ebook img

B-Spline-Based Monotone Multigrid Methods Markus Holtz, Angela Kunoth no. 252 PDF

25 Pages·2006·0.37 MB·English
by  
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 B-Spline-Based Monotone Multigrid Methods Markus Holtz, Angela Kunoth no. 252

B-Spline-Based Monotone Multigrid Methods Markus Holtz, Angela Kunoth no. 252 Diese Arbeit ist mit Unterstützung des von der Deutschen Forschungs- gemeinschaft getragenen Sonderforschungsbereiches 611 an der Univer- sität Bonn entstanden und als Manuskript vervielfältigt worden. Bonn, Dezember 2005 B{SPLINE{BASED MONOTONE MULTIGRID METHODS(cid:3) MARKUS HOLTZ AND ANGELA KUNOTHy Abstract. Forthee(cid:14)cientnumericalsolutionofellipticvariationalinequalitiesonclosedconvex sets,multigridmethodsbasedonpiecewiselinear(cid:12)niteelementshavebeeninvestigatedoverthepast decades. Essentialfortheirsuccessistheappropriateapproximationoftheconstraintsetoncoarser gridswhichisbasedonfunctionvaluesforpiecewiselinear(cid:12)niteelements. Ontheotherhand,there are a number of problems which pro(cid:12)t from higher order approximations. Among these are the problem of prizing American options, formulated as a parabolic boundary value problem involving Black{Scholes’ equation with a free boundary. In addition to computing the free boundary, the optimal exercise prize of the option, of particular importance are accurate pointwise derivatives of thevalueofthestockoptionuptoordertwo,theso{calledGreekletters. Inthispaper,weproposeamonotonemultigridmethodfordiscretizationsintermsofB{splines ofarbitraryordertosolveellipticvariationalinequalitiesonaclosedconvexset. Inordertomaintain monotonicity (upper bound) and quasi{optimality (lower bound) ofthe coarse gridcorrections, we proposeanoptimizedcoarsegridcorrection(OCGC)algorithmwhichisbasedonB{splineexpansion coe(cid:14)cients. WeprovethattheOCGCalgorithmisofoptimalcomplexityofthedegreesoffreedomof thecoarsegridand,therefore,theresultingmonotonemultigridmethodisofasymptoticallyoptimal multigridcomplexity. Finally, the method is applied to a standard model for the valuation of American options. In particular,itisshownthat adiscretization based onB{splinesoforderfourenables ustocompute thesecondderivativeofthevalueofthestockoptiontohighprecision. Keywords. Variationalinequality,linearcomplementaryproblem,monotonemultigridmethod, cardinalhigherorderB{spline,systemoflinearinequalities,optimizedcoarsegridcorrection(OCGC) algorithm,optimalcomplexity,convergence rates,Americanoption,Greekletters,highprecision. AMS subject classi(cid:12)cations. 65M55,35J85,65N30,65D07. 1. Introduction. The motivation for this paper stems from an application in Mathematical Finance, the fair prizing of American options. In a standard model, this problem can be formulated as a parabolic boundary value problem involving Black{Scholes’equation[BS]with afreeboundary. Inadditiontocomputingthe free boundary(theoptimalexerciseprizeoftheoption),pointwisehigherorderderivatives of the solution (the value of the stock option) are particularly important. These so{ calledGreeklettersareneededwithhighprecisionastheyplayacrucialroleashedge parameters in the analysis of market risks. Thus, a discretization in terms of higher order basis functions is