HAL Id: inria-00113372
https://hal.inria.fr/inria-00113372v2
Submitted on 13 Nov 2006
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Symplectic integrators in sub-Riemannian geometry: the Martinet case
Monique Chyba, Ernst Hairer, Gilles Vilmart
To cite this version:
Monique Chyba, Ernst Hairer, Gilles Vilmart. Symplectic integrators in sub-Riemannian geometry:
the Martinet case. [Research Report] RR-6017, INRIA. 2006, pp.10. �inria-00113372v2�
inria-00113372, version 2 - 13 Nov 2006
a p p o r t
d e r e c h e r c h e
ISSN0249-6399ISRNINRIA/RR--6017--FR+ENG
Thème NUM
INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE
Symplectic integrators in sub-Riemannian geometry:
the Martinet case
Monique Chyba — Ernst Hairer — Gilles Vilmart
N° 6017
Novembre 2006
Unité de recherche INRIA Rennes
IRISA, Campus universitaire de Beaulieu, 35042 Rennes Cedex (France)
Téléphone : +33 2 99 84 71 00 — Télécopie : +33 2 99 84 71 71
the Martinet ase
Monique Chyba
∗
,Ernst Hairer
†
, GillesVilmart
† ‡
ThèmeNUMSystèmesnumériques
ProjetIpso
Rapport dereherhe n°6017Novembre200610pages
Abstrat: Weompare theperformanes ofsympletiand non-sympletiintegratorsfortheomputation
of normalgeodesisand onjugate points in sub-Riemanniangeometry at the example of the Martinetase.
For this ase study we onsider rst the atmetri, and then a one parameter perturbation leading to non
integrablegeodesis. Fromouromputationswededuethatneartheabnormaldiretionsasympletimethod
ismuhmoreeientforthisoptimalontrolproblem. Theexplanationreliesonthetheoryofbakwarderror
analysis ingeometrinumerialintegration.
Key-words: sub-Riemanniangeometry,Martinet,abnormalgeodesi,sympletiintegrator,bakwarderror
analysis.
Thisworkwas partially supportedbythe Fonds NationalSuisse,projet No. 200020-109158 and bythe NationalSiene
FoundationgrantDMS-030641.
∗
UniversityofHawaii,DepartmentofMathematis,2565MCarthytheMall,Honolulu96822,Hawaii
†
UniversitédeGenève,Setiondemathématiques,2-4rueduLièvre,CP64,1211Genève4,Switzerland
‡
INRIARennes,projetIPSO,CampusdeBeaulieu,35042RennesCedex,Frane
le problème de Martinet
Résumé: Onomparelesperformanesd'intégrateurssympletiquesetnonsympletiquespourlealulde
géodésiquesnormales et depointsonjuguésdans unexemplesous-riemannien, le problèmede Martinet. On
étudie leproblèmed'abordave unemétrique plate,puisave une perturbation àunparamètreonduisantà
des géodésiquesnon intégrables. Deetteétude,ondéduit queprohedesdiretionsanormales, uneméthode
sympletiqueestbienpluseaepoureproblèmedeontrleoptimal. L'expliationreposesurlathéoriede
l'analyse rétrogradeenintégrationnumériquegéométrique.
Mots-lés : géométriesous-riemannienne,Martinet,geodésiqueanormale,intégrateursympletique,analyse
rétrograde
Introdution
This paper is a follow-up to a series of artiles that were published in the past deades, see [1, 2℄ and the
referenes therein. There, the authors provide an analyti study of the singularity of the sub-Riemannian
sphere in the Martinet ase. They omplementtheir analysis with numerial omputations to represent the
geodesisandthesphere,andtoloatetheonjugatepoints. Theintegratorusedforthese omputationsisan
expliit Runge-Kuttamethod oforder 5(4). Thegoalof thepresentpaperis to omparetheperformanesof
asympletiintegratorversusanon-sympletioneforthis optimalontrol problem. Tobemorepreise, let
(U,∆, g)beasub-Riemannianstruture whereU isanopenneighborhoodofR3,∆ adistribution ofonstant rank 2and g a Riemannian metri. When ∆ is aontatdistribution, there are noabnormalgeodesis,and anon-sympletiintegratorisaseient asasympletione. However,whenthe distributionis takenasthe
kerneloftheMartinetone-form,weshowthatasympletiintegratorismuhmoreeientfortheomputation
ofthenormalgeodesisandtheironjugatepointsneartheabnormaldiretions.
