Contents lists available atScienceDirect
Journal
of
Computational
Physics
www.elsevier.com/locate/jcp
Structure
preserving
schemes
for
the
continuum
Kuramoto
model:
Phase
transitions
José
A. Carrillo
a,
∗
,
Young-Pil Choi
b,
Lorenzo Pareschi
c aDepartmentofMathematics,ImperialCollegeLondon,LondonSW72AZ,UnitedKingdombDepartmentofMathematicsandInstituteofAppliedMathematics,InhaUniversity,Incheon,402-751,RepublicofKorea cDepartmentofMathematicsandComputerScience,UniversityofFerrara,ViaN.Machiavelli35,44121,Ferrara,Italy
a
r
t
i
c
l
e
i
n
f
o
a
b
s
t
r
a
c
t
Articlehistory:
Received10March2018
Receivedinrevisedform6September2018 Accepted27September2018
Availableonline2October2018
Keywords:
Structurepreserving Kuramotooscillators Kineticequations Phasetransitions
TheconstructionofnumericalschemesfortheKuramotomodelischallengingduetothe structural properties of the system whichare essentialin orderto capture the correct physicalbehavior,likethedescriptionofstationarystatesandphasetransitions.Additional difficulties are represented by the high dimensionality of the problem in presence of multiplefrequencies.Inthispaper, wedevelopnumericalmethods whichare capableto preservethesestructuralpropertiesoftheKuramotoequationinthepresenceofdiffusion and tosolve efficiently themultiple frequenciescase.The novelschemes are thenused tonumericallyinvestigatethephasetransitionsinthecaseofidenticalandnonidentical oscillators.
©2018TheAuthors.PublishedbyElsevierInc.ThisisanopenaccessarticleundertheCC BYlicense(http://creativecommons.org/licenses/by/4.0/).
1. Introduction
Synchronization phenomena of large populations of weakly coupled oscillators are very common in naturalsystems, andit has beenextensively studied invarious scientific communities such asphysics, biology,sociology, etc. [2,4,29,52]. Synchronization arises due to the adjustment of rhythms of self-sustained periodic oscillators weakly connected [2,47], anditsrigorousmathematicaltreatmentispioneeredbyWinfree[53] andKuramoto[4].In[4,53],phasemodelsforlarge weaklycoupledoscillatorsystemswereintroduced,andthesynchronizedbehaviorofcomplexbiologicalsystemswasshown to emerge from the competingmechanisms of intrinsicrandomness andsinusoidal couplings. Since then, the Kuramoto modelbecomesaprototypemodelforsynchronizationphenomenaandvariousextensionshavebeenextensivelyexplored invariousscientific communitiessuch asapplied mathematics,engineering, controltheory,physics,neuroscience,biology, andsoon[28,47,52].
Givenan ensembleof sinusoidallycoupled nonlinear oscillators, whichcan be visualized as active rotorson the unit circle
S
1,let zj
=
eiϑj be thepositionofthe j-throtor.Then, thedynamicsof zj iscompletelydetermined bythat ofits phaseϑ
j.Letusdenotethephaseandfrequencyofthe j-thoscillatorbyϑ
jand˙ϑ
j,respectively.Then,thephasedynamics ofKuramotooscillatorsaregovernedbythefollowingfirst-orderODEsystem[4]:˙ϑ
i=
ω
i−
K N N j=1 sin(ϑ
i− ϑ
j),
i=
1,
· · · ,
N,
t>
0,
(1.1)*
Correspondingauthor.E-mailaddresses:carrillo@imperial.ac.uk(J.A. Carrillo),ypchoi@inha.ac.kr(Y.-P. Choi),lorenzo.pareschi@unife.it(L. Pareschi).
https://doi.org/10.1016/j.jcp.2018.09.049
0021-9991/©2018TheAuthors.PublishedbyElsevierInc.ThisisanopenaccessarticleundertheCCBYlicense (http://creativecommons.org/licenses/by/4.0/).
subject to initial data
ϑ
i(
0)
=: ϑ
i0, i=
1,
· · · ,
N, where K is the uniform positive couplingstrength, andω
i denotes the naturalphase-velocity(frequency)andisassumedtobearandomvariableextractedfromthegivendistribution g=
g(
ω
)
satisfying
R
g
(
ω
)
dω
=
1.
Wenoticethatthefirsttermandsecondtermintherighthandsideoftheequation(1.1) representtheintrinsicrandomness andthenonlinearattraction–repulsioncoupling,respectively.
