Optimal a priori error estimates of parabolic optimal control problems with a moving point control DmitriyLeykekhmanandBorisVexler 7 1 0 2 r a M 6 AbstractInthispaperweconsideraparabolicoptimalcontrolproblemwithaDirac ] typecontrolwithmovingpointsourceintwo spacedimensions.We discretizethe A problemwithpiecewiseconstantfunctionsintimeandcontinuouspiecewiselinear N finite elements in space. For this discretization we show optimal order of conver- . gencewithrespecttothetimeandthespacediscretizationparametersmodulosome h t logarithmicterms.Erroranalysisforthesameproblemwascarriedoutintherecent a paper[17],however,theanalysistherecontainsaseriousflaw.Oneofthemaingoals m ofthispaperistoprovidethecorrectproof.Themainingredientsofouranalysisare [ theglobalandlocalerrorestimatesonacurve,thathaveanindependentinterest. 2 v 5 4 1 Introduction 0 3 0 Inthispaperweprovidenumericalanalysisforthefollowingoptimalcontrolprob- . lem: 1 1 T α T 0 minJ(q,u):= ku(t)−uˆ(t)k2 dt+ |q(t)|2dt (1) 7 q,u 2 0 L2(Ω) 2 0 Z Z 1 subjecttothesecondorderparabolicequation : v i u(t,x)−∆u(t,x)=q(t)δ , (t,x)∈I×Ω, (2a) X t γ(t) u(t,x)=0, (t,x)∈I×∂Ω, (2b) r a u(0,x)=0, x∈Ω (2c) DmitriyLeykekhman Department of Mathematics, University of Connecticut, Storrs, CT 06269, USA, e-mail: [email protected] BorisVexler Technical University of Munich, Chair of Optimal Control, Center for Mathematical Sciences, Boltzmannstraße3,85748GarchingbyMunich,Germany,e-mail:[email protected] 1 2 DmitriyLeykekhmanandBorisVexler andsubjecttopointwisecontrolconstraints q ≤q(t)≤q a.e.inI. (3) a b HereI=(0,T),Ω⊂R2 isaconvexpolygonaldomainandδ istheDiracdelta γ(t) functionatpointx =γ(t)ateacht.Wewillassume: t Assumption1 • γ∈C1(I¯)andmaxt∈I¯|γ′(t)|≤Cγ. Assumption2 • γ(t)⊂Ω ⊂⊂Ω ,foranyt∈I,withΩ ⊂⊂Ω. 0 1 1 The parameter α is assumed to be positive and the desired state uˆ fulfills uˆ ∈ L2(I;L∞(Ω)). The control bounds q ,q ∈R∪{±∞} fulfill q <q . The precise a b a b functional-analyticsettingisdiscussedinthenextsection. For the discretization,we consider the standard continuouspiecewise linear fi- nite elements in space and piecewise constant discontinuous Galerkin method in time. This is a special case (r=0, s=1) of so called dG(r)cG(s) discretization, seee.g.[14]fortheanalysisofthemethodforparabolicproblemsande.g.[25,26] forerrorestimatesinthecontextofoptimalcontrolproblems.Throughout,wewill denotebyhthespatialmeshsizeandbyk thesizeoftimesteps,seeSection3for details. Themainresultofthepaperisthefollowing. Theorem1.Letq¯beoptimalcontrolfortheproblem(1)-(2)andq¯ betheoptimal kh dG(0)cG(1) solution. Then there exists a constantC independentof h and k such that kq¯−q¯ k ≤C |lnh|3(k+h2)+C |lnh|k kq¯k +kuˆk . kh L2(I) γ L2(I) L2(I;L∞(Ω)) (cid:0) (cid:1)(cid:16) (cid:17) We