ebook img

Fixed-Endpoint Optimal Control of Bilinear Ensemble Systems PDF

0.32 MB·
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Fixed-Endpoint Optimal Control of Bilinear Ensemble Systems

FIXED-ENDPOINTOPTIMALCONTROLOFBILINEARENSEMBLESYSTEMS∗ SHUOWANG† AND JR-SHIN LI‡ Abstract. Optimal control ofbilinear systems has been awell-studied subject in the area ofmathematical control. However,techniquesforsolvingemergingoptimalcontrolproblemsinvolvinganensembleofstructurally identical bilinear systems are underdeveloped. In this work, we develop an iterative method to effectively and 6 systematicallysolvethesechallengingoptimalensemblecontrolproblems,inwhichthebilinearensemblesystem 1 is represented as a time-varying linear ensemble system at each iteration and the optimal ensemble control law 0 isthenobtainedbythesingularvalueexpansionoftheinput-to-state operatorthatdescribes thedynamicsofthe 2 linear ensemble system. Weexamine the convergence ofthe developed iterative procedure and pose optimality conditionsfortheconvergentsolution.Wealsoprovideexamplesofpracticalcontroldesignsinmagneticresonance n todemonstratetheapplicabilityandrobustnessofthedevelopediterativemethod. a J Keywords. Ensemblecontrol, iterative methods,sweepmethod,fixed-endpoint problems, bilinearsystems, 3 optimalityconditions,magneticresonance. 1 AMSsubjectclassifications. ] C 1. Introduction. Newly emerging fields in science and engineering, such as systems O neuroscience,synchronizationengineering,andquantumscienceandtechnology,giveriseto . h newclassesofoptimalcontrolproblemsthatinvolveunderactuatedmanipulationofindivid- t ualand collective behaviorof dynamicunits in a largeensemble. Representativeexamples a m include neural stimulation for alleviating the symptoms of neurological disorders such as Parkinson’sdisease,whereapopulationofneuronsinthebrainisaffectedbyasmallnumber [ ofelectrodes[1]; pulse designsforexcitingandtransportingquantumsystemsbetweende- 1 siredstates, whereanensembleofquantumsystemsisdrivenbyasingleormultiplepulses v inapulsesequence[2,3];andtheengineeringofdynamicalstructuresforcomplexoscillator 9 networks, where sequential patterns of a network of nonlinear rhythmic elements are cre- 2 3 atedandalteredbyamildglobalwaveform[4]. Solvingthesenontraditionalandlarge-scale 3 underactuatedcontrolproblemsrequiresthedevelopmentofsystematicandcomputationally 0 tractableandeffectivemethods. . 1 Amongtheseemergingcontrolproblems,inthispaper,wewillstudyfixed-endpointop- 0 timalcontrolproblemsinvolvingbilinearensemblesystems,whicharisefromthedomainof 6 quantumcontrol[5]andappearinavarietyofotherdifferentfields,suchascancerchemother- 1 apy[6] androbotics[7]. Thecontrolofbilinearsystemshasbeena well-studiedsubjectin : v theareaofmathematicalcontrol. FromPontryagin’smaximumprincipletospectralcolloca- i X tionmethods,awidevarietyoftheoreticalandcomputationalmethodshavebeendeveloped to solve optimal control problems of bilinear systems [8, 9]. In particular, the numerical r a methodsare in principlecategorizedinto direct, e.g., pseudospectralmethods[10, 