Thesystem(1.1) hasbeenextensivelystudied,anditstillremainsapopularsubjectinnonlineardynamicsandstatistical physics.Wereferthereaderto[2] andreferencesthereinforgeneralsurveyoftheKuramotomodelanditsvariants.In[4], Kuramoto firstobserveda continuousphase transitioninthecontinuum Kuramotomodel(N
→ ∞
,see(1.2) below)with a symmetric distribution function g(
ω
)
by introducing an orderparameter r∞ whichmeasures thedegree of the phase synchronizationinthemean-fieldlimit.Moreprecisely,theorderparameterisgivenbyrN
(
t)
eiϕN(t):=
1 N N j=1 eiϑj(t),
r∞:=
lim t→∞Nlim→∞r N(
t).
In fact,Kuramoto showedthat theorderparameter r∞ asafunction ofthecoupling strength K changesfromzero (dis-ordered state)toanon-zerovalue(orderedstate)whenthecouplingstrength K exceedsacriticalvalue Kc
:=
2/(
π
g(
0))
, i.e., r∞(
K)
=
0 (incoherentstate)for K∈ [
0,
Kc)
,1≥
r∞(
K)
>
0 (coherentstate)forK>
Kc,andr∞(
K)
increaseswith K . Later,itisalsoobservedthatthecontinuumKuramotomodelcanexhibitcontinuous ordiscontinuous phasetransitionby takingintoaccountdifferenttypesofnaturalfrequencydistributionfunctions[11,12,45].Here,continuousphasetransitions refertothecontinuityatK=
Kc oftheorderparameterr∞(
K)
.ItisknowntobecontinuousfortheGaussiandistributed infrequencyoscillators[52] anddiscontinuousfortheuniformlydistributedinfrequencyoscillators[45].As the numberofoscillatorsgoes toinfinity
(
N→ ∞)
,a continuum descriptionofthe system(1.1) canbe rigorously derivedbyemployingbynowstandardmean-fieldlimittechniquesfortheVlasovequation[19,21,40,42].Letρ
=
ρ
(ϑ,
ω
,
t)
betheprobabilitydensityfunctionofKuramotooscillatorsin
ϑ
∈ T
:= R/(
2π
Z)
withanaturalfrequencyω
extractedfrom adistributionfunctiong=
g(
ω
)
attimet.ThenthecontinuumKuramotoequationisgivenby∂
tρ
+ ∂
ϑ(
u[
ρ
]
ρ
)
=
0,
(ϑ,
ω
)
∈ T × R,
t>
0,
u[
ρ
](ϑ,
ω
,
t)
=
ω
−
KT×R
sin
(ϑ
− ϑ
∗)
ρ
(ϑ
∗,
ω
,
t)
g(
ω
)
dϑ
∗dω
,
(1.2)subjecttotheinitialdata:
ρ
(ϑ,
ω
,
0)
=:
ρ0
(ϑ,
ω
),
T
ρ0
(ϑ,
ω
)
dϑ
=
1.
(1.3)Theorderparameterr andtheaveragephase
ϕ
associatedto(1.2) aregivenasr
(
t)
eiϕ(t)=
T×R eiϑρ
(ϑ,
ω
,
t)
g(
ω
)
dϑ
dω
,
leadingto r(
t)
=
T×R cos(ϑ
−
ϕ
(
t))
ρ
(ϑ,
ω
,
t)
g(
ω
)
dϑ
dω
.
(1.4)Forthecontinuum equation(1.2), globalexistence anduniqueness ofmeasure-valued solutionsarestudiedin[21,40] and important qualitative properties such as theLandau damping towards the incoherentstationary stateswere analyzed in [30].
