NonamemanuscriptNo. (willbeinsertedbytheeditor) Numerical algorithms for the forward and backward fractional Feynman-Kac equations WeihuaDeng ,1,MinghuaChen1,EliBarkai2 ∗ 4 1 0 2 Received:date/Accepted:date n a J Abstract TheFeynman-Kacequationsareatypeofpartialdifferentialequationsdescribing thedistributionoffunctionalsofdiffusivemotion.Theprobabilitydensityfunction(PDF)of 3 Brownianfunctionals satisfiestheFeynman-Kacformula, beingaSchro¨dinger equationin ] imaginarytime.Thefunctionalsofno-Brownianmotion,oranomalousdiffusion,followthe h fractionalFeynman-Kacequation[J.Stat.Phys.141,1071-1092,2010],wherethefractional p substantialderivativeisinvolved.Basedonrecentlydevelopeddiscretizedschemesforfrac- - p tionalsubstantialderivatives[arXiv:1310.3086],thispaperfocusesonprovidingalgorithms m fornumericallysolvingtheforwardandbackwardfractionalFeynman-Kacequations;since o thefractionalsubstantialderivativeisnon-localtime-spacecoupledoperator,newchallenges c areintroducedcomparingwiththegeneralfractionalderivative.Twoways(finitedifference s. and finite element) of discretizing the space derivative are considered. For the backward c fractionalFeynman-Kacequation,thenumericalstabilityandconvergenceofthealgorithms i s withfirstorderaccuracyaretheoreticallydiscussed;andtheoptimalestimatesareobtained. y Foralltheprovidedschemes,includingthefirstorderandhighorderones,ofbothforward h andbackwardFeynman-Kac equations,extensivenumerical experiments areperformed to p showtheireffectiveness. [ Keywords fractional Feynman-Kacequation fractionalsubstantialderivative optimal 1 · · v convergent order numerical stabilityandconvergence numerical inversionofLaplace · · 4 transforms 5 6 0 1 Introduction . 1 0 Letting x(t) be a trajectory of a Brownian particle andU(x) be a prescribed function, the 4 Brownian functional can be defined as A= tU[x(t )]dt [18], which has many physical 1 0 applications. In 1949, inspiring by Feynman’s path integrals Kac derives a Schro¨dinger- : R v likeequationforthedistributionofthefunctionalsofdiffusivemotion[15].Withtherapid i development on the study of non-Brownian motion, or anomalous diffusion [19,21], the X r Correspondingauthor.E-mail:[email protected] a ∗ 1SchoolofMathematicsandStatistics,LanzhouUniversity,Lanzhou730000,P.R.China 2DepartmentofPhysicsandAdvancedMaterialsandNanotechnologyInstitute,Bar-IlanUniversity,Ramat Gan52900,Israel.E-mail:[email protected] 2 functionalsofanomalousdiffusionnaturallyattracttheinterestsofphysicists.Inparticular, Carmi, Turgeman, and Barkai derive the forward and backward fractional Feynman-Kac