fUFY
be a restricted lineari/o
map.By
anabstractrealization of
f
we refertoatriple(X, g, h)
whereX
is aK[z ]-module
andg"
IUX
and h’X-FY
areK[z]-module
homomorphisms such that thediagram
fU
---
FY
X
commutes.
For
thesystemtheoreticinterpretationof anabstract realization, the reader is referred toKalman(1968)
and Kalman et al.(1969,
Chap.10).
The moduleX
iscalledthestatespace(and
issometimesregarded onlyas a K-linearspace).
If
(X,
g,h)
isagivenabstract realization ofa restrictedlineari/o
mapf,
one canconstructfrom it a concrete realization off
(see
Kalmanetal.(1969,
Chap.10)).
Suchaconcrete realization isuniquelydeterminedby t[e
abstractrealization(X,
g,h)
andweshall henceforthcall(X, g, h)
simplyarealizationoff.
In
keepingwith standard systems terminology we then call a realization reachable if g is
suriective
andobservable ifh
isinjective.A
realization iscalled canonicalif it isbothreachable and observable.
Consider now a restricted linear
i/o
map/:U-
FY.
Let X
be aK[z]-module and letg"
IU
-
X
beaK[z]-module
homomorphism.Thepair(X, g)
willbecalledasemirealization
of/
ifitcanbeextendedtoarealization(X,
g,h)
with someK[z
]-homomorphism h"X
-
F
Y. (X, g)
willbe calledareachable semireali-zation ifg issurjective, andcanonicalifevery
extension realization(X,
g,h)
of(X, g)
is canonical. Thefollowingcharacterization of reachable semirealizationsisTHEOREM 3.1.
Let
f:fU-
FY
be a restricted lineari/o
map, letX
be aK[z ]-module,
andletg"fU-X
beasurjectiveK[z,
]-homomorphism. Then (i)(X,
g)
is a(reachable)
semirealizationof
f
ff
andonlyif
ker gcker]
.
(ii)
If
(X,
g)
isareachablesemirealizationoff
thereisa uniqueh such that(X,
g,h)
is arealizationoff.
(iii)
(X, g)
is acanonicalsemirealizationof
f
if
andonlyif
kerg kerf.
Considernow aK[z
]-homomorphismg"IU- X
whereX
is aK[z ]-module.
Ifwereferonly totheK-linear structureof
X and
regardg as aK-linearmap then, as in(2.16)
and(2.17),
we can construct the restricted and extended lineari/o
maps(3.2)
,
IU
+FX,
(3.3)
g"A U
AX
associated with g.
In
view of the fact thatX
is aK[z]-module
and notjust aK-linearspaceand gis a
K[z ]-homomorphism
andnotjustaK-linearmap,themaps and havepropertiesthatdistinguishthemfromordinary
i/o
maps. Forthisreasonwe will call and
,
respectively,arestricted andanextended lineari/s
map
(where i/s
stands forinput-state).More
generally,arestricted lineari/o
mapf:
l)U-FY
andthecorrespondingextendedi/o
mapf"
AU- A Y
arecalledi/s
maps if
Y
can be endowed with aK[z]-module
structure, compatible with itsK-vector spacestructure, such that the associated
i/o
valuemapf
:=pao
f
oj=plofis a
K[z
]-module
homomorphism. Itiseasilyverified thatif)z
is arestrictedlineari/s
mapandf
the associated valuemapthen(3.4)
kerf
kerf.
Indeed, for w
U,
]7(w)
0if and only iff(ztw)-O
for all =>0(see
(2.16)).
However,
in view of the fact thatf
is aK[z]-homomorphism,
the latter isequivalentto
f(w)
0sincef(z’w)
z’f(w).
Assume
now thatf’fUoFY
is a (restricted) lineari/s
map and that itsassociatedvaluemap
f
issurjective.Then(3.4)
andTheorem 3.1implythat(Y,
f)
is areachable(infacteven acanonical) semirealizationof
j
We
shall henceforth callarestricted lineari/s
mapfreachable
wheneverthepair(Y,
f)
is areachable semirealization ofjr
or, equivalently,whenever the associatedi/o
valuemapf
issurjective.
(The
namei/s
wasadopted preciselyfor the reason thatinthesespeciali/o
mapstheoutputvaluespace qualifiesas statespace.)
Assume
nextthatf"
flUFY
isalineari/o
mapsuch that its associatedi/o
valuemap
f
issurj.ective
and that(3.4)
holds. ThenY
isisomorphicasaK-linearspaceto
flU/ker
f
which is aK[z ]-module.
Thisisomorphisminducesa compati-bleK[z]-module
structure onY
and it is easily seen thatf
is then aK[z]-homomorphism.