Bothproblems,theMartinetataseandanonintegrableperturbation,areintroduedinSet.1together
with the orresponding dierentialequations. Numerial experiments with an expliit RungeKutta method
and withthesympletiStörmerVerletmethod arepresentedin Set.2andillustratedwith gures. Closeto
abnormalgeodesis,theresultsarequitespetaular. Forarelativelylargestepsize,thesympletiintegrator
providesasolution withthe orretqualitative behaviorand asatisfatory auray, whilefor thesame step
size the non-sympleti integrator gives a ompletely wrong numerial solution with an inorret behavior,
partiularly for the non integrable ase. The explanation relies on the theory of bakward error analysis
(Set.3). Itis relatedto thegeometristruture oftheproblemanditssolutions.
1 A Martinet type sub-Riemannian struture
In this setion, we briey reall someresults of [1℄ for a sub-Riemannianstruture (U,∆, g). Here, U is an
openneighborhoodoftheorigininR3 withoordinatesq= (x, y, z),andgisaRiemannianmetriforwhiha
graduatednormalform,atorder0,isg= (1 +αy)dx2+ (1 +βx+γy)dy2. Thedistribution∆is generatedby
thetwovetoreldsF1= ∂x∂ +y22∂z∂ andF2=∂y∂ whih orrespondto∆ = kerω whereω =dz−y22dxis the
Martinetanonialone-form. Tothisdistributionweassoiatetheaneontrolsystem
˙
q=u1(t)F1(q) +u2(t)F2(q),
where u1(t), u2(t)aremeasurableboundedfuntionswhihatasontrols.
Weonsidertwoases,theMartinet ataseg =dx2+dy2, anintegrablesituation,andaoneparameter
perturbationg=dx2+ (1 +βx)2dy2forwhih thesetofgeodesisisnonintegrable.
1.1 Geodesis
It follows from the Pontryagin maximum priniple, see [1, 2℄, that the normal geodesis orresponding to
g=dx2+ (1 +βx)2dy2aresolutionsofanHamiltoniansystem
˙ q=∂H
∂p (q, p), p˙=−∂H
∂q (q, p), (1)
where q= (x, y, z)isthestate,p= (px, py, pz)istheadjointstate,andtheHamiltonianis
H(q, p) =1 2
px+pz
y2 2
2
+ p2y (1 +βx)2
.
Inotherwords,thenormalgeodesisaresolutionsofthefollowingequations:
˙
x = px+pz
y2 2
˙
y = py
(1 +βx)2
˙
z =
px+pzy2 2
y2 2
˙
px = β p2y (1 +βx)3
˙
py = −
px+pzy2 2
pzy
˙
pz = 0.
(2)
Notiethatthevariableszandpzdonotinuenetheotherequations(exeptviatheinitialvaluepz(0)),so
thatweareatuallyonfrontedwithaHamiltoniansystemindimensionfour. FortheMartinetatase(β= 0),
theinterestingdynamistakesplaein thetwo-dimensionalspaeofoordinates(y, py). TheHamiltonianis
H(y, py) = p2y 2 +1
2
px+pz
y2 2
2 ,
where pxandpz havetobeonsideredasonstants. Thisisaone-degreeoffreedommehanialsystemwitha
quartipotential. Forpx<0< pz,theHamiltonianH(y, py)hastwoloalminimaat(y=±p
−2px/pz, py= 0),
whihorrespondtostationarypointsofthevetoreld. Inthisase,theorigin(y= 0, py= 0)isasaddlepoint.
