Bilinear
H
2
-optimalModel Order Redu tion
with appli ations
to thermal parametri systems
Dissertation
zurErlangungdesakademis henGrades
do torrerumnaturalium
(Dr.rer.nat.)
vonDipl.-Math. AngelikaSusanneBruns
geb. am06.01.1986in Freudenstadt
genehmigtdur hdieFakultätfürMathematik
derOtto-von-Gueri ke-UniversitätMagdeburg
Guta hter: Prof. Dr. PeterBenner
Prof. Dr. TobiasDamm
eingerei htam20.03.2015
Danksagung v
Zusammenfassung vii
Summary ix
ListofFigures xi
ListofTables xiii
ListofAlgorithms xv Notations xvii Chapter1. Introdu tion 1 1.1. Motivation 1 1.2. Dissertationoverview 2 1.3. Thesis ontributions 3
Chapter2. Mathemati alprerequisites 5
2.1. LinearAlgebra 5
2.2. Dierentialgeometry 8
2.3. Systemstheory 10
Chapter3. Modelingofheattransferproblems 25
3.1. ThermalModeling 26
3.2. Theheatequation 29
3.3. BoundaryandInterfa e onditions 30
3.5. Thermalmodelingofanele tri almotor 32
Chapter4. Modelparametrization 35
4.1. Dis retizationoftheheatequation 35
4.2. Physi alparametrization 37
4.3. Geometri variations 37
Chapter5. ModelOrderRedu tion 49
5.1. Proje tion-basedMORandtheerrorsystem 50
5.2. MORoflinearsystems 52
5.3. Parametri ModelOrderRedu tion(pMOR) 60
5.4. BilinearModelOrderRedu tion 69
5.5.
H
2
-optimalbilinearModelOrderRedu tion 73Chapter6. Challengeswhen applyingBIRKAto thermalindustrial
models 101
6.1. Krone kerprodu tappproximation 101
6.2. Stability 105
6.3. Singularstinessmatrix
A
andlargenormmatri esN
k
115Chapter7. Redu tionofphysi allyparametrizedthermalmodels 121
7.1. Resultsforthe
H
2
-optimalredu tiononGrassmannmanifolds1217.2. Resultsfortheredu tionusingBIRKA 131
Chapter8. Redu tionofthermalmodelswithgeometri variations 145
8.1. Reformulationofthelinearparametri asbilinearsystems 146
8.2. Methodsfortheinterpolationoftheredu edmodels 149
8.3. Redu tionandinterpolationusingreformulationone 154
8.4. Redu tionandinterpolationusingthese ondreformulation 159
Chapter9. Con lusionsandOutlook 171
9.1. SummaryandCon lusions 171
9.2. Futureresear h 173
AppendixA. Derivationofthebilinear
H
2
-optimal onditions 175A.1. Wilson onditions 175
A.2. Derivationoftheoptimality onditionsbyBennerandBreiten177
Bibliography 187
Diese Arbeit wäre ohne die Beteiligung vielerPersonenin dieser Art
undWeiseni htmögli hgewesen. Natürli hgiltmeinDankanersterStelle
meinemBetreuerProf. Dr. PeterBenner,dersi hbeimeinenBesu henin
MagdeburgimmerZeitfürmi hnahm,mirhilfrei heHinweiseundTippsgab
undinsbesonderedasWagnisIndustrie-undNumerik-fremdeDoktorandin
einging.Herzli henDank!
Auÿerdembedankei h mi hherzli hbeiProf. Dr. TobiasDamm fürdie
ÜbernahmederZweitkorrektur.
DieCSCGruppeamMPIinMagdeburghatmi hbeimeinenBesu hen
im-merfreundli haufgenommen. Danke füreureHilfsbereits haftunddiverse
s höneAbendeinMagdeburg,S hlossRingberg,imHarzundin Kroatien.
EinspeziellerDankgehtanmeineMentorinDr. UlrikeBaur.
MeinePromotionszeitbeiBos hwäreni htdasGlei hegewesenohnemeine
KollegenausderMathematikere ke. I hdankehierinsbesondereDr.
Ka-trinS huma herundDr. RudyEidfürdiehervorragendeBetreuung-ni ht
nurfa hli h,au hpersönli hhabei hvielvoneu hgelernt.Ebensobedanke
i hmi hbeiallenanderenKollegenausderCR/ARH,diemirvorallembei
Fragenzuthermis henSimulationenweitergeholfenhaben,zusätzli hdanke
i hDr. KilianKrienerundThomasHeidvonED/ESY3.
I h danke auÿerdemmeinen Elternund meiner S hwesterfür ihre
Unter-stützungwährendnuns honüber10JahrenMathematik. DankeanSophie
fürüber6JahreHorizonterweiterung.
WirdinderIndustrieeineneueKomponenteentwi kelt,sospielen
Com-putersimulationenmittlerweileeinewi htigeRolle. Immers hnellereund
im-mergenauereSimulationsmodellewerdengewüns ht,damitZeitundKosten
gespart werden können. Mit Hilfe von Modellordnungsreduktion(MOR)
kann man aus groÿen, mit der FiniteElemente Methode erstellten
Mod-ellenkleineundgenaueModelleerhalten,diedann inkurzer Zeitsimuliert
werdenkönnen. Immerhäugerwirdau hgefordert,dieVariationvon
Pa-rameternimgroÿenFiniteElementeModellaufdiekleinenreduzierten
Mod-ellezuübertragen. DieseParameterbes hreibenbeispielsweisevers hiedene
Randbedingungen,dieimModellabgebildetwerden,genausowie
Änderun-geninderGeometrie(z.B.VariationvonLängen). MitHilfevonMethoden
ausderparametris henModellordnungsreduktion(pMOR)könnendiese
Pa-rameterabhängigkeitenau himreduziertenModellerhaltenundzur
Simu-lationvonunters hiedli henSzenariengenutztwerden.
Anstattdie heute übli hen Verfahren zur pMOR zu benutzen, werden in
dieserArbeit die parametris henModelle, die eine spezielle
Parameterab-hängigkeitzeigen, in bilineareModelle umges hrieben. Nun können au h
Verfahrenzur bilinearenModellordnungsreduktionangewandt werden,
ins-besondereVerfahrenzur
H
2
-optimalenReduktion.ZieldieserH
2
-optimalen Verfahren ist es, den Fehler zwis hen dem Ausgangsmodellund demre-duziertenModellinder
H
2
-Normzuminimieren. Wirverwendenzumeinen densogenanntenBilinearInterpolatoryRationalKrylovAlgorithm(BIRKA)vonBennerundBreiten[12℄. Auÿerdementwi kelnwirneuebilineare
H
2
-optimaleAlgorithmen,dieaufOptimierungsverfahrenaufGrassmann-Man-nigfaltigkeitenberuhen.
Dietheoretis henGrundlagenderthermis henModellierungwerdenerklärt
ModellekönnenausdenFiniteElementeModellendur heineAnalyse der
Glei hungenabgeleitetwerden. DieParametersindeinerseitsGröÿen, die
das thermis heVerhalten währenddes Betriebserklärenundandererseits
Gröÿen, dieVariationenin der Geometriedes Motors bes hreiben. Diese
ParametersollenindenreduziertenModellenerhaltenbleiben.
Währenddieneuentwi keltenAlgorithmenno hni htreiffürdieReduktion
vongroÿenModellensind,wirdinderArbeitgezeigt,dassdieReduktionmit
BIRKAzuguten reduziertenModellenführt. Allerdingsmüssen dazu
ver-s hiedeneNa hbesserungenanderReduktionsmethodikvorgenommen
wer-den, beispielsweisemüssen Methoden zur Stabilitätserhaltung angewandt
werden. InModellen mitVariationen in derGeometrie,werden zusätzli h
zumursprüngli henBIRKAna hderReduktionno hInterpolationsverfahren
verwendet,umreduzierteModellemitderParameterabhängigkeitdes
The design pro ess of a new omponent in industry is nowadays
al-mostalwaysa ompaniedby omputersimulations. In order to savetime
andmoney,fastanda uratemodelsforthesimulationofthe omponent
arerequired. UsingModelOrderRedu tion(MOR)largemodelsobtained
by Finite Elementsimulations an beredu edto smallmodels possessing
thesame behavioras the original. Oftenit isrequiredto obtain redu ed
models,wherethe dependen e in one orseveralparameters (for example
thelengthorwidth ofapart)oftheoriginalmodel ispreserved. Usingso
alledparametri ModelOrder Redu tion(pMOR) the parameters in the
redu edmodel anbevariedandthemodels anbeusedforfastsimulation
ofseverals enarios.