We
summarizethe situationinthefollowingTHEOREM 3.5. Let
f:
I)U-FY
be a restricted lineari/o
map such thattheassociated
i/o
value mapf pl:
IU-
Yissur]ective. Thenis
ani/s
mapif
andIf
H: X
Yis a K-linearmap thenit induces in anaturalwayaAK-linearmap
(3.6)
H:
AX
A Y:
Y
xtz-t(Hxt)z
-tand also a
K[z ]-homomorphism
(3.7)
H:
FX
FY:
x,z-’
-
(Hxt)z-’
t>l
wherewehave identified
A+
Y
withFX
andA+
Y
withF Y.
Themapsdefinedby(3.6)
and(3.7)
arecalledstaticmapsandit willbeconvenient todenote thembythe samesymbol H.
We
nowhave thefollowingTI-IEORE3.8.Let
f
IqU-FYbearestricted lineari/o
map,letg"fU-Xbe aK[z ]-homomorphism
andlet,
be therestricted lineari/s
mapassociatedwithg.(i)
If
(X, g)
is asemirealizationoff,
then thereexistsaK-linearmapH:X-Y
such that
f
H.
g (correspondinglyf
H.,).
(ii) Conversely,
if
gissurjectiveand thereexistsH:X-
Y
such that=
H.,,
then
(X,
g)
is asemirealizationoff.
Proof.
(i)Let
h"X-
F
Y
be aK[z]-homomorphism
such that(X,
g,h)
is a realization offi
and defineH:X-,
Y:x-p h(x),
where paisdefinedin
(2.15).
Then,if[
:= pf,
wehavef
=Plof
=plohog=Hog
andconsequentlyfor w6fluwehaveby
(2.16)
and(3.7)
Z
f(z’w)z
Y.
Hg(z’w)z
-t-’t>=o t>o
=t-/o
E
t=>0andtheproofof(i)iscomplete.
(ii) Theexistenceof
H
suchthat/r=
H. impliesthatkerg kerHence,
sincegissurjective the result follows from Theorem 3.1.The following result gives another characterization of the linear
i/s
mapsamongst alllinear
i/o
maps.THEOREM 3.9. Let
(,"
fU-FY
bea restricted lineari/o
map. Then,
is areachablerestricted linear
i/s
map if
andonlyif
the followingcondition,
holds"for
every restricted lineari/o
mapf:
IU- FS
satisfying kerff
ckerf
(where
Sis aK-linear
space),
thereexists aunique K-linearmapH" Y
S such thatH
Proof.
Assume
that is a reachable restricted lineari/s
map and that satisfies ker ckerf.
Then theexistence ofH
such thatf
H. followsfromTheorems 3.1 and 3.8.Also,
f
H.g wheref
po.1
rand g pif,
and fromthesurjectivityofg
(the
reachabilityof)
itfollows thatHisunique.Conversely, assume the condition holds. Define S lqU/ker and let
f:
fU Sbe thecanonicalp.rojection.
Let
]r:
flU-.FS
bethe restrictedlineari/o
thereexists auniqueK-linearmap
H" Y-
Ssuch that1
H. or,equivalently,1
H.g.Hence
ker gcker]’
ker and, since ker kerg isobvious, it fol-lows that ker kerg. FromtheuniquenessofH
wehave thatgissurjectiveand the result follows from Theorem 3.5.We
conclude this section withthefollowing result(due
toWyman
(1972))
whichsummarizesin a
tra.nsparent
wayourcurrentpointofview.TI4EOREM 3.10. Let
[:IIU-. FY
be arestricted lineari/o
map. Then every reachable realization(X,
g,h)
of
inducesauniquecommutativediagramAU
.AX
.
AYIIU, FX J,FY
4. Finite dimensionality and matrixrepresentations. Let
):
U-
FY
be arestricted linear
i/o
map.Two
reachablerealizations(X1,
gl,h1)
and(X2,
g2,h2)
of
1
arecalledisomorphic ifthere existsaK[z
]-isomorphisma"X1
X.-suchthat thediagramllU
.
FYcommutes.
Let
Acker"
beasubmodule. Thereis associated with A(uniquely)areachable realization
(Xa,
ga,ha)
of]"
whereXA
:=fU/A,
gA’fU-Xa is thecanonical projection and h
a:Xa-
F Y
is defined via Theorem 3.1(ii).Suppose
nowthat
(X,
g,h)
is anyreachablerealization of1
and letA
kerg. Itis easily verifiedthat(X,
g,h)
isthenisomorphicto(XA,
ga,ha).
Itfollows that the class of reachablerealizationsof1
is inone-onecorrespondence(modulo
isomorphisms)withtheclassof submodules of ker
]
Specifically,witheverysubmoduleA ker]r
realizations
(X,
g,h)
forwhich A ker gareessentially completelydeterminedbyA.
Thisfact will beparticularlyusefulin ourstudyoffeedback whereitwill turn out tobe easiertodealwith(the submodule)
Athanwith(the
quotientmodule)
ag/a.
Sofarwehaveimposednofinitedimensionalityconditionsoneither
U, X
orY.
In
this section we shallspecializeourresultstothefinitedimensionalcaseand shall henceforth assume thatU
andY
are finite dimensional linearspaces. More specifically,weletU K
m,
Y
Kpsothat fU=
K"[z]
andFY=
KP((z-1))/KP[z].