11], and indirectapproaches,e.g.,indirecttranscriptionmethod[12]andshootingmethods[13]. Im- plementingtheseexistingnumericalmethodstosolveoptimalcontrolproblemsinvolvingan ensemble,i.e.,alargenumber(finitelyorinfinitelymany)oraparameterizedfamily,ofbilin- earsystemsmayencounterlowefficiency,slowconvergence,andinstabilityissues,because most of these methodsrely on suitable discretization of the continuous-timedynamicsinto alarge-scalenonlinearprogram(LSNLP).Inaddition,theglobalconstraintforsuchanop- ∗ThisworkwassupportedinpartbytheNationalScienceFoundationundertheawardsCMMI-1301148,CMMI- 1462796,andECCS-1509342. †DepartmentofElectricalandSystemsEngineering,WashingtonUniversity,St. Louis,Missouri,63130,USA ([email protected]). ‡DepartmentofElectricalandSystemsEngineering,WashingtonUniversity,St. Louis,Missouri,63130,USA ([email protected]).Questions,comments,orcorrectionstothisdocumentmaybedirectedtothisemailaddress. 1 2 S.WANGANDJ.-S.LI timalensemblecontrolproblem,inwhicheachindividualsystemreceivesthe samecontrol input,makesthediscretizedLSNLPveryrestrictiveandintractabletosolveoreventofinda feasiblesolution[14]. On the other hand, optimalcontrol problemsinvolving a linear system, or a linear en- semblesystem,areoftencomputationallytractableandanalyticallysolvableformanyspecial cases,suchasthelinearquadraticregulator(LQR)[15]andtheminimum-energycontrolof harmonicoscillatorensembles[16].Thissuggestsabypasstosolveoptimalcontrolproblems ofbilinearensemblesystemsthroughsolvingthatoflinearensemblesystemsandmotivates thedevelopmentoftheiterativemethodinthiswork.Thecentralideaistorepresentthebilin- earensemblesystemasalinearensemblesystemateachiteration,andthenfeasiblycalculate theoptimalcontrolandtrajectoryforeachiterationuntilaconvergentsolutionisfound.Itera- tivemethodshavebeenintroducedandadoptedtodealwithdiversecontroldesignproblems, including the free-endpointquadratic optimal control of bilinear systems [17] and optimal state trackingfornonlinearsystems[18], while the fixed-endpointproblemsalongwith the emergingproblemsthatinvolvecontrollingabilinearensemblesystemremainunexplored. Inthispaper,wecombinetheideaoftheaforementionediterativemethodwithourpre- viousworkonoptimalcontroloflinearensemblesystemstoconstructaniterativealgorithm forsolvingoptimalcontrolproblemsinvolvingatime-invariantbilinearensemblesystemof theform, d X(t,b )=A(b )X(t,b )+B(b )u(t)+ (cid:229)m u(t)B(b ) X(t,b ), i i dt i=1 (cid:16) (cid:17) whereX =(x ,...,x )T ∈M⊂Rn denotesthe state, b ∈K ⊂Rd with K compactand d a 1 n positiveinteger,u(t)=(u (t),...,u (t))T ∈Rmisthecontrol,andthematricesA(b )∈Rn×n, 1 m B(b )∈Rn×m,andB(b )∈Rn×n,i=1,...,m,forb ∈K. i This paper is structuredas follows. In the nextsection, we presentthe developediter- ative method for fixed-endpointoptimal control of a time-invariant bilinear system, where weintroduceasweepmethodthataccountsfortheterminalconditionbasedonthenotionof flowmappingfromtheoptimalcontroltheory. InSection3,weexaminetheconvergenceof theiterativemethodusingthefixed-pointtheorem. InSection4, weproposetheconditions forglobaloptimalityoftheconvergentsolution.Then,inSection5,weextendthedeveloped iterativemethodtosolveoptimalcontrolproblemsinvolvingbilinearensemblesystemsand showtheconvergenceofthemethod. Finally,examplesandsimulationsofpracticalcontrol designproblemsareillustrated in Section6 to demonstratethe applicabilityandrobustness ofthedevelopediterativeprocedure. 