Insteadof using the ommonlyemployedmethods from pMOR,methods
frombilinearModelOrderRedu tionwillbeusedwithinthiswork,as
para-metri modelswitha ertainformofparameterdependen e anberewritten
asbilinearmodels. We fo uson methodsfrombilinear
H
2
-optimalModel OrderRedu tion,astheirobje tiveistominimizetheerrorbetweentheorig-inaland theredu edmodel measuredinthe
H
2
-norm. First, theBilinear InterpolatoryRationalKrylovAlgorithm(BIRKA)developedbyBennerandBreiten[12℄isused. Se ond,wederivenewbilinear
H
2
-optimalalgorithms basedonoptimizationonGrassmannmanifolds.Thefoundationsofthermalmodelingandtheirappli ationtothermal
sim-ulationsofele tri almotorsusingFiniteElementsoftwarewillbeexplained.
Parametri modelssuitableforpMOR anbederivedfromaFiniteElement
softwareanalyzingtheunderlyingequations. Two lassesofparameterswill
be onsidered:Constantsinuen ingthethermalbehaviorofthemodeland
hangesinthegeometryofthemodel.
we ndthat they are not yet readyforthe redu tionof large parametri
modelsasen ounteredinourthermalsimulations. In ontrast,theBIRKA
performswellfortheredu tionofthesemodels. However,several
modi a-tionsontheredu tionmethodsneedtobeperformedtoassure,forexample,
thepreservationofstabilityduringtheredu tion. Fortheredu tionof
mod-els with parameters resulting from hangesin the geometry, interpolation
pro eduresneedtobeappliedaftertheredu tiontotransfertheparameter
3.1 Temperatureontheinterfa ebetweentwosolidsin onta t. 27
3.2 Drawingofasli ethroughanele tri almotor. 31
3.3 Example: Bos hgenerators. 32
3.4 Modelforsimulatingtheheattransferinastator. 33
4.1 Simplelinears alingofare tangle. 38
4.2 S alingsforthegeometryvariationofanele tri almotor. 39
4.3 Dierents alingsshowninthemotormodel. 39
4.4 Modelparametrizedingeometry,topview. 40
4.5 S alingofatriangularmeshelementintheannulus. 41
4.6 Simulationoflargemodelnos alingfun tionwasapplied. 46
4.7 Simpliedmotormodel. 47
4.8 Simpliedmotormodelafterthes alingandashortsimulation. 48
6.1 Proposedworkowforstabilization. 112
6.2 Redu tionwithstabilizationviamirroringofpoles. 114
7.1 Redu tionwithbilGFA,bilFGFAandbilSQA. 124
7.2 Des entinfun tion
J (U)
. 1257.3 ResultsbilGFA,(I1),dierentstopping riteria. 127
7.4 ResultsbilFGFA,(I1),dierentstopping riteria. 128
7.5 ResultsbilSQA,(I1),dierentstopping riteria. 129
7.6 Comsol R
7.7 Temperatureproleoriginalandredu edordermodel. 133
7.8 Resultsfordierentheattransfer oe ients. 135
7.9 One-sidedmethods. 137
7.10Temperatureprolesfordierentheattransfer oe ients. 140
7.11Redu tionwithdierentapproa hes. 141
8.1 Interpolationofredu edordermodelrstoutput. 156
8.2 Interpolationofredu edordermodelfourthoutput. 157
8.3 Interpolationwithdierentmethods,modelwithfourparameters.160
8.4 Interpolationwithdierentmethods,vesamplingpoints,
p
new
1
. 1628.5 Interpolationwithdierentmethods,vesamplingpoints,
p
new
2
. 1638.6 Interpolationwithdierentmethods,vesamplingpoints,
p
new
3
. 1648.7 Interpolationusingtheapproa hes(P1)and(P2). 167
5.1 Exponentialandlogarithmmappingsfordierentmanifolds. 66
7.1 ResultsforbilGFA,bilFGFAandbilSQA,initialization(I1). 126
7.2 ResultsforbilGFA,bilFGFAandbilSQA,initialization(I2). 126
7.3 Comparisonofsimulationandredu tiontimes. 136
8.1 Tworeformulationmethodsshortsummary. 150
8.2 One-stepmethodsfortheinterpolationofredu edordermodels. 153
8.3 Two-stepmethodsfortheinterpolationofredu edordermodels. 154
8.4 Costsmodelwithoneaneparameter. 158
1 IRKAasgivenin[6℄. 60
2 GeneralizedSylvesteriteration( f. [12℄). 77
3 BilinearIRKAforsystemswith
E
6= I
,E
nonsingular( f. [12℄). 794 GFAforbilinearsystems(bilGFA). 91
5 FGFAforbilinearsystems(bilFGFA). 93
R
eldofrealnumbersC
eldof omplexnumbersRe(z )
realpartofz
∈ C
Im(z )
imaginarypartofz
∈ C
A
T
transposeofthematrixA
∈ C
n×m
A
∗
onjugate transpose of the matrixA
∈ C
n×m
A
−1
inverseofthematrix
A
∈ C
n×n
A
−∗
inverse, onjugatetranspose ofthe
matrix
A
∈ C
n×n
I
n
identitymatrixofdimensionn
× n
σ
i
(A)
i
-thsingularvalue ofamatrixA
∈
C
n×m
, singular values ordered by magnitudeλ
i
(A)
,λ
i
(A, E)
i
-theigenvalueofamatrixA
∈ C
n×n
insomeordering,generalized
eigen-valueofthepen il
A
− λE
insomeordering
||A||
2
= σ
max
(A)
spe tralnormofamatrix ve(A) =
a
11
, . . . , a
1n
, a
21
, . . . , a
mn
T
the ve tor formed by sta king the
olumnsofthematrix
A
∈ C
n×m
det(A)
determinantofamatrixA
tr(A) =
P
n
k=1
a
kk
tra eofamatrixA
span(A)
spa espannedbythe olumnsofthematrix
A
A
⊗ B =
a
11
B
. . .
a
1m
B
. . . . . .a
n1
B
a
nm
B
Krone ker produ tof two matri esA
∈ C
n×m
andB
∈ C
k×l
diag(a
11
, . . . , a
nn
) =
a
11
0
. . .0
a
nn
matrixwithdiagonal(a
11
, . . . , a
nn
)
O
n
⊂ R
n×n
orthogonalrealmatri esi.e.matri- eswith
A
T
A
= AA
T
= I
n
P
rea habilityGramianQ
observabilityGramianκ
2
(A) = ||A||
2
||A
−1
||
2
=
σ
max(A)
σ
min(A)
Introdu tion
1.1. Motivation
Inindustry,simulationsareanimportant toolin thedesignpro essof
a new omponent. In order to save time and money, fast and a urate
modelsforsimulationareneeded. ModelOrderRedu tion(MOR)isa
pow-erfulmethodtoobtainsmallanda uratemodelsfromlargeFiniteElement
models.Moreandmoreoften,FiniteElementmodelsareused,whi h
on-tainseveralparameters. Su h parameters anbe lengths and heights as
wellasphysi albehavior. Theseparametrizedmodelswilloftenbeusedto
ndoptimaldesignsbyusingoptimizationw.r.t. thegiven parameters. As
theFiniteElementmodelsarelarge,optimizationruns aneasilyex eedthe
omputation apa ities.Itishen edesirabletoredu emodelswhile
preserv-ingthe parameterdependen y. This istheobje tive ofparametri Model
OrderRedu tion(pMOR). Re ently,Bennerand Breiten[11℄ presenteda methodtorewritelinearparametri modelsintobilinearmodels.Thisallows
bilinearModel OrderRedu tionmethods tobeused forparametri Model
OrderRedu tion. Theresultingredu edordermodelshouldbeagood
ap-proximation of the original model. Within the framework of
H
2
-optimal ModelOrder Redu tion,the error anbemeasuredand minimizedin theH
2
-norm. In this work, we will examinebilinearH
2
-optimalmethods for theredu tionoflinearparametri systems,whi hhavebeenappliedtoand1.2. Dissertationoverview
InChapter2,wereviewresultsfromLinearAlgebra,Dierential
Geom-etryandSystemsTheory. The on eptswillbestatedforlinearandbilinear
systems.
