Article
Reference
Optimized domain decomposition methods for the spherical Laplacian
LOISEL, Sébastien, et al.
LOISEL, Sébastien, et al . Optimized domain decomposition methods for the spherical Laplacian. SIAM Journal on Numerical Analysis , 2010, vol. 48, no. 2, p. 524–551
DOI : 10.1137/080727014
Available at:
http://archive-ouverte.unige.ch/unige:6259
Disclaimer: layout of this document may differ from the published version.
Laplaian
S. Loisel
∗
J. Cté
†
M. J.Gander
∗
L.Laayouni
‡
A. Qaddouri
†
September 21, 2009
Abstrat
The Shwarz iteration deomposes a boundary value problem overa large domain
Ω
into smallersubproblems by iteratively solving Dirihlet problems on a over
Ω 1 , ..., Ω 𝑝
ofΩ
. In this paper, wedisussalternatetransmissiononditionsthatleadtofasteronvergenefortheLaplaianonthesphere
Ω
. We look at Robin onditions, seond order tangential onditions and a disretized version of anoptimalbutnonloaloperator.
1 Introdution
AttheheartofnumerialweatherpreditionalgorithmslieaLaplaeandpositivedeniteHelmholtzprob-
lemsonthesphere[35℄. Reently,therehasbeeninterestinusingniteelements[5℄anddomaindeomposi-
tionmethods[4,26℄. TheShwarziteration[23,24,25℄anditsoptimizedvariantshavebeenverysuessful
andthesubjetofmuhreentresearh. WenowoutlinethehistoryofOptimizedShwarzMethods(OSM)
andreferto[14℄forfurtherdetails,aswellasaompletederivationforHelmholtzandLaplaeproblems in
theplane.
The optimizedShwarzmethod was introduedin [2,31, 32℄under various names. SinetheOSMhas
oneormorefreeparameters,whihneedtobehosenarefullytoguaranteethebestonvergenerate,there
followedaneorttondthebestpossibleparameterhoies[13,21,20℄. Despitethesepositivedevelopments,
aproofofonvergeneforageneralsituationhasprovenelusive[22,27℄. Onewayofobtainingonvergene
istodenearelaxationofthemethod[7℄.
TheusualmethodforoptimizingthefreeparametersistotakeaFouriertransformofthepartialdier-
entialequation, obtaininganexpliitreurrenerelationfortheiteration. This works onlyinspeialases
(e.g., thedomain is aretangle and the dierentialoperator is the Laplaianwith homogeneousDirihlet
onditions). In addition to theLaplae and Helmholtzproblems [19℄, the method an beused for various
otheranonialproblems(f. [9,18,17,29,28℄foronvetion-diusionproblems,[16℄forthewaveequation,
[8℄ for uid dynamis, [33℄ for the shallow water equation). The urrent paper is a ontinuation of this
researh.
OurpaperprovidesanOSMforthespherialLaplaian,adierentialoperatorwhihhadpreviouslynot
beenanalyzed,usingFourieranalysis. However,webrieymentionthattherehasbeenmuhreentresearh
onvariousother aspetsoftheOSM.Forinstane, domainswithornersan giverisetosingularities,and
this isanalyzed in [3℄. Inasimilar vein, dierentialoperators with disontinuousoeientsoften leadto
ill-onditionedproblems, and anOSMforthis situation isprovidedin [12℄. A ompletelydierent issueis
to phrasethe OSMin analgebrai way. In atypialappliation, one hasaode to omputethe stiness
matrix
𝐴
,butonewouldideallyprefernottohavetorewritetheentireodeinordertoimplementanOSM.Inthatvein,[34℄providesaframeworkforexpressingtheOSMin thelanguageofmatries.
∗
Setion de Mathématiques, Université de Genéve, CP 64, CH-1211 Geneva, Switzerland (sloiselgmail.om,
gandermath.unige.h).
†
Reherheenprévisionnumérique,EnvironnementCanada,2121Route Transanadienne,Dorval, Québe,Canada H9P
1J3(jean.otee.g.a, abdessamad.qaddourie. g. a)
‡
Shool of Siene and Engineering Al AkhawaynUniversity Avenue Hassan II, 53000 P.O. Box 2165, Ifrane, Moroo
(L.Laayouniaui.ma)
deompositionmethods areoftennaturallyparallel,andheneanbeusedonlargelustersandsuperom-
puters. Thealgorithmthat runsoneahproessorisasmall,loalversionoftheglobalproblem,soproven
sequentialodesan beadapted and beome loal solverswhih will thenrun in parallel onsuh lusters.
Insuhasenario,animportantquestionisthat ofsaling.
In the simplest saling senario, one xes the number
𝑝
of subdomains, and one piks subdomainsΩ 1 , . . . , Ω 𝑝
insuhawaythattheoverlapbetweenadjaentsubdomainsisexatlyoneelementthik. Inthissituation,itiswellknown[36℄thatthe(1-level)Shwarzmethodhasaonvergenefatorof
𝜌 𝐶𝑆 = 1 − 𝑂(ℎ)
.Inotherwords,ateahstepoftheiterativemethod,theerrorismultipliedbytheoeient
𝜌 𝐶𝑆 < 1
. Beausethis oeientdepends on
ℎ
, and tends to 1whenℎ
tendsto zero,wesee that moreand moreiterationswillbeneededtoreahagiventoleraneaswemake
ℎ
smaller. It ispreferableto ndanalgorithmwhoseonvergenefatoreitherdoesnotapproah
1
,oratleastapproahes1
nottooquiklywhenℎ
tendstozero.In this ontext, themain advantageof theOptimized Shwarzmethods is that they havemuh better
onvergenefatorsthanthelassialShwarzmethod. Forexample,theone-sidedoverlappingOO0method
hasaonvergenefatorof
𝜌 𝑂𝑂0 = 1 − 𝑂(ℎ 1/3 )
,andtheoverlappingOO2methodhasaonvergenefator of𝜌 𝑂𝑂2 = 1 − 𝑂(ℎ 1/5 )
.Inadditiontooverlappingmethods,Lions[25℄proposedtouseRobintransmissiononditionsalongthe
interfae to obtain nonoverlappingdomain deomposition methods. There is no nonoverlapping lassial
Shwarzmethodwithwhihtoompare,buttheOO0andOO2methodsanbeusedinthenonoverlapping
ontext. Theonvergenefators are
1 − 𝑂(ℎ 1/2 )
and1 − 𝑂(ℎ 1/4 )
, forthenonoverlappingOO0 andOO2 methods,respetively.Although ouranalysis islimitedtolatitudinalsubdomains,wewillseein Setion5thatweobtainvery
good performane,evenwhensubdomainshavemorevariedshapes.
Wenotethat,ifthenumberofsubdomains
𝑝
inreasesasℎ
tendstozero,itisneessarytodesigna2-levelalgorithm to maintaingood performane. Ina 2-levelalgorithm, detailed informationsfrom neighbors are
ombinedwithlesspreise informationstoredonalow-resolutionoarsegrid. Thedesignand analysisof
anOptimizedShwarzMethodwithaoarsegridorretionisahallengingproblemwhihwillbeanalyzed
inanupomingpaper[10℄.
1.1 Our ontributions
Inthispaper,weintrodueimprovedtransmissionoperators fortheLaplaeproblemonthe unitsphere
Ω
in
ℝ 3
. We modify the Shwarz iteration by replaing theDirihlet onditions onthe interfaes by Robinor seondordertangentialonditions. Thisallowsusto signiantlyimprovetheonvergenefatorofthe
iteration. Toanalyzetheonvergenerateandoptimizetheoeients,weusetheFouriertransformonthe
sphere. Inadditionto theusual Robinand seondordertangentialonditions,wealsodisusstheuseofa
nonloaloperatoralongtheinterfaes. ThisnonloaloperatorisrelatedtothesquarerootoftheLaplaian
oftheinterfaesandwasrstintroduedfornumerialalulationsin[4,26℄. Whiletheontinuousanalysis
showsthat theseoperators resultin aniterationthatonvergesin twosteps,in pratiewedonotseethis
exatbehavior. However,theresultingiterationseemstoonvergeataveryfastratethatisindependentof
themeshsize
ℎ
orthethiknessoftheoverlap𝐿
.Wehighlightfour ontributionsofthispaper. First,wehaveaninnovativeuseoftheenvelopetheorem
toestablishanequiosillationproperty. Seond,wehighlightourontinuousanddisreteanalysesandtheir
omparison. Third,ouranalysistakesplaeonthesphere,whihisidealforweatherandlimatesimulation.
Fourth, we provide a nite element method whih is appropriate for spherial disretizations. Sine the
spherialLaplaianissingular,thehoieoftheniteelementspaeisimportantandapoorhoieleadsto
divergentiterationsandpoorlyonditionedsystems. Weshowthatwiththeproperhoieofbasisfuntions,
theonvergenefatoroftheoptimizedShwarzmethodsareunaeted.