Averyrelevantissuefromtheapplicationviewpointishowstablethesestationarystatesandphasetransitionsareby addingnoisetothesystem[4,50].Themean-fieldlimitequationassociatedto(1.1) withstandardGaussiannoiseofstrength
√
2D,withD>
0,is∂
tρ
+ ∂
ϑ(
u[
ρ
]
ρ
)
=
D∂
ϑ2ρ
,
(ϑ,
ω
)
∈ T × R,
t>
0,
u[
ρ
](ϑ,
ω
,
t)
=
ω
−
K T×R sin(ϑ
− ϑ
∗)
ρ
(ϑ
∗,
ω
,
t)
g(
ω
)
dϑ
∗dω
,
(1.5)subjecttotheinitial data(1.3). Phasetransitionshavealsobeenfound intermsoftheorder parameterr∞ asafunction of K for D
>
0. In fact, it was proven in [50] that foridentical oscillators there is a continuous phase transition. This phasetransitionisalsocontinuousforGaussiananduniformlycouplednoisy oscillators[2] anddiscontinuous forbimodal distributions[16,25].StabilityofthecoherentandincoherentstatesandtheasymptoticdynamicsfortheKuramotomodel withnoise(1.5) wereanalyzed in[31,32,37,38].Noise-drivenphasetransitionsininteractingparticlesystemsareahottopic ofresearchfromthenumericalandtheoreticalviewpoints,see[5–10,25,33] andthereferencestherein.Inthepresentwork,wefocusontheconstructionofeffectivenumericalschemesforthecontinuumKuramotosystem withdiffusionobtainedinthelimit ofaverylarge numberofoscillatorsandinpresenceofnoise.Thenumericalsolution of such system is challenging due to the high dimensionality of the problem and the intrinsic structural properties of the system which are essential in order to capture the correct physical behavior. The literature dealing with the study ofnumericalschemes to thecontinuum Kuramoto model(1.2) ismainly focused ondeterministic spectral andstochastic methods.Wereferto[2,SectionVI.B] andthereferencesthereinforgeneraldiscussionsonthestochasticanddeterministic numerical methods used in Kuramoto models, see also [10,35,36,47] for Monte Carlo related methods and [1,3,46] for spectralmethods.
Structurepreservingnumericalschemesformean-fieldtypekineticequationshavebeenpreviouslyconstructedin[2,14,
17,18,20,22,44,51].Wereferalsoto[34] forrelatedapproachesforgeneralsystemsofbalancelawsandto[27] foran intro-ductiontonumericalmethodsforkineticequations.Here,followingtheapproachin[18,22,44] weintroduceadiscretization ofthephasevariablesuch thatnonnegativity of thesolution,physicalconservations,asymptoticbehavior andfree energy dissipationarepreservedatadiscretelevelforidenticaloscillators.Thisapproachisthencoupledwithasuitablecollocation methodforthefrequencyvariablebasedonorthogonalpolynomialswithrespecttothegivenfrequencydistribution.Similar collocationstrategieshavebeenusedforkineticequationsinthecaseoflineartransportandsemiconductors[39,41,49].
Therestofthemanuscriptisorganizedasfollows.First,inSection2werecallsomebasicpropertiesofthecontinuum Kuramoto model, in particular concerning some relevant existence theory and some useful estimates underpinning our numericalstrategy.WealsodiscusstheasymptoticbehavioroftheKuramotomodelwithnoise,stationarystatesandrelated freeenergyestimatesinthecaseofidenticaloscillators.ThediscretizationofKuramotosystemsisdiscussedinSection 3. Structure preserving schemes are developed in combination with a collocation method for the frequency variable. The asymptoticbehavioroftheschemesaswellastheotherrelevantphysicalpropertiesarediscussed.InSection4wepresent severalnumericaltests,witha particularemphasisonthestudyofphasetransitions.We willalsocompareourstructure preserving methodsto MonteCarloandspectral methods.Futureresearch directionsandconclusions are reportedinthe lastSection.
2. The continuum Kuramoto model
2.1. StabilityoftheKuramotomodelwithrespecttothefrequencydistribution
Inthissubsection,webriefly reviewsome relevantissuesrelatedtothewell-posedness theoryandseveraluseful esti-matesforthecontinuumKuramotoequation(1.2) anditsversionwithnoise(1.5).Wereferto[19,21,40] forfurtherdetails. Wefirstprovidea priori estimatesfortheequation(1.2).
Lemma 2.1. Let
ρ
beasmoothsolutionto(1.2)–(1.3).Thenwehave(
i)
d dt Tρ
(ϑ,
ω
,
t)
dϑ
=
0 i.e., Tρ
(ϑ,
ω
,
t)
dϑ
=
Tρ0
(ϑ,
ω
)
dϑ
=
1 for w∈ R,
(
ii)
d dt T×Rρ
(ϑ,
ω
,
t)
log(
ρ
(ϑ,
ω
,
t))
g(
ω
)
dϑ
dω
=
K r2(
t),
fort
≥
0,wherer isgivenin(1.4).Proof. The proofof
(
i)
isstraightforward.Fortheidentity(
ii)
,byusingthedefinitionofr in(1.4) werewriteu[
ρ
]
asu
[
ρ
](ϑ,
ω
,
t)
=
ω
−
K r sin(ϑ
−
ϕ
(
t)).