Chapter3 providesthereaderwiththefoundationsofheattransfer
mod-eling. Theunderlyingphysi alee ts(heat ondu tan e, onve tiveheat
transfer, radiation) will be reviewedand the mode of operation and the
thermalmodelingofanele tri almotorwillbedes ribed. Threedierent
ele tri almotor modelshavebeen builtand will bepresented. Chapter4
gives anoverview over theequations that aresolved duringheat transfer
modeling,andthepro eduretoobtainparametri modelsby arefulanalysis
oftheseequations.
InChapter5,methodsforModelOrderRedu tion(MOR)willbedis ussed.
First, methods for linearMOR will be reviewed, followed by a dis ussion
ofmethods for theredu tionofparameterdependentmodels(parametri
MOR).Itispossibletorewriteparametri modelswitha ertainparameter
dependen yasbilinearmodels,andhen emethodsfrombilinearMORwill
be onsidered. Of parti ularinterest are methods from the lassof
H
2
-optimal bilinearMOR,as their obje tiveis to minimizetheerrorbetweenoriginalandredu edmodel. First,wereviewexistingmethodsandstatethe
Bilinear Interpolatory Rational Krylov Algorithm(BIRKA) [12℄. Se ond, wedevelopalgorithmsfortheredu tionofbilinearsystemsviaoptimization
on Grassmannmanifolds. Thesemethodsareofinterest,as theypreserve
stabilityduringtheredu tionpro ess.
The obje tive of Chapter 6 is the dis ussionof several issues that were
en ountered while applying BIRKA to thermal models. Theseissues are
examined,and strategiesfortheirmitigationwillbedeveloped. Espe ially
preservationofstabilityduringthe al ulationis ru ial. ResultsforBIRKA
andthenew
H
2
-optimalmethodswillbegiveninChapters7and8.Whereas thenewmethodsarenotyetappli ableto largesystems,BIRKAperformswellonbilinearsystemsthathavebeenobtainedfromlinearparametri
sys-tems. First, onlyphysi alparametersare onsidered. Se ond, wepresent
resultsforsystemswithaparameterdependen yresultingfrom hangesin
geometry, whi h anonly be rewritten partiallyas bilinear systems. For
su hsystems,parametri redu edordermodels anthenbeobtainedbyan
1.3. Thesis ontributions
Themain ontributionsofthisthesisare:
•
Oneobje tive ofthisthesis isMORof thermalele tri almotor models. Hen e,itisshownhowmatri essuitableforpMOR anbeobtainedfromComsol R
,aFiniteElementSoftware. Todoso,
theequationswhi haresolvedbytheSoftwareareusedto
theo-reti allyre onstru tthedependen einparametersofthemodel
( f.Chapter4).
•
In ontrasttootherworksaboutpMOR,inthisthesisthe redu -tionoftheparametri modelsisdoneusingBIRKA[12℄. Several issueswhereen ounteredwhen thealgorithmwas applied: Onelassofparametersleadstoanon-singularstinessmatrix,in
sev-eral asesthereistheneedtos aleothersystemmatri estofulll
aKrone kerprodu tapproximationandinaddition,BIRKAdoes
notpreservestability.Alltheseissueshavebeenresolved,andwe
showresultsfortheredu tionofamotormodelfrom
n
= 41, 199
degreesoffreedomtoaredu edorderof
r
= 300
. Thishasbeendonefor
13
physi alparameters.•
In addition, models with geometri al variations are onsidered. Aftertheredu tionwithBIRKA,severalinterpolationstrategiesbetweentheredu edordermodelsobtainedin severalparameter
pointshavebeen ompared.
•
Finally, we develop newH
2
-optimal bilinearmethods for MOR usingoptimizationonGrassmannmanifolds. Thesemethods anpreservestabilityforsymmetri systemsmatri es,andtheir
Mathemati al prerequisites
2.1. LinearAlgebra 5
2.2. Dierentialgeometry 8
2.3. Systemstheory 10
In this rst theoreti al hapter, some results from dierent areasof
mathemati sarereviewed.First,generalresultsfromLinearAlgebrawillbe
presented, followedby a loser look on some denitions fromDierential
Geometry. The last se tion provides the reader with an introdu tion to
linearandbilinearsystemstheory.
2.1. LinearAlgebra
Withinthisse tionwereviewthede ompositionofmatri es,the
prop-ertiesoftheKrone kerprodu tandprovidethereaderwithbasi knowledge
onmatrixpertubationtheory.
2.1.1. Matri es and their de ompositions. Most of thematri es in
thisworkaresymmetri ,whi hiswhywestatethedenitionhere.
Denition 2.1.1. A matrix
A
∈ R
n×n
is alled symmetri if
A
= A
T
. A
symmetri matrixispositive(semi)denite,denotedby
A >
(≥)0
,ifx
T
Ax >
(≥)0
forall ve tors0 6= x ∈ R
n
. Itisnegative(semi)denite,denotedby
A <
(≤)0
,ifx
T
Ax <
(≤)0
forallve tors
0 6= x ∈ R
n
We will often refer to the following two matrixde ompositions, the
eigenvalueandthesingularvaluede omposition.
Denition2.1.2 (Generalizedeigenvaluede omposition[38,Se tion7.7℄). If
A, B
∈ C
n×n
,thenthesetofallmatri esoftheform
A
− λB
withλ
∈ C
isapen il. Thegeneralizedeigenvaluesof
A
− λB
areelementsofthesetλ(A, B)
denedasλ(A, B) = {z ∈ C : det(A − z B) = 0}.
Ifλ
∈ λ(A, B)
and0 6= x ∈ C
n
satisesAx
= λBx,
(2.1)then
x
is an eigenve tor ofA
− λB
. The problem of nding nontrivialsolutionsto(2.1)isthegeneralizedeigenvalueproblem.If
B
isnonsingular,λ(A, B) = λ(B
−1
A)
holds.
Theorem 2.1.3 (The singularvalue de omposition (SVD) [38, Theorem 2.4.1℄). If
A
∈ R
m×n
,thenthereexistorthogonalmatri es
U
= [u
1
, . . . , u
m
] ∈
R
m×m
andV
= [v
1
, . . . , v
n
] ∈ R
n×n
su hthat
U
T
AV
= diag(σ
1
, . . . , σ
p
) ∈ R
m×n
,
(2.2) withp
= min (m, n)
whereσ
1
≥ σ
2
≥ · · · ≥ σ
p
≥ 0
.The
σ
j
will be alled singularvalues. If it shallbe lariedthat they resultfromasingularvaluede ompositionofthematrixA
,wedenotethemby
σ
j
(A)
. Letr
besu hthatσ
1
≥ σ
2
≥ · · · ≥ σ
r
> σ
r +1
= · · · = σ
p
= 0
. Thenrk(A) = r
andA
anbede omposedinthefollowingway:A
=
r
X
i =1
σ
i
u
i
v
i
T
.