This paperis organizedas follows. InSetion 2, weintrodue ourmodel problem andstate our main
results. InSetion 3, we review the Laplae operator on the sphereand reall theShwarz iteration and
its onvergene estimates, previously published in [4℄; we also give a new semidisrete estimate whih is
substantiallysimilartotheontinuousone. Insetion4,weperformtheFourieranalysisofthenewOptimized
ShwarzMethods. Insetion5,wepresentnumerialresultsthat agreewiththetheoretialpreditions.
Ω 2
Ω 1
a b
Ω 2
Ω 1
a
Ω 3
2
b 2
a 3 b 1
Figure1: Latitudinal domaindeomposition. Left: twosubdomains;right: multiplesubdomains.
2 OSM on the sphere
Considerthe followingiteration. Let
𝑏 < 𝑎
. Begin with random andidate solutions𝑢 0
and𝑣 0
. Dene𝑢 𝑘+1
and𝑣 𝑘+1
iterativelyby:⎧
⎨
⎩
Δ𝑢 𝑘+1 = 𝑓
inΩ 1 = { (𝜑, 𝜃) ∣ 0 ≤ 𝜑 < 𝑎 } , 𝑢 𝑘+1 (𝑎, 𝜃) = 𝑣 𝑘 (𝑎, 𝜃) 𝜃 ∈ [0, 2𝜋),
Δ𝑣 𝑘+1 = 𝑓
inΩ 2 = { (𝜑, 𝜃) ∣ 𝑏 < 𝜑 ≤ 𝜋 } , 𝑣 𝑘+1 (𝑏, 𝜃) = 𝑢 𝑘 (𝑏, 𝜃) 𝜃 ∈ [0, 2𝜋);
(1)
(see gure 1.) This is the Shwarz iteration for two latitudinal subdomains. The idea of the Optimized
Shwarz Methods is to replae the Dirihlet onditions on the interfae by Robin or other transmission
onditions. Themodiediterationis:
⎧
⎨
⎩
Δ𝑢 𝑘+1 = 𝑓
inΩ 1
((𝜓 + ∂𝜑 ∂ )𝑢 𝑘+1 )(𝑎, 𝜃) = ((𝜓 + ∂𝜑 ∂ )𝑣 𝑘 )(𝑎, 𝜃) 𝜃 ∈ [0, 2𝜋),
Δ𝑣 𝑘+1 = 𝑓
inΩ 2
((𝜉 + ∂𝜑 ∂ )𝑣 𝑘+1 )(𝑏, 𝜃) = ((𝜉 + ∂𝜑 ∂ )𝑢 𝑘 )(𝑏, 𝜃) 𝜃 ∈ [0, 2𝜋);
(2)
where
𝜓
and𝜉
are either real oeients or trae operators, andΩ 1 , Ω 2
are as previously dened. If𝜓
and
𝜉
aredierentialtraeoperatorsoforder𝑘
,wesaythattheoptimizedShwarzmethodisOptimizedofOrder
𝑘
, orOOk[15℄. Choiesinlude:1.
(𝜓𝑤)(𝜃) = 𝑐𝑤(𝜃)
where𝑐
isarealoeient. Thisresultsin aRobinor OO0transmissionondition.2.
(𝜓𝑤)(𝜃) = 𝑐𝑤(𝜃)+𝑑𝑤 ′′ (𝜃)
,where𝑐
and𝑑
arerealoeients. Thisresultsinaseondordertangential, or OO2transmissionondition.3. Anonloalhoieof
𝜓
leadingtoaniterationthatonvergesintwosteps.Wewill showin Setion 3thattheonvergenefator depends onthethiknessof theoverlap
𝐿
, givenby
𝐿 := − 2 log 1 + cos 𝑎 sin 𝑎
sin 𝑏 1 + cos 𝑏 ≥ 0.
OurmainresultforRobin(orOO0)transmissiononditionsisas follows.
Theorem2.1(OptimizedRobinorOO0oeients). Considerthe iteration(2)with
(𝜓𝑤)(𝜃) = 𝑐𝑤(𝜃)
and(𝜉𝑤)(𝜃) = 𝑑𝑤(𝜃)
. Set𝑥 = − 𝑑 sin 𝑏
,𝑦 = 𝑐 sin 𝑎
. Letℎ = 1/𝑁
bethe grid size.∙
The one-sidedonditionis𝑥 = 𝑦
. Withthisondition,andfurtherassumingthat𝐿 < 4/3
and𝐿 > ℎ 𝛼
for some
𝛼 < 3/2
, there is a unique optimized hoie of𝑥
leading to the best possible onvergenefator. Using
𝑥 = 𝑦 = ˆ 𝑥 1 = 𝐿 − 1 3
, theonvergenefatorevery other stepis𝜌 = 1 − 4𝐿 1 3 + 𝑂(𝐿 2 3 )
.∙
If𝑥 ∕ = 𝑦
is allowed, the optimizedhoie(𝑐, 𝑑)
is two-sided. Assumethat𝐿 > ℎ 𝛼
,𝛼 < 5/4
. Thereisa unique
(𝑐, 𝑑)
leading tothe bestonvergenefator. Using𝑥 = 8 − 1 5 𝐿 − 1 5
and𝑦 = 8
2 5
2 𝐿 − 3 5
leads toaonvergenefator of
𝜌 = 1 − 2 ⋅ 2 3 5 𝐿 1 5 + 𝑂(𝐿 2 5 )
everyother step.Theonvergenefator
𝜌 < 1
multipliesthe𝐿 2
errorofthetraeoneahsubdomaineveryotherstep. Weseethatusingasymmetri Robinonditionsleadstoanimprovementonthenumberofiterationsrequired
for a given tolerane, from
𝑂(𝐿 − 1 3 )
to𝑂(𝐿 − 1 5 )
. The one-sided estimate however is easier to use with𝑝
subdomains,andsowewouldbeinterestedinahievinga
𝑂(𝐿 − 1 5 )
algorithmusingonlyone-sidedestimates.Thisispossibleifweuseseondordertangentialtransmissiononditions.
Theorem 2.2 (One-sidedseond order tangential,or OO2,optimized oeients). Consider the iteration
(2)with
(𝜓𝑤)(𝜃) = 𝑐𝑤(𝜃) + 𝑑𝑤 ′′ (𝜃)
,(𝜉𝑤)(𝜃) = ˜ 𝑐𝑤(𝜃) + ˜ 𝑑𝑤 ′′ (𝜃)
. Let𝑠 = 𝑐 sin 𝑎 = − 𝑐 ˜ sin 𝑏
and𝑡 = 𝑑 sin 𝑎 =
− 𝑑 ˜ sin 𝑏
, whih is the one-sided assumption. Assumethat𝐿 < ℎ 𝛼
,𝛼 < 5/4
. There isa uniquehoie of𝑠, 𝑡
leadingtothebestpossibleonvergenefator. Using𝑠 = 1 2 2 3 5 𝐿 − 1 5
and𝑡 = 1 2 2 1 5 𝐿 3 5
leadstoaonvergenefatorof
𝜌 = 1 − 4 ⋅ 2 2 5 𝐿 1 5 + 𝑂(𝐿 2 5 )
everyother step.Theonvergerateestimateshold evenif
𝐿
issmallerthanℎ 𝛼
(e.g., whenthegridishighlyanisotropi, with𝛼 = 3/2
or5/4
,asabove),butthentheoeientsgivenmaynolongerbeoptimized. Anextremeaseiswhen
𝐿 = 0
,anonoverlappingase,andweanalyzethisaseinSetion3.1. Themain aseis𝐿 = 𝑂(ℎ)
andishandledbytheTheorems2.1and2.2.
Our numerialexamplesarebased ontwodierentodes. Many odesin meteorologyuse thedisrete
Fouriertransformin thelongitudevariable
𝜃
toobtainasemispetral solver. This solvermakestheimple- mentationofourthreetypesoftransmissiononditionsequallyeasy,forlatitudinalinterfaes. However,suhsolversdonoteasilyhandlearbitrarilyshapedsubdomains,andthegridshavesingularitiesat thepoles.