2. Iterativemethodforoptimalcontrolofbilinearsystems. Westartwithconsider- ingafixed-endpoint,finite-time,quadraticoptimalcontrolprobleminvolvingatime-invariant bilinearsystemoftheform min J= 1 tf xT(t)Qx(t)+uT(t)Ru(t) dt, 2 0 Z h m i (cid:229) (P1) s.t. x˙=Ax+Bu+ uB x, i i i=1 h i x(0)=x , x(t )=x , 0 f f wherex(t)∈Rnisthestateandu(t)∈Rmisthecontrol;A∈Rn×n,B ∈Rn×n,andB∈Rn×m i areconstantmatrices;R∈Rm×m≻0ispositivedefiniteandQ∈Rn×n(cid:23)0ispositivesemi- definite; and x ,x ∈Rn are the initial andthe desired terminalstate, respectively. We first 0 f FIXED-ENDPOINTCONTROLFORBILINEARENSEMBLES 3 representthetime-invariantbilinearsystemin(P1)asatime-varyinglinearsystem, n (cid:229) (2.1) x˙(t)=Ax+Bu+ x (t)N u, j j "j=1 # inwhichwewritethebilinearterm (cid:229) m uB x= (cid:229) n x N uwithx the jthelementofx, i=1 i i j=1 j j j N ∈Rn×m for j=i,...,n,andu=(u ,...,u )T ∈Rm. Then,wesolvethisoptimalcontrol j (cid:0) 1 m(cid:1) (cid:0) (cid:1) problembyPontryagin’smaximumprinciple.TheHamiltonianofthisproblemis (2.2) H(x,u,l )= 1(xTQx+uTRu)+l T Ax+ B+((cid:229) n x N ) u , j j 2 j=1 n h i o wherel (t)∈Rnistheco-statevector.Theoptimalcontrolisthenobtainedbythenecessary condition, ¶ H =0,givenby ¶ u (2.3) u∗=−R−1 B+(cid:229) n x N Tl , j j j=1 (cid:16) (cid:17) andtheoptimaltrajectoryofthestatexandtheco-statel satisfy,fort∈[0,t ], f n n (2.4) x˙ = Ax − B+(cid:229) x N R−1 B+(cid:229) x N Tl , i i j j j j i j=1 j=1 (cid:2) (cid:3) h(cid:0) (cid:1) (cid:0) (cid:1) i n n (2.5) l˙ =− Qx − ATl +l T NR−1 B+(cid:229) x N T+ B+(cid:229) x N R−1NT l i i i i j j j j i j=1 j=1 (cid:2) (cid:3) (cid:2) (cid:3) n (cid:0) (cid:1) (cid:0) (cid:1) o withtheboundaryconditionsx(0)=x andx(t )=x ,wherex,l and[·],i=1,...,n,are 0 f f i i i theith componentoftheassociatedvectors.Bythefollowingchangeofvariables, n n (2.6) A˜ =A − (N R−1 B+(cid:229) x N T+ B+(cid:229) x N )R−1NT l , ij ij j j j j j j i j=1 j=1 h (cid:0) (cid:1) (cid:0) (cid:1) i n n (2.7) B˜R−1B˜T =BR−1BT− (cid:229) x N R−1 (cid:229) x N T, j j j j j=1 j=1 (cid:0) (cid:1) (cid:0) (cid:1) (2.8) Q˜ =Q, wecanrewrite(2.4)and(2.5)intotheform (2.9) x˙=A˜x−B˜R−1B˜Tl , x(0)=x , x(t )=x , 0 f f (2.10) l˙ =−Q˜x−A˜Tl , whichcoincideswiththecanonicalformofthestateandco-stateequationscharacterizingthe optimal trajectories for the analogousoptimalcontrol problem involvingthe time-invariant linearsystemx˙=A˜x+B˜u[19]. Inthisway,theoptimalstateandco-statetrajectoriesforthe optimalcontrolproblem(P1)involvingatime-invariantbilinearsystemarenowexpressedin termsoftheequationsrelatedtoatime-varyinglinearsystemasin(2.9)and(2.10). Usingthis“linear-systemrepresentation”togetherwiththeSweepmethod[19,20],we will solve the optimal control problem (P1) in an iterative manner. Specifically, we will considerateachiterationthefixed-endpointlinearquadraticoptimalcontrolproblem, min J= 1 tf (x(k+1))T(t)Qx(k+1)(t)+(u(k+1))T(t)Ru(k+1)(t) dt, 2 0 Z h i (P2) s.t. x˙(k+1)(t)=A˜(k)x(k+1)+B˜(k)u(k+1) x(k+1)(0)=x , x(k+1)(t )=x , 0 f f 4 S.WANGANDJ.