(2.1)Directcomputationsyield
d dt
T×Rρ
(ϑ,
ω
,
t)
log(
ρ
(ϑ,
ω
,
t))
g(
ω
)
dϑ
dω
= −
T×R g(
ω
)∂
ϑ(
u[
ρ
]
ρ
)
logρ
dϑ
dω
=
T×R g(
ω
)
u[
ρ
]∂
ϑρ
dϑ
dω
= −
T×R(∂
ϑu[
ρ
])
ρ
(ϑ,
ω
)
g(
ω
)
dϑ
dω
=
K r T cos(ϑ
−
ϕ
(
t))
ρ
(ϑ,
ω
)
g(
ω
)
dϑ
dω
=
K r2,
completingtheproof.
2
We next discussglobalexistence anduniqueness ofmeasure-valuedsolutions. Letusdenoteby
M(T
× R)
the setof nonnegative Radonmeasures onT
× R
,which canbe regardedasnonnegativeboundedlinearfunctionals onC
0(
T
× R)
. Thenournotionofmeasure-valuedsolutionsto(1.2) isdefinedasfollows.Definition 2.1. For T
∈ [
0,
∞)
,letμ
∈
C
w(
0,
T;
M(T
× R))
be ameasure-valuedsolutionto (1.2) withan initial measureμ
0∈
M(T
× R)
ifandonlyifμ
satisfiesthefollowingconditions:•
μ
isweaklycontinuous,i.e.,μ
t,
his a continuous as a function of t
,
∀
h∈
C
0(
T × R),
whereμ
t,
h:=
T×R h(ϑ,
ω
)
μ
t(
dϑ,
dω
).
•
Forh∈
C
10(
T
× R
× [
0,
T))
,μ
satisfiesthefollowingintegralequation:μ
t,
h(
·, ·,
t)
−
μ0
,
h(
·, ·,
0)
=
t 0μ
s, ∂
sh+
u[
μ
]∂
ϑhds
,
whereu[
μ
]
isgivenby u[
μ
](ϑ,
ω
,
t)
:=
ω
−
K T×R sin(ϑ
− ϑ
∗)
g(
ω
)
μ
t(
dϑ
∗,
dω
).
WealsointroducethedefinitionoftheboundedLipschitzdistance.
Definition 2.2. Let
μ
,
ν
∈
M(T
× R)
betwoRadonmeasures.ThentheboundedLipschitzdistanced(
μ
,
ν
)
betweenμ
andν
isgivenbyd
(
μ
,
ν
)
:=
suph∈S
|
μ
,
h−
ν
,
h| ,
wheretheadmissibleset
S
oftestfunctionsisdefined asS
:= {
h: T × R → R :
max{
hL∞,Lip(
h)
} ≤
1}
with Lip(
h)
:=
sup (ϑ,ω) =(ϑ∗,ω∗)|
h((ϑ,
ω
))
−
h((ϑ
∗,
ω
∗))
|
|(ϑ,
ω
)
− (ϑ
∗,
ω
∗)
|
.
Wenowpresenttheresultsontheglobalexistenceandstabilityofmeasure-valuedsolutionsto(1.2).Wereferthereader to[19,21,40] fordetailsoftheproof.
Proposition 2.1. Forany
μ
0∈
M(T
×R)
,(1.2) hasauniquesolutionμ
∈
C
w(
0,
T;
M(T
×R))
withtheinitialdataμ
0.Furthermore,if
μ
andν
aretwosuchsolutionsto(1.2),thenthereexistsaconstantC>
0 dependingonT suchthatRemark 2.1. Onecanalsocharacterizethesolution
μ
oftheinitial-valueproblemfor(1.2) asthepush-forwardoftheinitial dataμ
0throughtheflowmapgeneratedbyu[
μ
]
,i.e.,foranyh∈
C
c1(
T
× R)
andt,
s≥
0 T×R h(ϑ,
ω
)
μ
t(
dϑ,
dω
)
=
T×R h( (
0;
t, ϑ,
ω
),
ω
)
μ0
(
dϑ,
dω
),
wheresatisfies d dt
(
t;
s, ϑ,
ω
)
=
u[
μ
]( (
t;
s, ϑ,
ω
),
ω
,
t)
with(
s;
s, ϑ,
ω
)
= ϑ.
