ebook img

Lie group integrators PDF

0.25 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 Lie group integrators

Lie group integrators BrynjulfOwren 6 1 0 2 n a J 8 1 Abstract In this survey we discuss a wide variety of aspects related to Lie group integrators.Thesenumericalintegrationschemesfordifferentialequationsonman- ] A ifolds have been studied in a general and systematic manner since the 1990s and N theactivityhassincethenbranchedoutinseveraldifferentsubareas,focussingboth . ontheoreticalandpracticalissues.Fromtwoalternativesetups,usingeitherframes h or Lie groupactions on a manifold,we here introducethe most importantclasses t a ofschemesusedtointegratenonlinearordinarydifferentialequationsonLiegroups m andmanifolds.Wedescribeanumberofdifferentapplicationswherethereisanatu- [ ralactionbyaLiegrouponamanifoldsuchthatourintegratorscanbeimplemented. Anissuewhichisnotwellunderstoodistheroleofisotropyandhowitaffectsthe 1 v behaviourofthenumericalmethods.TheordertheoryofnumericalLiegroupinte- 4 gratorshasbecomeanadvancedsubtopicinitsownright,andherewegiveabrief 6 introduction on a somewhat elementary level. Finally, we shall discuss Lie group 6 integrators having the property that they preserve a symplectic structure or a first 4 integral. 0 . 1 0 6 1 Introduction 1 : v [19].LeonhardEulerisundoubtedlyoneofthemostaccomplishedmathematicians i X ofalltimes,andthemodernthemeNumericalmethodsforLiegroupscanbetraced r backtoEulerinmorethanonesense.Indeed,thesimplestandpossiblymostlyused a numerical approximation method for ordinary differential equations was first de- scribed in Euler’swork InstitutionumCalculi Integralis(1768),VolumenPrimum, Ch VII) and bears the name Euler’s method.And undoubtedly,the most used test BrynjulfOwren NorwegianUniversityofScienceandTechnology,AlfredGetzvei1,N-7491Trondheim,Norway, e-mail:[email protected] 1 2 BrynjulfOwren caseforLiegroupintegratorsistheEuler’sfreerigidbodysystem,whichwasde- rivedinhisamazingtreatiseonMechanicsin1736. Intheliteratureonstructuralmechanics,Liegroupintegratorshavebeenaround fora longtime, butgeneral,systematic studiesofnumericalintegratorsfor differ- entialequationsonLiegroupsandhomogeneousmanifoldsbeganasrecentasthe 1990s.SomenotableearlycontributionswerethoseofCrouchandGrossman[24] andLewisandSimo[40].A seriesofpapersbyMunthe-Kaas[49, 50,51] caused anincreasedactivityfromthelateninetieswhenalargenumberofpapersappeared overashortperiodoftime.Manyoftheseearlyresultsweresummarisedinasurvey paperbyIserlesetal.[34].TheworkonLiegroupintegratorshasbeeninspiredby many subfields of mathematics. Notably, the study of order conditions and back- ward error analysis uses results from algebraic combinatorics,Hopf algebras, and hasmorerecentlybeenconnectedtopostLiealgebrasbyMunthe-Kaasandcoau- thors,seee.g.[52].Inconnectionwiththesearchforinexpensivecoordinaterepre- sentations of Lie groupsas well as their tangentmaps, the classical theory of Lie algebrashas been put to use in many differentways. The theory of free Lie alge- bras [63] has been used to find optimal truncationsof commutatorexpansionsfor generalLiealgebras,seee.g.