would also like to point out that in addition to the optimal order estimate, modulologarithmicterms, our analysis does notrequire any relationshipbetween thesizesofthespacediscretizationhandthetimestepsk. The problem with fixed location of the point source (i.e. with δ (x) for some x0 fixed x ∈Ω) starting with the work of Lions [23], was investigated in a number 0 of publications,see [2, 3, 10, 12, 28] for the continuousproblemand[16, 21, 22] for the finite element approximation and error estimates. There is also a closely relatedproblemofmeasuredvaluedcontrols,whichreceivedalotofattentionlately [5,6,7,8,20]. TheproblemwithmovingDiracwasconsideredin[9,27]onacontinuouslevel. The erroranalysis was carried outin the recentpaper [17]. However,the analysis there contains a serious flaw. The last inequality in the estimate (3.33) in [17] is not correct. One of the main goals of this paper is to provide the correct proof. The main ingredientsof our analysis are the globaland local error estimates on a curve,Theorem2andTheorem3,respectively.Theseresultsarenewandhavean independentinterest. Throughout the paper we use the usual notation for Lebesgue and Sobolev spaces. We denoteby(·,·)Ω theinnerproductin L2(Ω) andby (·,·)I˜×Ω the inner productinL2(I˜×Ω)foranysubintervalI˜⊂I. Parabolicoptimalcontrolproblemswithamovingpointcontrol 3 Therestofthepaperisorganizedasfollows.InSection2wediscussthefunc- tionalanalyticsettingoftheproblem,statetheoptimalitysystemandproveregular- ityresultsforthestateandfortheadjointstate.InSection3weestablishimportant global and local best approximationresults along the curve for the heat equation. FinallyinSection4weproveourmainresult. 2 Optimal control problem andregularity Inordertostatethefunctionalanalyticsettingfortheoptimalcontrolproblem,we firstintroducetheauxiliaryproblem v(t,x)−∆v(t,x)= f(t,x), (t,x)∈I×Ω, t v(t,x)=0, (t,x)∈I×∂Ω, (4) v(0,x)=0, x∈Ω, witharight-handside f ∈L2(I;Lp(Ω))forsome1<p<∞.Thisequationpossesses auniquesolution v∈L2(I;H1(Ω))∩H1(I;H−1(Ω)). 0 DuetotheconvexityofthepolygonaldomainΩthesolutionvpossessesanaddi- tionalregularityfor p=2: v∈L2(I;H2(Ω)∩H1(Ω))∩H1(I;L2(Ω)), 0 withthecorrespondingestimate kvk +kvk ≤Ckfk , (5) L2(I;H2(Ω)) t L2(I;L2(Ω)) L2(I;L2(Ω)) see, e.g.,[15].From theSobolevembeddingH2(Ω)֒→W1,s(Ω) foranys<∞in twospacedimensionsandthepreviouslemmawecanestablishthefollowingresult fors>2, kvk ≤Cskvk ≤Cskfk . (6) L2(I;W1,s(Ω)) L2(I;H2(Ω)) L2(I;L2(Ω)) The exact form of the constant can be traced, for example, from the proof of [1, Thm.10.8].Inaddition,thereholdsthefollowingregularityresult(see[21]). Lemma1.If f ∈L2(I;Lp(Ω))foranarbitrary p>1,thenv∈L2(I;C(Ω))and kvk ≤C kfk , L2(I;C(Ω)) p L2(I;Lp(Ω)) whereC ∼ 1 ,as p→1. p