preferable. Ontheotherhand,forthefastnumericalsolutionoftheresulting(semi{discrete) ellipticvariationalinequality,themethodofchoiceisthemonotonemultigridmethod developed in [Ko1, Ko2]. Multigrid methods have been proposed previously for such problems using second order discretizations (i.e., standard (cid:12)nite di(cid:11)erence stencils or piecewise linear (cid:12)nite elements) in di(cid:11)erent variants [BC, HM, Ho, Ma] where, however,notallof themhaveassuredconsequentlythat the obstaclecriterionismet. Usingpiecewiselinear(cid:12)niteelementansatzfunctions,geometricconsiderationsbased on point values are used in [Ko1] to represent the problem{inherent obstacles on coarsergridsin suchawaythat aviolationof the obstacleisexcluded. The di(cid:14)culty tocorrectlyidentifyingcoarsegridapproximationshasalsobeen the motivationfora (cid:3) Thisworkhasbeen supportedinpartbytheDeutsche Forschungsgemeinschaft SFB611,Uni- versita(cid:127)tBonn. yInstitut fu(cid:127)r Numerische Simulation, Universita(cid:127)t Bonn, Wegelerstr. 6, 53115 Bonn, Germany, fholtz,[email protected], www.ins.uni-bonn.de/(cid:24)kunoth. 1 2 MarkusHoltzandAngelaKunoth cascadic multigrid algorithm for variationalinequalities in [BBS] for which, however, no convergencetheory is yet available. In this paper, we generalize the monotone multigrid (MMG) method from [Ko1, Ko2] to discretizations involving higher order B{splines. One of the key ingredients of an MMG method are restrictions of the obstacle to coarsergrids which satisfy the (upper) bound imposed by the obstacle (monotonicity) as well as a lower one which corresponds to the condition of quasi{optimality in [Ko1]. We formulate the con- struction of coarse grid approximationsas a linear constrained optimization problem with respect to the B{spline expansion coe(cid:14)cients. Our construction heavily pro(cid:12)ts from properties of B{splines [Bo, Sb]. In particular, we present with our optimized coarsegridcorrection(OCGC)algorithmamethodtoconstructmonotoneandquasi{ optimal coarsegrid approximationsto the obstaclefunction in optimal complexity of the coarse grid for B{spline basis functions of any degree. Building the OCGC scheme into the MMG method, our higher-order MMG method is shown to be of optimal multigrid complexity. Moreover, following the argumentsin[Ko1],wecanprovethatourmethodisgloballyconvergentandreduces asymptotically to a linear subspace correction method once the contact set has been identi(cid:12)ed [HzK]. Hence, we can expect particular robustness of the scheme and full multigrid e(cid:14)ciency in the asymptotic range in the numerical experiments. This is con(cid:12)rmedbycomputationsforanAmericanoptionpricingproblemin terms of cubic B{splines. Details about the derivation of the problem of fair prizing American op- tionsandits formulationasafreeboundaryvalueproblem andcorrespondingresults can be found in [WHD, Hz2]. Of course, once higher-oder MMG methods are avail- able, they may be applied to other obstacle problems like Signorini’s problem which has been solved using piecewise linear hat functions in [Kr]. Thispaperisstructuredasfollows. InSection2weintroducemonotonemultigrid methods (MMG), recollect the main features of B{splines and specify a B{spline{ based projected Gauss{Seidel relaxation as smoothing component of the scheme. In