-S.LI by treating the previoustrajectory x(k) as a known quantity, where k∈N denotesthe itera- tion. Inthefollowingsections,wewillintroducetheSweepmethodandpresenttheiterative procedure. 2.1. Sweep method for fixed-endpoint problems. Observe that in (2.9) and (2.10) there are two boundaryconditionsfor the state x while none for the co-state l . It requires implementingspecializedcomputationalmethods,suchasshootingmethods,tosolvesucha two-pointboundaryvalue problem, which in generalinvolveintensive numericaloptimiza- tions. Here,weadopttheideaoftheSweepmethodbyletting (2.11) l (t)=K(t)x(t)+S(t)n , withl (t )=n ,whereK(t), S(t)∈Rn×n andn isthemultiplier,aconstantassociatedwith f theterminalconstrainty ,whichinthiscaseisy (x(t ))=x(t )=x . Fromthetransversality f f f conditionin Pontryagin’smaximumprinciple, we know that K(t )=0 because there is no f ¶y terminalcostandS(t )= =I.Moreover,ifKischosentosatisfytheRiccatiequation f ¶ x x(tf) (2.12) K˙(t)=−(cid:12)(cid:12) Q−A˜TK(t)−K(t)A˜+K(t)B˜R−1B˜TK(t), withtheterminalconditionK(t )=0,thenSsatisfiesthematrixdifferentialequation f (2.13) S˙(t)=−(A˜T−K(t)B˜R−1B˜T)S(t), withtheterminalconditionS(t )=I,bytakingthetimederivativeof(2.11)andusing(2.9), f (2.10)and(2.12). Inaddition,inordertofulfilltheterminalconditiony (x(t ))=x attime f f t ,themultipliern associatedwithy mustsatisfy f (2.14) x =ST(t)x(t)+P(t)n f forallt∈[0,t ],whereP(t)∈Rn×nobeysthematrixdifferentialequation f (2.15) P˙(t)−ST(t)B˜R−1B˜TS(t)=0, withtheterminalconditionP(t )=0. Itfollowsfrom(2.14)usingt=0that f (2.16) n = P(0) −1 x −ST(0)x , f 0 providedP(t)isinvertiblefort∈[0(cid:2),tf]. M(cid:3)or(cid:2)edetailsabou(cid:3)ttheSweepmethodbasedonthe notionofflowmappingareprovidedinAppendix8.1.1. 2.2. Iterationprocedure. Theoptimalsolutionoftheproblem(P1)ischaracterizedby thehomogeneoustime-varyinglinearsystemdescribedin(2.9)and(2.10),andwewillsolve forxandl viaaniterativeprocedure,whichisbasedonanalyticalexpressionsandrequiresno numericaloptimizations.Toproceedthis,wewrite(2.9)and(2.10)astheiterationequations, (2.17) x˙(k+1)=A˜(k)x(k+1)−B˜(k)R−1(B˜(k))Tl (k+1), (2.18) l˙(k+1)=−Q˜(k)x(k+1)−(A˜(k))Tl (k+1), with identicalboundaryconditionsx(k+1)(0)=x andx(k+1)(t )=x forallk=0,1,2,..., 0 f f whereA˜(k),B˜(k)R−1(B˜(k))T,andQ˜(k) aredefinedaccordingto(2.6),(2.7),and(2.8),by n n (2.19) A˜(k)=A − (N R−1 B+(cid:229) x(k)N T+ B+(cid:229) x(k)N )R−1NT l (k) , ij ij j j j j j j i j=1 j=1 h (cid:0) (cid:1) (cid:0) (cid:1) i n n (2.20) B˜(k)R−1(B˜(k))T =BR−1BT− (cid:229) x(k)N R−1 (cid:229) x(k)N T, j j j j j=1 j=1 (cid:0) (cid:1) (cid:0) (cid:1) (2.21) Q˜(k)=Q. FIXED-ENDPOINTCONTROLFORBILINEARENSEMBLES 5 ApplyingtheSweepmethodintroducedinSection2.1,weletl (k+1)(t)=K(k+1)(t)x(k+1)(t)+ S(k+1)(t)n (k+1)fort∈[0,t ],whereK(k) satisfiestheRiccatiequation f (2.22) K˙(k+1)=−Q(k)−K(k+1)A˜(k)−(A˜(k))TK(k+1)+K(k+1)B˜(k)R−1(B˜(k))TK(k+1), withtheboundaryconditionK(k+1)(t )=0,andS(k)follows f (2.23) S˙(k+1)=− (A˜(k))T−K(k+1)B˜(k)R−1(B˜(k))T S(k+1), S(k+1)(t )=I. f h i Moreover,themultipliern (k) satisfies (2.24) n (k+1)= P(k+1)(0) −1 x −(S(k+1))T(0)x , f 0 where P(·)(t)∈Rn×n is invertibl(cid:2)e (see Lem(cid:3)ma(cid:2) 3.1 in Section 3) a(cid:3)nd satisfies the dynamic equation (2.25) P˙(k+1)=(S(k+1))TB˜(k)R−1(B˜(k))TS(k+1) with the terminalconditionP(k+1)(t )=0. Then, the optimalcontrol(2.3) for the original f Problem(P1)canbeexpressedas (2.26) u∗(t)=−R−1 B+(cid:229) n x∗(t)N T[K∗(t)x∗(t)+S∗(t)n ∗], j j j=1 h i ifthisiterativeprocedureisconvergent,wherex(k)→x∗,K(k)→K∗,andS(k)n (k)→S∗n ∗. REMARK 1. The iterative method can be initialized by convenientlyusing the optimal controlofthesysteminvolvingonlythelinearpartofthebilinearsystemin(P1),i.e.,theLQR control. Thatis,thesolution(x(0)(t),l (0)(t))tothehomogeneoussystem x˙(0)=Ax(0)−BR−1BTl (0), x(0)(0)=x , x(0)(t )=x , 0 f f l˙(0)=−ATl (0). However, the linearsystem x˙=Ax+Bu may be uncontrollableso that the desired transfer between x and x is impossible and the LQR solution does not exist. In such a case, any 0 f statetrajectorywiththeendpointsx andx canbeafeasibleinitialtrajectoryx(0)(t)ofthe 0 f iterativeprocedure. 2.3. Aspecialcase: minimum-energycontrolofbilinearsystems. Beforeanalyzing the convergence of the iterative method, we illustrate the procedure using the example of minimum-energycontrolof bilinear systems, which is a special case of Problem (P1) with Q=0. Considerthefollowingfixed-endpointoptimalcontrolproblem, min J= 1 tf uT(t)Ru(t)dt, 2 0 Z n (cid:229) (P3) s.t. x˙(t)=Ax+Bu+ x (t)N u, j j "j=1 # x(0)=x , x(t )=x . 0 f f TheHamiltonianofthisproblemisH(x,u,l )= 1uTRu+l T[Ax+Bu+((cid:229) n x N )u],where 2 j=1 j j l (t)∈Rnistheco-statevector.Theoptimalcontrolisoftheformasin(2.3)andtheoptimal 6 S.WANGANDJ.-S.LI stateandco-statetrajectoriessatisfy(2.9)and(2.10),respectively,withQ=0.Therespective iterationequationsfollow(2.17)and(2.18)withQ(k)=0forallk=0,1,2,.... FollowingtheiterativemethodpresentedinSection2.2,werepresentthecostatel (k+1)(t)= K(k+1)(t)x(k+1)(t)+S(k+1)(t)n (k+1),t ∈[0,t ],andthematrixK(k+1)(t)∈Rn×n satisfiesthe f Riccatiequation, (2.27) K˙(k+1)=−K(k+1)A˜(k)−(A˜(k))TK(k+1)+K(k+1)B˜(k)R−1(B˜(k))TK(k+1), with the terminal condition K(k+1)(t )= 0, which has the trivial solution, K(k+1)(t)≡ 0, f ∀k=0,1,2,...,andfort∈[0,t ]. Thisgives f (2.28) l (k+1)(t)=S(k+1)(t)n (k+1), andS(k+1)satisfies (2.29) S˙(k+1)=−(A˜(k))TS(k+1), S(k+1)(t )=I. f In addition, the multiplier associated with the terminal constraintis expressed as in (2.24). Combining(2.28)with(2.3)givestheminimum-energycontrolatthe(k+1)th iteration, (2.30) (u∗)(k+1)(t)=−R−1 B+(cid:229) n x(k+1)N TS(k+1)n (k+1). j j j=1 h i NotethattheauxiliaryvariableP(k)(t)∈Rn×nateachiterationsatisfies(2.25),andthus P(k+1)(0)=−F (t ,0)W(k+1)F T (t ,0), A˜(k) f A˜(k) f where F T (t ,t)=F (t,t ) is the transition matrix for the homogeneous equation A˜(k) f −(A˜(k))T f (2.29)and W(k+1)= tf F (0,s )B˜(k)R−1(B˜(k))TF T (0,s )ds