Ourseond odeisaniteelementsolveronthesphere
Ω
. Thevariationalformulationofthespherial Laplaianleadsto integralsonthesphereΩ
,and sowemust hoose pieewise𝐶 1
basis funtionsfor someniteelementspaeonthesphere
Ω
. Weuseastandardprojetiontehniquetogeneratespherialelementsfrompieewiselinearelementsonapolyhedralapproximationof thesphere
Ω
.3 The Laplae operator on the sphere
Ω
WetaketheLaplaeoperatorin
ℝ 3
,givenbyΔ𝑢 = 𝑢 𝑥𝑥 + 𝑢 𝑦𝑦 + 𝑢 𝑧𝑧 ,
rephraseitinspherialoordinatesandset
∂𝑢
∂𝑟 = 0
to obtainΔ𝑢 = 1
sin 2 𝜑
∂ 2 𝑢
∂𝜃 2 + 1 sin 𝜑
∂
∂𝜑 (
sin 𝜑 ∂𝑢
∂𝜑 )
,
where
𝜑 ∈ [0, 𝜋]
istheolatitudeand𝜃 ∈ [ − 𝜋, 𝜋]
thelongitude.3.1 The solution of the Laplae problem
WetakeaFouriertransformin
𝜃
butnotin𝜑
;this letsusanalyzedomaindeompositionswithlatitudinal boundaries. TheLaplaianbeomesΔˆ ˆ 𝑢(𝜑, 𝑚) = − 𝑚 2
sin 2 𝜑 𝑢(𝜑, 𝑚) + ˆ 1 sin 𝜑
∂
∂𝜑 (
sin 𝜑 ∂ 𝑢(𝜑, 𝑚) ˆ
∂𝜑 )
, 𝜑 ∈ [0, 𝜋], 𝑚 ∈ ℤ.
(3)Forboundaryonditions,theperiodiityin
𝜃
istakenareofbytheFourierdeomposition.Thepolesimpose that𝑢(0, 𝜃)
and𝑢(𝜋, 𝜃)
donotvaryin𝜃
. For𝑚 ∕ = 0
thisisequivalenttoˆ
𝑢(0, 𝑚) = 𝑢(𝜋, 𝑚) = 0, 𝑚 ˆ ∈ ℤ, 𝑚 ∕ = 0.
(4)For
𝑚 = 0
,therelation𝑢 𝜑 (0, 𝜃) = − 𝑢 𝜑 (0, 𝜃 + 𝜋)
leadsto∫ 2𝜋
0 𝑢 𝜑 (0, 𝜃) 𝑑𝜃 = − ∫ 2𝜋
0 𝑢 𝜑 (0, 𝜃) 𝑑𝜃
,i.e.,ˆ
𝑢 𝜑 (0, 0) = 𝑢 ˆ 𝜑 (𝜋, 0) = 0.
(5)If
𝑢
is asolution ofΔ𝑢 = 𝑓
then so is𝑢 + 𝑐
(𝑐 ∈ ℝ
), henethe ODE for𝑚 = 0
is determined up to anadditiveonstant.
With
𝑚 ∕ = 0
xed,thetwoindependentsolutionsofΔ𝑢 = 0
are𝑔 ± (𝜑, 𝑚) =
( sin(𝜑) cos(𝜑) + 1
) ±∣ 𝑚 ∣
= 𝑒 ±∣ 𝑚 ∣ log cos(𝜑)+1 sin(𝜑) , 𝑚 ∈ ℤ ∖ { 0 } .
For
𝑚 = 0
thetwoindependentsolutionsgiveˆ
𝑢(𝜑, 0) = 𝐶 1 + 𝐶 2 log 1 − cos 𝜑 sin 𝜑 .
Ifweinsist,forinstane, that
𝑢 ∈ 𝐻 1 (Ω)
,weeliminatethelogarithmitermandweobtainthat𝑢(𝜑, ˆ 0)
isaonstantin
𝜑
.All theeigenvaluesof
Δ
areoftheformof− 𝑛(𝑛 + 1)
for𝑛 = 0, 1, ...
;in partiular,theyarenon-positive (andΔ
isnegativesemi-denite).3.2 The Shwarz iteration for
Δ
with two latitudinal subdomainsConsidertheShwarziteration(1). Weareinterestedin studyingtheerrorterms
𝑢 0 − 𝑢
and𝑣 0 − 𝑢
whereΔ𝑢 = 𝑓
. Observethat𝑢 𝑘 − 𝑢
and𝑣 𝑘 − 𝑢
arelassialShwarziterates asperequation(1),butwith𝑓 = 0
.For this reason,it sues to analyzethe lassialShwarziteration with
𝑓 = 0
, whih weassume fortheremainderofouranalysis.
UsingtheFouriertransformin
𝜃
,weanwrite𝑢 ˆ 𝑘+2 (𝑏, 𝑚)
expliitlyin termsof𝑢 ˆ 𝑘 (𝑏, 𝑚)
. Thisallowsustoobtainaonvergenefatorestimate,whihwereallfrom [4℄.
Lemma3.1. The Shwarziteration onthesphere
Ω
partitionedalong twolatitudes𝑏 < 𝑎
onverges(exeptfor the onstantterm). The onvergene fator
∣ 𝑢 ˆ 𝑘+2 (𝑏, 𝑚)/ˆ 𝑢 𝑘 (𝑏, 𝑚) ∣
is𝐶(𝑚) =
( sin(𝑏) cos(𝑏) + 1
) 2 ∣ 𝑚 ∣ (
sin(𝑎) cos(𝑎) + 1
) − 2 ∣ 𝑚 ∣
< 1.
(6)This onvergenefatordependsonthefrequeny
𝑚
of𝑢 𝑘
onthelatitude𝑏
.AnanalysisthatislosertothenumerialalgorithmwouldbetoreplaetheontinuousFouriertransform
in
𝜃
byadisreteone.Lemma3.2 (Semidisreteanalysis.). The Laplaian disretizedin
𝜃
with𝑛
sample points:Δ 𝑛 𝑢 = 𝑛 2 4𝜋 2 sin 2 𝜑
( 𝑢
(
𝜑, 𝑗 + 1 2𝜋𝑛
)
− 2𝑢(𝜑, 𝑗) + 𝑢 (
𝜑, 𝑗 − 1 2𝜋𝑛
))
+ cot 𝜑𝑢 𝜑 + 𝑢 𝜑𝜑
(7)leads toaShwarz iterationwhose onvergenefatoris
( sin(𝑏) cos(𝑏) + 1
) 2 ∣ 𝑚 ˜ ∣ (
sin(𝑎) cos(𝑎) + 1
) − 2 ∣ 𝑚 ˜ ∣
< 1
every twoiterations, where
˜
𝑚 2 = 𝑛 2
4𝜋 2 (1 − cos(2𝜋𝑘/𝑛))
for the
𝑘
thfrequeny.Proof. WedoadisreteFouriertransformin
𝜃
andtheequationtosolvebeomes𝐿 𝑛 𝑢(𝜑, 𝑘) = ˆ 𝑛 2
4𝜋 2 sin 2 𝜑 (cos(2𝜋𝑘/𝑛) − 1)ˆ 𝑢(𝜑, 𝑘) + cot 𝜑ˆ 𝑢 𝜑 (𝜑, 𝑘) + ˆ 𝑢 𝜑𝜑 (𝜑, 𝑘) = 0.
(8)Bywriting
𝑚
as notedin thestatementof theLemma,wehavereduedthis problemto the previousoneandweanreuse Lemma3.1.
The two ontration onstants are verysimilar. For small valuesof
𝑚
(ignoring𝑚 = 0
beause thatmode need notonvergeat all),the speedof onvergene is very poor. The overall onvergene fator as
measuredinthe
𝐿 2
normalongtheinterfaesisgivenbysup
𝑚 ≥ 1
𝐶(𝑚) = 𝐶(1)
= 1 − 𝑂(𝑎 − 𝑏),
andsotheonvergenefatoroftheShwarziterationdeterioratesrapidlyas
𝑎 − 𝑏
beomessmall.4 Optimized Shwarz iteration for
Δ
with latitudinal boundaries Inthissetion,weproveTheorems2.1and2.2. Thisrequiresmanytehnialresultsandsoweproeedwithasequeneofsimplelemmas.
Toanalyze the onvergene fator, as we did in Setion 3, weassume, without lossof generality, that
𝑓 = 0
. Beause theLaplaianof aonstantis zero, we will seethat theerrorterms onvergeto onstantfuntions,whihisto say,theloalsolutionsonvergeupto anadditiveonstant.
Onekeytehniqueisto usethefat that,although
𝑢 0
and𝑣 0
maybearbitraryin someSobolevspaes,𝑢 1
and𝑣 1
areharmoni inΩ 1
andΩ 2
, respetively. This keyobservationallowsus to nditerations that onvergefaststarting from the seond step. Thisobservation hasanimpaton thenumerialsimulations.Often, the rstiterations
𝑢 1
and𝑣 1
are notmuh better than the initial guesses𝑢 0
and𝑣 0
, andthey aresometimesevenworse.