Usingmatri es,wewillwritethisde ompositionasfollows:
A
= U
r
Σ
r
V
r
T
,
(2.3) withU
r
∈ R
m×r
,Σ
r
∈ R
r ×r
andV
r
∈ R
n×r
andreferto it asthe ompa t
2.1.2. Properties of the Krone ker produ t. The following matrix
produ tisreferredtoastheKrone kerprodu t:
Denition2.1.4. Fortwomatri es
A
∈ C
n×m
and
B
∈ C
k×l
,theKrone ker
produ t isdenedas:
A
⊗ B =
a
11
B
. . .
a
1m
B
. . . . . .a
n1
B
a
nm
B
.
The Krone ker produ thasthe followingproperties(see for example
[38℄,Se tion12.3):
(A ⊗ B)
T
= A
T
⊗ B
T
,
withA
∈ C
n×m
, B
∈ C
k×l
,
(A ⊗ B)
−1
= A
−1
⊗ B
−1
,
withA
∈ C
n×m
, B
∈ C
k×l
,
(A ⊗ B) ⊗ C = A ⊗ (B ⊗ C),
withA
∈ C
n×m
, B
∈ C
k×l
andC
∈ C
s×q
,
(AC ⊗ BD) = (A ⊗ B)(C ⊗ D),
withA
∈ C
n×m
, B
∈ C
k×l
, C
∈ C
m×s
andD
∈ C
l ×q
,
butingeneral
A⊗B 6= B ⊗A
! Inadditiononeobtaines(withA
∈ C
n×m
, B
∈
C
k×l
):
rk(A ⊗ B) = rk(A) · rk(B),
det(A ⊗ B) = det(A)
n
· det(B)
m
forA
∈ R
m×m
andB
∈ R
n×n
,
tr(A ⊗ B) = tr(A) · tr(B),
||A ⊗ B||
2
= ||A||
2
· ||B||
2
.
IfC
= AXB
forC
∈ R
n×m
, A
∈ R
n×k
, X
∈ R
k×l
andB
∈ R
l ×m
thenoneobtainsfortheKrone kerprodu tandtheve operator:
ve
(C) = (B
T
⊗ A)
ve
(X).
(2.4)2.1.3. Matrixpertubationtheory. The onne tionbetweenthe
eigen-valuesoftwomatri eswillbeneededwithinthiswork. Thefollowingresults
havebeenestablishedin the ontextofmatrixpertubationtheory, the
re-lationoftheeigenvaluesofapertubedMatrix
M
+ S
andtheunpertubedTheorem2.1.5(Bauer-Fike,[38,Theorem7.2.2℄). If
µ
isaneigenvalueofM
+ S ∈ C
n×n
andX
−1
MX
= diag(λ
1
, . . . , λ
n
)
,thenmin
i =1,...,n
|λ
i
− µ| ≤ κ
2
(X)||S||
2
.
(2.5) Corollary2.1.6. LetX
−1
MX
= diag(λ
1
, . . . , λ
n
)
,andM
+ S ∈ C
n×n
. Foreveryeigenvalue
λ(M + S)
aneigenvalueλ
i
(M)
existssu hthat|λ
i
(M) −
λ(M + S)| ≤ κ
2
(X)||S||
2
.
Thenextresultsshowthe onne tionbetweentheeigenvaluesoftwo
realsymmetri matri es
A
andB
.Proposition 2.1.7 (Weyl,[60, Theorem 4.8, Corollary 4.9℄). Let
A, B
∈
R
n×n
betwosymmetri matri es. Let
λ
i
(A)
andλ
i
(B)
fori
= 1, . . . , n
be theeigenvaluesofA
andB
withλ
1
(A) ≥ · · · ≥ λ
n
(A)
andλ
1
(B) ≥ · · · ≥
λ
n
(B)
. Thenitholds:λ
i
(A + B) ∈ [λ
i
(A) + λ
n
(B), λ
i
(A) + λ
1
(B)]
fori
= 1, . . . , n.
(2.6)Corollary2.1.8 ([60, Corollary4.10℄). Undertheassumptionsof Proposi-tion2.1.7itholds
|λ
i
(A + B) − λ
i
(A)| ≤ ||B||
2
fori
= 1, . . . , n.
(2.7)2.2. Dierentialgeometry
InSe tion5.5.4,severalalgorithmsbasedonoptimizationonmanifolds
willbederived. Foramoredetailedpresentationofthistopi , wereferto
[1℄and[30℄. Let
O
r
denotethesetoftheorthogonalmatri esinR
r ×r
.
Denition2.2.1(Stiefelmanifold[1,Se tion3.3.2℄). For
r
≤ n
,theStiefel manifoldisdenedasthesetofalln
× r
orthonormalmatri es:St
(r, n) := {X ∈ R
n×r
|X
T
X
= I
r
}.
Clearly,St
(r, n) ⊂ R
n×r
. It anbeshownthat St
(r, n)
is a ompa tsubmanifoldof
R
n×r
( f.[1,Se tion3.3.2℄). Thetangentspa eofaStiefel manifoldat
X
∈
St(r, n)
isdenedasfollows( f. [1,Example3.5.2℄):T
X
St(r, n) = {Z ∈ R
n×r
|X
T
Z
+ Z
T
X
= 0}.
Headingfor analgorithmforthegradientow, thegradientofafun tion
on themanifoldhastobe al ulated. Therefore, werst haveto provide
denitionofaninnerprodu tonthetangentspa e.ForaStiefelmanifold,
theinnerprodu t isdenedas
hξ, ηi = tr(ξ
T
η)
withξ, η
∈ T
X
St(r, n).
(2.8)Thegradientin
X
ofafun tionF
onaStiefelmanifoldisdenedtobethetangentve tor
∇F
su hthattr(F
X
T
Y
) = tr((∇F )
T
(I −
1
2
XX
T
)Y ),
(2.9)
holdsforalltangentve tors
Y
∈ T
X
St(r, n)
. Here,F
X
isthematrixofall partialderivativesofF
withrespe ttoX
,i.e.:(F
X
)
i j
=
∂F
∂X
i j
.
(2.10)Solvingequation(2.9)leadstothefollowingexpressionforthegradient:
∇F = F
X
− XF
X
T
X.
(2.11)TheGrassmann manifold
Gr(r, n)
,r
≤ n
, is dened as the set of allr
-dimensionalsubspa esof
R
n
. Following[30℄,it anbeseenasaquotient manifoldinthefollowingway: Twomatri es
U
1
andU
2
inSt(r, n)
are equiv-alent,iftheyspanthesamer
-dimensionalsubspa e. Thisholdsifandonlyif
U
1
= U
2
Q
foranorthogonalmatrixQ
∈ R
r ×r
. Theequivalen e lass
[U]
ofapoint
U
∈
St(r, n)
anbedenedas:[U] = {UQ|Q ∈ O
r
} .
Themap
G
: Gr(r, n) →
St(r, n)/O
r
isabije tion. We will therefore onsiderthe Grassmannmanifoldas this
quotientmanifold ofSt
(r, n)
. A matrixU
∈ St(r, n)
represents a wholeequivalen e lassin
Gr(r, n)
. Thetangentspa eoftheGrassmannmanifoldanbedes ribedasfollows[30,Se tion2.5℄:
T
X
Gr(r, n) = {Z ∈ R
n×r
|X
T
Z
= 0}.
(2.12)Onamanifold,theshortest onne tionbetweentwopointsis alleda
ge-odesi . Let
X(0) = X
andX(0) = H
˙
. LetH
= W ΣV
T
bethe ompa t
singularvalue de omposition ( f. equation(2.3)) of
H
withW
∈ R
n×r
,
Σ, V ∈ R
r ×r
.
Thegeodesi anbedes ribedas[30,Se tion2.5.1℄:
X(t) =
XV
W
cos Σt
sin Σt
V
T
.
(2.13)ForaGrassmannmanifold,theinnerprodu tisdenedas
hξ, ηi = tr(ξ
T
η),
withξ, η
∈ T
X
Gr(r, n).