Thus,inthiscase
IIU
is afreeK[z]-module
andhasabasis of m elements.Since
K[z]
is aprincipalideal domain, everysubmoduleA ofIIU
hasatmost mfree generators and their numberd,whichwecall rank
A,
isindependentoftheir choice(see
e.g. Hartleyand Hawkes(1970,
Thm.7.8)).
Our
maininterestisin therepresentationof finitedimensional linearsystems, andwe areconcernedwiththecasein which forasubmoduleA fU,)U/A
is afinite dimensionalK-linearspace,orequivalently,a torsionmodule
(see
e.g.Lang
(1965,
p.388)).
Thusweshall makeuseof thefollowingstandard butimportantresult
(the
proofofwhich canbe found in Fuhrmann(1976)).
THEOREM4.1.LetAcfU beasubmodule. Then
fU/A
is a torsionmodule(or
equivalentlyafinite
dimensional K-linearspace)
if
andonlyif
rankA m.LetAc
D.U
beasubmodule of rankm and letdl,,
d,, be a basisforA.
Define theK[z
]-homomorphismD" U
-
A by Dei di, 1,,
m, where e,,
e,,denotesthenatural basis inKm,
(as
well as infU). We
canviewD
also asan mx
mpolynomialmatrix (i.e. amatrixwith entries inK[z])
by regardingdi
K"
[z
asthe th column ofD. Conversely,ifD D(z)
is anx
mpolynomialmatrix, we can regard
D
as aK[z]-homomorphismK"[z]-,K[z]:e-d,
i=1,...,m, where
di
di(z)Kl[z
is the th column ofD.
If inparticular m, then dl,"" ,d,,, are elements ofIIU=Km[z]
and are thus generators of asubmodule A fU definedby
A=DU
:={Dwlw
ClearlyrankA rankD
(rank
D beingthe matrix rank ofD),
and rankA m ifandonlyif
D
isnonsingular (i.e.0 detD
K[z ]).
Consider now the K[z]-homomorphism
IIUIIU
defined by an mx
mnonsingular
polyn_omial
matrixD.
It iseasily verified that thereexists a uniqueAK-linear
map D" A U
-*A
U
such thatthe diagramAU
AU
D
fU ---*- fU
commutes, where], asusual, denotes thecanonicalinjection. Itis readily noted thatthe transferfunctionofD isgiven bythe(polynomial) matrixD.
[We
shallrefer to a AK-linear map whose transfer function is apolynomial matrix as a
polynomial
map.]
SinceD
isnonsingular,itis obviously invertibleoverAK,
and thusD is aninvertiblemap.We
shall henceforth denote both themaps D andD
and their associatedpolynomialmatrixbythe singlesymbol D.Themeaningwill
alwaysbe clear from the context.
An
m rn polynomialmatrixR iscalled unimodularif itsdeterminantis anonzero constant, i.e. if its inverse is also a polynomial matrix. The following
theorem whose elementary proof is omitted and the correspondingimmediate
corollarywillbe useful:
THEOREM 4.2. Let
A,
A*c12U be submodules given by A DfU and A*D*OU
whereDandD*
are nonsingular. Then A*cAif
andonlyif
thereexists apolynomialmatrixR such that
D*=
DR.COROLLARY4.3. Under theconditions
of
Theorem 4.2,A* Aif
andonlyif
thereexists aunimodularmatrixRsuch that
D*
DR.We
nowturn to the study of representations of lineari/o
maps(and
theirrealizations). Let
)z.
A U- A Y
be an extended lineari/o
map, let)z
be the associated restricted lineari/o
mapand let(X,
g,h)
beareachable realization ofwithXafinite dimensionalK-vector space (i.e.rank ker g
rn).
Letker g A DfU anddefinethemap(4.4)
N"AU
AY:
w-f(Dw).
SinceNis acompositeoftwo AK-linearmaps, it isclearlyalsoAK-linear.
Let
z,(z)
denote the transferfunctionofN.
We
observe thatZu
isapolynomialmatrix(i.e.Nt
0for all> 0).
Indeed,forw7r
(Nw
7tof
(Dw
f
(Dw
O,
withthelastequality holdingsinceDw ker gcker
f.
Itfollows thatNwKP[7. ].
Since w
IU
waschoosen arbitrarily,wesee(compare
(2.7))
thatz,(z)= Y,
Ntz-’
t=<o
asclaimed.
We
shallnow(justas wehave done forD)
identify (notationally)Zu
withNsothatNwilldenote both themapand its associatedpolynomialmatrix.
In
view of the relation
Nw
f(Dw)
for every wA U,
and the fact thatD
is invertibleweobtainthefollowing expression"(w)
ND-’w
(w
AU).