equationsfordescribingthedistributionofthefunctionalsofanomalousdiffusion[4,5,25], whichinvolvesthefractionalsubstantialderivative[12].BeingthesameformofBrownian functional,thefunctionalofanomalousdiffusioncanalsobedefinedas t A= U[x(t )]dt , (1.1) Z0 wherex(t) isatrajectory of non-Brownian particle; and therearealot of different choice to prescribeU(x). For example, we can takeU(x)=1 in a given domain and to be zero otherwise, which characterizes the time spent by a particle in the domain; this functional canbeusedinkineticstudiesofchemicalreactionsthattakeplaceexclusivelyinthedomain [2,5]. For inhomogeneous disorder dispersive systems, the motion of the particles is non- Brownian,andU(x)istakenasxorx2[5]. In recent decades, the numerical methods for fractional partial differential equations (PDEs)are well developed, including finitedifference methods [8,9,20,23,26], finite ele- ment[10,11,14],spectralmethod[16,17],etc.However,itseemsthattherearenopublished worksfornumericallysolvingfractionalPDEswithfractional substantialderivative.Frac- tional substantial derivative isanon-local time-spacecoupled operator; discretizing itand numericallysolvingthecorrespondingequationsundoubtedlyintroducesomenewdifficul- ties comparing with the fractional derivative. We detailedly discuss the properties and ef- fectivelynumericaldiscretizationsofthefractionalsubstantialderivativesin[6].Thispaper focuses on numerically solving the forward and backward fractional Feynman-Kac equa- tionswiththefractionalsubstantialderivativebeingdiscretizedbythewaysgivenin[6]and theclassicalspatialderivativeistreatedbyfinitedifference andfiniteelementmethod, re- spectively.ForthebackwardFeynman-Kacequation,wetheoreticallyprovethenumerical stabilityandconvergenceofitsfirstorderscheme.Foralltheproposedschemes,including thefirstorderandhighorderones,ofbothforwardandbackwardfractionalFeynman-Kac equations,theextensivenumericalexperimentsareperformedtoshowtheireffectiveness. Thedefinitionsoffractionalsubstantialcalculusaregivenasfollows[6]. Definition1 Let n >0, r be a constant, and P(t) be piecewise continuous on (0,¥ ) and integrableonanyfinitesubinterval[0,¥ ).ThenthefractionalsubstantialintegralofP(t)of ordern isdefinedas sItn P(t)= G (1n )Z0t(t−t )n −1e−rU(x)(t−t)P(t )dt , t>0, whereU(x)isaprescribedfunctionin(1.1). Definition2 Letm >0,r beaconstant,andP(t)be(m-1)-timescontinuouslydifferentiable on(a,¥ )anditsm-timesderivativebeintegrableonanyfinitesubintervalof[a,¥ ),where misthesmallestintegerthatexceedsm .ThenthefractionalsubstantialderivativeofP(t)of orderm isdefinedas sDtm P(t)=sDtm[sItm−m P(t)], where ¶ m sDm= +r U(x) . t ¶ t (cid:18) (cid:19) 3 TheforwardandbackwardfractionalFeynman-Kacequationderivedin[4,5,25]are ¶ ¶ 2 ¶ tP(x,r ,t)=k a ¶ x2sDt1−a P(x,r ,t)−r U(x)P(x,r ,t), (1.2) and ¶ ¶ 2 ¶ tP(x,r ,t)=k a sDt1−a ¶ x2P(x,r ,t)−r U(x)P(x,r ,t), (1.3) whereP(x,r ,t):= ¥ P(x,A,t)e r AdA;for(1.2),P(x,A,t)denotesthejointprobabilityden- 0 − sityfunction(PDF)offindingtheparticleon(x,A)attimet;whilefor(1.3),P(x,A,t)isthe R jointPDFoffindingtheparticleonAattimet withtheinitialpositionoftheparticleatx; the functional A is defined as (1.1); the diffusion coefficient k a is apositive constant and a (0,1);whenU(x)=0,both(1.2)and(1.3)reducestothecelebratedfractionalFokker- ∈ Planckequation[3,21].Infact,fromthedefinitionoffractionalsubstantialderivative, Eq. (1.3)canberewrittenas ¶ 2 sDtP(x,r ,t)=sDt1−a k a ¶ x2P(x,r ,t) ; (1.4) (cid:20) (cid:21) thenwecanfurthergetitsequivalentform(seetheAppendix) scDta P(x,t)=sDta [P(x,t)−e−rU(x)tP(x,0)] =sDta P(x,t)−t−Ga(e1−rUa(x))tP(x,0)=k a ¶¶x22P(x,t), (1.5) − hereandinthefollowingP(x,r ,t)isreplacedbyP(x,t)sincer istakenasafixedconstant. For(1.2),fromthedefinitionoffractionalsubstantialderivative,wecanalsorecastitas ¶ 2 sDtP(x,t)=k a ¶ x2sDt1−a P(x,t); (1.6) butitshouldbenotedthatthetwooperators ¶ 2 andD1 a donotcommute. ¶ x2 t− The outline of this paper is as follows. In Section 2, for (1.2) and (1.3)we derive the numericalschemeswithfinitedifferencemethodtodiscretizethespacederivative;andthe- oreticallyprovethatthefirstordertimediscretizationschemeisunconditionallystableand convergent for (1.3). In Section 3, for (1.3) the time semi-discretized and full discretized schemes of finite element method are provided; stability and convergence of the schemes arerigourously established; moreover, theoptimal convergent rateisobtained. Toconfirm thetheoreticalresultsandshowtheeffectivenessofthefirstorderandhighorderschemes, theextensivenumericalresultsareprovidedinSection4.Weconcludethepaperwithsome remarksinthelastsection. 2 FinitedifferenceforfractionalFeynman-Kacequation In this section wefocuses on deriving the difference schemes for the backward fractional Feynman-Kacequation (1.4)andtheoretically provethattheprovided firstorder timedis- cretizationschemeof(1.4)isunconditionallystableandconvergent;thedifferenceschemes fortheforwardfractionalFeynman-Kacequation(1.2)aregivenasaremark. 