(2.14)Thegradientin
X
ofafun tionF
ontheGrassmannmanifoldisdenedtobethetangentve tor
∇F
su hthattr(F
X
T
Y
) = tr((∇F )
T
Y
),
(2.15)holdsforalltangentve tors
Y
∈ T
X
Gr(r, n)
. Solvingequation(2.15)leads tothefollowingexpressionforthegradient[30,Se tion2.5.3℄:∇F = F
X
− XX
T
F
X
.
(2.16)Wewillalsoneedthefollowingdenition:
Denition 2.2.2 ([1,Denition4.2.1℄). Givenafun tion
F
on St(r, n)
or Gr(r, n)
,asequen e{η
k
}
,η
k
∈ T
x
k
St(r, n)
orη
k
∈ T
x
k
Gr(r, n)
is gradient-related if,forany subsequen e{x
k
}
k∈K
of{x
k
}
that onvergesto a non- riti alpointofF
,the orrespondingsubsequen e{η
k
}
k∈K
isboundedand satiseslim
k→∞
sup
k∈K
h∇F (x
k
), η
k
i < 0.
(2.17)
2.3. Systemstheory
Manyphysi alphenomena, hemi alrea tions,biologi alpro essesor
models for the fore ast of nan ial pro esses anbe mathemati ally
de-s ribedbythesame lassofsystems,so alleddynami alsystems. External
inuen esthathaveadire timpa tonthebehaviorofthesystemare alled
inputs. Thebehaviorofthesystemswillbemonitoredwithina ertaintime
rangeandat ertainpoints,thesystem'soutputs. The onne tionbetween
the inputsandthe outputswill oftenbemeasuredand referredto as the
system'sinput-output-relationship.Adynami alsystem anbedes ribedby
adierentialequation.Inthiswork,twokindsofdynami alsystemswillbe
2.3.1. LinearSystems. Inthefollowingse tionsomebasi knowledge
onlineardynami alsystemswillbereviewed,su hasstability,observability,
ontrollability, balan ed systems, norms of systems and the input-output
relationship.
Denition 2.3.1. A linear system
Σ
lin
of ordern
is a system ofordinary dierentialequationsofthefollowingform:Σ
lin
:
E
˙x (t) = Ax(t) + Bu(t),
y
(t) = Cx(t), x(0) = x
0
,
(2.18) whereE, A
∈ R
n×n
,B
∈ R
n×m
,C
∈ R
p×n
. Theinput
u(t) ∈ R
m
anbe
time-dependentjustasthestates
x(t) ∈ R
n
andtheoutput
y
(t) ∈ R
p
are.
Thevalueof
x(0) = x
0
is alled initialvalue. Thespa eX
ontainingall statesx(t)
is alledstatespa e.2.3.1.1. Stability. Systemswithboundedsolutiontraje tories
x(t)
areof spe ial importan e. This hara teristi of a system is referredto as
stability.Forlinearsystems( .f. system(2.18) )withnonsingular
E
,stability isdenedasfollows:Denition2.3.2( .f. [63℄Chapter2.7,[5℄Chapter5.8,[61℄Chapter3.2.1). Thesystem
E
˙x (t) = Ax(t), E
nonsingular, isasymptoti allystableif(i) Forall
x
0
∈ R
n
theinitialvalueproblem
E
˙x (t) = Ax(t), x(0) =
x
0
,
hasasolutionandforevery
ε >
0
thereexistsaδ >
0
su hthat
||x(t)||
2
< ε
forallt
≥ 0
andforall||x(0)||
2
< δ
(Lyapunov stability).(ii) Thereexists
δ >
0
su hthatx(t) → 0
ast
→ ∞
if||x(0)||
2
< δ.
Theorem2.3.3([63℄Corollary2.11,[61℄Theorem3.7). ThesystemE
˙x (t) = Ax(t), E
nonsingular,isasymptoti allystableifandonlyifalltheeigenvaluesof
λE
− A
lieintheopenlefthalf-plane.
We will therefore speak of a stable system, if all the eigenvalues of
λE
− A
,E
nonsingular, lie in the open left half-plane. In this ase, the eigenvaluesofthepen ilλE
− A
arethoseofthematrixE
−1
A
2.3.1.2. Controllability, Observability and Balan ed Systems. During
the analysis of a linear system (2.18) one might ask how the system is
ae tedbytheinput
u(t)
. Thefollowingtwo hara terisationswillbeon-sidered.
Denition 2.3.4 ([5℄).
x
∗
∈ R
n
isrea hable(from theorigin
x
(0) = 0
)ifthereexistan admissibleinputfun tionand
t
e
<
∞
su hthatx
(t
e
) = x
∗
holds (andhen e
x(t
e
) = x
∗
belongstothestatespa eofalinearsystem
(2.18)).
Denition2.3.5([5,46℄). Anonzerostate
x(0) = x
0
is ontrollableifthere existsanadmissibleinputfun tionsu hthatthesystem anbetransformedfrom
x
0
toanygivenendstatex
(t
e
)
withinanitetime[0, t
e
]
.Forlinear ontinuoustimesystemsthe on eptsof ontrollabilityand
rea hability oin ide( f. [5℄,Theorem4.18). Hen e,thefollowing on epts willbedevelopedforthe ontrollabilityofalinearsystem. Inthefollowing
hapterswewillneedthe on eptofthe ontrollabilityGramian.
Denition2.3.6([61℄Lemma4.57). Considerastablelinearsystem(2.18) with
E
nonsingular. The ontrollabilityGramian anbedenedasfollows:P
=
1
2π
Z
∞
−∞
(iωE − A)
−1
BB
∗
(iωE − A)
−∗
d ω.
(2.19)If one onsiders the eigenvaluede omposition of
P
, the eigenvaluesmeasurethedegreeof ontrollability,whereastheeigenve tors
orrespond-ingtothelargesteigenvalues anbeunderstoodasthedire tionsinwhi h
thesystemiseasyto ontrol.
Proposition 2.3.7 ([61℄ Corollary 4.58). Consider a stable linear system (2.18)with
E
nonsingular.The ontrollabilityGramianP
(2.19)existsand istheuniqueHermitiansolutiontothefollowingLyapunovequation:AXE
T
+ EXA
T
+ BB
T
= 0.
(2.20)Inaddition,
P
ispositivedeniteifandonlyifthesystemis ontrollable.Inpra ti e,wewilloftenbeabletomeasuretheoutput
y
(t)
ofalinearsystem (2.18) . Iftheinput
u(t)
andtheoutputy
(t)
areknown,wewant tore onstru tthestatesx
(t)
. Thisleadstothe on eptofobservability.Denition2.3.8([46℄). Alinearsystem(2.18)is ompletelyobservable,if theinitialstate
x
0
anbere onstru tedfromthebehavioroftheinputu(t)
andtheoutputy
(t)
withinanitetimeinterval[0, t
e
]
.Again,wewillneedthe on eptofthesystemsobservabilityGramian.
Denition2.3.9. Considerastablelinearsystem(2.18)with
E
nonsingular. TheobservabilityGramianQ
isdenedasfollows:Q
= E
T
QE,
˜
with˜
Q
=
1
2π
Z
∞
−∞
(iωE − A)
−∗
C
∗
C
(iωE − A)
−1
d ω.
(2.21)Theinterpretationissimilartothe ontrollability ase: Ifone onsiders
theeigenvaluede ompositionof
Q
,theeigenvaluesmeasurethedegreeofobservability, whereas the largest eigenve tors anbe understood as the
dire tionsinwhi hthesystemiseasyto observe.
Proposition 2.3.10([61℄ Corollary4.58). Considera stablelinearsystem (2.18)with
E
nonsingular. ThematrixQ
˜
(seeDenition2.3.9)existsand istheuniqueHermitiansolutionto thefollowingLyapunovequation:A
T
XE
+ E
T
XA
+ C
T
C
= 0.
(2.22)In addition,
Q
˜
and therefore also the observabilityGramianQ
is positivedeniteifandonlyifthesystemisobservable.
Abalan edrepresentationofalineardynami alsystemisa
representa-tionofthesysteminwhi heverystateisequally"rea hableandobservable.