Section3thecrucialingredientsofthehigher-orderMMGschemes,suitablerestriction operators for the obstacle function, are presented for B{spline functions of arbitrary degreein the univariatecase. Theirconstructionforhigherspatialdimensionsis pre- sentedinSection4usingtensorproducts. InSection5someshortremarksconcerning the convergence theory for B{spline{based monotone multigrid schemes are made. Finally, in Section 6 we present a numerical example of prizing American options. Theconvergencebehaviorofthe projectedGauss{Seidelandthe multigridschemesis compared for basis functions of di(cid:11)erent orders. We conclude with an estimation of asymptotic multigrid convergencerates which exhibit full multigrid e(cid:14)ciency for the truncated version. 2. Monotone Multigrid Methods. 2.1. Elliptic Variational Inequalities and Linear Complementary Prob- lems. Let (cid:10) be a domain in Rd and J(v) := 1a(v;v) (cid:0) f(v) a quadratic func- 2 tional induced by a continuous, symmetric and H1{ elliptic bilinear form a((cid:1);(cid:1)) : 0 H1((cid:10))(cid:2)H1((cid:10))!R and a linear functional f :H1((cid:10))!R. As usual, H1((cid:10)) is the 0 0 0 0 subspace of functions belonging to the Sobolev space H1((cid:10)) with zero trace on the boundary. We consider the constrained minimization problem (cid:12)nd u2K: J(u)(cid:20)J(v) for all v 2K (2.1) B{Spline{BasedMonotoneMultigridMethods 3 on the closed and convex set K:=fv 2H1((cid:10)): v(x)(cid:20)g(x) for all x2(cid:10)g(cid:26)H1((cid:10)): 0 0 The function g 2 H1((cid:10)) represents an upper obstacle for the solution u 2 H1((cid:10)). 0 0 Lower obstacles can be treated in the obvious analogous way. If g satis(cid:12)es g(x) (cid:21) 0 for all x 2 @(cid:10), problem (2.1) admits a unique solution u 2 K by the Lax{Milgram theorem. It is well{knownthat (2.1)can be rewritten as a variationalinequality, see, e.g., [EO, KS]: (cid:12)nd u2K: a(u;v(cid:0)u)(cid:21)f(v(cid:0)u) for all v 2K or, equivalently, as a linear complementary problem Lu (cid:21) f; u (cid:20) g; (2.2) (u(cid:0)g)(Lu(cid:0)f) = 0 almost everywhere in (cid:10). Here L : H1((cid:10)) ! H(cid:0)1(= (H1((cid:10)))0) is the Riesz operator 0 0 de(cid:12)ned by hLu;vi:=a(u;v) for all v 2H1((cid:10)). 0 Discretizingin a(cid:12)nite dimensionalsplinespaceS of piecewise polynomialsona L grid (cid:1) with uniform grid spacing h leads to the discrete formulation of (2.1), L L (cid:12)nd u 2K : J(u )(cid:20)J(v ) for all v 2K (2.3) L L L L L L on the closed and convex set K :=fv 2 S : v (x) (cid:20)g (x) for all x 2 (cid:10)g(cid:26)S ; L L L L L L or, equivalently, L u (cid:21) f ; L L L u (cid:20) g ; (2.4) L L (u (cid:0)g )(L u (cid:0)f ) = 0: L L L L L In[BHR]regularityu2H5=2(cid:0)(cid:15)((cid:10))ofthesolutionuto(2.2)isshownforarbitrary (cid:15) > 0. Moreover, error estimates ku(cid:0)u k = O(h ) and ku(cid:0)u k = L H1((cid:10)) L L H1((cid:10)) O(h3=2(cid:0)(cid:15))areprovedinthe caseof piecewiselinear,respectivelypiecewise quadratic, L functions, provided the functions f;g are su(cid:14)ciently regular. 