(2.2)Note that the characteristic system (2.2) is well definedsince the velocity field u
[
μ
]
is globally Lipschitz in(ϑ,
ω
)
and continuousont evenforgeneralmeasuresμ
∈
C
w(
0,
T;
M(T
× R))
.Remark 2.2. AsasimpleextensionofProposition2.1,see[21],wealsoobtainthatthesolution
μ
t canbeapproximatedas asumofDiracmeasuresofthefollowingformμ
N t=
1 N N i=1δ
ϑi(t)⊗ δ
ωi,
i.e., d
(
μ
t,
μ
Nt)
→
0 as N→ ∞
. This can be accomplished by choosing a particle approximation of the initial data. One proceduretoconstructparticleapproximationsisthefollowing.Letuschoosethesmallestsquarecontainingthesupportof theinitialdataμ
0,i.e.,supp(
μ
0)
⊂
R.Foragivenn,wedividethesquare R inton2subsquares Ri,i.e.,R
=
n2i=1
Ri
.
Let
ϑ
i0,
ω
i bethecenterofRi fori=
1,
· · · ,
n2.Thenweconstructtheinitialapproximationμ
n0 asμ
n0=
n2 i=1 miδ
ϑ0 i⊗ δ
ωi with mi=
Riμ0
(ϑ,
ω
)
dϑ
dω
.
Thenwecaneasilyshowthatd
(
μ
n0,
μ
0)
≤
C/
n→
0 asn→ ∞
.Fromnow on,weassumethat theinitialmeasureisa smoothabsolutely continuous withrespecttoLebesguedensity withconnectedsupport.ThisassumptioncanberemovedbythestabilitypropertyinProposition2.1.Wenextshowstability oftheorderparameterr withrespecttoapproximatedfrequencydistributionsandinitialdataby againusingthestability estimatepresentedinProposition2.1.
Proposition 2.2. Let
μ
bethemeasure-valuedsolutiontotheequation(1.2)–(1.3) inthetimeinterval[
0,
T]
withg∈
S
andtheinitial dataμ
0∈
M(T
× R)
.Letusdefinernasrn
(
t)
eiϕn(t)=
T×R
eiϑ
ρ
n(ϑ,
ω
,
t)
gn(
ω
)
dϑ
dω
,
where
ρ
nisthemeasure-valuedsolutionto(1.2) withtheinitialdataρ
n0
∈
S
satisfyingd(
ρ
n
0
,
μ
0)
→
0.Thenforany gn
∈
M(R)
satisfyingd(
gn,
g)
→
0 asn→ ∞
,wehavelim
n→∞r
n
(
t)
=
r(
t)
for t∈ [
0,
T].
Proof. We firstestimate the1-Wassersteindistancebetween
ρ
n andμ
.Observethatρ
n(
t)
∈
S
forall t≥
0 bythe push-forwardcharacterizationinRemark2.1.Moreover,italsofollowsfromRemark2.1that T×R h(ϑ,
ω
)
ρ
n(ϑ,
ω
,
t)
dϑ
dω
−
T×R h(ϑ,
ω
)
μ
t(
dϑ,
dω
)
=
T×R h(
n(
0;
t, ϑ,
ω
),
ω
)
ρ
n 0(ϑ,
ω
)
dϑ
dω
−
T×R h( (
0;
t, ϑ,
ω
),
ω
)
μ0
(
dϑ,
dω
)
,
foranyh
∈
C
c1(
T
× R)
andt,
s≥
0,wheren satisfies d dt
n
(
t;
s, ϑ,
ω
)
=
un[
ρ
n](
n(
t;
s, ϑ,
ω
),
ω
,
t)
=
ω
−
K T×R sin(
n(
t;
s, ϑ,
ω
)
− ϑ
∗)
ρ
n(ϑ
∗,
ω
,
t)
gn(
ω
)
dϑ
∗dω
,
with
n
(
s;
s,
ϑ,
ω
)
= ϑ
forall n≥
1 andsatisfies (2.2). Fornotationalsimplicity, we denote
n
(
t;
0,
ϑ,
ω
)
=
n(
t)
and(
t;
0,
ϑ,
ω
)
= (
t)
.Thenastraightforwardcomputationyieldsd dt
|
n(
t)
− (
t)
| ≤
K T×R sin(
n(
t)
− ϑ
∗)
ρ
n(ϑ
∗,
ω
,
t)
dϑ
∗(
gn−
g)(
dw)
+
K T×R sin(
n(
t)
− ϑ
∗)
g(
w) (
ρ
n t−
μ
t)(
dϑ
∗,
dω
)
+
K T×R sin(
n(
t)
− ϑ
∗)
−
sin( (
t)
− ϑ
∗)
g(
w)
μ
t(
dϑ
∗,
dω
)
≤
K d(
gn,
g)
+
K d(
ρ
tn,
μ
t)
+
K|
n(
t)
− (
t)
|.