[53].Forcoordinatemapstakingadvantageproperties ofaparticularLiealgebra,toolssuchasrootspacedecomposition[61]andgener- alisedpolardecompositions[37]havebeenapplied.Alsothereisofcourseastrong connectionbetween numericalmethodsfor Lie groupsand the area of Geometric Mechanics.Thisconnectionisoftenusedinthesetuporformulationofdifferential equationsinLiegroupsorhomogeneousmanifoldswhereitprovidesanaturalway of choosing a group action, and in order to construct Lie group integratorswhich aresymplecticorconserveaparticularfirstintegral. Inthispaperwe shalldiscussseveralaspectsofLiegroupintegrators,we shall however not attempt to be complete. Important subjects related to Lie group in- tegrators not covered here include the case of linear differential equations in Lie groupsandthemethodsbasedonMagnusexpansions,Ferexpansions,andZassen- haussplittingschemes. These aremethodsthatcouldfit wellinto a surveyonLie groupintegrators,butforinformationonthesetopicswereferthereadertoexcel- lentexpositionssuchas[4,33].Anothertopicweleaveouthereisthatofstochastic Liegroupintegrators,seee.g.[42].Weshallalsofocusonthemethodsandthethe- ory behind them rather than particularapplications, of which there are many.The interestedreadermaycheckoutthereferences[8,9,18,35,65,66]. In the next section we shall define a compact setup of notation for differential equationsondifferentialmanifoldswith a Lie groupaction.Thenin Section3 we discusssomeofthemostimportantclassesofLiegroupintegratorsandgiveafew examplesofmethods.Section4brieflytreatsaselectionofgroupactionswhichare interestinginapplications.TheninSection5weshalladdresstheissueofisotropy inLiegroupintegrators,inparticularhowthefreedomofferedbytheisotropygroup caneither beusedto reducethe computationalcostofthe integratororbe usedto improvethe qualityofthe solution.We take ourownlookat ordertheoryandex- pansionsintermsofageneralisedformofB-seriesinSection6.Finally,insections Liegroupintegrators 3 7 and 8 we consider Lie groupintegratorswhich preserve a symplectic form or a firstintegral. 2 The setup Let M be a differentiable manifold of dimension d <¥ and let the set of smooth vector fields on M be denoted X(M). Nearly everything we do in this paper is concerned with the approximation of the h-flow of a vector field F ∈X(M) for somesmallparameterhusuallycalledthestepsize.Inotherwords,weapproximate thesolutiontothedifferentialequation d y˙= y= F| , F ∈X(M). (1) dt y Crouch and Grossman [24] used a set of smooth frame vector fields E1,...,En , n ≥donM,assuming span(E1|x,E2|x,...,En |x)=TxM, foreachx∈M. Itcanbeassumedthattheframevectorfieldsarelinearlyindependentasderivations oftheringF(M)ofsmoothfunctionsonM,andwedenotetheirR-spanbyV.Any smoothvectorfieldF ∈X(M)canberepresentedbyn functions f ∈F(M) i n (cid:229) F| = f(x)E| (2) x i i x i=1 wherethe f arenotnecessarilyunique.Wethenhaveanaturalaffineconnection i (cid:209) G=(cid:229) F(g)E F i i i which is flat with constant torsion t ((cid:229) f E ,(cid:229) gE)=(cid:229) f g[E,E ]. For later, j j j i i i i,j j i i j we shall need the notion of a frozen vector field relative to the frame. The freeze operatorFr:M×X(M)→V isdefinedas (cid:229) Fr(x,F):=F = f(x)E x i i i We notethatthetorsioncanbe definedbyfreezingthevectorfieldsandthentake theLie-Jacobibracket,i.e. t (F,G)| =[Fr(x,F),Fr(x,G)] x Another setup is obtained by using a Lie group G acting transitively from the leftonM [51]. TheLie algebraof G is denotedg. AnyvectorfieldF cannowbe representedviaamap f :M→gandtheinfinitesimalactionr :g→X(M) 4 BrynjulfOwren d F| =r ◦f(x)| , r :g→X(M), r (x )| = exp(tx )·x (3) x x x dt (cid:12)t=0 (cid:12) Wenotethatthemap f isnotnecessarilyunique. (cid:12) (cid:12) 3 Types ofschemes There is now a large variety of numericalintegrationschemes available, typically formulated with either of the setups of the previous section. In what follows, we assumethatafinitedimensionalLiegroupGactstransitivelyonamanifoldMand theLiealgebraofGisdenotedg. 3.1 Schemes ofMunthe-Kaas type Using the secondsetup, a powerfulway of derivingnumericalintegratorsdevised byMunthe-Kaas[51]isto 1. InaneighborhoodU⊂gof0introducealocaldiffeomorphismy :U→G,such that y (0)=1∈G, T y =Id 0 g 2. Observe that the map l (v)=y (v)·y is surjective on a neighborhoodof the y0 0 initialvaluey ∈M. 0 3. ComputethepullbackofthevectorfieldF =r ◦f alongl y0 4. ApplyonestepofastandardnumericalintegratortotheresultingproblemonU 5. MaptheobtainedapproximationbacktoMbyl y0 Eventhoughtheideaisverysimple,thereareseveraldifficultiesthatneedtobe resolved in order to obtain fast and accurate integrationschemes from this proce- dure. Observethatthederivativeofy canbetrivialisedbyrightmultiplicationofthe Liegroup,suchthat Tuy =TRy (u)◦dy u, dy u:g→g. WiththisinmindwesetouttocharacterisethevectorfieldonU⊂g,thisisasimple generalisationofaresultin[51] Lemma1.LetMbeasmoothmanifoldwithaleftLiegroupactionL :G×M→M andletgbetheLiealgebraofG.Lety :g→Gbeasmoothmap,y (0)=1.Fixa pointm∈M,andsetL =L (·,m)sothat m r (x )| =T L (x ) m 1 m Liegroupintegrators 5 SupposeF ∈X(M)isoftheform F| =r (x (m))| , forsomex :M→g. m m Definel (u)=L (y (u),m).ThenthereisanopensetU ⊆gcontaining0suchthat m thevectorfieldh ∈X(U)definedas h | =dy −1(x ◦l (u)) u u m isl -relatedtoF. m The original proof in [51] where y =exp was adapted to general coordinate maps in [61]. One step of a Lie group integrator is obtained just by applying a classical integrator, such as a Runge–Kutta method, to the corresponding locally definedvectorfieldh ong.ARunge–Kuttamethodwithcoefficients(A,b)applied totheproblemy˙=h (y)inalinearspaceisofthefollowingform s (cid:229) Y =y +h a k , r=1,...,s, r 0 rj j j=1 k = h | , r=1,...,s, r Yr s (cid:229) y =y +h b k 1 0 r r r=1 Here y is the initial value, h is the step size, and y ≈y(t +h) is the approxi- 0 1 0 matesolutionattimet +h.Theparameters≥1iscalledthenumberofstagesof 0 the method.If the