2.2. The MMG{algorithm. For solving (2.3) numerically, a by now popular methodisthemonotonemultigridmethod(MMG)[Ko1]. Byaddingaprojectionstep and employing speci(cid:12)c restriction operators, it can be implemented as a variant of a standardmultigridscheme. LetS (cid:26)S (cid:26):::(cid:26)S (cid:26)H1((cid:10))beanestedsequenceof 1 2 L 0 (cid:12)nite{dimensionalspaces,andletu(cid:23) 2S betheapproximationinthe(cid:23){thiteration L L of the MMG method. The basic multigrid idea is that the error v := u (cid:0)u(cid:23);1 L L L between the smoothed iterate u(cid:23);1 := S(u(cid:23)) (S always being the standard Gauss{ L L Seideliteration)andtheexactsolutionu canbeapproximatedwithoutessentialloss L of information on a coarsergrid (cid:1) . We explain how this is realized in the case of L(cid:0)1 a linear complementary problem for two grids (cid:1) and (cid:1) . Introducing the defect L L(cid:0)1 d :=f (cid:0)L u(cid:23);1, (2.4) can be written as L L L L L v (cid:21) d ; L L L v (cid:20) g (cid:0)u(cid:23);1; (2.5) L L L (v (cid:0)g +u(cid:23);1)(L v (cid:0)d ) = 0: L L L L L L 4 MarkusHoltzandAngelaKunoth On a coarsergrid (cid:1) the defect problem can now be approximated by L(cid:0)1 L v (cid:21) d ; L(cid:0)1 L(cid:0)1 L(cid:0)1 v (cid:20) g ; L(cid:0)1 L(cid:0)1 (v (cid:0)g )(L v (cid:0)d ) = 0; L(cid:0)1 L(cid:0)1 L(cid:0)1 L(cid:0)1 L(cid:0)1 where d := rd and g := r~(g (cid:0)u(cid:23);1) with (di(cid:11)erent) restriction operators L(cid:0)1 L L(cid:0)1 L L r;r~ : S ! S . The solution v of the coarse grid problem is then used as L L(cid:0)1 L(cid:0)1 an approximation to the error v . It is (cid:12)rst transported back to the (cid:12)ne grid by a L prolongationoperatorpand isthen addedto the approximationu(cid:23);1. It isimportant L that the restriction r~is chosen such that the new iterate satis(cid:12)es the constraint u(cid:23);2 :=u(cid:23);1+pv (cid:20)g (2.6) L L L(cid:0)1 L on the (cid:12)ne grid. Applying this idea recursivelyon severaldi(cid:11)erent grids,one obtains the monotone multigrid method (MMG) for linear complementary problems. Algorithm 2.1. MMG ((cid:23){ th cycle on level ‘(cid:21)1) ‘ Let u(cid:23) 2S be a given approximation. ‘ ‘ 1. A priori smoothing and projection : u(cid:23);1 :=(P(cid:14)S(u(cid:23)))(cid:17)1. ‘ ‘ 2. Coarse grid correction: d :=r(f (cid:0)L u(cid:23);1), ‘(cid:0)1 ‘ ‘ ‘ g :=r~(g (cid:0)u(cid:23);1), ‘(cid:0)1 ‘ ‘ L :=rL p. ‘(cid:0)1 ‘ If ‘=1, solve exactly the linear complementary problem L v (cid:21) d ; ‘(cid:0)1 ‘(cid:0)1 v (cid:20) g ; ‘(cid:0)1 (v(cid:0)g )(L v(cid:0)d ) = 0 ‘(cid:0)1 ‘(cid:0)1 ‘(cid:0)1 and set v :=v. ‘(cid:0)1 If ‘>1, do (cid:13) steps of MMG with initial value u0 :=0 and solution v . ‘(cid:0)1 ‘(cid:0)1 ‘(cid:0)1 Set u(cid:23);2 :=u(cid:23);1+pv . ‘ ‘ ‘(cid:0)1 3. A posteriori smoothing and projection : u(cid:23);3 :=(P(cid:14)S(u(cid:23);2))(cid:17)2: ‘ ‘ Set u(cid:23)+1 :=u(cid:23);3: ‘ ‘ The number of a priori and a posteriori smoothing steps is denoted by (cid:17) and (cid:17) , 1 2 respectively. For (cid:13) = 1 one obtains a V{cycle, for (cid:13) = 2 a W{cycle. P denotes a projection operator de(cid:12)ned in (2.7) and (2.11) below. Condition(2.6)leadstoaninnerapproximationofthesolutionsetK andensures L that the multigrid scheme is robust [Ko1]. Striving for optimal multigrid e(cid:14)ciency, satisfaction of the constraint should not be checked by interpolating v back to the ‘ (cid:12)nestgrid. Instead,specialrestrictionoperatorsr~areneededfortheobstaclefunction. A corresponding construction for B{splines of general order k will be introduced in Sections 3 and 4. Next we discuss the projection step for general order B{splines. 