Thusweconclude|
n(
t)
− (
t)
| ≤
C(
K,
T)
⎛
⎝
d(
gn,
g)
+
t 0 d(
ρ
sn,
μ
s)
ds⎞
⎠ ,
whereC
>
0 isindependentofn.Thistogetherwiththeboundednessofn and
in
[
0,
T]
gives T×R h(
n(
0;
t, ϑ,
ω
),
ω
)
ρ
0n(ϑ,
ω
)
dϑ
dω
−
T×R h( (
0;
t, ϑ,
ω
),
ω
)
μ0
(
dϑ
dω
)
≤
C⎛
⎝
d(
gn,
g)
+
t 0 d(
ρ
n s,
μ
s)
ds⎞
⎠ +
Cd(
ρ
n 0,
μ
0).
Hencewehave d(
ρ
tn,
μ
t)
≤
C⎛
⎝
d(
gn,
g)
+
t 0 d(
ρ
sn,
μ
s)
ds⎞
⎠ +
Cd(
ρ
n0,
μ0
),
i.e., d(
ρ
n t,
μ
t)
≤
Cd(
gn,
g)
+
Cd(
ρ
0n,
μ
0),
whereC
>
0 isindependentofn.Wenowshowthestrongconvergenceofrn tor.Forthis,weusethefollowingidentitiesrn
(
t)
=
T×R eiϑρ
n(ϑ,
ω
,
t)
gn(
ω
)
dϑ
dω
and r(
t)
=
T×R eiϑg(
ω
)
μ
t(
dϑ,
dω
)
,
toobtain|
rn(
t)
−
r(
t)
| ≤
T×R cosϑ
ρ
tngn−
μ
tg(
dϑ,
dω
)
+
T×R sinϑ
ρ
tngn−
μ
tg(
dϑ,
dω
)
=:
I1+
I2.
Here I1 canbeestimatedas I1
≤
T×R cosϑ
ρ
nt−
μ
t g(
dϑ,
dω
)
+
T×R cosϑ
gn−
gρ
tn(
dϑ,
dω
)
≤
d(
ρ
nt,
μ
t)
+
d(
gn,
g),
duetocos
ϑ
g(
ω
),
cosϑ
ρ
nt
(ϑ,
ω
)
∈
S
.Similarly,wealsofind I2≤
d(
ρ
tn,
μ
t)
+
d(
gn,
g)
,concludingthat|
rn(
t)
−
r(
t)
| ≤
I1+
I2≤
2d(
ρ
tn,
μ
t)
+
2d(
gn,
g)
≤
Cd(
gn,
g)
+
Cd(
ρ
n0
,
μ
0)
→
0 as n→ ∞.
Thiscompletestheproof.2
Remark 2.3. AsimilarresulttoProposition2.2canbeobtainedifweapproximatetheinitialdata
μ
0∈
M(T
× R)
andthe frequencydistribution gn(
ω
)
∈
L∞(
R)
asfollows:μ
n0
∈
M(T
× R)
and gn(
ω
)
∈
S
satisfying d(
μ
n0,
μ
0)
→
0 and gn(
·)
−
g
(
·)
L∞(R)→
0 asn→ ∞
.Remark 2.4. AsimilarresulttoProposition2.2canalsobeprovedfortheKuramotomodelwithnoise(1.5).Thestrategyof theproofisanalogousbutitusesstochasticprocessestechniquestowritethecorrespondingstochasticdifferentialequation systems.Moreover,one needstoresorttoWasserstein distancesinsteadoftheboundedLipschitzdistanceabovetomake thestabilityargumentofthesolutions.Wedonotincludethedetailsoftheproofsincethetechnicalitieslieoutsideofthe scopeofthepresentwork,see[15] forrelatedproblems.