A˜(k) A˜(k) 0 Z isthecontrollabilityGramianforthetime-varyinglinearsystemasinProblem(P3),or,equiv- alently,asin(2.9)and(2.10)withQ=0. Moreover,theclosed-loopexpressionin(2.30)is consistentwiththeopen-loopexpressionoftheminimum-energycontrolintermsofthecon- trollabilityGramian,thatis, n (2.31) (u∗)(k+1)=R−1 B+(cid:229) x(k+1)N TF T (0,t) W(k+1) −1x (k), j j A˜(k) j=1 (cid:2) (cid:3) (cid:0) (cid:1) wherex (k)=F (0,t )x −x . A˜(k) f f 0 3. ConvergenceoftheIterativeMethod. Followingtheiterativealgorithmdescribed inSection2.2,weexpecttofindtheoptimalcontrolforProblem(P1),providedtheiterations areconvergent.Inthissection,weshowthattheconvergenceofthisalgorithmispertinentto thecontrollabilityofthelinearsystemconsideredateachiterationanddependsonthechoice of the weightmatrix R. In Section 5, we will extendthis iterative methodto solve optimal controlproblemsinvolvingbilinearensemblesystems. Tofacilitatetheproof,weintroducethefollowingmathematicaltools. Consideringthe . . . Banach spaces, X =C([0,t ];Rn), Y =C([0,t ];Rn×n), and Z =C([0,t ];Rn) with the f f f FIXED-ENDPOINTCONTROLFORBILINEARENSEMBLES 7 norms (3.1) kxka = sup kx(t)kexp(−a t) , for x∈X, t∈[0,tf] (cid:2) (cid:3) (3.2) kyka = sup ky(t)kexp(−a (tf−t)) , for y∈Y, t∈[0,tf] (cid:2) (cid:3) (3.3) kzka = sup kz(t)kexp(−a (tf−t)) , for z∈Z, t∈[0,tf] (cid:2) (cid:3) in which kvk=(cid:229) n |v| for v∈Rn and kDk=max (cid:229) n |D | for D∈Rn×n, and the i=1 i 1≤j≤n i=1 ij parameter a serves as an additional degree of freedom to control the rate of convergence [17], we define the operators T :X ×Y ×Z →X, T :X ×Y ×Z →Y, and T : 1 2 3 X ×Y ×Z →Z thatcharacterizethedynamicsofx∈X,K∈Y,andSn ∈Z asdescribed inSection2.2,givenby d T [x,K,Sn ](t)=A˜(x(t),K(t),S(t)n )T [x,K,Sn ](t)−B˜(x(t))R−1B˜T(x(t))T [x,K,Sn ](t) 1 1 3 dt (3.4) −B˜(x(t))R−1B˜T(x(t))T [x,K,Sn ](t)T [x,K,Sn ](t), 2 1 T [x,K,Sn ](0)=x 1 0 d T [x,K,Sn ](t)=−Q+T [x,K,Sn ](t)B˜(x(t))R−1B˜T(x(t))T [x,K,Sn ](t) 2 2 2 dt (3.5) −T [x,K,Sn ](t)A˜(x(t),K(t),S(t)n )−A˜T(x(t),K(t),S(t)n )T [x,K,Sn ](t) 2 2 T [x,K,Sn ](t )=0, 2 f d T [x,K,Sn ](t)=− A˜T(x(t),K(t),S(t)n )−T [x,K,Sn ](t)B˜(x(t))R−1B˜T(x(t)) · 3 2 dt (3.6) T [hx,K,Sn ](t), i 3 T [x,K,Sn ](t )=n (T [x,K,Sn ],T [x,K,Sn ],T [x,K,Sn ]) 3 f 1 2 3 wheren (T [x,K,Sn ],T [x,K,Sn ],T [x,K,Sn ])isthemultipliersatisfying(2.24). Withthese 1 2 3 definitionsand the followinglemma, the convergenceof the iterativemethodcan be devel- opedusingthefixed-pointtheorem. LEMMA 3.1. The matrix P(k+1)(t) as in (2.25) is nonsingular over t ∈[0,tf] at each iterationk ifandonlyifthetime-varyinglinearsysteminProblem(P2)iscontrollableover [0,t ][19]. f Proof: SeeAppendix8.1.3. (cid:3) THEOREM3.2. Considertheiterativemethodwiththeiterationsevolvingaccordingto (3.7) x(k+1)(t)=T [x(k),K(k),S(k)n (k)](t), 1 (3.8) K(k+1)(t)=T [x(k),K(k),S(k)n (k)](t), 2 (3.9) S(k+1)(t)n (k+1)=T [x(k),K(k),S(k)n (k)](t), 3 wheretheoperatorsT ,T ,andT aredefinedin(3.4),(3.5),and(3.6),respectively.Ifateach 1 2 3 iteration k the linear system as in (P2) is controllable, then T , T , and T are contractive. 