2.3. A B{Spline{Based Projected Gauss{Seidel Scheme. Since the op- erator L is symmetric positive de(cid:12)nite and continuous piecewise linear functions are used for discretization,the discrete form (2.4) canbe solvedbythe projected Gauss{ Seidel scheme, see, e.g., [Cr]. Given an iterate u(cid:23), a standard Gauss{Seidel sweep L u(cid:22)(cid:23) :=S(u(cid:23))issupplemented byaprojectionu(cid:23)+1 =Pu(cid:22)(cid:23) intothe convexsetK . If L L L L L B{Spline{BasedMonotoneMultigridMethods 5 S consists of hat functions, the projection can be de(cid:12)ned for given grid points f(cid:18) g L i i by Pv ((cid:18) ):=minfv ((cid:18) ); g ((cid:18) )g: (2.7) L i L i L i Forhigher{orderfunctions v , the di(cid:14)cultyarisesalreadyinthe univariatecasethat L for given x2[(cid:18) ;(cid:18) ] the estimate i i+1 minfv ((cid:18) );v ((cid:18) )g(cid:20)v (x)(cid:20)maxfv ((cid:18) );v ((cid:18) )g (2.8) L i L i+1 L L i L i+1 is not valid any more. Thus, controlling function values on grid points is not a su(cid:14)cient criterion in this case. We propose here instead a construction using higher order B{splines, which compares B{spline expansion coe(cid:14)cients instead of function values and heavily pro(cid:12)ts from the fact that B{splines are nonnegative. We begin with the univariate case. For readers’ convenience, we recall the relevant facts about B{spline bases from [Bo]. Definition 2.2 (B{Spline Basis Functions). For k 2 N and n 2 N let T := f(cid:18) g be an expanded knot sequence with uniform grid spacing h in the i i=1;:::;n+k L interior of the interval I :=[a;b] of the form (cid:18) =:::=(cid:18) =a<(cid:18) <:::<(cid:18) <b=(cid:18) =:::=(cid:18) : (2.9) 1 k k+1 n n+1 n+k Then the B{spline basis functions N of order k are recursively de(cid:12)ned for i = i;k 1;:::;n by 1; if x2[(cid:18) ;(cid:18) ) N (x)= i i+1 ; i;1 0; else (cid:26) (2.10) x(cid:0)(cid:18) (cid:18) (cid:0)x i i+k N (x)= N (x)+ N (x) i;k i;k(cid:0)1 i+1;k(cid:0)1 (cid:18) (cid:0)(cid:18) (cid:18) (cid:0)(cid:18) i+k(cid:0)1 i i+k i+1 for x2I. It is known that suppN (cid:18) [(cid:18) ; (cid:18) ] (local support), N (x) (cid:21) 0 for all x 2 I i;k i i+k i;k (nonnegativity) and N 2Ck(cid:0)2(I) (di(cid:11)erentiability) holds. Moreoverthe set (cid:6) := i;k L fN ;:::;N g constitutes a locally independent and unconditionally stable basis 1;k n;k with respect to k(cid:1)k , 1 (cid:20) p (cid:20) 1, for the (cid:12)nite dimensional space S = N := Lp L k;T span(cid:6) of the splines of order k. L Lemma 2.3. If the B{spline coe(cid:14)cients of v ;g 2N =S satisfy v (cid:20)g for L L k;T L i i all i=1;:::;n, then v (x)(cid:20)g (x) holds for all x2I. L L Proof. Usingtherepresentationv = n v N andg = n g N andthe L i=1 i i;k L i=1 i i;k nonnegativity N (x) (cid:21) 0 for all x 2 I, we deduce that g (x)(cid:0)v (x) = n (g (cid:0) i;k P L PL i=1 i v )N (x)(cid:21)0 for all x2I: i i;k P Here and below in Section 5, we use the subscript i in v = (v ) to denote i L i B{spline expansion coe(cid:14)cients. Theprojectioncannowbede(cid:12)nedforB{splinefunctionsofgeneralorderksimilar to (2.7) but now involving expansion coe(cid:14)cients by setting Pv :=minfv ; g g: (2.11) i i i