NoticethatProposition2.2andRemark2.4allowustoworkwithcontinuousfrequencydistributionsforbothKuramoto models(1.2) and(1.5) byapproximation.Moreprecisely,wecanapproximateGaussiananduniformfrequencydistributions bysumsofDiracDeltasatafinitenumberoffrequencieswhileapproximatingtheinitialdatabysmoothfunctionsifthey arenotregularenough.ThisfactwillbeusedinthenumericalschemesinSection3andthesimulationsinSection4.
2.2. TheKuramotomodelwithdiffusion:stationarystatesandfreeenergies
Wefirstderiveanexplicitcompatibilityconditionforsmoothstationarystates
ρ
∞ oftheequation(1.5).Wefirsteasily findfrom(1.5) and(2.1) thatD
∂
ϑ2ρ
∞= ∂
ϑ((
ω
−
K r∞sin(ϑ
−
ϕ
∞))
ρ
∞) ,
wherer∞ and
ϕ
∞ aregivenbyr∞eiϕ∞
=
T×R
eiϑ
ρ
∞(ϑ,
ω
)
g(
ω
)
dϑ
dω
.
Wenoticethatwecanset
ϕ
∞=
0 inthestationarystatewithoutlossofgeneralitybychoosingtherightangularreference system.Thisyields∂
ϑρ
∞(ϑ,
ω
)
−
ω
−
K r∞sinϑ
D
ρ
∞=
c0(
w),
forsomefunctionc0
(
w)
.Solvingtheabovedifferentialequation,wefindρ
∞(ϑ,
ω
)
=
ρ
∞(
0,
ω
)
exp−
K r∞+
ω
ϑ
+
K r∞cosϑ
D×
⎛
⎝
1+
c0(
ω
)
e K r∞ Dρ
∞(
0,
ω
)
ϑ 0 exp−
ω
ϑ
∗+
K r∞cosϑ
∗ D dϑ
∗⎞
⎠ ,
for
ϑ
∈ [
0,
2π
)
,whereρ
∞(
0,
ω
)
isfixedbythenormalization,i.e.,T
Ontheotherhand,since
ρ
∞(
2π
,
ω
)
=
ρ
∞(
0,
ω
)
,weget 1=
exp 2π ω
D⎛
⎝
1+
c0(
ω
)
e K r∞ Dρ
∞(
0,
ω
)
2π 0 exp−
ω
ϑ
∗+
K r∞cosϑ
∗ D dϑ
∗⎞
⎠ ,
andsubsequently,thisimplies
c0
(
ω
)
e K r∞ Dρ
∞(
0,
ω
)
=
exp(
−
2π ω
/
D)
−
1 2π 0 exp(
−(
ω
ϑ
∗+
K r∞cosϑ
∗)/
D)
dϑ
∗.
Hencewehaveρ
∞(ϑ,
ω
)
=
ρ
∞(
0,
ω
)
exp−
K r∞+
ω
ϑ
+
K r∞cosϑ
D×
1+
(
exp(
−
2π ω
/
D)
−
1)
ϑ 0 exp(
−(
ω
ϑ
∗+
K r∞cosϑ
∗)/
D)
dϑ
∗ 2π 0 exp(
−(
ω
ϑ
∗+
K r∞cosϑ
∗)/
D)
dϑ
∗.
(2.3)LetusdiscussfurtherpropertiesinthecaseofnoisyidenticalKuramotooscillators,whicharegovernedbytheequation (1.5) with g
= δ
0.Thenwecansetω
=
0 withoutlossofgeneralityin(2.3) toconcludethatitsstationarystateisgivenbyρ
∞(ϑ)
=
ρ
∞(
0)
exp−
K r∞+
K r∞cosϑ
D,
where
ρ
∞(
0)
isagainfixedbythenormalizationT
ρ
∞(ϑ)
dϑ
=
1.
Notethatforbothidenticalandnon-identicaloscillatorsifr∞
=
0,i.e.,incoherencestate,weobtainρ
∞(ϑ,
ω
)
=
ρ
(
0,
ω
)
=
12
π
,
for allω
∈ R,
duetothenormalization.Inthiscasetheequation(1.5) hasthestructureofagradientflow.Moreprecisely,ifweset
ξ
[
ρ
](ϑ,
t)
:=
K(
Wρ
)(ϑ,
t)
−
D logρ
(ϑ,
t),
withW
(ϑ)
=
cosϑ
,thenitiseasytocheckthat∂
tρ
+ ∂
ϑ(
ρ
∂
ϑξ
[
ρ
]) =
0.