1 2 3 Furthermore,startingwithatriple offeasibletrajectories(x(0),K(0),S(0)n (0)), theiteration procedure is convergent, and the sequences x(k), K(k) and S(k)n (k) converge to the unique fixedpoints,x∗,K∗,and(Sn )∗,respectively. Proof:Becausethelinearsystemin(P2)iscontrollableateachiterationk,byLemma3.1 thematrixP(k+1) definedin(2.25)isinvertibleandhencethemultipliern (k+1) expressedin 8 S.WANGANDJ.-S.LI (2.24)iswell-defined.Then,wehave,attimet ,S(k+1)(t )n (k+1)=T [x(k),K(k),S(k)n (k)](t )= f f 3 f n (k+1),sinceS(k+1)(t )=I. f From(2.19)and(2.20),foreachfixedt∈[0,t ],weobtainthebounds f kA˜(k+1)−A˜(k)k≤ (cid:229) n kGk2 1/2kl (k+1)−l (k)k i i=1 h i +k (cid:229) n kH k2 1/2 kl (k+1)kkx(k+1)−x(k)k+kx(k)kkl (k+1)−l (k)k , ij i,j=1 h i n o kB˜(k+1)R−1(B˜(k+1))T−B˜(k)R−1(B˜(k))Tk≤k (cid:229) n kH k2 1/2k(x(k+1))2−(x(k))2k, ij i,j=1 h i whereG =NR−1BT+BR−1NT andH =NR−1NT+N R−1NT,andfrom(2.21),wehave i i i ij i j j i Q˜(k+1) =Q˜(k) for all k=0,1,2,.... Substituting (2.18) into the aboveinequalities, we can writetheseboundsintermsofkx(k+1)−x(k)k,kK(k+1)−K(k)kandkS(k+1)n (k+1)−S(k)n (k)k, givenby kA˜(k+1)−A˜(k)k≤ (cid:229) n kGk2 1/2+kx(k)k (cid:229) n kH k2 1/2 · i ij i=1 i,j=1 nh i h i o (3.10) kK(k+1)kkx(k+1)−x(k)k+kK(k+1)−K(k)kkx(k)k+kS(k+1)n (k+1)−S(k)n (k)k (cid:8)+ (cid:229) n kH k2 1/2kK(k+1)kkx(k+1)k+kS(k+1)n (k+1)kkx(k+1)−x(k)k, (cid:9) ij i,j=1 h i kB˜(k+1)R−1(B˜(k+1))T−B˜(k)R−1(B˜(k))Tk≤ (cid:229) n kH k2 1/2· ij i,j=1 h i (3.11) kx(k+1)k+kx(k)k kx(k+1)−x(k)k. n o Inaddition,thesolutionto(2.23)isgivenby S(k+1)(t)=F (t,t )S(k+1)(t ) −[(A˜(k))T−K(k+1)B˜(k)R−1(B˜(k))T] f f (3.12) =F T (t ,t), [A˜(k)−B˜(k)R−1(B˜(k))TK(k+1)] f whereF denotesthetransitionmatrixassociatedwiththehomogeneoussystem(2.23)and (.) S(k+1)(t )=I. Then,wehave f T kS(k+1)(t)−S(k)(t)k≤ k (S(k+1))T(t) −1kkS(k+1))T(s )k kA˜(k)(s )−A˜(k−1)(s )k t Z (3.13) +kB˜(k)R(cid:2)−1(B˜(k))T(s (cid:3))−B˜(k−1)R−1(B˜(k−h1))T(s )kkK(k+1)(s )k +kB˜(k−1)R−1(B˜(k−1))T(s )kkK(k+1)(s )−K(k)(s )k kS(k)(s )kds . i FromtheRiccatiequationforK(k) describedin(2.22),wecanwritethedifferentialequation forthedifferenceK(k+1)−K(k)as d (K(k+1)−K(k))=−(K(k+1)−K(k)) A˜(k)−B˜(k)R−1(B˜(k))TK(k+1) dt h T i (3.14) − A˜(k)−B˜(k)R−1(B˜(k))TK(k+1) (K(k+1)−K(k)) −Kh(k)(A˜(k)−A˜(k−1))−(A˜(k)−Ai˜(k−1))TK(k+1) +K(k)(B˜(k)R−1(B˜(k))T−B˜(k−1)R−1(B˜(k−1))T)K(k+1), FIXED-ENDPOINTCONTROLFORBILINEARENSEMBLES 9 with the terminal condition K(k+1)(t )−K(k)(t )=0. Applying the variation of constants f f formula,backwardintimefromt=t ,to(3.14)andemploying(3.12)yield f K(k+1)(t)−K(k)(t)=(S(k))T(t) tf (S(k))T(s ) −1 K(k)(s ) A˜(k)(s )−A˜(k−1)(s ) t Z + A˜(k)(s )−A˜(k−1)n(s ) ThK(k+1)(s )−iK(hk)(s )· (cid:0) (cid:1) −1 B˜(cid:0)(k)R−1(B˜(k))T−B˜(k−1(cid:1))R−1(B˜(k−1))T K(k+1)(s ) S(k+1)(s ) ds S(k+1)(t), whichresult(cid:0)sin (cid:1) ih i o kK(k+1)(t)−K(k)(t)k≤ tf b kA˜(k)(s )−A˜(k−1)(s )k 1 t Z h (3.15) +b kB˜(k)R−1(B˜(k))T(s )−B˜(k−1)R−1(B˜(k−1))T(s )k ds , 2 i whereb andb arebothfinitetime-varyingcoefficients(seeAppendix8.2). 