4 LettingT >0,W =(0,l),rewriting(1.5)andmakingitsubjecttothegiveninitialand boundaryconditions,wehave ¶ 2 scDta P(x,t)=sDta [P(x,t)−e−rU(x)tP(x,0)]=k a ¶ x2P(x,t), 0<t≤T, x∈W , (2.1) withinitialandboundaryconditions P(x,0)=f (x), x W , ∈ (2.2) P(0,t)=y (t), P(l,t)=y (t), 0<t T. 1 2 ≤ 2.1 Derivationofthedifferencescheme Letthemeshpointsx =ihfori=0,1,...,M,andt =nt ,n=0,1,...,N,whereh=l/M i n andt =T/Naretheuniformspacestepsizeandtimesteplength,respectively.DenotePnas i thenumericalapproximationtoP(x,t ).Toapproximate(2.1),weutilizethesecondorder i n centraldifferenceformulaforthespatialderivative;thatis ¶ 2P(x,t) P(x ,t ) 2P(x,t )+P(x ,t ) ¶ x2 (xi,tn)= i+1 n − h2i n i−1 n +O h2 . (cid:12) From(3.8)of[6],wek(cid:12)nowthatthefractionalsubstantialderivativehasq(cid:0)-th(cid:1)orderapproxi- (cid:12) mations,i.e., n sDta P(x,t)|(xi,tn)=t −a (cid:229) diq,k,a P(xi,tn−k)+O(t q), q=1,2,3,4,5, (2.3) k=0 with diq,k,a =e−rUiktlkq,a , Ui=U(xi), q=1,2,3,4,5, wherel1,a ,l2,a ,l3,a ,l4,a andl5,a aredefinedby(2.2),(2.4),(2.6),(2.8)and(2.10)in[7], k k k k k respectively.Inthefollowing,wedothedetailedtheoreticalanalysisforthefirstordertime discretizationschemeof(2.1).Forthesimplification,wedenoted1,a byda ;then i,k i,k n sDta P(x,t)|(xi,tn)=t −a (cid:229) dia,kP(xi,tn−k)+O(t ); k=0 (2.4) n sDta [e−rU(x)tP(x,0)](xi,tn)=t −a (cid:229) dia,ke−rUi(n−k)t P(xi,0)+O(t ), k=0 wherethecoefficients a dia,k=e−rUikt gk, gk=(−1)k k , (2.5) (cid:18) (cid:19) with a +1 g =1, g = 1 g , k 1. 0 k − k k−1 ≥ (cid:18) (cid:19) ThenEq.(2.1)canberewrittenas n n t −a (cid:229) dia,kP(xi,tn−k)−t −a (cid:229) dia,ke−rUi(n−k)tP(xi,0) k=0 k=0 (2.6) P(x ,t ) 2P(x,t )+P(x ,t ) =k a i+1 n − h2i n i−1 n +rin, 5 with rn C (t +h2), (2.7) | i|≤ P whereC isaconstantdependingonlyonP. P Multiplying(2.6)byt a ,wehavethefollowingequation n 1 n 1 (cid:229)− dia,kP(xi,tn−k)−(cid:229)− dia,ke−rUi(n−k)tP(xi,0) k=0 k=0 (2.8) =k a t a P(xi+1,tn)−2P(hx2i,tn)+P(xi−1,tn)+Rni, with Rn = t a rn C t a (t +h2). (2.9) | i| | i|≤ P From(2.5)and(2.8),theresultingdiscretizationof(2.1)canberewrittenas k t a n 1 n 1 Pin− ah2 Pin+1−2Pin+Pin−1 =k(cid:229)=−0dia,ke−rUi(n−k)tPi0−k(cid:229)=−1dia,kPin−k, n≥1; (2.10) (cid:0) (cid:1) or k t a n 1 n 1 Pin− ah2 Pin+1−2Pin+Pin−1 =k(cid:229)=−0e−rUint gkPi0−k(cid:229)=−1e−rUiktgkPin−k, n≥1, (2.11) (cid:0) (cid:1) with i=1,2,...,M 1. It is worthwhile to noting that the second term on the right hand − sideof(2.10)or(2.11),respectively,automaticallyvanisheswhenn=1. Remark2.1 Ifweutilizetheq-th order approximation of(2.3)todiscretizethetimefrac- tionalsubstantialderivativeof(2.1),theresultingdiscretizationof(2.1)is k t a n 1 n 1 diq,0,a Pin− ah2 Pin+1−2Pin+Pin−1 =k(cid:229)=−0diq,k,a e−rUi(n−k)tPi0−k(cid:229)=−1diq,k,a Pin−k, (2.12) (cid:0) (cid:1) whichgivesalocaltruncationerrorofO t q+h2 ,q=2,3,4,5. Remark2.2 Using the second order