Thus,the transferfunctionof
f
has therepresentation(4.5)
Z(z)=
Az-=N(z)D-’(z).
t=l
A
AK-linear mapf"
A U- A Y
and its L-Laurent seriesZ(z)
are called rational ifand onlyifthereexists a (nonsingular) polynomial matrixD(z)
suchthat
ZD
is apolynomial matrix. Thuswehaveprovedthenecessity partof thefollowing
THEOREM 4.6.
A
restricted lineari/o
mapf:IU-
FY
has a reachablerealization with
finite
dimensionalstatespaceif
andonlyif
f
isrational.Proo[(sufficiency). If
f
is rational, there existpolynomialmatricesNandD
such that
f
ND-1.
Then(XA,
gA,hA),
whereA
DU,
isareachable realizationof
f
withfinite dimensionalstatespace.TwomatricesNandD arecalled rightcoprime ifevery m m polynomial
matrixR satisfying N
NR
andD
DR
forsomepolynomialmatricesN
andD,
isnecessarilyunimodular.THEOREM 4.7.Let
(X,
g,h)
beareachable realizationof
arestricted lineari/o
map
f.
Let ker gA
DI)Uand assumeD
isnonsingular. Let NZD.
Then(X,
g,h)
isobservable (i.e.ker]=
kerg)
if
andonlyif
N,
andDare right coprime.Proof.
Let
D
be chosen such thatkerf=Dl)U.
Thenker g
=
kerf
D
ll)U, andby
Theorem4.2 there exists apolynomialmatrixR
such that
D =D1R.
D
is then nonsingular and ifN
is defined such thatN1D
-1
Z,
itisreadilyseenthatNNR.
IfweassumethatNandD
areright,
coprime,then thematrixRmustbeunimodularandbyCorollary4.3ker g
kerf
and observability holds. Conversely, let
N-NaR
and DD1R.
Then R isnonsingularand
Z
ND-
ND-X.
Itfollows thatDlieU
ker)
andwehavekerg Dl)UC
DllU
ckerfi
If
(X,g,h)
is observable, then kerg=kerf so thatDfU=DIU
and by Corollary4.3R isunimodular, whenceNandD
areright coprime. [3Remark 4.8. Therepresentation
Zf
ND-1ofatransferfunction iswidely
known asa
fraction
representation.We
havejustseen howfractionrepresentationsarisenaturallyfromreachable
(abstract)
realizations(X,
g,h)
ofaninput-output map. In fact,thereisessentiallyaone-onecorrespondencebetween
the class of allfinitedimensionalreachablerealizationsof
f
and the class offractionrepresenta-tions. It is important to observe thatby writing
Zf
ND-
weessentially have(modulo
isomorphism)aspecificreachable realizationinmind. Theobservability alongwiththe associatedright coprimenessisessentially ofnoconsequencein ourstudyof feedback asweshallseelater.
Let
X
Kn.
Assume
X is also endowed with a module structure and let g"IIU-*X be a surjectiveK[z]-homomorphism. Let
"
IIU- FX
be the restricted(reachable)
lineari/s
mapassociated withg.Then(X,
g)
isacanonical semirealizationof andthereexistsh" X-
FX
suchthat(X,
g,h)
is a(canonical)realizationof
.
Accordingto theforegoingwe canwrite(4.9)
Z
SD-,
where
Ze
isthe transfer function of,
withD being a(nonsingular) polynomialmatrixsuchthat ker
DIIU,
andS beingan nxm polynomialmap. Inviewof theobservabilityof(X,
g,h)
itfollows from Theorem 4.7 thatD andS arerightcoprime, and moreover, degdetD n.
If
Z
isthe transfer function ofanextendedlineari/o
mapf,
then thestrictcausalityof
f
is equivalenttoZ
beingastrictly properrationalmatrix; that is, eachthan thedegreeofthe numerator.
We
nowspecializeTheorem 3.9to the finite dimensionalcase.TUEOREM 4.10. Let S and Dbe right coprimepolynomial matrices with
D
nonsingular, andassumethat
Z
:= SD-aisstrictly proper(or
equivalently strictlycausal).
ThenZ
isthetransfer
function
of
areachable extendedlineari/s
map,
if
andonly
if for
everypolynomialmatrixN satisfying theconditionND-a isstrictly proper, thereexists auniqueconstant matrixH
(withentries inK)
suchthat N HS.Proof.
Assume
thatZ
Zg
where isareachable extendedlineari/s
map.By
thecoprimenessofSandDwehave ker DI2U and
ifjrN
is therestricted lineari/o
map associated withZN=ND
-a(ND
-a strictlyproper),
thenker=
Dl)Uckerfu.
By
Theorem 3.9 there exists a unique H:X-Y
such thatjrn
H. so that ifH
also denotes the corresponding matrix we haveN-(ND-a)D
(HSD-a)D
HS. Conversely,letjr
bearestricted lineari/o
mapwithker cker
jr.
ThenDI)U kerjr
and there existsapolynomialmatrixNsuch thatf=ND
-a(
denoting the extended lineari/o
map associated withjr).