Usingthesameargumentsasin[Cr],theresultingprojectedGauss{Seidelschemestill convergessincethediscretesolutionsetfv2IRn : v (cid:20)g fori=1;:::;ngdescribesa i i cuboidinIRn. Moreover,iftheproblemisnon{degenerate,thecontactset,de(cid:12)nedby 6 MarkusHoltzandAngelaKunoth allcoe(cid:14)cientsforwhichequalityholds,isidenti(cid:12)edaftera(cid:12)nitenumberofiterations [Cr, EO]. We treat the multivariate case by takingtensor products. Specifying the domain (cid:10) as (cid:10):= d [a ;b ](cid:26)Rd, the i-th d-dimensional tensor product B{spline of order ‘=1 ‘ ‘ k on a tensorized extended knot sequence T(d) is de(cid:12)ned by Q d N(d)(x):= N (x ); x2(cid:10); (2.12) i;k i‘;k ‘ ‘=1 Y wherei:=(i ;:::;i )denotesamulti-index. De(cid:12)ningS inanalogytotheunivariate 1 d L case, the result of Lemma 2.3 immediately carriesover to the d-dimensional setting. 3. Constructionof Monotoneand Quasi{optimalObstacle Approxima- tions. In this section, the second essential ingredient for our B{spline{based mono- tonemultigridmethodsisprovided,theconstructionofso{calledmonotoneandquasi{ optimal coarse grid approximations of the obstacle function, which lead to suitable restriction operators r~. We begin with the univariate case; the extension to d di- mensions follows in Section 4. We consider in the following only two grids, as the generalization to several grids is obvious. Given an obstacle function S~ which is de- (cid:12)ned on a (cid:12)ne grid (cid:1)(cid:26)I, we provide an approximation S with respect to a coarser grid T which satis(cid:12)es 1. S(x)(cid:20)S~(x) for all x2I; 2. S(x) (cid:21) L (x) for all x 2 I and a still to be speci(cid:12)ed lower barrier L (x) k k provided below in Section 3.2; 3. S (cid:25)S~ with respect to a target functional F de(cid:12)ned below in (3.10). k The (cid:12)rst condition ensuresthe monotonicity androbustnessof the multigridscheme, thesecondanasymptoticalreductionofthemethodtoalinearrelaxationandthethird an e(cid:14)cient coarsegrid correction. As the construction is used as a component of the monotonemultigridscheme,strivingforoptimalcomputationalmultigridcomplexity, it also has to satisfy 4. the number of arithmetic operations must be of order O(n) where n denotes the number of degrees of freedom on the coarse grid. Speci(cid:12)cally, let T be an extended knot sequence with grid spacing H as in (2.9) and let (cid:1):=f(cid:18)~g be a (cid:12)ner knot sequence i i=1;:::;n~+k (cid:18)~ =:::=(cid:18)~ =a<(cid:18)~ <:::<(cid:18)~ <b=(cid:18)~ =:::=(cid:18)~ (3.1) 1 k k+1 n~ n~+1 n~+k with grid spacing h= 1H. It is de(cid:12)ned such that (cid:18) =(cid:18)~ for i=k;:::;n+1 and 2 i 2i(cid:0)k 1((cid:18) +(cid:18) )=(cid:18)~ for i=k+1;:::;n+1. Then it holds 2 i(cid:0)1 i 2i(cid:0)k(cid:0)1 n~ =2n+1(cid:0)k: (3.2) The corresponding spline spaces are N and N with member functions N k;(cid:1) k;T i;k;(cid:1) andN ,respectively. Letnowthe obstaclefunction onthe (cid:12)negrid S~2N and i;k;T k;(cid:1) its approximationS 2N be expanded as k;T n~ n S~= c~ N =:~cT N ; S = c N =:cT N : (3.3) i i;k;(cid:1) k;(cid:1) i i;k;T k;T i=1 i=1 X X There is a natural prolongation operator p from N to N for B{splines N in k;T k;(cid:1) i;k;T terms of their re(cid:12)nement or mask coe(cid:14)cients [Bo, Sb]. In the special case H = 2h B{Spline{BasedMonotoneMultigridMethods 7 considered here the re(cid:12)nement relation is given by k N = a N (3.4) i;k;T j 2i(cid:0)k+j;k;(cid:1) j=0 X with the subdivision or mask coe(cid:14)cients k a :=21(cid:0)k for j =0;:::; k: (3.5) j j (cid:18) (cid:19) In Step 2 of Algorithm 2.1, we choose the restriction r as the adjoint of p, following [Ha]. However, for the obstacle function the restriction operator r cannot be used since it does not satisfy condition (2.6). 