cen(cid:0)tral diff(cid:1)erence formula for the spatial derivative leadsto ¶ 2 ¶ x2sDt1−a P(xi,tn) = sDt1−a P(xi+1,tn)−2sDt1−a P(xi,tn)+sDt1−a P(xi−1,tn)+O h2 . h2 (cid:0) (cid:1) Furtherapplying(2.3)toapproximatethetimefractionalsubstantialderivatives,thenweget thediscretizationschemesof(1.6): k t a diq,0,1Pin− ah2 (diq+,11−,0a Pin+1−2diq,0,1−a Pin+diq−,11−,0a Pin−1) n k t a n (2.13) =−k(cid:229)=1diq,k,1Pin−k+ ah2 k(cid:229)=1(diq+,11−,ka Pin+−1k−2diq,k,1−a Pin−k+diq−,11−,ka Pin−−1k), withthelocaltruncationerrorO t q+h2 ,q=1,2,3,4,5. (cid:0) (cid:1) 6 2.2 Stabilityandconvergence Inthissubsection,weprovethatthescheme(2.10)isunconditionallystableandconvergent indiscreteL2normandL¥ normundertheassumptionthat0 r U h .First,weintroduce i ≤ ≤ some relevant notations and properties of discretized inner product given in [13]. Denote un= un 0 i M,n 0 andvn= vn 0 i M,n 0 ,whicharegridfunctions.And { i| ≤ ≤ ≥ } { i| ≤ ≤ ≥ } (un) =(un un)/h, (un) =(un un )/h; i x i+1− i i x i − i−1 M 1 (un,vn)= (cid:229) − unvnh, un =(un,un)1/2; i i || || i=1 M (un,vn]=(cid:229) unvnh, un] =(un,un]1/2. i i || | i=1 Inparticular,ifun=0andun =0,thereexists 0 M l2 (un,(vn) )= (un,vn] and un 2 un]2, (2.14) x x − x x || || ≤ 8|| x | wherelmeanstheoneappearedinW =(0,l). Lemma2.1 Thecoefficientsg definedin(2.5)satisfy k n 1 ¥ g =1; g <0, (k 1); (cid:229)− g >0; (cid:229) g =0; (2.15) 0 k k k ≥ k=0 k=0 and 1 n(cid:229)−1 (cid:229)¥ 1 < g = g , for n 1. (2.16) na G (1 a ) k − k≤ na ≥ − k=0 k=n Proof From [22, p.208], it is easy to get (2.15). Next we prove (2.16). Denoting v = n −na (cid:229) ¥k=ngk=na (cid:229) nk=−01gk,n≥1,accordingto[9],thereexists v <v , i.e., (cid:229) n g < na n(cid:229)−1g , for n 1, (2.17) n+1 n k (n+1)a k ≥ k=0 k=0 and 1 n(cid:229)−1 (cid:229)¥ < g = g , for n 1. na G (1 a ) k − k ≥ − k=0 k=n Nextweprovethefollowinginequalitybymathematicalinduction n(cid:229)−1g = (cid:229)¥ g 1 , for n 1. (2.18) k − k≤ na ≥ k=0 k=n Itisobviousthat(2.18)holdswhenn=1orn=2.Supposingthat s(cid:229)−1g = (cid:229)¥ g 1 , s=1,2,...,n 1, i − i≤ sa − i=0 i=s andusing(2.17),weobtain n(cid:229)−1 (n 1)a n(cid:229)−2 (n 1)a 1 1 gk< −na gk≤ −na (n 1)a = na , for n≥2. k=0 k=0 − Thenthedesiredinequality(2.16)holds. 7 Theorem2.1 When0 r U h ,thedifferencescheme(2.11)isunconditionallystable. i ≤ ≤ Proof LetPn betheapproximatesolutionofPn,whichistheexactsolutionofthescheme i i (2.11). Taking en =Pn Pn, i=1,2,...,M 1, then from (2.11) we get the following i i − i − perturbationeequation e eni −t a k a eni+1−2he2ni +eni−1 =n(cid:229)−1e−rUint gke0i −n(cid:229)−1e−rUiktgkein−k, (2.19) k=0 k=1 withen=en =0.Multiplying(2.19)byhenandsummingupforifrom1toM 1,then 0 M i − hM(cid:229) −1(eni)2−t a k a hM(cid:229) −1eni+1−2he2ni +eni−1eni i=1 i=1 (2.20) M 1n 1 M 1n 1 =h (cid:229) − (cid:229)− e−rUintgke0ieni −h (cid:229) − (cid:229)− e−rUiktgkein−keni. i=1 k=0 i=1 k=1 Nextweestimate(2.20).Firstly,wehave M 1 h (cid:229) − (en)2= en 2, (2.21) i || || i=1 andfrom(2.14),itleadsto −t a k a hM(cid:229) −1eni+1−2he2ni +eni−1eni i=1 M 1 =−t a k a h (cid:229) − ((eni)x)xeni =−t a k a (en,(enx)x] (2.22) i=1 =t a k a (enx,exn]=t a k a ||enx]|2≥ 8t la2k a ||en||2≥0. Sincee−rUint [e−h T,1]andfrom(2.5)and(2.15),weobtain ∈ gk≤dia,k=e−rUikt gk≤e−h Tgk<0, k≥1, (2.23) and n 1 n 1 n 1 0<e−rUint (cid:229)− gk= (cid:229)− e−rUint gk (cid:229)− gk, n 1. (2.24) ≤ ≥ k=0 k=0 k=0 Therefore,accordingto(2.24)and(2.23),weobtain hM(cid:229) −1n(cid:229)−1e−rUint gke0ieni ≤hM(cid:229) −1n(cid:229)−1e−rUint gk(e0i)2+2(eni)2 i=1 k=0 i=1 k=0 (2.25) 1n(cid:229)−1g e0 2+ en 2 , ≤ 2 k || || || || k=0 (cid:0) (cid:1) and −hM(cid:229) −1n(cid:229)−1e−rUiktgkeni−keni ≤−hM(cid:229) −1n(cid:229)−1e−rUiktgk(ein−k)22+(eni)2 i=1 k=1 i=1 k=1 (2.26) ≤−21n(cid:229)−1gk ||en−k||2+||en||2 . k=1 (cid:16) (cid:17) 8 From(2.20-2.26),thereexists en 2 1n(cid:229)−1gk e0 2+ en 2 1n(cid:229)−1gk en−k 2+ en 2 || || ≤ 2 || || || || −2 || || || || k=0 k=1 (cid:0) (cid:1) (cid:16) (cid:17) (2.27) = 1+21nk(cid:229)=−11gk!||e0||2−21nk(cid:229)=−11gk||en−k||2. Nextweprovethat en 2 e0 2 bymathematicalinduction.Forn=1,(2.27)holdsob- || || ≤|| || viously.Supposing es 2 e0 2, for s=1,2,...,n 1, || || ≤|| || − andusing(2.27),thenweget ||en||2≤ 1+12nk(cid:229)=−11gk!||e0||2−21nk(cid:229)=−11gk||en−k||2 1+1n(cid:229)−1g e0 2 1n(cid:229)−1g e0 2= e0 2. ≤ 2k=1 k!|| || −2k=1 k|| || || || Hence,theproofiscomplete. Lemma2.2 LetR 0;e k 0,k=0,1,...,Nandsatisfy ≥ ≥ n 1 e n (cid:229)− gke n−k+R, n 1, (2.28) ≤− ≥ k=1 thenwehavethefollowingestimates: (a)when0<a <1, 1 e n n(cid:229)−1g − R na G (1 a )R; (2.29) k ≤ k=0 ! ≤ − (b)whena 1, → e n nR. (2.30) ≤ Proof Itisworthtonotingthatthefirsttermontherighthandsideof(2.28)automatically vanisheswhenn=1. (1)Case0<a <1:Weprovethefollowingestimatebythemathematicalinduction, 1 e n n(cid:229)−1g − R. k ≤ k=0 ! Eq.(2.28)holdsobviouslyforn=1.Supposingthat 1 e s s(cid:229)−1g − R, s=1,2,...,n 1, i ≤ i=0 ! − 9 thenform(2.28)wehave 1 1 e n n(cid:229)−1gke n−k+R n(cid:229)−1gk n−(cid:229) k−1gi − R+R n(cid:229)−1gk n(cid:229)−1gi − R+R ≤−k=1 ≤−k=1 i=0 ! ≤−k=1 i=0 ! 