By
hypothesis there exists a unique constant matrix
H
such thatN--HS,
andconsequently
f
H. andbyTheorem3.9 theproofis complete.COROLLARY4.11.Let
D
beanonsingularrnxtnpolynomialmatrixandlet detD
have degree n. The setof
m-row polynomial vectorsv(z)
such thatv(z)D-a(z)
is strictly proper, is an n dimensionalvector space overK.
If
S is apolynomial matrix such that SD-a is strictly proper, then SD-a is the
transfer
function
of
areachablelineari/s
mapif
andonlyif
therowsof
Sform
abasisof
X.Proof.
That is aK-vectorspace
is obvious.Suppose
SD-1isthe transfer function ofalinear
i/s
map.If v(z)
6X
isanyvector, thenbyTheorem 4.10 there existsa(row)
matrixH
suchthat v HS andthus,the rows ofS generate’.
The uniqueness condition ofH
implies
that therows of S arelinearlyindependent
(over K)
andthus formabasis forX. Conversely,iftherowsofSformabasis forX
then SD-a is a strictly proper matrix and for each matrix Nwhose rows are
elementsof
X
there exists auniqueK-matrixH
such thatNHS
whence, byTheorem4.10, SD-aisthetransferfunctionofalinear
i/s
map.Sincethe number of rows inS isequalto n,itfollows that dim n.5. Feedback.Considerarestrictedlinear
i/o
mapjr:
U
FY,
let(X, g)
beareachablesemirealization of
f
andlet"
A
UAX
be the extendedlineari/s
mapassociatedwithg.
Suppose
wemodify inthe followingway.For
each w6AU
instead oflettingg
act on w we letg
act on q :=V-
w-L sc,
whereV: U- U
and
L’X- U
areK-linear (static) mapswithV
invertible, and wheresc issuch that:
(qg).
Then w,q, and:
arerelatedbytheequations(5.)
o
v.
w-L
;
=
In
linewiththeaboveblock diagramwecallthe pair(L, V)
afeedback
pair. Themap
L
will be called afeedback
map andV
will be called a(static)
inputtransformation.
From
(5.1)
wecaneliminateso,
to obtain(5.2)
(I
+z;g).0
v.
w.SinceL
g
isstrictly causal,itiseasilyverifiedthat ord(I
+
Lg)0
=ord0foreach qA U.
This implies that ker(I +Lg)=
0, from which it follows that I+Lg
isinvertible.Thus,
(5.2)
hasaunique solution for0 givenby qg=l.w,where
(5.3)
-
:=
t\.
:=(I
+L)
-
v.
Upon
substitutingforo
inthe second formula of(5.1),
weobtain theexpression(5.4)
=fL, v(W),
where
(5.5)
f,,,
:=g
t.
The schematicinterpretationof
(5.5)
is givenby
the blockdiagramwhere l"
A
U
A U
isregarded as a dynamic inputtransformation.
It is readilynoted
(see
alsoLemma5.6below)
that themap hasthefollowing properties: (i) is an invertible AK-linearmap.(ii) Both
-
and--1
arecausalmaps.We
shall henceforth call a mapI’AU-AU
which satisfies both (i) and (ii) abicausalisomorphism on
AU.
Itis obviousthat
fL,
Vis anextended lineari/o
map.Thisfollowsimmediatelyfrom the fact that the compositeofAK-linearmaps isAK-linear, and that the
compositeofacausalmapwithastrictlycausaloneisstrictlycausal.
In
fact,it willbeseen laterthat
fL,
vis even areachablelineari/s
map.We
have seen that ifwe can constructf,v
fromg
byfeedback(as
in our firstinterpretation) then we can also construct it by cascading with a bicausal
isomorphism
l,v
(whichis an"open
loop"construction).We
shallnow turn to the more difficult question" when can a AK-linear mapI’AU-AU
beexpressedas in
(5.3)
for someL andV.
If this is the case we call afeedback
transformation
(correspondingto(L,
V)).
LEMMA
5.6. Let’AU-
AU
beacausal AK-linearmapandletZr(z-1)
.k=O
Ak(
)Z
-kbeits
transfer
function.
Then hasacausalinverseif
andonlyif
Ao(1-)"
U-
Uisinvertible, inwhichcaseAo(--1)
(Ao(-))-1.
THEOREM 5.7.
Let
g’AU
:A U
-
A
U
isaAK-linearmap, then thereexists afeedback
pair(L,
V)
such thatl,v (as
in(5.3))
if
andonly
if
(i) isabicausalisomorphism,
(ii)
for
wIU,
g(w)
IX
implies--l(w)
U
(equivalently,--1
(ker
)
OU).
Proof.
Assume
firstthatlt,
v.Thebicausalityof hasbeen noted beforeso that(i)holds. Also,since--
V-
+
V-1L,
it isreadilyseenthatif w6fU,then,(w)6lX
impliesthat{-(w)OU
sothat(ii)
alsoholds.Conversely, assume satisfies (i)and(ii). Let V:U--)Ubedefined by V :=
Ao(1).