Whereasnormalgeodesisorrespondtoosillatingmotion,itisshownin[1,2℄thattheabnormalgeodesis
are thelines z=z0 ontainedin the planey = 0. Fortheonsidered metris,theabnormalgeodesisanbe
obtained asprojetions of normal geodesis, we say that they are not stritly abnormal. In [2℄, the authors
introdue a geometri framework to analyze the singularities of the sphere in the abnormal diretion when
β 6= 0. Seealso[3,4℄forapreisedesriptionoftheroleoftheabnormalgeodesisinsub-Riemanniangeometry in the generalnon-integrablease,i.e., when theabnormalgeodesisanbestrit. Themajorresultof these
papersistheproofthatthesub-Riemanniansphereis notsub-analytibeauseoftheabnormalgeodesis.
Interestingphenomenaarisewhen thenormalgeodesisareloseto theseparatriesonnetingthe saddle
point. Therefore,weshallonsiderinSet.2theomputationofnormalgeodesiswithy(0) = 0andpy(0)>0
but small.
1.2 Conjugate points
FortheHamiltoniansystem(1) weonsidertheexponentialmapping
expq0,t:p07−→q(t, q0, p0)
whih, for xed q0 ∈ R3, is the projetionq(t, q0, p0) onto the state spae of the solution of (1) starting at t= 0from(q0, p0). Followingthedenition in[1℄wesaythatthepointq1 isonjugateto q0 alongq(t)ifthere
exists (p0, t1), t1 >0, suh that q(t) = expq0,t(p0) with q1 = expq0,t1(p0), and themappingexpq0,t1 isnotan
immersion at p0. We say that q1 is the rst onjugate point if t1 is minimal. Firstonjugate points play a
majorrolewhenstudyingoptimalontrolproblemssineitisawellknownresultthatageodesiisnotoptimal
beyondtherstonjugatepoint.
For the numerial omputation of the rst onjugatepoint, weompute the solution of the Hamiltonian
system(1)togetherwithitsvariationalequation,
˙
y=J−1∇H(y), Ψ =˙ J−1∇2H(y)Ψ. (3)
Here,y= (q, p)andJ istheanonialmatrixforHamiltoniansystems. Itanbeshownthat forRunge-Kutta methods, thederivativeofthenumerialsolutionwithrespetto theinitial value,Ψn=∂yn/∂y0,istheresult
ofthesamenumerialintegratorappliedtotheaugmentedsystem(3),see[5,LemmaVI.4.1℄. Here,thematrix
Ψ =
∂q/∂q0 ∂q/∂p0
∂p/∂q0 ∂p/∂p0
hasdimension6×6. Theonjugatepointsareobtainedwhen∂q/∂p0 beomessingular,i.e.,det(∂q/∂p0) = 0.
2 Comparison of sympleti and non-sympleti integrators
For the numerial integration of the Hamiltonian system (1), where we rewrite
∂H
∂q(q, p) = Hq(q, p) and
∂H
∂p(q, p) =Hp(q, p),weonsidertwointegratorsofthesameorder2:
• anon-sympleti,expliitRungeKuttadisretization,denoted rk2 (see[5,Set.II.1.1℄), qn+1/2 = qn+h
2Hp(qn, pn)
qn+1 = qn+hHp(qn+1/2, pn+1/2)
pn+1/2 = pn−h
2Hq(qn, pn)
pn+1 = pn−hHq(qn+1/2, pn+1/2)
(4)
• thesympletiStörmerVerletsheme(seee.g. [5,Set.VI.3℄),
pn+1/2 = pn−h
2Hq(qn, pn+1/2) qn+1 = qn+h
2
Hp(qn, pn+1/2) +Hp(qn+1, pn+1/2)
(5)
pn+1 = pn+1/2−h
2Hq(qn+1, pn+1/2)
where qn = (xn, yn, zn)andpn= (px,n, py,n, pz,n). Here,qn≈q(nh), pn≈p(nh)andhis thestepsize.