Lemma 4.1 (Nonloal operator). If, for eah
𝑚
,𝜓(𝑚) = ˆ ∣ 𝑚 ∣ / sin 𝑎
and𝜉(𝑚) = ˆ −∣ 𝑚 ∣ / sin 𝑏
,Δ𝑢 1 = 0
inΩ 1
andΔ𝑣 1 = 0
inΩ 2
, then𝑢 2 = 0
and𝑣 2 = 0
.Proof. Let
𝑢 1 (𝜑, 𝜃) = ˆ 𝑢 1 (𝑎, 0)+ ∑
𝑘 ∕ =0 𝑢 ˆ 1 (𝑎, 𝑘)𝑔 + (𝜑, 𝑘)𝑒 𝑖𝑘𝜃
and𝑣 1 (𝜑, 𝜃) = ˆ 𝑣 1 (𝑎, 0)+ ∑
𝑘 ∕ =0 𝑣 ˆ 1 (𝑎, 𝑘)𝑔 − (𝜑, 𝑘)𝑒 𝑖𝑘𝜃
,thentheboundaryvalueproblemforthenorthhemisphereyieldsforeah
𝑚 ∕ = 0
:ˆ
𝑢 2 (𝑎, 𝑚) (
𝜓(𝑚) + ˆ ∣ 𝑚 ∣ sin 𝑎
)
= ˆ 𝑣 1 (𝑎, 𝑚) (
𝜓(𝑚) ˆ − ∣ 𝑚 ∣ sin 𝑎
)
= 0,
(9)and so
𝑢 ˆ 2 (𝑎, 𝑚) = 0
. Likewise, we nd that𝑣 ˆ 2 (𝑏, 𝑚) = 0
for eah𝑚 ∕ = 0
. For𝑚 = 0
, we have thatˆ
𝑢 2 (𝑎, 𝑚) = ˆ 𝑣 1 (𝑎, 𝑚)
: the iteration does not aet the onstant mode𝑚 = 0
. Sine the solution of theLaplaeproblemisonlydeneduptoaonstant,theiterationhasonverged.
Theseondtangentialderivative
𝐷 2 𝜃
alonglatitude𝜑
hasFouriertransform− 𝑚 2
. Theoptimaloperatorsare therefore the symmetri and positive denite square root of
− 𝐷 𝜃 2
, multiplied by1/ sin 𝑎
or1/ sin 𝑏
aordingto thelatitude. This normalizationonstantisrelatedtothelengthofthelatitudeat
𝜑 = 𝑎
and𝜑 = 𝑏
,respetively,sinealatitude𝜑
isairle ofradius𝑟 = sin 𝜑
.Next,westatetheonvergenefatoroftheoptimizedShwarziteration,if
𝜓
and𝜉
arenotthesenonloaloperators.
Lemma4.2. If
Δ𝑢 1 = 0
inΩ 1
thenˆ
𝑢 3 (𝑏, 𝑚) = 𝑢 ˆ 1 (𝑏, 𝑚) 𝜉(𝑚) sin ˆ 𝑏 + ∣ 𝑚 ∣ 𝜉(𝑚) sin ˆ 𝑏 − ∣ 𝑚 ∣
𝜓(𝑚) sin ˆ 𝑎 − ∣ 𝑚 ∣ 𝜓(𝑚) sin ˆ 𝑎 + ∣ 𝑚 ∣
𝑔 − (𝑚, 𝑎) 𝑔 − (𝑚, 𝑏)
𝑔 + (𝑚, 𝑏) 𝑔 + (𝑚, 𝑎) .
Theproof isasimpleomputation usingequation(9). Wenowlookforgoodonvolutionkernels
𝜉
and𝜓
so thatthetheonvergenefator𝜅(𝜓, 𝜉, 𝑚, 𝑎, 𝑏) =
𝜅 1
z }| {
𝜉(𝑚) sin𝑏 ˆ + ∣𝑚∣
𝜉(𝑚) sin𝑏 ˆ − ∣𝑚∣
𝜅 2
z }| {
𝜓(𝑚) sin ˆ 𝑎 − ∣𝑚∣
𝜓(𝑚) sin ˆ 𝑎 + ∣𝑚∣
𝜅 3
z }| { 𝑔 − (𝑚, 𝑎) 𝑔 − (𝑚, 𝑏)
𝜅 4
z }| { 𝑔 + (𝑚, 𝑏)
𝑔 + (𝑚, 𝑎) ,
(10)with
𝑚 ∈ ℤ ∖ { 0 }
,isassmallaspossible. Foragivenhoieof𝜓 ˆ
and𝜉 ˆ
,theonvergenefatorasmeasuredinthe
𝐿 2
normalongtheinterfae𝜑 = 𝑏
will begivenbysup 𝑚 ∈ℤ∖{ 0 } ∣ 𝜅(𝜓, 𝜉, 𝑚, 𝑎, 𝑏) ∣ .
Theonvergenefatorestimate(10)alsoleadstothefollowingresult:
Corollary 4.3. The iteration (2) is onvergent (modulo the onstant mode)if
𝜓
and− 𝜉
are onvolutionoperatorsthatarepositive denite,regardless ofoverlap.
Proof. Let
𝜓(𝑚) ˆ > 0
and𝜉(𝑚) ˆ < 0
,andlet𝜅 1 , 𝜅 2 , 𝜅 3 , 𝜅 4
beasin (10). Sinesin 𝑎 > 0
andsin 𝑏 > 0
,weseethat
∣ 𝜅 1 ∣ < 1
and∣ 𝜅 2 ∣ < 1
. Furthermore,𝜅 3 𝜅 4 =
( sin(𝑏) cos(𝑏) + 1
) 2 ∣ 𝑚 ∣ (
sin(𝑎) cos(𝑎) + 1
) − 2 ∣ 𝑚 ∣
=
⎛
⎜ ⎜
⎜ ⎝
≤ 1
z }| { ( sin(𝑏)
sin(𝑎) )
≤ 1
z }| { ( cos(𝑎) + 1
cos(𝑏) + 1 )
⎞
⎟ ⎟
⎟ ⎠
2 ∣ 𝑚 ∣
≤ 1,
wherewehaveusedthat
𝑏 < 𝑎
. Hene,∣ 𝜅(𝜓, 𝜉, 𝑚, 𝑎, 𝑏) ∣ < 1
forevery𝑚 ∕ = 0
.By omparison, in [27℄, it was shown that if
𝜉
and𝜓
are Robin transmission onditionsfor a domain deomposition with relativelyuniform overlap,then the iterationonvergesso long as𝜉
and𝜓
are bothpositivedenite.
Our analysis willproeed by onsideringthe ontrationonstant (10)forthevarious hoiesof trans-
missiononditions
𝜉
and𝜓
. WeonsiderrsttheuseofRobintransmissiononditionsand proveTheorem 2.1. Seond,weproveTheorem2.2forseondordertangentialtransmissiononditions.4.1 Robin or OO0 transmission onditions
Let
𝜓𝑢 = 𝑐𝑢
and𝜉𝑢 = 𝑑𝑢
,then𝜅 𝑂𝑂0 (𝜓, 𝜉, 𝑚, 𝑎, 𝑏) = 𝑑 sin 𝑏 + ∣ 𝑚 ∣ 𝑑 sin 𝑏 − ∣ 𝑚 ∣
𝑐 sin 𝑎 − ∣ 𝑚 ∣ 𝑐 sin 𝑎 + ∣ 𝑚 ∣
𝑔 − (𝑚, 𝑎) 𝑔 − (𝑚, 𝑏)
𝑔 + (𝑚, 𝑏) 𝑔 + (𝑚, 𝑎)
= 𝜅(𝑐, 𝑑, 𝑚, 𝑎, 𝑏).
For
𝑎, 𝑏
xed,wenowneedtoomputethehoiesof𝑐, 𝑑
minimizing𝜑(𝑐, 𝑑) = sup
𝑚 𝜅 𝑂𝑂0 (𝑐, 𝑑, 𝑚, 𝑎, 𝑏) 2 ,
where wehaveintrodued asquare to simplify somealulations; theatual ontration onstantwill be
givenby
√ 𝜑(𝑐, 𝑑)
. Tosimplifyournotation,itisonvenienttodoahangeofvariables. Weset𝑥 = − 𝑑 sin 𝑏
,𝑦 = 𝑐 sin 𝑎
and𝐿 = − 2 log
( 1 + cos 𝑎 sin 𝑎
sin 𝑏 1 + cos 𝑏
)
≥ 0,
andrewritingthe
𝑔 ±
termsasexponentials,theoptimizationproblembeomes Minimize𝜑(𝑥, 𝑦)
where𝜑(𝑥, 𝑦) = sup
𝑚 ∈{ 1,2,... }
𝐹 (𝑥, 𝑦, 𝑚) 𝐹 (𝑥, 𝑦, 𝑚) =
𝑥 − 𝑚 𝑥 + 𝑚
𝑦 − 𝑚 𝑦 + 𝑚 𝑒 − 𝐿𝑚
2
(11)
= 𝐹 𝐿 (𝑥, 𝑦, 𝑚).