By (i)
and Lemma 5.6, V is invertible andV-a=Ao(/-a).
Hence,
the map-a-
V-1(where
V-ais regardedas a staticmap)
isstrictly causal and(5.8)
v-’
is an extendedlinear
i/o
mapA
U-)A U. We
claimthat ker ker]r
(where
1
istherestricted linear
i/o
map associated withf).
Indeedif(w)=
0 implies(w)
YX
and by (ii)--l(w)6fU.
By (5.8)
it then follows thatf(w)612U
so thatiV(w)
0.Upon
employing,
Theorem3.9,weconclude_
theexistenceofaK-linearmap H:X--)Usuch that
f
H.(or
equivalently F H.).
LettingL"X--)U
be definedby L VHweobtainupon substitutinginto
(5.8)
f--l__
V-1
..[_V-1Lg
and theproofiscomplete.COROLLARY
5.9. Let," A
U-*AX
be areachable extended lineari/s
map.Then
for
everyfeedback
pair(L,
V)
the mapfL,
Vdefined
by(5.5)
isareachableextended linear
i/s
map.Proof.
We
shall provethe corollary by showingthatfL,
V satisfies Theorem3.9._
Letf:
IU
FSbeanyrestricted lineari/o
map
such thatker_ jL,V
CkerI
7.
Letf
be the extendedlineari/o
map associated withf
anddefinefI’AU
AS
byI
A U
-
A
S w-
f{-
wIf we can show that ker c
kerfa,
then by Theorem 3.9 there exists a uniqueH’X-_S
such_that jl
H.(or
equi.valentl,
f-1
H- g),
whence,
f=
fl
-=
H.,
l=H.f,v (or
equivalentlyf=H’f,v)
showing thatf,v
satisfiesTheorem 3.9.
Hence,
the proof will be complete upon showing that kerker]’a.
Ifw61Uand
(w)
0 then(w)
6lXandbyTheorem5.7(ii)w
:=-a(w)6U.
Itfollows that
fL,
v(W1)=o
f(W1)=
io
f-I(w)=(W)E’X
andconsequentlyfL,
v(Wl)
0 fromwhichf(wl)
0byassumption.Thus,f(wl)
S
andL(w)= o
We
shall nowspecializeourresultstothe finite dimensional case and assume,justas in 4, that
U
K"
andY
Kp. Moreover,
if:
U
FY
is a restrictedlinear
i/o
map,weshall assume that]
is rational,i.e.thatithasarealization withfinitedimensionalstatespace.
Firstwehave thefollowing specializationof Theorem5.7.
THEOREM5.10. Let g"
A U- AX
beareachable extendedfinite
dimensional(i.e.
X
K)
lineari/s
map,and letDIU
ker.
If
l"AU- AU
is a AK-linearmap, therd exists apair
(L, V)
such thatl;v (see (5.3))
if
andonlyif
is abicausal isomorphism and
Z;ID
is apolynomial matrix(where
Zr
denotes thetransfer
function
of
).
Remark 5.11. Thecondition that
ZflD
is a polynomial matrix is obviousupon examination of
(5.3). In
the finite dimensional case we haveZg
SD-1(compare
Theorem4.10),
andconsequently(5.3)
impliesthatZ-aD
V-I(I+LSD-1)D
V-I(D
+LS).
It is interestingtoobserve thatif is givenby
,
(where
l,v),
thenZir
S(D +LS)
-
V
and consequently(5.12)
kerf’r
V-’(D
+LS)nU.
Alternatively,
(5.12)
canalso bewritten inthe form kerf’t-
V-’(D
+
Q)IU,where
O
is anarbitrary polynomialmatrixsuch thatQD-
isstrictlyproper(the
matrixL beingofcourseuniquely specified throughTheorem
4.10).
Theinterest-ingpointinthisalternate formulationisthe fact thatwehave eliminated
L
fromexplicitconsideration.
It is also interesting to make explicit the possible
i/o
maps(or
transferfunctions) obtainable byfeedback. If ND
-
where it is understood that the factorizationdefines a realizationoff,
then the transferfunctionsobtainableusingfeedbackaregiven by
ft,v=N(D+O)-lv,
where Qisany polynomialmatrixsuch that QD-1isstrictlyproper
(L
being, asbefore, specified throughTheorem
4.10).
The reader should noticethesimilarityofthesituation withthesingleinput
single
out,
putcase.Let
f"
IU
FY
bea rationalrestrictedlineari/o
map, and let l"AU-
AU
bearationalcausalAK-linearmap. Let
(X,
g,h)
beareachable finitedimensional(X
K")
realizationoff
andlet be the extended lineari/s
mapassociatedwithg. Theorem 5.10 then states that there exists a feedback pair
(L, V)
such that-=
(I
+
L)
-aV
provided-
is abicausal isomorphismandZ{ID
is apolynomialmatrix, where
Zr
is the transfer function of[,
and D is a polynomial matrixsatisfying
DIU
ker.