matrix A=(a ) is strictly lowertriangular,then the methodis rj calledexplicit.Theapplicationofsuchamethodtothetransformedvectorfieldof Lemma1cabewrittenoutasfollows s u =h(cid:229) a k˜ , k =x ◦l (u ), k˜ =dy −1(k ), r=1,...,s, (4) r ij j r y0 r r ur r j=1 s v=h(cid:229) b k˜ y =l (v). (5) r r 1 y0 r=1 AmajoradvantageofthisapproachcomparedtoothertypesofLiegroupschemes isitsinterpretationasasmoothchangeofvariableswhichcausestheconvergence ordertobe(atleast)preservedfromtheunderlyingclassicalintegrator.Asweshall see later, one generally needs to take into account additional order conditions to accountforthefactthatthephasespaceisnotalinearspace. 3.1.1 Choosingtheexponentialmapascoordinates,y =exp ThefirstpapersbyMunthe-KaasonLiegroupintegrators[49,50,51]allusedy = expascoordinateson the Liegroup.In thiscase, thereare severaldifficultiesthat 6 BrynjulfOwren need to be addressed in order to obtain efficient implementationsof the methods. Oneisthe computationofthe exponentialmapitself. Formatrixgroups,thereare a largenumberofalgorithmsthatcanbeapplied,see e.g.[47, 48].In[15, 16]the authorsdevelopedapproximationstotheexponentialwhichexactlymapmatrixLie subalgebrastotheircorrespondingLiesubgroups.Anotherissuetobedealtwithis thedifferentialoftheexponentialmap dexp :=T R ◦T exp u exp(−u) exp(−u) u Lemma2.(Tangentmapofexp:g→G).Letu∈g,v∈T g≃g.Then u d T exp(v)= exp(u+sv)=TR ◦dexp v=dexp (v)·exp(u) u ds exp(u) u u (cid:12)s=0 (cid:12) where (cid:12) (cid:12) 1 ez−1 dexp (v)= exp(rad )(v)dr= (v) u u z Z0 (cid:12)z=adu (cid:12) (cid:12) (cid:12) Proof. Lety (t)=exp(t(u+sv))suchthat s d T exp(v)= y (1) u s ds (cid:12)s=0 (cid:12) Butfornowwedifferentiatewithrespecttot(cid:12)toobtain (cid:12) d y˙ := y (t)=(u+sv)y (t) (6) s s s dt Wealsonotethaty (t)=exp(tu)+O(s)ass→0.From(6)wethenget s y˙−uy =svetu+O(s2) s andtheintegratingfactore−tu yields d e−tuy =se−tuvetu+O(s2) s dt (cid:0) (cid:1) Integrating,andusingthaty (0)=Id,weget s t y (t)=etu+s eruve−ruetudr+O(s2) s 0 Z andso d 1 1 y (1)= eruve−rudreu= eradu(v)dreu s ds(cid:12)s=0 Z0 Z0 (cid:12) wherewehaveuse(cid:12)dthewell-knownidentityAd =exp(ad )inthelastequality. (cid:12) exp(u) u Formally,wecanwrite Liegroupintegrators 7 1 1 ez−1 eradu(v)dr= erz| (v)dr= (v) z=adu z Z0 Z0 (cid:12)z=adu (cid:12) (cid:12) Itisoftenusefultoconsiderthedexp-mapasaninfinites(cid:12)eriesofnestedcommu- tators 1 1 1 1 dexp (v)=(I+ ad + ad2+···)(v)=v+ [u,v]+ [u,[u,v]]+··· u 2! u 3! u 2 6 In(4)itistheinverseofdexpwhichisneeded.Notethatthefunction ez−1 f (z)= 1 z isentire,thismeansthatitsreciprocal z ez−1 is analytic where f (z) 6=0. In particular this means that 1 has a converging 1 f1(z) Taylor series about z=0 in the open disk |z|<2p . This series expansion can be showntobe ¥ z =1− z +(cid:229) B2k z2k ez−1 2 (2k)! k=1 whereB aretheBernoullinumbers,thefirstfewofthemare:B = 1,B =− 1 , 2k 2 6 4 30 B = 1 ,B =− 1 ,B = 5 .Themap 6 42 8 30 10 66 v=dexp−1(w) (wheneverw=dexp (v)) u u isgivenpreciselyas z 1 B dexp−1(w)= (w)=w− [u,w]+ 2[u,[u,w]]+··· (7) u ez−1 2 2! (cid:12)z=adu (cid:12) (cid:12) Weobservefrom(4)thatonen(cid:12)eedstocomputedexp−1(k )andthateachu =O(h). ur r r Thismeansthatonemayapproximatetheseriesin(7)byafinitesum, dexpinv(u,w,m)=w−1[u,w]+(cid:229)m B2k ad2k(w). 