1 2 Similarly, from (2.17) and (2.18), we can write the differential equation for (x(k+1)− x(k)),thatis, d (x(k+1)−x(k))= A˜(k)−B˜(k)R−1(B˜(k))TK(k+1) (x(k+1)−x(k)) dt h i + (A˜(k)−A˜(k−1))− B˜(k)R−1(B˜(k))T−B˜(k−1)R−1(B˜(k−1))T K(k+1) n (cid:16) (cid:17) (3.16) −B˜(k−1)R−1(B˜(k−1))T(K(k+1)−K(k)) x(k) o − B˜(k)R−1(B˜(k))T−B˜(k−1)R−1(B˜(k−1))T S(k+1)n (k+1) −B(cid:16)˜(k−1)R−1(B˜(k−1))T S(k+1)n (k+1)−S(k)(cid:17)n (k) , with the terminal condition x(k+1)(t )−x(cid:0)(k)(t )=0. Applying (cid:1)the variation of constants f f formulato(3.16)yields, −1 t x(k+1)(t)−x(k)(t)= (S(k+1))T(t) (S(k+1))T(s ) A˜(k)(s )−A˜(k−1)(s ) 0 Z − hB˜(k)R−1(B˜(k)i)T(s )−B˜(k−1)R−1(nB˜h(k(cid:0)−1))T(s ) K(k+1)(s )(cid:1) −(cid:0)B˜(k−1)R−1(B˜(k−1))T(s ) K(k+1)(s )−K(k)(s )(cid:1)x(k)(s ) − B˜(k)R−1(B˜(k))T−B˜(k−(cid:0)1)R−1(B˜(k−1))T S(k+1(cid:1))in (k+1) (cid:16) (cid:17) −B˜(k−1)R−1(B˜(k−1))T(S(k+1)n (k+1)−S(k)n (k)) ds . o Itfollowsthat t kx(k+1)(t)−x(k)(t)k≤ b kA˜(k)(s )−A˜(k−1)(s )k+b kK(k+1)(s )−K(k)(s )k 3 4 0 Z h (3.17) +b kB˜(k)R−1(B˜(k))T(s )−B˜(k−1)R−1(B˜(k−1))T(s )k 5 +b kS(k+1)n (k+1)(s )−S(k)n (k)(s )k ds , 6 i whereb ,b , b andb areallfinitetime-varyingcoefficients(seeAppendix8.2). Further- 3 4 5 6 more,sincen (k+1) isaconstantwithineachiterationk,from(2.23)wecanwrite d (S(k+1)n (k+1))=− (A˜(k))T−K(k+1)B˜(k)R−1(B˜(k))T S(k+1)n (k+1), dt h i 10 S.WANGANDJ.-S.LI withtheterminalconditionS(k+1)(t )n (k+1)=n (k+1). Thisallowsustowrite f d T (S(k+1)n (k+1)−S(k)n (k))=− A˜(k)−B˜(k)R−1(B˜(k))TK(k+1) (S(k+1)n (k+1)−S(k)n (k)) dt h i − (A˜(k)−A˜(k−1))−B˜(k−1)R−1(B˜(k−1))T(K(k+1)−K(k)) n T − B˜(k)R−1(B˜(k))T−B˜(k−1)R−1(B˜(k−1))T K(k+1) S(k)n (k), (cid:16) (cid:17) o withtheterminalconditionS(k+1)(t )n (k+1)−S(k)(t )n (k)=n (k+1)−n (k),andthen f f −1 S(k+1)(t)n (k+1)−S(k)(t)n (k)= (S(k+1))T(t) (n (k+1)−n (k)) − tf(S(k+1))T(hs ) (A˜(k)−A˜i(k−1n))−B˜(k−1)R−1(B˜(k−1))T(K(k+1)−K(k)) t Z h T − B˜(k)R−1(B˜(k))T−B˜(k−1)R−1(B˜(k−1))T K(k+1) S(k)n (k)ds . (cid:16) (cid:17) i o From(2.24)wemayobtain kn (k+1)−n (k)k≤k(P(k+1))−1kkP(k+1)(t)−P(k)(t)kkn (k)k (3.18) +k(P(k+1))−1kkS′(k+1)(t)−S′(k)(t)kkx k, 0 inwhich,byevolving(2.25)backwardintimefromt=t ,thedifferenceP(k+1)(t)−P(k)(t) f satisfies P(k+1)(t)−P(k)(t)=− tf S(k+1)+S(k) B˜(k−1)R−1(B˜(k−1))T S(k+1)(s )−S(k)(s ) t Z (cid:20) (cid:0) (cid:1) (cid:0) (cid:1) (3.19) +S(k+1) B˜(k)R−1(B˜(k)(s ))T−B˜(k−1)R−1(B˜(k−1)(s ))T S(k) ds . (cid:16) (cid:17) (cid:21) Using(3.18)and(3.19),weobtain kS(k+1)(t)n (k+1)−S(k)(t)n (k)k≤ tf b kA˜(k)(s )−A˜(k−1)(s )k 7 t Z h (3.20) +b kB˜(k)R−1(B˜(k))T(s )−B˜(k−1)R−1(B˜(k−1))T(s )k 8 +b kK(k+1)(s )−K(k)(s )k ds , 9 i whereb ,b andb arefinitetime-varyingcoefficients(seeAppendix8.2). 7 8 9 Combining the bounds in (3.10), (3.11), (3.15), (3.17), and (3.20), and using the def- initions of the operators T , T , and T in (3.4), (3.5) and (3.6), respectively, we reach the 1 2 3 inequalitythatholdscomponent-wise,givenby kT1[x(k),K(k),S(k)n (k)]−T1[x(k−1),K(k−1),S(k−1)n (k−1)]ka kT2[x(k),K(k),S(k)n (k)]−T2[x(k−1),K(k−1),S(k−1)n (k−1)]ka   kT3[x(k),K(k),S(k)n (k)]−T3[x(k−1),K(k−1),S(k−1)n (k−1)]ka  kx(k)−x(k−1)ka  (3.21) ≤M kK(k)−K(k−1)ka ,   kS(k)n (k)−S(k−1)n (k−1)ka  

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.