p−1 Wewillalsoneedthefollowinglocalregularityresult(see[21]). Lemma2.LetΩ ⊂⊂Ω ⊂⊂Ω and f ∈L2(I;L2(Ω))∩L2(I;Lp(Ω )) for some 0 1 1 2≤p<∞.Thenv∈L2(I;W2,p(Ω ))∩H1(I;Lp(Ω ))andthereexistsaconstantC 0 0 4 DmitriyLeykekhmanandBorisVexler independentof psuchthat kvk +kvk ≤Cp(kfk +kfk ). t L2(I;Lp(Ω0)) L2(I;W2,p(Ω0)) L2(I;Lp(Ω1)) L2(I;L2(Ω)) To introduce a weak solution of the state equation (2) we use the method of transposition,(cf.[24]).Foragivencontrolq∈Q=L2(I)wedenotebyu=u(q)∈ L2(I;Lp(Ω)) with 2≤ p<∞ a weak solution of (2), if for allϕ∈L2(I;Lp′(Ω)) with 1+ 1 =1thereholds p p′ hu,ϕiL2(I;Lp(Ω)),L2(I;Lp′(Ω))= w(t,γ(t))q(t)dt, I Z where w∈L2(I;W2,p′(Ω)∩H1(Ω))∩H1(I;Lp′(Ω)) is the weak solution of the 0 adjointequation −w(t,x)−∆w(t,x)=ϕ(t,x), (t,x)∈I×Ω, t w(t,x)=0, (t,x)∈I×∂Ω, (7) w(T,x)=0, x∈Ω. Theexistenceofthisweaksolutionu=u(q)followsbydualityusingtheembedding L2(I;W2,p′(Ω))֒→L2(I;C(Ω))forp′>1.UsingLemma1wecanproveadditional regularityforthestatevariableu=u(q). Proposition2.1 Without lose of generality we assume 2 ≤ p <∞. Let q∈Q = L2(I) be given and u=u(q) be the solution of the state equation (2). Then u∈ L2(I;Lp(Ω)) for any p <∞ and the following estimate holds for p→∞ with a constantCindependentof p, kuk ≤Cpkqk . L2(I;Lp(Ω)) L2(I) Proof. Toestablishtheresultweuseadualityargument.Thereholds 1 1 kuk = sup (u,ϕ) , where + =1. L2(I;Lp(Ω)) I×Ω p p′ kϕkL2(I;Lp′(Ω))=1 Let w be the solution to (7) forϕ∈L2(I;Lp′(Ω)) with kϕkL2(I;Lp′(Ω)) =1. From Lemma1,w∈L2(I;C(Ω))andthefollowingestimateholds C C kwkL2(I;C(Ω))≤ p′−1kϕkL2(I;Lp′(Ω))= p′−1 ≤Cp, as p→∞. Thus, Parabolicoptimalcontrolproblemswithamovingpointcontrol 5 kuk = sup (u,ϕ) L2(I;Lp(Ω)) I×Ω kϕkL2(I;Lp′(Ω))=1 = q(t)w(t,γ(t))dt≤kqk kwk ≤Cpkqk . L2(I) L2(I;C(Ω)) L2(I) I Z Remark1.WewouldliketonotethattheaboveregularityrequiresonlyAssumption 2onγ.Higherregularityofγisneededforoptimalordererrorestimatesonly. Afurtherregularityresultforthestateequationfollowsfrom[13]. Proposition2.2 Let q∈Q=L2(I) be given and u=u(q) be the solution of the stateequation(2).Thenforeach1<s<2thereholds u∈L2(I;W1,s(Ω)) and u ∈L2(I;W−1,s(Ω)). 0 t Moreover,thestateufulfillsthefollowingweakformulation hu,ϕi+(∇u,∇ϕ)= q(t)ϕ(t,γ(t))dt forall ϕ∈L2(I;W1,s′(Ω)), t 0 I Z where 1 + 1 = 1 and h·,·i is the duality product between L2(I;W−1,s(Ω)) and s′ s L2(I;W1,s′(Ω)). 0 Proof. For s<2 we have s′ >2 and thereforeW1,s′(Ω) is embedded intoC(Ω¯). 0 Thereforetheright-handsideq(t)δ ofthestateequationcanbeidentifiedwithan γ(t) elementin L2(I;W−1,s(Ω)). Using the result from[13, Theorem5.1]on maximal parabolic regularity and exploiting the fact that −∆: W1,s(Ω)→W−1,s(Ω) is an 0 isomorphism,see[19],weobtain u∈L2(I;W1,s(Ω)) and u ∈L2(I;W−1,s(Ω)). 