3.1. Monotone Coarse Grid Approximations. There is a vast amount of literature, see, e.g., [DV, Mv, Pi] especially from approximation theory, dealing with monotone approximations to a given function g. The function g^ is a monotone (or one{sided) lower approximation to g if g^(x) (cid:20) g(x) for all x 2 I. There the number n of degrees of freedom of the function g^ is chosen such that a given approximation accuracy can be reached. In contrast to these studies, the question here is di(cid:11)erent, since the number n of degrees of freedom is given by the mesh size H. Definition 3.1 (Monotone CoarseGrid Approximation). For knot sequences T and (cid:1) from (2.9) and (3.1), respectively, we call S 2 N a monotone lower coarse k;T grid approximation to S~2N if S(x)(cid:20)S~(x) holds for all x2I. k;(cid:1) For hat functions such approximations are constructed in [Ma, Ko1]. A cor- responding construction for higher{order functions has to our knowledge not been providedsofar. In viewof Lemma2.3weproposeheretocontrolB{splineexpansion coe(cid:14)cients. Theorem 3.2 (Monotone Coarse Grid Approximation). Let S~ 2 N be an k;(cid:1) upper obstacle with S~ = ~cT N for a given order k and the knot sequence (cid:1) from k;(cid:1) (3.1). Then S 2 N with S = cT N de(cid:12)ned on the knot sequence T from (2.9) k;T k;T is a monotone lower coarse grid approximation to S~ if the inequality system A c(cid:20)~c (3.6) k is satis(cid:12)ed. The two{slanted matrix A is de(cid:12)ned by k a a 0 ak(cid:0)1 ak(cid:0)3 ... 1 k k(cid:0)2 BB ak(cid:0)1 a0 CC B C BB ak a1 CC B .. C B . a C Ak :=BBB ...2 CCC2Rn~(cid:2)n BBB a ... CCC B k(cid:0)1 C B a C B k C BB ... a CC B 0 C B C @ a1 A with the subdivision coe(cid:14)cients a from (3.5) and has maximal rank. j 8 MarkusHoltzandAngelaKunoth Proof. The proof relies on the subdivision property (3.4) and on the nonnega- tivity of B{splines. We only consider the case k even as the other case is analogous. Substituting (3.4) into (3.3) and sortingaccordingto the basisfunctions N leads i;k;(cid:1) to n~ S(x) = a c +a c +:::+a c N (x) k(cid:0)1 (i+1)=2 k(cid:0)3 (i+3)=2 1 (i+k(cid:0)1)=2 i;k;(cid:1) i=1 iXodd(cid:0) (cid:1) n~(cid:0)1 + a c +a c +:::+a c N (x); k i=2 k(cid:0)2 (i+2)=2 0 (i+k)=2 i;k;(cid:1) i=2 iXeven(cid:0) (cid:1) where all c with j <1 or j >n are treated as zero. De(cid:12)ning the coe(cid:14)cients j c~ (cid:0) a c +a c +:::+a c ; if i is odd; i k(cid:0)1 (i+1)=2 k(cid:0)3 (i+3)=2 1 (i+k(cid:0)1)=2 d := i ( c~i(cid:0)(cid:0)akci=2+ak(cid:0)2c(i+2)=2+:::+a0c(i+k)=2 ; (cid:1) if i is even; which can be wr(cid:0)itten in compact matrix/vectorform as (cid:1) d =c~ (cid:0)(A c) (3.7) i i k i (involving the ith component of the vector A c), we obtain k n~ S~(x)(cid:0)S(x)= d N (x): (3.8) i i;k;(cid:1) i=1 X By Lemma 2.3 we have S~(x)(cid:0)S(x) (cid:21) 0 for all x 2 I, provided d (cid:21) 0 holds for all i i=1;:::;n~. By(3.7),weobtaintheinequalitysystem(3.6). SincetheB-splinesform bases for N and N , the matrix A has full rank for each k. k;T k;(cid:1) k Example 3.3. In the special case of continuous, piecewise linear functions (k =2), C1{smooth, piecewise quadratic (k =3) and C2{smooth, piecewise cubic (k =4) splines one has 1 3 1 0 1 1 1 0 1 3 1 2 2 A2 =BBBBBBBBBBBBBB .112.. 