Thisse tionintrodu esthe on eptswhi hwillbeneededfortheBalan ed
Trun ation Model Order Redu tionin Se tion 5.2.1. The reader should
notethat thereexistseveralotherbalan edrepresentationsbesidetheone
presentedhere. They anbefoundintheworkbyGuger inandAntoulas
[40℄andthereferen estherein.
Denition 2.3.11 ([61, Denition7.5℄). The Hankel singularvalues, de-noted by
ς
j
, of a stablelinearsystem (2.18) withE
nonsingularare the square-rootsoftheeigenvaluesofP Q
.Proposition2.3.12([61,Corollary7.7℄). Astablelinearsystem(2.18)with
E
nonsingularis ontrollableandobservableifandonlyifitsHankelsingular valuesarenon-zero.Denition2.3.13([61,Denition7.10℄). Astablelinearsystem(2.18)with
E
nonsingularis alledbalan ed,ifthe ontrollabilityandtheobservability Gramiansareequalanddiagonal.Everystable, ontrollableandobservablelinearsystem with
E
nonsin-gular anbe transformed intoa balan ed representation. To do so, one
omputestheCholeskyfa torizationoftheGramians
P
= RR
T
andQ
˜
= L
T
L,
whi hexistsdueto thepositivedenitenessof
P
andQ
˜
( f. Propositions2.3.7and2.3.10). ComputingtheQRde ompositionoftheCholesky
fa -tors
L
andR
leadstothefollowingde ompositionwithorthogonalmatri esQ
c
andQ
o
:R
T
= Q
c
R
˜
T
andL
= Q
o
˜
L.
It isobvious thatP
= RR
T
= ˜
R ˜
R
T
andQ
˜
= L
T
L
= ˜
L
T
L
˜
. TheHankelsingularvalues annowbe omputedviathesingularvaluesof
LE ˜
˜
R
:ς
j
2
= λ
j
(P E
| {z }
T
QE
˜
Q
) = λ
j
( ˜
R ˜
R
T
E
T
L
˜
T
LE) = λ
˜
j
( ˜
R
T
E
T
L
˜
T
LE ˜
˜
R) = σ
2
j
(˜
LE ˜
R),
withthesingularvaluede omposition
˜
LE ˜
R
= U
b
ΣV
b
T
,
andorthogonal
U
b
, V
b
andΣ =
diag(ς
1
, . . . , ς
n
).
Thematri esofthelinear system annowbetransformedtoabalan edsystemrepresentation:W
b
T
ET
b
, W
b
T
AT
b
, W
b
T
B, CT
b
,
where
W
b
= ˜
L
T
U
b
Σ
−1/2
,
T
b
= ˜
RV
b
Σ
−1/2
,
W
b
−1
= T
b
T
E
T
,
T
b
−1
= W
b
T
E.
TheGramian(astheobservabilityandthe ontrollabilityGramian oin ide
f. Denition2.3.13)ofthebalan edsystemisobtainedfromthoseofthe
originalsysteminthefollowingway:
T
b
−1
P T
b
−T
= Σ = W
b
−1
QW
˜
b
−T
= T
b
T
QT
b
.
2.3.1.3. Systemsnormsandspa esandinput-outputrelationship. As
theobje tiveisto approximatethegiven originalmodels,oneneedstobe
abletoquantifythedieren ebetweentheoriginalandtheredu edsystem,
or generallyspeaking, between twodynami alsystems. To do so,several
dierent spa es and their norms, both in the time and in the frequen y
Denition2.3.14([5, Se tion5.1.2℄). Let
f
: I → R
n
,with
I ∈ {R, R
−
,
R
+
,
[a, b]}
beave torvaluedfun tion. TheLebesguespa eL
n
2
(I)
isdened as:L
n
2
(I) =
(
f
: I → R
n
:
Z
t∈I
||f (t)||
2
2
1
2
<
∞
)
.
(2.23)Inourmodels,inputandoutputwillbe onsideredasfun tionsinthese
spa es:
u(t) ∈ L
m
2
(I)
andy(t) ∈ L
p
2
(I)
witht
∈ I
( f. thedenitionof alinearsystem(2.18) ). Usually,oneisinterestedinarelationshipbetween inputand output. As su ha relationshipin the timedomainis des ribedby a onvolutionwhi his oftendi ultto al ulate,the relationisoften
examinedinthefrequen ydomain. There,it aneasilybedeterminedbya
produ tofmatri es,as wewillseeinthisse tion. Forthetransformation
fromtimetofrequen ydomaintheLapla etransformationisused.
Denition2.3.15([18,Se tion15.2℄). TheLapla etransformofafun tion
f
: R
+
→ R
isdenedasF
(s) = L{f (t)}(s) =
Z
∞
0
f
(t)e
−st
dt,
(2.24) withL{f
′
(t)}(s) = sF (s) − f (0).
(2.25) Forave tor,theLapla etransformhasto beseenelementwise. Wetransformthelinearsystem(assuming
x(0) = x
0
= 0
):L{E ˙x (t)}(s) = L{Ax(t) + Bu(t)}(s)
⇒
EL{ ˙x (t)}(s) = AL{x(t)}(s) + BL{u(t)}(s)
⇒
sEX(s) = AX(s) + BU(s)
⇒
X(s) = (sE − A)
−1
BU
(s),
and
Y
(s) = CX(s)
. Thisleads to thefollowing onne tion between theinputandtheoutput:
Y
(s) = C(sE − A)
−1
BU(s).
Denition2.3.16. Thetransferfun tion
H
: C → C
p×m
ofthelinearsystem
(2.18)isdenedas
Fun tions in frequen ydomain will oftenbe interpreted as fun tions
of a omplexvariable. A detaileddes riptionof frequen ydomainspa es
forlinearsystems anbefound in[5℄. Hereweuse Hardyspa es
H
2
andH
∞
. Thefollowingsystemnorms anthenbeestablishedusingthetransfer fun tionH(s)
andthe orrespondingHardyspa enorms:Denition2.3.17([5,Se tion5.1.3℄). The
H
2
normofastablesystemis denedas||Σ
lin
||
H
2
:=
Z
∞
−∞
tr(H
∗
(−iy )H(iy ))dy
1
2
.
(2.27)The
H
∞
normofastablesystemisdenedas||Σ
lin
||
H
∞
:= sup
y ∈R
(σ
max
(H(iy ))) ,
(2.28)withmaximalsingularvalue
σ
max
. Proposition2.3.18([5℄). Itholds:||Σ
lin
||
H
2
=
p
tr(B
∗
QB) =
p
tr(CP C
∗
),
(2.29)
forthesystemsGramiansasdenedin(2.21)and (2.19).
2.3.2. BilinearSystems. These ond lassofdynami alsystemswhi h
willbe onsideredinthisthesisarebilinearsystems.Anoverviewand
exam-ples anbefoundin[49℄.
Denition 2.3.19. A bilinearsystem of order
n
is asystem ofdierentialequationsofthefollowingform:
Σ
bil
:
E
˙x (t) = Ax(t) +
m
X
k=1
N
k
u
k
(t)x(t) + Bu(t),
y
(t) = Cx(t),
x(0) = x
0
,
(2.30) whereE, A, N
k
∈ R
n×n
,B
∈ R
n×m
,C
∈ R
p×n
. Theinput
u(t) ∈ R
m
anbe
time-dependentjustasthestates
x(t) ∈ R
n
andtheoutput
y
(t) ∈ R
p
are.
Thevalueof
x
(0) = x
0
is alledinitialvalue.Inthis se tion,onlysystemswith
E
6= I
n
,E
nonsingular,willbe on-sidered.2.3.2.1. Volterraseriesrepresentation. A onne tionbetweenthe
sys-tems inputand output anbeestablished by using thefollowing Volterra
seriesrepresentationforthestatesofbilinearsystemsestablishedbyMohler
[49℄. Wewill onsidersystemswith
E
nonsingular.x(t) =
∞
X
i =1
Z
∞
0
· · ·
Z
∞
0
m
X
k
1
,k
2
,...,k
i
=1
e
E
−1
A(τ
1
)
E
−1
N
k
1
·
· e
E
−1
A(τ
2
−τ
1
)
E
−1
N
k
2
e
E
−1
A(τ
3
−τ
2
)
· · · E
−1
N
k
i−1
e
E
−1
A(τ
i
−τ
i−1
)
E
−1
b
k
i
·
· u
k
1
(t − τ
1
) · · · u
k
i
(t − τ
i
)dτ
1
. . . d τ
i
.