0 t Given the above regularity the corresponding weak formulation is fulfilled by a standarddensityargument. Asthenextstepweintroducethereducedcostfunctional j: Q→Ronthecon- trolspaceQ=L2(I)by j(q)=J(q,u(q)), whereJ isthecostfunctionin(1)andu(q)istheweaksolutionofthestateequa- tion (2) as defined above. The optimal control problem can then be equivalently reformulatedas min j(q), q∈Q , (8) ad wherethesetofadmissiblecontrolsisdefinedaccordingto(3)by Q ={q∈Q|q ≤q(t)≤q a.e.inI}. (9) ad a b By standardargumentsthis optimizationproblempossesses a uniquesolution q¯∈ Q=L2(I)withthecorrespondingstateu¯=u(q¯)∈L2(I;Lp(Ω))forall p<∞,see 6 DmitriyLeykekhmanandBorisVexler Proposition 2.1 for the regularity of u¯. Due to the fact, that this optimal control problem is convex, the solution q¯ is equivalently characterized by the optimality condition j′(q¯)(∂q−q¯)≥0 forall∂q∈Q . (10) ad The(directional)derivative j′(q)(∂q)forgivenq,∂q∈Qcanbeexpressedas j′(q)(∂q)= (αq(t)+z(t,γ(t)))∂q(t)dt, I Z wherez=z(q)isthesolutionoftheadjointequation −z(t,x)−∆z(t,x)=u(t,x)−uˆ(t,x), (t,x)∈I×Ω, (11a) t z(t,x)=0, (t,x)∈I×∂Ω, (11b) z(T,x)=0, x∈Ω, (11c) andu=u(q)ontheright-handsideof(11a)isthesolutionofthestateequation(2). Theadjointsolution,whichcorrespondstotheoptimalcontrolq¯isdenotedbyz¯= z(q¯). The optimality condition (10) is a variational inequality, which can be equiva- lentlyformulatedusingtheprojection P : Q→Q , P (q)(t)=min q ,max(q ,q(t)) . Qad ad Qad b a Theresultingconditionreads: (cid:0) (cid:1) 1 q¯(t)=P − z¯(t,γ(t)) . (12) Qad α (cid:18) (cid:19) Inthenextpropositionweprovideregularityresultsforthesolutionoftheadjoint equation. Proposition2.3 Letq∈Qbegiven,letu=u(q)bethecorrespondingstatefulfill- ing(2)andletz=z(q)bethecorrespondingadjointstatefulfilling(11).Then, (a)z∈L2(I;H2(Ω)∩H1(Ω))∩H1(I;L2(Ω))andthefollowingestimateholds 0 k∇2zk +kzk ≤C(kqk +kuˆk ). L2(I;L2(Ω)) t L2(I;L2(Ω)) L2(I) L2(I;L2(Ω)) (b) IfΩ ⊂⊂Ω, thenz∈L2(I;W2,p(Ω ))∩H1(I;Lp(Ω )) for all2≤ p<∞and 0 0 0 thefollowingestimateholds k∇2zk +kzk ≤Cp2(kqk +kuˆk ). L2(I;Lp(Ω0)) t L2(I;Lp(Ω0)) L2(I) L2(I;L∞(Ω)) Proof. (a) Theright-handsideoftheadjointequationfulfillsu−uˆ∈L2(I;Lp(Ω)) forall1<p<∞,seeProposition2.1.DuetotheconvexityofthedomainΩwe directlyobtainz∈L2(I;H2(Ω)∩H1(Ω))∩H1(I;L2(Ω))andtheestimate 0 Parabolicoptimalcontrolproblemswithamovingpointcontrol 7 k∇2zk +kzk ≤Cku−uˆk . L2(I;L2(Ω)) t L2(I;L2(Ω)) L2(I;L2(Ω)) TheresultfromProposition2.1leadsdirectlytothefirstestimate. (b)FromLemma2for p≥2wehave k∇2zk +kzk ≤Cpku−uˆk . L2(I;Lp(Ω0)) t L2(I;Lp(Ω0)) L2(I;Lp(Ω)) Hence,bythetriangleinequalityandProposition2.1weobtain ku−uˆk ≤C pkqk +kuˆk . L2(I;Lp(Ω)) L2(I) L2(I;L∞(Ω)) (cid:16) (cid:17) Thatcompletestheproof. 