The onvergene fator every other step for partiular parameter hoies
𝑥
and𝑦
is√
𝜑(𝑥, 𝑦)
. The ase𝐿 = 0
onlyours when𝑎 = 𝑏
,in whih ase𝜑(𝑥, 𝑦) = 1
forall𝑥, 𝑦 ≥ 1
and thenwewill needtotakeintoaountertaindisretizationparameterstoobtainaonvergenefatorestimate. Thisasewillbetreated
separately.
WenowusetwodierentanalysestoomputeoptimizedRobinparameters
𝑐
and𝑑
. First,wemakethesimplifyingassumption that
𝑥 = 𝑦
,allowingustooptimizeinsteadtheobjetivefuntionmax 𝑚 ∣ (𝑥 − 𝑚)/(𝑥 + 𝑚)𝑒 − 𝐿 2 𝑚 ∣ 2 .
This situation is interesting beause this an lead to an algorithm where all subdomains do the same
thing. Thisisthemethodwhihmostobviouslygeneralizesto manysubdomains. Seond,weonsiderthe
expression
𝐹 (𝑥, 𝑦, 𝑚)
withbothparameterssimultaneously. Thisisalsointeresting,beauseinthissituation, the iteration does something dierent on eah subdomain. We will see that, although this approah isdiulttogeneralizetomanysubdomains,itleadsto substantiallyimprovedonvergenefators.
4.2 OO0 one-sided analysis,
𝐿 > 0
.Weanwrite
𝐹 𝐿 (𝑥, 𝑦, 𝑚) = 𝐹 𝐿 (𝑥, 𝑚)𝐹 𝐿 (𝑦, 𝑚)
(f. (11)),wherewedene𝐹 𝐿 (𝑥, 𝑚) = 𝐹(𝑥, 𝑚) =
𝑥 − 𝑚 𝑥 + 𝑚 𝑒 − 𝐿 2 𝑚
2
.
Werstlookattheproblemofomputingtheoptimalparameter
𝑥 0
minimizing𝜑 𝐿 (𝑥) = sup
𝑚 ∈{ 1,2,... }
𝐹 𝐿 (𝑥, 𝑚),
sine it is simpler. The solution
𝑥 0
minimizing𝜑 𝐿 (𝑥)
leads to the hoies𝑥 = 𝑦 = 𝑥 0
, whih minimizes𝜑 𝐿 (𝑥, 𝑦 )
subjettothesymmetryonstraint𝑥 = 𝑦
. Theonvergenefatorforapartiularparameterhoie𝑥
is√
𝜑(𝑥, 𝑥) = 𝜑 𝐿 (𝑥)
.We now deal with the various spetra over whih we an optimize. Our ontinuous problem has a
frequenyvariable
𝑚 ∈ ℤ
,butsinethe𝑚 = 0
frequenyneednotonverge,andsine𝐹 𝐿 (𝑥, − 𝑚) = 𝐹 𝐿 (𝑥, 𝑚)
,wehavealreadysimpliedourproblembyonsideringthefrequenies
𝑚 ∈ { 1, 2, . . . }
. Beauseofoureventualdisretization,wewillwanttoonsiderthefrequenydomain
{ 1, 2, ..., 𝑁 }
,where𝑁
isthehighestfrequenythatan beresolvedontheinterfae,givenourgrid. However,beauseofthediulties inoptimizingover
disretesets, wealso onsider thepossibilityof allowing
𝑚
tovary in theintervals[1, ∞ )
aswellas[1, 𝑁]
.For tehnialreasons,itisalsoonvenienttoinlude
𝑚 = ∞
inouralulations.Let
2 < 𝑁 < ∞
anddene𝜑 (𝑐) 𝐿 (𝑥) = sup
𝑚 ∈ [1, ∞ ]
𝐹 𝐿 (𝑥, 𝑚), (𝐼 (𝑐) = [1, ∞ ]), 𝜑 (𝑑) 𝐿 (𝑥) = sup
𝑚 ∈{ 1,2,... }
𝐹 𝐿 (𝑥, 𝑚), (𝐼 (𝑑) = { 1, 2, ..., ∞} ), 𝜑 (𝑐𝑏) 𝐿 (𝑥) = sup
𝑚 ∈ [1,𝑁]
𝐹 𝐿 (𝑥, 𝑚), (𝐼 (𝑐𝑏) = [1, 𝑁]), 𝜑 (𝑑𝑏) 𝐿 (𝑥) = sup
𝑚 ∈{ 1,2,...,𝑁 }
𝐹 𝐿 (𝑥, 𝑚), (𝐼 (𝑑𝑏) = { 1, 2, ..., 𝑁 } ).
Aordingtoourdenitions, forall
𝑥
,𝜑 (𝑑) 𝐿 (𝑥) = 𝜑 𝐿 (𝑥)
. Theother funtionsaresimilarbut maximizeoveradierentset offrequenies.
One of the main toolsfor solvingmin-max problems is theenvelopetheorem, whih westate herefor
onveniene.
Theorem 4.4 (The EnvelopeTheorem). Let
𝑈
be open inℝ 𝑛
and𝜑 : 𝑈 → 𝑅
,𝛾 ∈ ℝ 𝑛
,∣∣ 𝛾 ∣∣ = 1
. For𝑥 ∈ 𝑈
, denethe one-sided diretional derivative𝐷 𝛾 𝜑(𝑥)
by𝐷 𝛾 𝜑(𝑥) = lim
𝜖 ↓ 0 +
𝜑(𝑥 + 𝜖𝛾) − 𝜑(𝑥)
𝜖 ,
if the(one-sided)limit exists. Let
𝑉
beaompat metrispae. Let𝐹 : 𝑈 × 𝑉 → ℝ
(𝑥, 𝑦) 7→ 𝐹 (𝑥, 𝑦)
beontinuousandassumethatthepartialgradient
𝐹 𝑥 (𝑥, 𝑦)
existsforall𝑥, 𝑦 ∈ 𝑈 × 𝑉
andvariesontinuously in(𝑥, 𝑦)
. Dene𝜑(𝑥) = max
𝑦 ∈ 𝑉 𝐹(𝑥, 𝑦),
𝑌 (𝑥) = { 𝑦 ∈ 𝑉 ∣ 𝐹 (𝑥, 𝑦) = 𝜑(𝑥) } ∕ = ∅ .
Then,
1. Forall
𝑥 ∈ 𝑈
and𝛾 ∈ ℝ 𝑛 , ∣∣ 𝛾 ∣∣ = 1
,𝐷 𝛾 𝑢(𝑥)
existsandisgiven bythe formula𝐷 𝛾 𝜑(𝑥) = max
𝑦 ∈ 𝑌 (𝑥)
∑ 𝑛
𝑖=1
𝛾 𝑖 𝐹 𝑥 𝑖 (𝑥, 𝑦).
(12)2. Let
𝑥 ˜ ∈ 𝑈
. If𝑌 (˜ 𝑥) = { 𝑦(˜ 𝑥) }
isasinglepointthen𝐷 𝛾 𝜑(˜ 𝑥) =
∑ 𝑛
𝑖=1
𝛾 𝑖 𝐹 𝑥 𝑖 (˜ 𝑥, 𝑦)
(13)and
𝜑
is(fully) dierentiableat𝑥 ˜
. If furthermore𝑌 (𝑥) = { 𝑦(𝑥) }
isasingleton for all𝑥 ∈ 𝑂 ⊂ 𝑈
anopen set,then
𝜑
isontinuouslydierentiablein𝑂
.This immediatelyleadstoanequiosillationproperty.
Lemma4.5. For
𝑗 ∈ { 𝑐, 𝑑, 𝑐𝑏, 𝑑𝑏 }
, thereisaminimizer𝑥 𝑗
of𝜑 (𝑗) (𝑥)
,and∣ 𝑀 (𝑗) (𝑥 𝑗 ) ∣ ≥ 2
, where𝑀 (𝑗) (𝑥) = { 𝑚 ∈ 𝐼 (𝑗) ∣ 𝜑 (𝑗) (𝑥) = 𝐹 𝐿 (𝑥, 𝑚) }
.Proof. Wemustplaeourselvesinthehypothesesoftheenvelopetheorem. Eahset
[1, ∞ ], { 1, 2, ..., ∞} , [1, 𝑁 ], { 1, 2, ..., 𝑁 }
isaompatmetrispae(for
𝐼 (𝑐)
and𝐼 (𝑑)
onemayusethemetri𝑑(𝑥, 𝑦) = ∣ 𝑥 − 1 − 𝑦 − 1 ∣
). Sinethefuntion𝜑 (𝑗)
is a ontinuous funtion on the ompat metri spae𝐼 (𝑗)
, it must have a minimum. The optimalparameter
𝑥 𝑗
isin[0, ∞ )
sowetake𝑈 = ( − 1 2 , ∞ )
,andthehypothesesoftheenvelopetheoremaresatised.Assumethat
𝑀 (𝑗) (𝑥 𝑗 ) = { 𝑚 ˜ }
isasingleton(thereisnoequiosillation). Then,bytheenvelopetheorem, thetwo-sidedderivative𝐷 𝑥 𝜑 (𝑗) (𝑥 𝑗 )
existsandsine𝜑 (𝑗) (𝑥 𝑗 )
isminimal,thederivativemustbezero. Usingformula(13),weobtain
0 = 𝐹 𝑥 (𝑥 𝑗 , 𝑚) = ˜ − 4 ˜ 𝑚 𝑚 ˜ − 𝑥 𝑗
( ˜ 𝑚 + 𝑥 𝑗 ) 3 𝑒 − 2𝐿 𝑚 ˜ .