2 (2k)! u k=1 Onehasdexpinv(u,w,m)∈gforeverym≥0andfurthermore dexp−1(k )−dexpinv(u ,k ,m)=O(h2m+1) ur r r r Aslongastheclassicalintegratorhasorder p≤2m+1,theresultingMunthe-Kaas scheme will also have order p. Thereexists howevera clever way to substantially reducethenumberofcommutatorsthatneedtobecomputedineachstep.Munthe– 8 BrynjulfOwren KaasandOwren[53]realisedthatonecouldformlinearcombinationsofthestage derivativesk˜ in(4)suchthat r r Q = (cid:229) s k˜ =O(hqr) r r,j j j=1 forq aslargeaspossibleforeachr.ThenthesenewquantitiesQ wereeachgiven r r thegradeq andoneconsideredthegradedfreeLiealgebrabasedonthisset.The r resultwasasignificantreductioninthenumberofcommutatorsneeded.AlsoCasas andOwren[13]providedawaytoorganisethecommutatorcalculationstoreduce evenfurtherthecomputationalcost.HereisaRunge–KuttaMunthe–Kaasmethod oforderfourwithfourstagesandminimalsetofcommutators k =hf(y ), 1 0 k =hf(exp(1k )·y ), 2 2 1 0 k =hf(exp(1k −1[k ,k ])·y ), 3 2 2 8 1 2 0 k =hf(exp(k )·y ), 4 3 0 y =exp(1(k +2k +2k +k −1[k ,k ]))·y . 1 6 1 2 3 4 2 1 4 0 Forlaterreference,wealsogivetheLie-Eulermethod,afirstorderLiegroupinte- gratorgeneralisingtheclassicalEulermethod y =exp(hf(y ))·y (8) 1 0 0 3.1.2 Canonicalcoordinatesofthesecondkind Theexponentialmapisgenerallyexpensivetocomputeexactly.FormatrixLieal- gebrasg⊆gl(n,F)whereFiseitherRorC,standardsoftwareforcomputingexp numericallyhasacomputationalcostofn3totheleadingorder,andtheconstantin frontofn3 maybeaslargeas20−30.Another,yetcompletelygeneralalternative totheexponentialfunctionisconstructedasfollows:Fixabasisforg,saye ,...,e 1 d andconsiderthemap y :v e +···+v e 7→exp(v e )·exp(v e )···exp(v e ) (9) 1 1 d d 1 1 2 2 d d Although it might seem unnaturalto replace one exponentialby many, one needs to keep in mind that if the basis can be chosen such that its exponential can be computedexplicitly,itmaystillbeanefficientmethod.Forinstance,inthegeneral linearmatrixLiealgebragl(mF)onemayusethebasistobee =eeT wheree is ij i j i theithcanonicalunitvectorinRn.Then exp(a e )=1+a e , i6= j, exp(a e )=1+(ea −1)e ij ij ii ii Liegroupintegrators 9 So computing(9) takes approximatelynd operationswhich is much cheaper than computingtheexponentialofageneralmatrix. Thedifficultylies howeverin computingthe map dy −1 in an efficientmanner. u Amethodforthiswasdevelopedin[61].Themethodologyisslightlydifferentfor solvableandsemisimpleLiealgebras.Wehereoutlinethemainidea,fordetailswe refertotheoriginalpaper[61].Differentiate(9)toobtain d dy u(v)=v1e1+(cid:229) vi Adeu1e1◦···◦Adeui−1ei−1(ei) i=2 Themainideaistofindanequivalentexpressionwhichisacompositionofcheaply invertibleoperators.Forthis,weintroduceaprojectorontothespanofthelastd−k basisvectorsasfollows d d (cid:229) (cid:229) P : ve 7→ ve k i i i i i=1 i=k+1 where