3 Discretization andthe bestapproximationtype results 3.1 Space-timediscretizationand notation Fordiscretizationoftheproblemundertheconsiderationweintroduceapartitions of I =[0,T] into subintervalsI =(t ,t ] of length k =t −t , where 0= m m−1 m m m m−1 t <t <···<t <t =T.Weassumethat 0 1 M−1 M k ≤κk , m=1,...,M−1, forsome κ>0. (13) m+1 m The maximaltime step is denotedby k=max k . The semidiscrete space X0 of m m k piecewiseconstantfunctionsintimeisdefinedby X0={v ∈L2(I;H1(Ω)): v | ∈P (I ;H1(Ω)), m=1,2,...,M}, k k 0 k Im 0 m 0 where P (I;V) is the space of constant functions in time with values in Banach 0 spaceV.WewillemploythefollowingnotationforfunctionsinX0 k v+= lim v(t +ε):=v , v−= lim v(t −ε)=v(t ):=v , [v] =v+−v−. m m m+1 m m m m m m m ε→0+ ε→0+ (14) LetT denoteaquasi-uniformtriangulationofΩwithameshsizeh,i.e.,T = {τ}isapartitionofΩintotrianglesτofdiameterh suchthatforh=max h , τ τ τ diam(τ)≤h≤C|τ|12, ∀τ∈T hold.LetV bethesetofallfunctionsinH1(Ω)thatarelinearoneachτ,i.e.V is h 0 h theusualspaceofcontinuouspiecewiselinearfinite elements.We willrequirethe modifiedCle´mentinterpolanti : L1(Ω)→V andtheL2-projectionP : L2(Ω)→ h h h V definedby h (P v,χ) =(v,χ) , ∀χ∈V . (15) h Ω Ω h 8 DmitriyLeykekhmanandBorisVexler Toobtainthefullydiscreteapproximationweconsiderthespace-timefiniteelement space X0,1={v ∈X0: v | ∈P (I ;V ), m=1,2,...,M}. (16) k,h kh k kh Im 0 m h Wewillalsoneedthefollowingsemidiscreteprojectionπ :C(I¯;H1(Ω))→X0de- k 0 k finedby πv| =v(t ), m=1,2,...,M, (17) k Im m andthefullydiscreteprojectionπ :C(I¯;L1(Ω))→X0,1definedbyπ =i π. kh k,h kh h k TointroducethedG(0)cG(1)discretizationwedefinethefollowingbilinearform M M B(v,ϕ)= ∑hv,ϕi +(∇v,∇ϕ) + ∑([v] ,ϕ+ ) +(v+,ϕ+) , (18) t Im×Ω I×Ω m−1 m−1 Ω 0 0 Ω m=1 m=2 whereh·,·i isthedualityproductbetweenL2(I ;W−1,s(Ω))andL2(I ;W1,s′(Ω)). Im×Ω m m 0 Wenote,thatthefirstsumvanishesforv∈X0.Rearrangingtheterms,weobtainan k equivalent(dual)expressionforB: M M−1 B(v,ϕ)=− ∑hv,ϕi +(∇v,∇ϕ) − ∑(v−,[ϕ] ) +(v−,ϕ−) . (19) t Im×Ω I×Ω m k m Ω M M Ω m=1 m=1 Inthetwofollowingtheoremsweestablishglobalandlocalbestapproximation type results along the curve for the error between the solution v of the auxiliary equation(4)anditsdG(0)cG(1)approximationv ∈X0,1definedas kh k,h B(v ,ϕ )=(f,ϕ ) forall ϕ ∈X0,1. (20) kh kh kh I×Ω kh k,h Since dG(0)cG(1) method is a consistent discretization we have the following Galerkinorthogonalityrelation: B(v−v ,ϕ )=0 forall ϕ ∈X0,1. kh kh kh k,h 3.2 Discretizationofthecurve andthe weight function To definefully discrete optimizationproblemwe will also requirea discretization ofthecurveγ.Wedefineγ =πγby k k γ| =γ(t ):=γ ∈Ω , m=1,2,...,M, (21) k Im m k,m 0 i.e., γ is a piecewise constant approximation of γ. Next we introduce a weight k function σ(t,x)= |x−γ(t)|2+h2 (22) q andadiscretepiecewiseconstantintimeapproximation