Sine
𝑥 𝑗 , 𝑚 > ˜ 0
, itmust bethat𝑚 ˜ = 𝑥 𝑗
. Hene,𝜑 (𝑗) (𝑥 𝑗 ) = 𝐹 (𝑥 𝑗 , 𝑚) = 0 ˜
. Butwe knowthat𝜑 (𝑗) (𝑥 𝑗 ) ≥ 𝐹 (𝑥, 𝑥 + 1) > 0
.Nowweharaterizetheuniqueminimizerforthease
𝑗 = 𝑐
.Lemma4.6. Thereisaunique
𝑥 𝑐
minimizing𝜑 (𝑐) 𝐿 (𝑥)
anditisthe uniqueroot𝜓(𝑥 𝑐 ) = 0
of𝜓(𝑥) = 𝜓(𝑥, 𝐿) =
( 𝑚(𝑥) − 𝑥 𝑚(𝑥) + 𝑥
) 2
𝑒 − 𝐿𝑚(𝑥) − ( 1 − 𝑥
1 + 𝑥 ) 2
𝑒 − 𝐿 ,
where
𝑚(𝑥) =
√ 4𝑥 𝐿 + 𝑥 2
.Proof. Beause any minimizer
𝑥 𝑐
is in[1, ∞ )
, the onlyloal maxima of𝐹 𝐿 (𝑥, 𝑚)
as𝑚
runs in[1, ∞ ]
areat
𝑚 = 1
and𝑚 = 𝑚(𝑥) = √
4𝑥
𝐿 + 𝑥 2
and so𝑀 (𝑐) (𝑥 𝑐 ) ⊆ { 1, 𝑚(𝑥 𝑐 ) }
. By Lemma 4.5, we know that#𝑀 (𝑐) (𝑥 𝑐 ) ≥ 2
so𝑀 (𝑐) (𝑥 𝑐 ) = {
1,
√ 4𝑥 𝑐
𝐿 + 𝑥 2 𝑐 }
.
Hene,
𝜓(𝑥 𝑐 ) = 0
. Thisshowstheequiosillation𝜓(𝑥) = 0
.Wenowshowtheuniquenessofthesolutionoftheequiosillationproblem. Tothatend, let
𝑔(𝑥) = 𝐹 (𝑥, 𝑚(𝑥)),
𝑔 ′ (𝑥) = 𝐹 𝑥 (𝑥, 𝑚(𝑥)) + 𝐹 𝑚 (𝑥, 𝑚(𝑥))𝑚 ′ (𝑥)
= 𝐹 𝑥 (𝑥, 𝑚(𝑥))
= − 4𝑚(𝑥) 𝑚(𝑥) − 𝑥
(𝑚(𝑥) + 𝑥) 3 𝑒 − 𝐿𝑚(𝑥) < 0,
and
ℎ(𝑥) = 𝐹 (𝑥, 1), ℎ ′ (𝑥) = − 4 1 − 𝑥
(1 + 𝑥) 3 𝑒 − 𝐿 > 0.
Hene,
𝜓 ′ = 𝑔 ′ − ℎ ′ < 0
and𝜓
maynothavemultiplezeros.Thisresultgivessomemoreinformationforthease
𝑗 = 𝑑
(thatis,𝐼 (𝑑) = { 1, 2, ..., ∞}
).Wemustassumethat
𝐿
isnottoobig,butourfousistheasewherethereislittleoverlapsothisisnotaproblem. Thespeialpoint
𝑚(𝑥)
ontinuestoplayanimportantrole. Notiethat𝑚 ′ (𝑥) > 1
andthat𝑚(1) = √
4 + 𝐿/ √
𝐿 > 1
sothat, forall
𝑥 > 1
,𝑚(𝑥) > 𝑚(1) + (𝑥 − 1) > 𝑥
.Lemma 4.7. Let
𝐿 < 4/3
and let𝑥 𝑐
minimize𝜑 (𝑐) 𝐿 (𝑥)
and𝑥 𝑑
minimize𝜑 (𝑑) 𝐿 (𝑥)
. Then1 ∈ 𝑀 (𝑑) (𝑥 𝑑 ) ⊆ { 1, ⌈ 𝑚(𝑥 𝑑 ) ⌉ , ⌊ 𝑚(𝑥 𝑑 ) ⌋}
and𝑥 𝑐 − 1 ≤ 𝑥 𝑑 ≤ 𝑥 𝑐
.Proof. Beause
𝑚 = 𝑚(𝑥 𝑑 )
istheonlyloalmaximumof𝐹 𝐿 (𝑥, 𝑚)
,weknowthat𝑀 (𝑑) (𝑥 𝑑 ) ⊆ { 1, ⌈ 𝑚(𝑥 𝑑 ) ⌉ , ⌊ 𝑚(𝑥 𝑑 ) ⌋}
,and we also knowfrom Lemma 4.5 that
#𝑀 (𝑑) (ˆ 𝑥) ≥ 2
. Assume that𝑀 (𝑑) (ˆ 𝑥) = {⌈ 𝑚(ˆ 𝑥) ⌉ , ⌊ 𝑚(ˆ 𝑥) ⌋}
. Thehypothesis on
𝐿
ombinedwiththefat that𝑥 𝑑 ≥ 1
guaranteesthat𝑚(𝑥 𝑑 ) > 𝑥 𝑑 + 1
so𝑥 𝑑 < ⌊ 𝑚(𝑥 𝑑 ) ⌋
. Weannowusetheone-sidedderivativeoftheEnvelopeTheorem,obtainingthat
𝐷 (1) 𝜑 (𝑑) (𝑥 𝑑 ) = max
𝑚 ∈{⌈ 𝑚(𝑥 𝑑 ) ⌉ , ⌊ 𝑚(𝑥 𝑑 ) ⌋} 𝐹 𝑥 (𝑥 𝑑 , 𝑚)
= max
𝑚 ∈{⌈ 𝑚(𝑥 𝑑 ) ⌉ , ⌊ 𝑚(𝑥 𝑑 ) ⌋} − 𝑚 𝑚 − 𝑥 𝑑
(𝑚 + 𝑥 𝑑 ) 3 𝑒 − 𝐿𝑚 < 0.
Therefore,
1 ∈ 𝑀 (𝑑) (ˆ 𝑥)
.Clearly, wemusthave
𝜑 (𝑑) (𝑥 𝑑 ) ≤ 𝜑 (𝑐) (𝑥 𝑐 )
. If𝑥 𝑑 > 𝑥 𝑐
then𝜑 (𝑑) (𝑥 𝑑 ) ≥ 𝐹(𝑥 𝑑 , 1) = 𝜑 (𝑐) (𝑥 𝑑 ) > 𝜑 (𝑐) (𝑥 𝑐 )
,henewemust have that
𝑥 𝑑 ≤ 𝑥 𝑐
. If𝑥 𝑑 < 𝑥 𝑐 − 1
, then in viewof the estimate𝑚 ′ (𝑥) > 1
, wehave that𝑚(𝑥 𝑑 ) < 𝑚(𝑥 𝑐 ) − 1
. Let𝑚 0 = ⌊ 𝑚(𝑥 𝑐 ) ⌋
andndthe unique𝑥 0 ∈ (1, 𝑚 0 )
suhthat𝑚(𝑥 0 ) = 𝑚 0
. Wehavethat
𝑥 𝑑 ∈ (1, 𝑥 0 ) ⊂ (1, 𝑚 0 )
. For𝑥 ∈ (1, 𝑚 0 )
,wehavethat𝐹 𝑥 (𝑥, 𝑚 0 ) < 0
. Hene,𝜑 (𝑑) (𝑥 𝑑 ) ≥ 𝐹 (𝑥 𝑑 , 𝑚 0 ) >
𝐹 (𝑥 0 , 𝑚 0 ) = 𝜑 (𝑐) (𝑥 0 ) > 𝜑 (𝑐) (𝑥 𝑐 )
.As aresult of Lemma 4.7, we see that using thedisrete frequenyspetrum
{ 1, 2, ..., ∞}
or using theontinuousfrequenyspetrum
[1, ∞ ]
resultsinapproximatelythesameRobinparameter.We would liketo givea formula for the optimized Robin parameter, but the equiosillation property
annotbesolvedexpliitly. However,wean solveitapproximatelyfor small
𝐿
. Thisfurther allowsus toonsiderthespetrum
[1, 𝑁 ]
. If𝑥 𝑐𝑏
minimizes𝜑 (𝑐𝑏) (𝑥)
then𝑀 (𝑐𝑏) (𝑥 𝑐𝑏 ) ⊆ { 1, 𝑁, 𝑚(𝑥 𝑐𝑏 ) }
. When𝐿
isnottoosmallomparedto
𝑁 − 1
,wenowshowthat𝑀 (𝑐𝑏) (𝑥 𝑐𝑏 ) = { 1, 𝑚(𝑥 𝑐𝑏 ) }
. This impliesthat, forthosevaluesof𝐿
and𝑁
,theoptimizedparameterforthefrequenyspetrum[1, 𝑁]
isthesameastheoneforthefrequenyspetrum
[1, ∞ ]
.Lemma4.8. The optimizedhoie
𝑥 𝑐
for𝜑 (𝑐) (𝑥 𝑐 )
isasymptotito𝐿 − 1 3
:𝐿 lim → 0 + 𝑥 𝑐 𝐿 1 3 = 1.