Whilethebicausalityof isclearlyanecessaryconditionfor itsfeedbackimplementationin agivenrealization,it is not sufficient.
We
shall conclude this sectionby showingthat therealwaysexists areachablerealizationinwhich canbe implemented byfeedbackif and onlyif is a rational bicausal
THEOREM5.13.
Let:
fU- FYbearational restricted lineari/o
mapand letl"
A U- A U
be a rational causal AK-linear map.A
necessary andsufficient
conditionfor
theexistenceof
afinite
dimensionalreachablerealization(X,
g,h)
suchthat
Zr
(I
+
LSD-1)
-
V
for
somefeedback
pair(L, V) (where
SD-1isthe
transfer function
of
thelineari/s
mapassociatedwith
g)
isthat is abicausalisomorphism.Proof.
Thenecessity
ofthecondition is obviousfromTheorem5.10.We
willprove sufficiency.
Iff
isrational,_
thenZ[
ND-1forpolynomialmatricesNand
Dand
DU
kerf.
Alsoif is a bicausalisomorphism,thenit canbeexpressedas
[
(I
+
po-1)-i
V
(withPQ-I
strictlyproper).Let
R
:= detO
I.ThenZf
NR
(DR)-1
andconsidertherealization(XA,
gA,ha)
whereA
DR
fU. ClearlyZflDR
V-I(I
+
PQ-1)D
detO
Iisapolynomialmatrixsatisfying the condi-tionof Theorem 5.10.Remark 5.14. Theorem 5.13 gives a complete characterization of those
dynamical compensators which can be implemented by
"state
feedback" in a(possibly
unobservable)
reachable state-space. This characterization problemreceivedsome attentionin the literature and the readeris referredtoWolovich
(1974).
6. Chains and invariants. Throughout this section we shall assume that
U
K"
andY
Kp.
LetAcOU
beasubmoduleofrankrn andlet(XA, ga)
be the (finite dimensional) reachable semirealization given byXA
:=12U/A
andga"fU-
XA
the canonicalprojection.Let
a
be the(restricted)
lineari/s
mapassociated withga.Thenker Aand,aswehaveseenatthe beginningof 4,all reachable realizations and hence also all reachablelinear
i/s
mapswhose kernelisA areisomorphic. Thus, modulo state space isomorphism, is uniquely deter-mined by
A.
If feedback is applied to,
then is transformed into a new reachable lineari/s
mapfr
(see
(5.5))_
and.
correspondingly the submodule A istransformedinto asubmodule
A(1) :=
kerfr
(see (5.12)). Two
submodulesAand A’offUof rankrn willbe calledfeedback
equivalentifthere existsafeedback such thatifA ker,
then A’= kerft.
Clearlyfeedbackequivalenceis anequival-ence relation and as we have just seen, each feedback equivalence class of
submodulescharacterizes a class of feedback equivalent linear
i/s
maps. In the presentsection westudythis equivalencerelationof submodules.For
w fU define the(polynomial) degreeof wbydegw:=
-ord w.LetA and A’ be submodulesof2U.A K[z ]-homomorphism
q"A- A’ is calleddegreepreserving if for each w6
A,
deg q(w)
degw.LEMMA 6.i. Two
submodulesA
and A’of
O
U
of
rank m arefeedback
equivalent
if
andonly
if
thereexists adegree
preservingK[z
]-isomorphism
q:A
-
A’.
Proof.
IfA
andA’
are feedback equivalent thenA
ker andA’=
ker wheref--/-
-
forafeedbacktransformation-.
Let q denotetherestrictionof/--
toA. We
see thatq mapsA
intoA’.
Indeed,if w6A,
then(w)
6OX,
andbyTheorem 5.7(ii) we also have
v:=q(w)=-l(w)612U.
Thus/-/-(v)=
-o/-(w)
,(w)12X,
whence)r(w)
0andq(w)=
vcA’.
From
thebicaus-alityof
t-]
itisclear thatord--a(w)
ord w forall wA U
and consequently theConversely,assumethathandA’ aresubmodulesof fU ofrankmand that a
degree preserving q"
A-
A’
exists. Let d1,’’’,d,, be abasis for A and denoted :=q(di), i=l,...,m.
Let
D
andD
be the polynomial matricesD
:=[dl,.
,
d,,, andD’
:=[d ],.
,
d’,,,].
Thensurely Disnonsingularandforeach w6Awehave
q(w)
D’D-lw.
By
formally regardingthematrixD’D
-
as a transfer function,weseethatqextendsuniquelytoa AK-linear map"
AU->
AU;
Zq
:=D’D
-1.
We
willcompletetheproof byshowing that is a bicausalisomorphismand that themap-:=
]-1
isthedesiredfeedbacktransformationsuch thatA’
A().
Firstnote that
(and
hence alsol)
isabicausalisomorphismif andonlyif(6.2)
ordt](w)
ord wforall w6
A U.