Fortheomputationoftheonjugatepoints,weapplythenumerialmethodstothevariationalequation(3).
Notie that only the partial derivatives with respet to p0 haveto be omputed. Conjugate points are then
deteted whendet(∂qn/∂p0)hangessign. We approximate themby linearinterpolationwhih introduesan errorofsizeO(h2). Thisisomparabletotheaurayofthehosenintegratorswhiharebothofseondorder.
Remark 2.1 The StörmerVerlet sheme (5) is impliit in general. A few xed point iterations yield the
numerialsolutionwiththedesiredauray. NotiehoweverthatthemethodbeomesexpliitintheMartinet
at aseβ = 0. One simplyhas to omputethe omponentsin a suitableorder, forinstane px,n+1, pz,n+1,
py,n+1/2,yn+1,xn+1,zn+1,py,n+1.
2.1 Martinet at ase
Weonsiderrsttheataseβ = 0intheHamiltoniansystem(2). Asinitialvalueswehoose(f. [1℄)
x(0) =y(0) =z(0) = 0, px(0) = cosθ0, py(0) = sinθ0, pz(0) = 10, where θ0=π−10−3, (6)
−6 −4 −2
−.5 .5
−6 −4 −2
−.5 .5
−6 −4 −2
−.5 .5
−6 −4 −2
−.5 .5
−6 −4 −2
−.5 .5
−6 −4 −2
−.5 .5 x
y
non-sympletimethod,h= 0.1
x
y
StörmerVerlet,h= 0.1
x
y
non-sympletimethod,h= 0.05
x
y
StörmerVerlet,h= 0.05
x
y
non-sympletimethod,h= 0.02
x
y
StörmerVerlet,h= 0.02
Figure1: Trajetoriesin the(x, y)-planefortheataseβ= 0.
sothatwestartlosetoanabnormalgeodesis,andweintegratethesystemovertheinterval[0,9].
Figure1displaystheprojetionontothe(x, y)-planeofthenumerialsolutionobtainedwithdierentstep
sizeshbythetwointegrators. Theinitialvalueisattheorigin,andthenalstateisindiatedbyatriangle. The irlesrepresenttherstonjugatepointdetetedalongthenumerialsolution,whilethestarsgivetheposition
oftherstonjugatepointontheexatsolutionoftheproblem. Thereisanenormousdierenebetweenthetwo
numerial integrators. The sympleti (StörmerVerlet) method (5) providesa qualitatively orret solution
already with a large step size h = 0.1, and it gives an exellent approximation for step sizes smaller than
h = 0.05. On the other hand,the non-sympleti, expliitRungeKutta method (4) givesompletely wrong results, andstepsizes smallerthan10−3 areneededto provide anaeptablesolution. An explanationof the dierentbehaviorof thetwointegratorswillbegivenin Set.3below.
−.5 .5
−1 1
.5
−1 1
y py
non-sympletimethod h= 0.05
y py
StörmerVerlet h= 0.05
zoom×100
Figure2: Phaseportraitsinthe(y, py)-planefortheataseβ = 0.
As notied in Set.1, the normal geodesis in the at ase are determined by a one-degree of freedom
Hamiltonian systemin thevariables y and py. Wethereforeshowin Figure2the projetionontothe (y, py)-
spae of the solutions previouslyomputed with step size h = 0.05. The exat solution starts at (0,sinθ0)
above the saddle point, turns around the positive stationary point, rossesthe py-axisat (0,−sinθ0), turns
around the negative stationary point, and then ontinues periodially. The numerial approximation by the
non-sympleti method overs morethan one and a half periods, whereas the StörmerVerlet and the exat
solutionoverlessthanoneperiodforthetimeinterval[0,9]. Sinetheonjugatepointisnotverysensiblewith
respetto perturbationsin the initialvalue forpy, the(y, py)oordinates ofthe onjugatepoint obtainedby
thenon-sympletiintegratorareratheraurate,buttheorrespondingintegrationtimeisompletelywrong.