(2.31)Theinput-outputrelationshipofthesystem anthenbedenedas:
y
(t) =
∞
X
i =1
Z
∞
0
· · ·
Z
∞
0
m
X
k
1
,k
2
,...,k
i
=1
Ce
E
−1
A(τ
1
)
E
−1
N
k
1
·
· e
E
−1
A(τ
2
−τ
1
)
E
−1
N
k
2
e
E
−1
A(τ
3
−τ
2
)
· · · E
−1
N
k
i−1
e
E
−1
A(τ
i
−τ
i−1
)
E
−1
b
k
i
· u
k
1
(t − τ
1
) · · · u
k
i
(t − τ
i
)dτ
1
. . . d τ
i
,
·
(2.32) with olumnsb
k
i
ofB
andVolterrakernelsdenedas:h
(k
1
,...,k
i
)
i
(τ
1
, . . . , τ
i
) = Ce
E
−1
Aτ
1
E
−1
N
k
1
e
E
−1
A(τ
2
−τ
1
)
· . . .
(2.33). . .
· E
−1
N
k
i−1
e
E
−1
A(τ
i
−τ
i−1
)
E
−1
b
k
i
,
where
i
= 1, 2, . . . , k
i
= 1, . . . , m,
andτ
i +1
≥ τ
i
≥ 0
. Theinput-output relation annowbewrittenas:y
(t) =
∞
X
i =1
Z
∞
0
· · ·
Z
∞
0
m
X
k
1
,k
2
,...,k
i
=1
h
(k
1
,...,k
i
)
i
(τ
1
, . . . , τ
i
)
(2.34)·
i
Y
j=1
u
k
j
(t − τ
j
)
!
d τ
1
. . . d τ
i
.
Inpra ti e,theVolterrakernels
h
(k
1
,...,k
i
)
i
(τ
1
, . . . , τ
i
)
needto be exam-ined in the frequen y domain as well. Therefore we need a multivariateDenition2.3.20([24℄). Givenafun tion
f
(t
1
, . . . , t
n
)
denedonR
n
dene
itsLapla etransform
F
(s
1
, . . . , s
n
)
by:F
(s
1
, . . . , s
n
) =
Z
∞
−∞
· · ·
Z
∞
−∞
f
(t
1
, . . . , t
n
)exp
−
n
X
k=1
t
k
s
k
!
d t
1
. . . d t
n
.
(2.35)We annowtransformtheVolterrakernels.
Denition2.3.21. The
i
-thordertransferfun tionoftheVolterrakernelh
(k
1
,...,k
i
)
i
(τ
1
, . . . , τ
i
)
= Ce
E
−1
Aτ
1
E
−1
N
k
1
e
E
−1
A(τ
2
−τ
1
)
. . . E
−1
N
k
i−1
e
E
−1
A(τ
i
−τ
i−1
)
E
−1
b
k
i
,
isdenedasH
(k
1
,...,k
i
)
i
(s
1
, . . . , s
i
)
= C(s
i
E
− A)
−1
N
k
1
(s
i −1
E
− A)
−1
. . . N
k
i−1
(s
1
E
− A)
−1
b
k
i
.
(2.36) BytakingN
= [N
1
. . . N
m
],
thisdenition anberewritten simultane-ouslyforallN
k
byusingKrone kerprodu ts:H
i
(s
1
, . . . , s
i
) =C(s
i
E
− A)
−1
N[I
m
⊗ (s
i −1
E
− A)
−1
](I
m
⊗ N) . . .
· [I
|
m
⊗ · · · ⊗ I
{z
m
}
i −2
times⊗(s
2
E
− A)
−1
)](I
|
m
⊗ · · · ⊗ I
{z
m
}
i −2
times⊗N)
· [I
m
⊗ · · · ⊗ I
m
|
{z
}
i −1
times⊗(s
1
E
− A)
−1
)](I
m
⊗ · · · ⊗ I
m
|
{z
}
i −1
times⊗B).
(2.37)Inaddition,Brunietal. [19℄examinedthe onvergen eoftheVolterra seriesandestablishedthefollowingresult:
Proposition2.3.22. IftheVolterraseriesin (2.31) onverges,thenit
uni-formly onvergestothesolutionofthebilinearsystem(2.30) . Forbounded
inputstheVolterraseries(2.31) onvergesonanynitetimeinterval
[0, t
e
]
. The onvergen eoftheVolterraseriesis onne tedto thestabilityof2.3.2.2. Stability. The notion of stability for bilinear systems diers
fromthatforlinearsystems. Forboundedinputs,thefollowingdenitionof
stabilityapplies:
Denition2.3.23([72,59℄). Thebilinearsystem(2.30)is alled bounded-input-bounded-output(BIBO)stable,ifforanyboundedinput,theoutput
isboundedon
[0, ∞)
. Aninput/outputis alledbounded ifitsatisesthefollowing ondition:
||u||
∞
= max
j
sup
t∈[0,∞)
|u
j
(t)| < M
.SiuandS hetzen[59℄ ombined onvergen eoftheVolterraserieswith BIBOstability. They showed the followingsu ient ondition forBIBO
stability.
Theorem 2.3.24 ([59℄). Let a bilinearsystem (2.30)with nonsingular
E
begiven,and letthepen ilA
− λE
bestable,i.e. thereexistreals alarsβ, α
∈ R
withβ >
0
and0 < α ≤ − max
i
(Re(λ
i
((A, E))))
su hthat||e
E
−1
At
||
2
≤ βe
−αt
, t
≥ 0.
(2.38) Assume||u(t)|| =
pP
m
k=1
|u
k
(t)|
2
≤ M
uniformelyon[0, ∞)
withM >
0
anddenoteΓ =
P
m
k=1
||E
−1
N
k
||
2
. ThenthesystemisBIBOstableifΓ <
α
Mβ
.The bilinearsystem ishen estableifthematri es
N
k
are su iently bounded.2.3.2.3. Rea hability,observabilityandbalan edrepresentation. Asfor
linearsystems,the on eptsofrea hability,observabilityandbalan ed
rep-resentationexistforbilinearsystems. However, the on epts need to be
generalized,whi hwillbedoneinthefollowingse tion.
Denition 2.3.25([25, 56℄). A state
x(t
e
)
of a bilinearsystem (2.30)is rea hable (from the originx
(0) = 0
) if there exists an admissibleinputfun tionthat mapstheoriginofthestate spa eintothestate
x
(t
e
)
ina niteintervaloftime[0, t
e
]
.Denition2.3.26([56℄). Abilinearsystem(2.30)is alled(span)rea hable ifthespa eofallrea hablestates
X
reach
spans
R
n
.
For a bilinear system (2.30) with
E
6= I
nonsingular, the following statementsforrea hability anbederived. LetP
1
(t
1
) = e
E
−1
At
1
E
−1
B,
P
i
(t
1
, . . . , t
i
) = e
E
−1
At
i
E
−1
[N
Denition2.3.27([72℄). Ifitexists,therea habilityGramianisdenedas
P
=
∞
X
i =1
Z
∞
0
· · ·
Z
∞
0
P
i
P
i
∗
d t
1
. . . d t
i
.
(2.39)ZhangandLam[72℄establishedthefollowingtheoremfortheexisten e oftherea habilityGramian:
Theorem2.3.28([72℄). Therea habilityGramian(2.39)exists,if (i) thepen il
A
− λE
isstable,with||e
E
−1
At
||
2
≤ βe
−αt
, t
≥ 0,
(2.40)where
β >
0
and0 < α ≤ − max
i
(Re(λ
i
(A, E))
,β, α
∈ R.