Parabolicoptimalcontrolproblemswithamovingpointcontrol 9 σ(t,x)= |x−γ(t)|2+h2. (23) k k q Define σ :=σ| =σ(t ,x)=σ(t ,x). (24) k,m k Im k m m Onecaneasilycheckthatσandσ satisfythefollowingpropertiesforany(t,x)∈ k I×Ω, kσ−1(t,·)kL2(Ω),kσk−1(t,·)kL2(Ω)≤C|lnh|21, t∈I¯, (25a) |∇σ(t,x)|, |∇σ(t,x)|≤C, (25b) k |∇2σ(t,x)|≤C|σ−1(t,x)|, (25c) k k |σ(t,x)|≤|∇σ(t,x)|·|γ′(t)|≤CC , (25d) t γ maxσ(x,t)≤Cminσ(x,t), ∀τ∈T. (25e) x∈τ x∈τ 3.3 Globalerrorestimatealong thecurve Inthissectionweprovethefollowingglobalapproximationresult. Theorem2 (Global best approximation). Assume v and v satisfy (4) and (20) kh respectively.ThenthereexistsaconstantCindependentofkandhsuchthatforany 1≤p≤∞, |(v−v )(t,γ(t))|2dt≤C|lnh|2× kh k I Z χ∈inXf0,1 kv−χk2L2(I;L∞(Ω))+h−4pkπkv−χk2L2(I;Lp(Ω)) . k,h(cid:16) (cid:17) Proof. To establish the result we use a duality argument. First, we introduce a smoothed Delta function, which we will denote by δ˜ . This function on each I γk m isdefinedasδ˜ andsupportedinonecell,whichwedenotebyτ0,i.e. γk,m m (χ,δ˜ ) =χ(γ )=χ(γ(t )), ∀χ∈P1(τ0), m=1,2,...,M. γk,m τm0 k,m m m Inadditionwealsohave(see[31, Appendix]) kδ˜γkkWps(Ω)≤Ch−s−2(1−1p), 1≤p≤∞, s=0,1. (26) Thusinparticularkδ˜γkkL1(Ω)≤C,kδ˜γkkL2(Ω)≤Ch−1,andkδ˜γkkL∞(Ω)≤Ch−2. Wedefinegtobeasolutiontothefollowingbackwardparabolicproblem 10 DmitriyLeykekhmanandBorisVexler −g(t,x)−∆g(t,x)=v (t,γ(t))δ˜ (x), (t,x)∈I×Ω, t kh k γk g(t,x)=0, (t,x)∈I×∂Ω, (27) g(T,x)=0, x∈Ω. Thereholds M v (t,γ(t))δ˜ (x)ϕ (t,x)dtdx= ∑ v (t,γ(t)) δ˜ (x)ϕ (t,x)dx dt kh k γk kh kh k γk kh ZI×Ω m=1ZIm (cid:18)ZΩ (cid:19) M = ∑ v (t,γ(t))ϕ (t,γ(t))dt kh k kh k m=1ZIm = v (t,γ(t))ϕ (t,γ(t))dt. kh k kh k ZI Letg ∈X0,1bedG(0)cG(1)solutiondefinedby kh k,h B(ϕ ,g )=(v (t,γ(t))δ˜ ,ϕ ) , ∀ϕ ∈X0,1. (28) kh kh kh k γk kh I×Ω kh k,h ThenusingthatdG(0)cG(1)methodisconsistent,wehave T |v (t,γ(t))|2dt=B(v ,g )=B(v,g ) kh k kh kh kh 0 Z (29) M =(∇v,∇g ) − ∑(v ,[g ] ) , kh I×Ω m kh m Ω m=1 wherewehaveusedthedualexpression(19)forthebilinearformBandthefactthat thelasttermin(19)canbeincludedinthesumbysettingg =0anddefining kh,M+1 consequently[g ] =−g .Thefirstsumin(19)vanishesduetog ∈X0,1.For kh M kh,M kh k,h eacht, integrating by parts elementwise and using that g is linear in the spacial kh variable,bytheHo¨lder’sinequalitywehave 1 (∇v,∇gkh)Ω= 2∑(v,[[∂ngkh]])∂τ≤CkvkL∞(Ω)∑k[[∂ngkh]]kL1(∂τ), (30) τ τ where[[∂g ]]denotesthejumpsofthenormalderivativesacrosstheelementfaces. n kh FromLemma2.4in[29]wehave ∑k[[∂ngkh]]kL1(∂τ)≤C|lnh|21 kσk∆hgkhkL2(Ω)+k∇gkhkL2(Ω) , τ (cid:16) (cid:17) where∆ :V →V isthediscreteLaplaceoperator,definedby h h h −(∆ v ,χ) =(∇v ,∇χ) , ∀χ∈V . h h Ω h Ω h Toestimatetheterminvolvingthejumpsin(29),wefirstusetheHo¨lder’sinequality andtheinverseestimatetoobtain