12 11212 21CCCCCCCCCCCCCC2R(2n(cid:0)1)(cid:2)n; A3 = 14BBBBBBBBBBBBB 31...13 31...133 1 CCCCCCCCCCCCC2R(2n(cid:0)2)(cid:2)n; @ 1A @ 1 3 A 4 4 0 1 6 1 1 B 4 4 C B C B 1 6 1 C A4 = 81BBBB ... ... CCCC2R(2n(cid:0)3)(cid:2)n: B 1 6 1 C B C B 4 4 C B C B 1 6 1 C B C @ 4 4 A B{Spline{BasedMonotoneMultigridMethods 9 3.2. Quasi{optimal Coarse Grid Approximations. Now we can immedi- ately derive a monotone lower coarse approximation. Proposition 3.4. The spline L :=qT N 2N with coe(cid:14)cients k k;T k;T q :=minfc~ ; :::; c~ g for i=1;:::;n (3.9) i 2i(cid:0)k 2i (leaving out c~ in the right hand side if j < 1 or j > n~) is a monotone lower coarse j grid approximation to S~=~cT N 2N . k;(cid:1) k;(cid:1) Proof. AsallrowsumsofA areequaltoone,thevectorq:=(q ;:::;q )T de(cid:12)ned k 1 n in(3.9)obviouslysatis(cid:12)estheinequalitysystemA q(cid:20)~csothattheassertiondirectly k follows from Theorem 3.2. Remark3.5. Inthespecialcasek =2,therestrictionoperatorr^: N !N , 2;(cid:1) 2;T S~7!L inducedby Proposition 3.4 coincides with therestrictionoperator from [Ma]. 2 As it is illustrated in Figure 3.1 and 3.2 for the cases k = 2 and k = 3, the approximationL canbefurther improvedin manycases. Thiswill bethe subjectof k the next subsections: there q is interpreted as a componentwise lower barrier for the B{spline coe(cid:14)cients c of the desired coarse grid approximation. Definition 3.6 (Quasi{optimalCoarseGrid Approximation). We call a mono- tone lower coarse grid approximation S =cT N to the spline S~=~cT N quasi{ k;T k;(cid:1) optimal if it is an improvement over L in the sense that c(cid:21)q holds with q de(cid:12)ned k in (3.9). 3.3. A Linear Optimization Problem. Aiming at improving the coarse grid approximation L from Proposition 3.4, we de(cid:12)ne an optimal monotone and quasi{ k optimalcoarsegridapproximationS =cT N toagiven S~=~cT N byformulat- k;T k;(cid:1) ing a linear optimization problem. We choose a target functional F which estimates k thesumofthedistancesfromapproximationtoobstacleonallcoarsegridpoints,i.e., F (c):= jS~((cid:18))(cid:0)S((cid:18))j: (3.10) k (cid:18)2T X Lemma 3.7. The function F de(cid:12)ned in (3.10) is a linear function Rn ! R of k the form F (c)=(cid:24)Tc+(cid:17) (3.11) k where (cid:24) :=(cid:0)ATs 2Rn; s :=((cid:12) ;(cid:13) ;(cid:12) ;:::)T 2Rn~ and (cid:17) :=sT~c2R: (3.12) k k k k k k k The values (cid:12) and (cid:13) can be computed explicitly: for odd k we have (cid:12) =(cid:13) = 1, and k k k k 2 for even k=2;4;6;8 the values are displayed in Table 3.1. k 2 4 6 8 (cid:12) 1 2 17 166 k 3 30 315 (cid:13) 0 1 13 149 k 3 30 315 Table 3.1 The values (cid:12)k and (cid:13)k fororders k=2;4;6;8.

Description:
cardinal higher order B–spline, system of linear inequalities, optimized coarse grid {holtz,kunoth}@ins.uni-bonn.de, www.ins.uni-bonn.de/∼kunoth.
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.