1 1 n(cid:229)−1 n(cid:229)−1 − n(cid:229)−1 − 1 g g R+R g R. k i i ≤ −k=0 ! i=0 ! ≤ i=0 ! Accordingto(2.16)andtheaboveinequality,itleadsto 1 e n n(cid:229)−1g − R na G (1 a )R. k ≤ k=0 ! ≤ − (2) Now weconsider the casea 1. SinceG (1 a ) ¥ as a 1 inthe estimate → − → → (2.29). Therefore, we need to look for an estimateof other form. We prove the following estimatebythemathematicalinduction: e n nR. ≤ Eq.(2.28)holdsobviouslyforn=1.Supposingthat e s sR, s=1,2,...,n 1, ≤ − thus,from(2.28)weget n 1 n 1 n 1 e n (cid:229)− gke n−k+R (cid:229)− gk(n k)R+R (cid:229)− gk(n 1)R+R (n 1)R+R=nR. ≤− ≤− − ≤− − ≤ − k=1 k=1 k=1 Theorem2.2 Let Pn be the approximate solution of P(x,t ) computed by the difference i i n scheme(2.11)withtheassumption0 r U h .Then i ≤ ≤ P(x,t ) Pn C G (1 a )l1/2Ta (t +h2), 0<a <1, || i n − i ||≤ P − whereC isdefinedby(2.7)and(x,t ) (0,l) (0,T],i=1,2,...,M 1; n=1,2,...,N. P i n ∈ × − Proof Similartotheproofof[9],letP(x,t )betheexactsolutionof(2.1)atthemeshpoint i n (x,t ),ande n=P(x,t ) Pn.Subtracting(2.8)from(2.11)andusinge 0=0,weobtain i n i i n − i i k t a n 1 ein− ah2 ein+1−2ein+ein−1 =−k(cid:229)=−1e−rUiktgkein−k+Rni, n≥1, (2.31) (cid:0) (cid:1) whereRnisdefinedby(2.9). i Multiplying(2.31)byhe nandsummingupforifrom1toM 1,thereexists i − hM(cid:229) −1(ein)2−t a k a hM(cid:229) −1ein+1−2he2in+ein−1ein i=1 i=1 (2.32) M 1n 1 M 1 =−h (cid:229) − (cid:229)− e−rUiktgkein−kein+h (cid:229) − Rniein. i=1 k=1 i=1 10 ItfollowsfromtheproofofTheorem2.1that M 1 h (cid:229) − (e n)2= e n 2; i || || i=1 −t a k a hM(cid:229) −1ein+1−2he2in+ein−1ein≥0; (2.33) i=1 −hM(cid:229) −1n(cid:229)−1e−rUiktgkein−kein≤−12n(cid:229)−1gk ||e n−k||2+||e n||2 . i=1 k=1 k=1 (cid:16) (cid:17) Accordingto(2.9),(2.7)and(2.16),weobtain[9] hM(cid:229) −1 Rne n =t a hM(cid:229) −1 rne n t a hM(cid:229) −1 t a (rn)2+(cid:229) kn=−01gk(e n)2 i=1 | i i| i=1 | i i|≤ i=1 "2(cid:229) kn=−01gk i 2t a i # = t 2a h M(cid:229) −1(rn)2+1n(cid:229)−1g e n 2 2(cid:229) n 1g i 2 k|| || k=−0 k i=1 k=0 ≤ t 2a na G2(1−a )h(M−1)CP2(t +h2)2+21n(cid:229)−1gk||e n||2 (2.34) k=0 ≤ Ta G (12−a )lCP2t a (t +h2)2+21n(cid:229)−1gk||e n||2 k=0 =C2t a (t +h2)2+1n(cid:229)−1g e n 2, k 2 2 || || k=0 where(x,t ) (0,l) (0,T],and i n ∈ × C =lC2G (1 a )Ta . (2.35) 2 P − Accordingto(2.33)and(2.34),thereexists ||e n||2≤−12n(cid:229)−1gk ||e n−k||2+||e n||2 +C22t a (t +h2)2+21n(cid:229)−1gk||e n||2 k=1 k=0 (cid:16) (cid:17) (2.36) =−12n(cid:229)−1gk||e n−k||2+C22t a (t +h2)2+21||e n||2, k=1 thatis n 1 e n 2 (cid:229)− gk e n−k 2+C2t a (t +h2)2. (2.37) || || ≤− || || k=1 Accordingto(2.35)-(2.37)andLemma2.2,wehave e n 2 na G (1 a )C t a (t +h2)2=lC2G (1 a )Ta G (1 a )Ta (t +h2)2. || || ≤ − 2 P − − Hence P(x,t ) Pn = e n C G (1 a )l1/2Ta (t +h2). || i n − i || || i||≤ P −