If wehoose
𝑥 ˆ 𝑐 = 𝐿 − 1 3
asouroptimizedparameter, the onvergene fatorevery otherstepis𝐹 (ˆ 𝑥 𝑐 , 1) = 1 − 4𝐿 1 3 + 𝑂(𝐿 2 3 ).
(14)Let
𝛼 > − 3/2
. If𝐿 = 𝐿(𝑁) > 𝑁 𝛼
then for all𝑁
suiently large,𝑚(𝑥 𝑐 ) < 𝑁
, and for suhvalues of𝑁
and
𝐿
, wehave that𝑥 𝑐 = 𝑥 𝑐𝑏
.Proof. Wetrytosolve
𝐹 (𝑥, 𝑚(𝑥)) − 𝐹 (𝑥, 1) = 0
bywriting𝑥
asapowerof𝐿
. Let𝑥 = 𝐶𝐿 𝛽
,thenweget0 = 𝜓 =
( √
𝐶𝐿 𝛽 − 1 + 𝐶 2 𝐿 2𝛽 − 𝐶𝐿 𝛽
√ 𝐶𝐿 𝛽 − 1 + 𝐶 2 𝐿 2𝛽 + 𝐶𝐿 𝛽 ) 2
𝑒 − 𝐿 √
4𝐶𝐿 𝛽− 1 +𝐶 2 𝐿 2𝛽
−
( 1 − 𝐶𝐿 𝛽 1 + 𝐶𝐿 𝛽
) 2
𝑒 − 𝐿 ,
(15)whih is dened for
𝐿 > 0
and𝐶 > 0
. We want to take a series expansion for𝜓
, and we nd that for𝛽 = − 1/3
,𝜓 = ( 4
𝐶 − 4 √ 𝐶
)
𝐿 3 1 + 𝑂(𝐿 2 3 ).
Wesee that
𝐺(𝑥) = 𝐹 (𝑥, 𝑚(𝑥)) − 𝐹(𝑥, 1) < 0
for suientlysmall𝐿
ifwehoose𝑥 = 𝐶𝐿 − 1 3
with𝐶 > 1
,and
𝐺(𝑥) > 0
forallsuientlysmall𝐿
if𝐶 < 1
. Hene,𝑥 𝑐 𝐿 1 3 → 1
as𝐿 → 0 +
.Toshowthat
𝑥 𝑐 = 𝑥 𝑐𝑏
,itsuestoshowthat𝑚(𝑥 𝑐 ) < 𝑁,
(16)forsuientlysmall
𝐿
. Wesubstitute𝑥 = 𝐿 − 1 3
andthen𝐿 = 𝑁 𝛼
into𝑚(𝑥)
obtaining𝑚(𝑥) =
√
4𝑁 − 4 3 𝛼 + 𝑁 − 2 3 𝛼 .
Hene if
𝛼 > − 3/2
,𝑚(𝑥)/𝑁 → 0
as𝑁 → ∞
and so𝑚(𝑥) < 𝑁
for large𝑁
. Conversely, if𝛼 < − 3/2
,𝑚(𝑥)/𝑁 → ∞
and𝑚(𝑥) > 𝑁
forlarge𝑁
.Thease
{ 1, 2, ..., 𝑁 }
followsimmediately.Corollary 4.9. Let
𝛼 > − 2 3
. Let𝐿 = 𝐿(𝑁 ) > 𝑂(𝑁 𝛼 )
,𝐿 < 4 3
and let𝑥 𝑐𝑏
minimize𝜑 (𝑐𝑏) (𝑥)
. If𝑥 𝑑𝑏
minimizes
𝜑 (𝑑𝑏) (𝑥)
thenfor all𝑁
suiently large,1 ∈ 𝑀 (𝑥 𝑑𝑏 ) ⊂ { 1, ⌊ 𝑚(𝑥 𝑐 ) ⌋ , ⌈ 𝑚(𝑥 𝑐 ) ⌉}
.Proof. Say that
𝑥 𝑑
minimizes𝜑 (𝑑) (𝑥)
. By lemma 4.7, we have that𝑀 (𝑑) (𝑥 𝑑 ) ⊆ { 1, ⌈ 𝑚(𝑥 𝑐 ) ⌉ , ⌊ 𝑚(𝑥 𝑐 ) ⌋}
,where
𝑥 𝑐
minimizes𝜑 (𝑐) (𝑥)
. Lemma 4.8 gives⌈ 𝑚(𝑥 𝑐 ) ⌉ < 𝑁
for large𝑁
, hene𝑀 (𝑑) (𝑥 𝑑 ) ⊂ { 1, ..., 𝑁 }
. If𝑥 𝑑𝑏 > 𝑥 𝑑
then𝐹 (𝑥 𝑑𝑏 , 1) > 𝐹 (𝑥 𝑑 , 1)
and𝑥 𝑑𝑏
is not a minimizer. If𝑥 𝑑𝑏 < 𝑥 𝑑
and if1 ∕ = 𝑚 𝑑 ∈ 𝑀 (𝑑) (𝑥 𝑑 )
then
𝜑 (𝑑𝑏) (𝑥 𝑑𝑏 ) ≥ 𝐹(𝑥 𝑑𝑏 , 𝑚 𝑑 ) ≥ 𝐹 (𝑥 𝑑 , 𝑚 𝑑 ) = 𝜑 (𝑑) (𝑥 𝑑 )
and𝑥 𝑑𝑏
is nota minimum. Hene𝑥 𝑑 = 𝑥 𝑑𝑏
for allsuientlylarge
𝑁
.Notethat theasymptotionvergenefator(14)isapproximatelyindependentofthespetrumweuse,
andsowemayusetheoptimizedparameter
𝑥 ˆ 𝑐 = 𝐿 − 1 3
in allsituations.This ompletestheproofoftherstpartofTheorem2.1.
4.3 Two-parameter OO0 analysis,
𝐿 > 0
.Returningtoformula(11),wenowomputetheoptimalparameters
𝑥
and𝑦
withoutassumingthat𝑥 = 𝑦
.Calulationsinthissettingaremoreompliatedthanintheone-parameterase,so wetreatonlytheases
of thespetrum
[1, ∞ ]
and[1, 𝑁]
. We an get arudimentary equiosillationpropertyout of theenvelope theorem.Lemma4.10. Let
𝐿 > 0
and𝐹 (𝑥, 𝑦, 𝑚) =
( 𝑥 − 𝑚 𝑥 + 𝑚
𝑦 − 𝑚 𝑦 + 𝑚 𝑒 − 𝐿𝑚
) 2
, 𝜑(𝑥, 𝑦) = sup
𝑚 ∈ [1, ∞ ]
𝐹 (𝑥, 𝑦, 𝑚),
𝑀 (𝑥, 𝑦) = { 𝑚 ∈ [1, ∞ ] ∣ 𝐹 (𝑥, 𝑦, 𝑚) = 𝜑 (𝑗) (𝑥, 𝑦) } .