we let P and P equal the identity operator and zero operator on g re- 0 d spectively. We may now define a modified version of the Ad-operator, for any u=(cid:229) ue ∈g,let i i Adeukek =(Id−Pk)+AdeukekPk Thisisalinearoperatorwhichactsastheidentityoperatoronbasisvectorse, i≤k, i c andonbasisvectorsei, i≥kitcoincideswithAdeukek Definition1. An ordered basis (e ,...,e ) is called and admissible ordered basis 1 d (AOB)if,foreachu=(cid:229) u e ∈gandforeachi=1,...,d−1,wehave j j Adeu1e1◦···◦AdeuieiPi=Adeu1e1◦···◦AdeuieiPi (10) Thisdefinitionisexactlywhatisneededtowcritedy uasaccompositionofoperators Proposition1.Ifthebasis(e ,...,e )isanAOB,then 1 d dy u=Adeu1e1◦···Adeuded Anotherimportantsimplificationcacnbeobtainecdif anabeliansubalgebrahof di- mensiond−d∗ canbeidentified.Inthiscasetheorderedbasiscanbechosensuch thath=span(ed∗+1,...,ed). Then Adeuiei|h fori>d∗ is the identityoperatorand thereforeAdeuiei istheidentityoperatoronallofg.Summarizing,wehavethefol- lowingexpression c dy u−1=Ad−eu1d∗ed∗ ◦···◦Ad−eu11e1 Choosing typically a basis consisting of nilpotent elements, the inversion of each c c Adeuiei canbedonecheaplybymakinguseoftheformula c Adeuiei =1+(cid:229)K ukki!adek, adKei+1=0. k=1 10 BrynjulfOwren For choosing the basis one may, for semisimple Lie algebras, use a basis known as the Chevalley basis. This arises from the root space decomposition of the Lie algebra g=h⊕ ga . (11) a ∈F a HereF isthesetofroots,andga istheone-dimensionalsubspaceofgcorrespond- ingtotheroota ∈h∗,seee.g.Humphreys[32].histhemaximaltoralsubalgebra of g and it is abelian. In the previous notation, the number of roots is d∗ and the dimension of h is d−d∗. The following result whose proof can be found in [61] providesatoolfordeterminingwhetheranorderedChevalleybasisisanAOB. Theorem1.Let{b ,...,b }, d =d−ℓ,bethesetofrootsF forasemisimpleLie 1 d∗ ∗ algebrag.SupposethataChevalleybasisisorderedas (eb1,...,ebd∗,h1,...,hℓ) whereebi ∈gbi,and(h1,...,hℓ)isabasisforh.SuchanorderedbasisisanAOBif kb +b =b , m<i<s≤d ,k∈N ⇒ b +b 6∈F , m<n≤i−1. (12) i s m ∗ m n HereF =F ∪{0}. Example1.Asanexample,weconsiderA =sl(ℓ+1,C),commonlyrealizedasthe ℓ setof(ℓ+1)×(ℓ+1)-matriceswithvanishingtrace.Themaximaltoralsubalgebra isthenthesetofdiagonalmatricesinsl(ℓ+1,C).Thepositiverootsaredenoted {b ,1≤i≤ j≤ℓ}. i,j Lettinge betheithcanonicalunitvectorinCℓ+1,therootspacecorrespondingto i b hasabasisvector i,j eieTj+1∈gbi,j, 1≤i≤ j≤ℓ whereasthenegativerootsareassociatedtothebasisvectors ej+1eiT ∈g−bi,j. As a basis for h, one may choose the matriceseeT −e eT , 1≤i≤ℓ. The re- i i i+1 i+1 mainingdifficultynowistochooseanorderingofthebasissothatanAOBresults. Asindicatedearlier,thebasisforhmaybeorderedasthelastones,i.e.withindices rangingfromd∗+1=ℓ2+ℓ+1tod=ℓ2+2ℓ.Withtheconventioneb ∈gb , b ∈F , h=span(e ,...,e ),let h1 hℓ B=(ebi1,j1,...,ebim,jm,e−bi1,j1,...,e−bim,jm,eh1,...,ehℓ), wherei ≤i ≤···≤i andm=ℓ(ℓ+1)/2.OnecanthenprovebyusingTheorem1 1 2 m thatBisanAOB.

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.