Table 1 lists the onjugate time obtained with the two integrators using various step sizes. There is a
signiantdierene between thetwo methods. We ansee that with theStörmerVerlet method (5) a step
sizeoforderh= 10−2providesasolutionwith4orretdigits. Astepsizea100timessmallerisneededtoget
thesamepreisionwiththenon-sympletimethod.
2.2 Non integrable perturbation
Forournextnumerialexperimentwehoosetheperturbationparameterβ=−10−4inthedierentialequation
(2). We onsider thesameinitial valuesand the sameintegrationintervalasin Set.2.1. The exatsolution
isnolongerperiodiand, duetothefat thatβ ishosennegative,itsprojetionontothe(y, py)-spaeslowly
spiralsinwardsaroundthepositivestationarypoint(seerightpitureinFigure4).
Figures3and4andTable1displaythenumerialresultsobtainedbythetwointegratorsforthedierential
equation (2) with β = −10−4. The interpretation of the symbols (triangles, irles, and stars) is the same as before. Theexellentbehaviorof thesympleti integratoris evenmorespetaular thanin theatase,
and thepitures obtainedfor the StörmerVerlet method agree extremely well with the exatsolution. The
Table1: Aurayfortherstonjugatetime.
Martinetatase
h rk2 Verlet
10−1 4.504945 8.504716 10−2 6.748262 8.416622 10−3 8.360340 8.416412 10−4 8.416349 8.416410
exatsolution: t1≈8.416409
Nonintegrablesituation
h rk2 Verlet
10−1 4.511294 4.883832 10−2 7.380322 4.877056 10−3 4.877183 4.876998 10−4 4.876997 4.876997
exatsolution: t1≈4.876997
−6 −4 −2
−.5 .5
−6 −4 −2
−.5 .5
−6 −4 −2
−.5 .5
−6 −4 −2
−.5 .5 x
y
non-sympletimethod,h= 0.05
x
y
non-sympletimethod,h= 0.02
x
y
non-sympletimethod,h= 0.009
x
y
StörmerVerlet,h= 0.1
Figure3: Trajetoriesinthe(x, y)-planeforthenonintegrableaseβ =−10−4.
−.5 .5
−1 1
.5
−1 1
y py
non-sympletimethod h= 0.05
y py
StörmerVerlet h= 0.05
zoom×10
Figure 4: Phaseportraitsin the(y, py)-planeforthenonintegrableaseβ=−10−4.
non-sympleti method givesqualitativelywrong solutionsforstep sizes largerthanh= 0.01. Inthe (y, py)-
spaeitalternativelyspiralsaroundtherightandleftstationarypointswhereastheexatsolutionspiralsonly
aroundthepositivestationarypoint. InontrasttotheMartinetatase,theonjugatepointobtainedbythe
non-sympletimethodisherewrongalsointhe(y, py)-spae.
2.3 An asymptoti formula on the rst onjugate time in the Martinet at ase
Now that we have shown the eieny of sympletiintegrators, we an make more preise the asymptoti
behaviourstudiedin [1℄. Fortheinitialvaluesof(6)andβ = 0,onsidertheratio R= t1√pz
3K(k),
where t1 istherstonjugatetimeforthenormalgeodesi,andK(k)isanelliptiintegraloftherstkind, K(k) =
Z π/2 0
p 1
1−k2sin2udu, k= sin(θ0/2).
Bystudyinganalytisolutionsforthenormalgeodesis,itisprovedin[1℄thatthisratiosatisestheinequality
2/3≤R≤1. It followsfromaresalingoftheequations(2)thatRis independentofpz.
InFigure5,werepresentthevaluesof1−R asafuntion ofε=π−θ0, forvariousinitialvaluesθ0. The
numerialresultsindiate thattheratioRdependsonθ0,andR−→1− slowlyforθ0−→π−.