Recallthat(6.2)
holds for wA
by assumption.SinceA
isof rank m(and
henceIU/A
is a torsionmodule)
there exists a polynomialq
such thatw
Aforall w6U.
Thus ord(w)
=ordw
forall w6IU
anditfollows that(6.2)
also holdsfor all w6flU.Next,
for wA U
let k be aninteger such thatw
:=(zkw)O.
Thenwa
61U and ordt](wa)=
ordw
=ordzkw.
Sinceforksufficiently largeord
](w)
=ordgl(zkw),
itfollows that(6.2)
holds forallwAU
and isabicausalisomorphism. Finally,toseethat
---
-
isthedesiredfeedback transformation we needtoshow(see
Theorem5.10)
thatZ-ID
is apolynomialmatrix. Indeedthisis true since
Z{D
Zr-D
D’D-aD
D’
and theproofiscomplete.
El
In
ordertoapply Lemma6.1 incheckingastowhethertwosubmodules A andA’
arefeedback equivalentonehastoverifythe existence(or
nonexistence)ofadegree preserving homomorphismq"A->
A’.
To
thisendweintroducethe concept ofdegreechain.DEFINITION6.3.
Let
A beasubmodule of12U and for each 0, 1,2,,
letA
Abe the submodulegenerated by theelements of A whosepolynomialdegreeis less than or equal to i. Let
v
:= vi(A):=rankAi.
The sequence(Ai)o
ofsubmodules iscalled the
degree
chain of A and thesequent(u)=o
is called thedegreelist.
It follows immediatelythat
Ao
cA1
cA2c. cA,
and Uo_-<Ul-<uz_-<" _-<u:= rank
A.
SinceIU
is Noetherian,itsatisfiestheascendingchainconditionand we haveAr
A
forsomefiniter.Hence
AoCAc
"CAr=A,
We
also introduce the following notation. Let a 0 be a polynomial or apolynomialvector.
We
denoteby the leading coefficient(vector)
ofa.That is,ifa
k=
oaz
andak 0 then ak.Also,if a 0wedefined := 0. Recall thatifdl,
,
dk
areelements ofaK[z ]-module
which generateasubmoduleM,
theyare called
free
generatorsof
M,
orsimplyfree
ifthey areK[z
]-linearlyindepen-dent.
We
now introducethe followingDEFINITION 6.4.
Let
dx,’’ ",d
be elements ofa finitely generatedK[z]-module (specifically of
U).
Thendl,",
dk
arecalledproperly
free
if1,"
,
dk
areK-linearlyindependent.LEMMA
6.5.ff
dl," ",dk
areproperlyfree
thentheyarefree.
Proof.
We
provethelemmabynegation. Ifdl,,
dk
arenotfree,there exist polynomials a1,’’" ,ak, not all zero, such thatEik=l
aidi=O.
Let r:=maxdeg
(aidi)
anddefine0 ifdeg(otidi
<
r,Ei
:-’-1 ifdeg(otidi) r.
Addingthetermsofdegree rin
ol.idi
yieldsk
E
Eiidi
O,
i=1
implyingthat
al,
,
k
are K-linearlydependentsince notall6ii
are zero.DEFINIaION 6.6.
Let
A be afreeK[z]-module.
A
basisd1,"’,dk
of A is called properifdl,.,
dk
areproperlyfree.The followingis an instrumentalresult.
LEMMA
6.7.Let
Abeasubmoduleof
12Uof
ranku(
>
0).
Thenthereexists aproperbasisdl, ",
d
of
A such that(6.8)
deg djfor
12i--1<j 12iand O, 1, 2,.where
(12i)i%0
isthedegreelistof
A
and12-- :’"O.Proof.
Firstobserve thatifk_->0is thefirstinteger such that 12k 0,we can choose 0d16 Ak
such that deg d k. Then d is clearly properly free and satisfies(6.8). We
proceed stepwise and assume that for >0, dl,...,dt
areproperly free elements of
A
satisfying(6.8)
and let S denote the submodulegenerated by dl,
,
dl.
If 12(and
hence SA),
then there existsanintegersuch that
At-1
S
At.Thus, thereexists an element
dl+l
At
such thatdt+
:
S anddegdl+l
t.Obvi-ously d,
,
dl+
satisfies(6.8)
andwecompletetheproof
by showingthat thissetis
also
propeOy
free.Assume
tothecontrarythatdl+
is a K-linear combina-tion ofdl,...,d"Consider the element
p ::
dl+l--
Aiz6’di
i=l
where 6i := t-degdi, 1,-..,
I.
Obviouslyp A and it is easilyverified thatdegp=<t- 1. Thus p At_lcS andweconclude that
dl+
S,
acontradictionto ourassumption, fqCOROLLARY 6.9.
Let
AOU
beasubmoduleof
rank uand let(Ai)i=o
ahd 12i)i=obe thedegreechain anddegree listof
A respectively. Thereexists aproper basis d,,
d
of
A such thatfor
each satisfyingui 0 thesetda,,
d,
isabasis