(ii)Γ
1
<
√
2α
β
,withΓ
2
1
= ||
P
m
k=1
E
−1
N
k
N
T
k
E
−T
||
2
.
The onne tionof
P
tothebilinearLyapuonvequationsandtherea h-abilityofthesystem annowbeestablished:
Theorem 2.3.29 ([72℄). Suppose
A
− λE
is stable, and the rea hability GramianP
exists. Then(i)
P
satisesthefollowingbilinearLyapunovequation:AXE
T
+ EXA
T
+
m
X
k=1
N
k
XN
k
T
+ BB
T
= 0.
(2.41)(ii) Thebilinearsystem(2.30)isrea hableifandonlyif
P
ispositive denite.Proposition2.3.30([72℄). If (2.41)hasauniquesolution,thenthesolution
P
issymmetri .Forlinearstablesystems,itisknownthatiftheLyapuonvequationhas
auniquesolutionitistherea hability( ontrollability)Gramian. Forbilinear
systems, however, it is possible that a unique solution to the Lyapunov
equationisnottherea habilityGramian.Considerforexamplethefollowing
bilinearsystem( f. [72℄):
˙x
=
−x + 2xu + u.
This leads to the solution of the Lyapunov equation
p
= −
1
2
. But the integralsp
˜
i
=
R
p
i
p
i
T
leadtop
˜
i
= 2
i −2
, whi hgivesp
=
P
∞
i =1
2
i −2
whi h doesnot onvergehen etherea habilityGramiandoesnotexist.Theorem2.3.31([72℄). Suppose
A
− λE
isstable.•
(2.41)hasapositive(semi)denitesolutionX
ifandonlyifthe rea hability Gramian (2.39) exists and onverges to a positivesemidenitematrix
X
ˆ
satisfying(2.41) .•
If (2.41)hasauniquepositive (semi)denite solutionX
, then (2.39) onvergestoX
andthereforeX
istherea habilityGramian. Forabilinearsystem(2.30)withE
nonsingularthefollowingstatements forobservability anbederived.LetQ
1
(t
1
) = Ce
E
−1
At
1
,
Q
i
(t
1
, . . . , t
i
) = [Q
i −1
E
−1
N
1
Q
i −1
E
−1
N
2
. . . Q
i −1
E
−1
N
m
]
T
e
E
−1
At
i
, i
= 2, 3, . . .
Denition2.3.32([72℄). Ifitexists,theobservabilityGramianisdenedasQ
=
∞
X
i =1
Z
∞
0
· · ·
Z
∞
0
Q
∗
i
Q
i
d t
1
. . . d t
i
.
(2.42)ZhangandLam[72℄establishedthefollowingtheoremfortheexisten e oftheobservabilitymatrix:
Theorem2.3.33([72℄). Theobservabilitymatrix(2.42)exists,if (i) thepen il
A
− λE
isstable,with||e
E
−1
At
||
2
≤ βe
−αt
, t
≥ 0,
(2.43)where
β >
0
and0 < α ≤ − max
i
(Re(λ
i
(A, E))
,β, α
∈ R.
(ii)Γ
1
<
√
2α
β
,withΓ
2
1
= ||
P
m
k=1
E
−1
N
k
N
T
k
E
−T
||
2
.
Theorem2.3.34. Suppose
A
− λE
isstable,andtheobservabilityGramianexists. Then
(i)
E
−T
QE
−1
satisesthefollowingbilinearLyapunovequation:
A
T
Y E
+ E
T
Y A
+
m
X
k=1
N
T
k
Y N
k
+ C
T
C
= 0.
(2.44)(ii) Thebilinearsystem(2.30)isobservableifandonlyif
Q
ispositive denite.Theorem2.3.35([72℄). Suppose
A
− λE
isstable.•
(2.44)hasapositive(semi)denitesolutionY
ifandonlyifthe observabilityGramian (2.42) exists and onvergesto a positivesemidenitematrix
Q
ˆ
satisfying(2.44)forE
−T
QE
ˆ
−1
.
•
If (2.44)hasa uniquepositive (semi) denitesolutionY
, then (2.42) onvergestoQ
= E
T
Y E
and
Q
istherea habilityGramian.A balan edrepresentationofabilinearsystem anbeobtainedin the
samewayasinthelinear ase. AssumethebilinearsystemisBIBOstable,
and the Gramians
P
andQ
existand are positive denite. They anbede omposedas
P
= RR
T
andQ
= L
T
L.
Byusingthesingularvaluede ompositonof
LER
= U
b
ΣV
b
T
,
oneobtainsW
b
T
ET
b
, W
b
T
AT
b
, W
b
T
N
k
T
b
, W
b
T
B, CT
b
,
whereW
b
= L
T
U
b
Σ
−1/2
,
T
b
= RV
b
Σ
−1/2
,
W
b
−1
= T
T
b
E
T
,
T
b
−1
= W
T
b
E.
Details anbefoundin[42,2℄andthereferen estherein. 2.3.2.4.
H
2
-normofabilinearsystem.Denition2.3.36. The
H
2
-normofabilinearsystemisdenedas||Σ
bil
||
2
H
2
= tr
∞
X
i =1
Z
∞
0
· · ·
Z
∞
0
m
X
k
1
,k
2
,...,k
i
=1
h
(k
1
,...,k
i
)
i
(s
1
, . . . , s
i
) ·
·(h
(k
1
,...,k
i
)
i
(s
1
, . . . , s
i
))
T
d s
1
. . . d s
i
,
(2.45)withVolterrakernels
h
(k
1
,...,k
i
)
i
(s
1
, . . . , s
i
)
denedin(2.33) .Zhang and Lam[72℄ showed,that thebilinear
H
2
-normsatises the samepropertyasthelinearnorm:Theorem2.3.37. Forabilinearsystem (2.30)if
A
− λE
isstableandthe rea hability GramianP
(or the observability GramianQ
) exists, then itsH
2
-norm anbe omputedfrom||Σ
bil
||
H
2
=
p
tr(CP C
T
)
(or=
p
tr(B
T
QB)
),
(2.46) whereP
(orE
−T
QE
−1
)satises (2.41)(or (2.44)).BennerandBreiten[12℄ showedthatthebilinear
H
2
-norm an equiv-alentlybewrittenas:Theorem2.3.38([12℄). Let
Σ
bil
beastablebilinearsystem. Thenitholds that||Σ
bil
||
2
H
2
= vec(I
p
)
T
(C ⊗ C)·
·
−A ⊗ E − E ⊗ A −
m
X
k=1
N
k
⊗ N
k
!
−1
(B ⊗ B)vec(I
m
).
(2.47)Modeling of heat transfer problems
3.1. ThermalModeling 26
3.2. Theheatequation 29
3.3. BoundaryandInterfa e onditions 30
3.4. Modeofoperationofanele tri almotor 31
3.5. Thermalmodelingofanele tri almotor 32
Thedesignofanewprodu tisa omplexpro esswithmanyexperts
involved. Fromtheideato thenal on ept,a lose ooperationbetween
designengineers,simulationexperts,testengineersandmanufa turing
spe- ialistsisrequired. Aftersettingup arstdesign,this designis examined
byateamofsimulationexperts. Dependingontherequirements,dierent
analysesneed tobe ondu ted. Severalphysi alaspe tsneed tobetaken
intoa ount,likeme hani aldeformations,uidows,ele tromagneti
ef-fe tsandthermalanalyses. Dependingontheevaluationofthesimulation
results,thedesignwillbeimproved. Aprototypeoftheoptimizedprodu t
isthen fabri atedand thoroughly testedin a seriesof experiments. Until
arrivingatthenalprodu t,allnewdesignswillbesimulatedhen e
sim-ulationplaysamajorrole. Inthenalstage oftheprodu tdevelopment,
simulationandexperimentshould oin ide. Themainpartisnowdesigning
themanufa turingpro ess,whi halsomightinvolve hangesinthedesign,