Then
𝜑(𝑥, 𝑦)
hasa minimum(𝑥 0 , 𝑦 0 )
. Inaddition,#𝑀 (𝑗) (𝑥 0 , 𝑦 0 ) ≥ 2
andmin(𝑥 0 , 𝑦 0 ) ≥ 1
.Proof. Note that
𝜑(1, 1) < 1
. Sine𝜑(𝑥, 𝑦) ≥ 𝐹 (𝑥, 𝑦, 1)
and sine𝐹 (𝑥, 𝑦, 1)
tends to1
as(𝑥, 𝑦)
tends to∞
, thereis an𝑀 < ∞
suh that, if(𝑥, 𝑦) ∈ / [1, 𝑀 ] × [1, 𝑀 ]
, then𝜑(𝑥, 𝑦) > 𝜑(1, 1)
. Hene,𝜑
must haveaminimumin
[1, 𝑀 ] × [1, 𝑀 ]
.We now plae ourselves under the hypotheses of the Envelope Theorem. We have that
(𝑥 0 , 𝑦 0 ) ∈ ( 1 2 , ∞ ) 2 = 𝑈
andthehoie𝑉 = [1, ∞ ]
doesthetrik.If
𝑀 (𝑗) (𝑥 0 , 𝑦 0 ) = { 𝑚 }
isasingleton,then𝜑(𝑥, 𝑦)
isdierentiableat𝑥 0 , 𝑦 0
and0 = ∇ 𝜑(𝑥 0 , 𝑦 0 )
= ( ∂𝐹
∂𝑥 (𝑥 0 , 𝑦 0 , 𝑚), ∂𝐹
∂𝑦 (𝑥, 𝑦, 𝑚) )
.
Itsuestolook atthepartialin
𝑥
,0 = − 4𝑚 𝑚 − 𝑥 (𝑚 + 𝑥) 3
( 𝑚 − 𝑦 𝑚 + 𝑦
) 2
𝑒 − 2𝐿𝑚 ,
toonludethateither
𝑚 = 𝑥
or𝑚 = 𝑦
. Ineitherase,𝜑(𝑥, 𝑦) = 𝐹 (𝑥, 𝑦, 𝑚) = 0
,whih isabsurd.In theone-sided ase,this result was suient beause theonly remainingpossibilitywas tohavetwo
pointsequiosillating. However,in thetwo-sidedase,weanequiosillatetwoorthree points,but lemma
4.10only saysthat there are at least twopointsequiosillating. Using thefat that
𝜑(𝑥, 𝑦)
hasone-sidedderivativeseverywhere,andthatforanydiretion
𝛾
thisone-sidedderivative𝐷 𝛾 𝜑(𝑥, 𝑦)
mustbenon-negative at(𝑥, 𝑦 )
wean obtainequiosillationof threepoints. First wemust desribetheshapeof𝐹 (𝑥, 𝑦, 𝑚)
asafuntionof
𝑚
.Lemma 4.11. Let
1 ≤ 𝑥 ≤ 𝑦
(without loss of generality). The two loal maxima𝑚 1 (𝑥, 𝑦) ≤ 𝑚 2 (𝑥, 𝑦)
of𝐹 (𝑥, 𝑦, 𝑚)
inside𝑚 > 1
aregiven by𝑚 2 1 (𝑥, 𝑦) = 2𝑥 + 2𝑦 + 𝐿𝑥 2 + 𝐿𝑦 2 − √ 𝑝 𝑥,𝑦 (𝐿)
2𝐿 ,
𝑚 2 2 (𝑥, 𝑦) = 2𝑥 + 2𝑦 + 𝐿𝑥 2 + 𝐿𝑦 2 + √ 𝑝 𝑥,𝑦 (𝐿)
2𝐿 ,
𝑝 𝑥,𝑦 (𝐿) = (𝑥 2 − 𝑦 2 ) 2 𝐿 2 + 4(𝑥 3 − 𝑥𝑦 2 − 𝑥 2 𝑦 + 𝑦 3 )𝐿 +4(𝑥 + 𝑦) 2 > 0.
Furthermore, if
𝑥 < 𝑦
then𝑚 1 (𝑥, 𝑦) ∈ (𝑥, 𝑦)
and𝑚 2 ∈ (𝑦, ∞ )
. If𝑥 = 𝑦
, then𝑚 1 (𝑥, 𝑥) = 𝑥
and𝑚 2 (𝑥, 𝑥) ∈ (𝑥, ∞ )
.Theritialpoints
𝑚 1
and𝑚 2
alsodependon𝐿
. Whenthatdependenemustbemadeexpliit,wewrite𝑚 1 (𝑥, 𝑦, 𝐿)
and𝑚 2 (𝑥, 𝑦, 𝐿)
.Proof. Wesetthederivative
∂𝐹/∂𝑚
tozero,obtaining0 = ∂𝐹
∂𝑚 (𝑥, 𝑦, 𝑚)
= 2 𝑥 − 𝑚 (𝑥 + 𝑚) 3
𝑦 − 𝑚
(𝑦 + 𝑚) 3 𝑞 𝑥,𝑦 (𝑚)𝑒 − 2𝐿𝑚 , 𝑞 𝑥,𝑦 (𝑚) = − 𝐿𝑚 4 + (2𝑥 + 2𝑦 + 𝐿𝑥 2 + 𝐿𝑦 2 )𝑚 2
− (2𝑦 2 𝑥 − 2𝑥 2 𝑦 − 𝐿𝑥 2 𝑦 2 ).
The hoies
𝑚 = 𝑥
and𝑚 = 𝑦
lead to𝐹 (𝑥, 𝑦, 𝑚) = 0
so they are minima, so we restrit our attentionto
𝑞 𝑥,𝑦 (𝑚) = 0
. Bysubstituting𝑚 2 = 𝑛
weobtaina quadratipolynomialin𝑛
whose roots are𝑚 2 1 (𝑥, 𝑦)
and
𝑚 2 2 (𝑥, 𝑦)
; we need to show that1 ≤ 𝑚 1 ≤ 𝑚 2
. First to show that they are real, we show thatthe disriminant
𝑝 𝑥,𝑦 (𝐿)
is non-negative. It sues to show that, as a quadrati polynomial over𝐿
, itsoeientsarenon-negative. Theoeientof
𝐿 2
andtheonstantoeientaresquares andhenenon-negative. Werewrite
𝑥 = 𝜆𝑦
with𝜆 ∈ (0, 1]
andsubstituteintotheoeientof𝐿
in𝑝 𝑥,𝑦 (𝐿)
to obtaintheexpression
𝑦 3 (4𝜆 3 − 4𝜆 2 − 4𝜆 + 4) > 0
sine𝑦 > 0
. Hene𝑝 𝑥,𝑦 (𝐿) > 0
and𝑚 1 < 𝑚 2
.Assume that
𝑥 < 𝑦
are xed. Sine𝐹(𝑥, 𝑦, 𝑥) = 0
and𝐹(𝑥, 𝑦, 𝑦) = 0
but𝐹 (𝑥, 𝑦, 𝑚) > 0
for any𝑚 ∈ (𝑥, 𝑦)
, there is a loal maximum in𝑚
in the interval(𝑥, 𝑦)
. Moreover, sine𝐹 (𝑥, 𝑦, 𝑦) = 0
andlim 𝑚 →∞ 𝐹 (𝑥, 𝑦, 𝑚) = 0
, there must be another loal maximum in theinterval(𝑦, ∞ )
. However,wehavealready shown that the onlyandidates for loal maxima are
𝑚 1
and𝑚 2
, and sine𝑚 1 ≤ 𝑚 2
it must bethat
𝑚 1 ∈ (𝑥, 𝑦)
and𝑚 2 ∈ (𝑦, ∞ )
.Lemma4.12. Let
(𝑥 0 , 𝑦 0 )
minimize𝜑(𝑥, 𝑦)
. Then𝑀 (𝑥, 𝑦) = { 1, 𝑚 1 (𝑥, 𝑦), 𝑚 2 (𝑥, 𝑦) }
with𝑥 ∕ = 𝑦
.Proof. Thederivativeof
𝜑
inthediretion𝛾 = (𝛼, 𝛽)
isgivenby𝐷 𝛾 𝜑(𝑥, 𝑦) = max
𝑚 ∈ 𝑀(𝑥,𝑦) 𝛼 ∂𝐹
∂𝑥 (𝑥, 𝑦) + 𝛽 ∂𝐹
∂𝑦 (𝑥, 𝑦)
= max
𝑚 ∈ 𝑀(𝑥,𝑦) (𝛼, 𝛽) ⋅ 𝐺(𝑚),
where
𝐺(𝑚) = 4𝑚𝑒 − 2𝐿𝑚 𝑥 − 𝑚 (𝑥 + 𝑚) 2
𝑦 − 𝑚 (𝑦 + 𝑚) 2
( 𝑦 − 𝑚 𝑥 + 𝑚 , 𝑥 − 𝑚
𝑦 + 𝑚 )
.
Inthease
𝑥 = 𝑦
,𝑚 1
isatuallyaminimumso𝑀 (𝑥, 𝑥) = { 1, 𝑚 2 (𝑥, 𝑥) }
andinpartiular𝑥 > 1
. Thevetors( 𝑥 − 1
𝑥 + 1 , 𝑥 − 1 𝑥 + 1
)
and