Usingthisobservation,wenowestimatethefollowingfreeenergy
E
(
t)
:= −
K 2 T(
W∗
ρ
)(ϑ,
t)
ρ
(ϑ,
t)
dϑ
+
D Tρ
(ϑ,
t)
logρ
(ϑ,
t)
dϑ.
Lemma 2.2. Let
ρ
beasmoothsolutiontotheequation(1.5) withg= δ
0.Thenwehaved
dt
E
(
t)
= −
D
E(
t),
wherethedissipationrate
D
E(
t)
isgivenbyD
E(
t)
:=
T
|∂
ϑξ
[
ρ
]|
2ρ
dϑ.
Proof. A straightforwardcomputationyields
d dt
E
(
t)
= −
T(
K(
Wρ
)
−
D logρ
) ∂
tρ
dϑ
= −
Tξ
[
ρ
]∂
tρ
dϑ
= −
T|∂
ϑξ
[
ρ
]|
2ρ
dϑ.
2
Lemma 2.3. Let
ρ
beasmoothsolutiontotheequation(1.5) withg= δ
0.Thenwehave˙
r=
K r T sin2(
ϕ
− ϑ)
ρ
dϑ
−
Dr.
Inparticular,ifthestrengthofnoiseD isstrongenoughsuchthatD
≥
K ,then˙
r≤
0 forallt≥
0.Proof. By definitionoftheorderparameterr,weget
˙
r= ˙
ϕ
T sin(ϑ
−
ϕ
)
ρ
(ϑ,
t)
dϑ
+
T cos(ϑ
−
ϕ
)∂
tρ
(ϑ,
t)
dϑ
=
I1+
I2,
whereI1vanishessinceI1
=
˙
ϕ
r T×T sin(ϑ
− ϑ
∗)
ρ
(ϑ,
t)
ρ
(ϑ
∗,
t)
dϑ
dϑ
∗=
0.
FortheestimateofI2,wefind
I2
=
T cos(ϑ
−
ϕ
)
D∂
ϑ2ρ
− ∂
ϑ(
u[
ρ
]
ρ
)
dϑ
= −
D T cos(ϑ
−
ϕ
)
ρ
dϑ
−
T sin(ϑ
−
ϕ
)
u[
ρ
]
ρ
dϑ
= −
Dr+
K T sin2(ϑ
−
ϕ
)
ρ
dϑ.
Combiningtheabovetwoestimatesconcludesthedesiredresult.
2
Beforepassingtotheconstructionofnumericalmethodsthefollowingremarkshouldbemade.
Remark 2.5. Severalworksforthecontinuum Kuramoto modelarebased onthe g-weighted kinetic density f
(ϑ,
ω
,
t)
:=
ρ
(ϑ,
ω
,
t)
g(
ω
)
(see[13] forexample).Inthisway,onecanrewritetheKuramotomodel(1.5) as∂
tf+ ∂
ϑ(
v[
f]
f)
=
D∂
ϑ2f,
with f(ϑ,
ω
,
t)
=
ρ
(ϑ,
ω
,
t)
g(
ω
),
(2.4) v[
f](ϑ,
ω
,
t)
=
ω
−
KT×R
sin
(ϑ
− ϑ
∗)
f(ϑ
∗,
ω
,
t)
dϑ
∗dω
.
Notethatthedistributionofthenaturalfrequenciesisnowgivenby
g
(
ω
)
=
T
f
(ϑ,
ω
,
t)
dϑ.
Evenif,inthecontinuation,wewillusetheform(1.5) fortheconstructionofourstructurepreservingnumericalmethods, theycanbeeasilyreformulatedtotheform(2.4).InSection4wewillshowboth
ρ
(ϑ,
ω
,
t)
and f(ϑ,
ω
,
t)
insomenumerical tests.3. Structure preserving methods
The goalnow isto propose finitevolume numericalschemes preserving the structureof gradientflow to the caseof identicaloscillatorsandthataregeneralizableforoscillatorswithnaturalfrequenciesgivenbyadistributionfunction g
(
w)
. Tostartwith,from(2.1) and(1.5),itfollowsthatthedensityρ
satisfiesthefollowingcontinuityequation∂
tρ
= ∂
ϑF[
ρ
],
(3.1)with
3.1. Semi-discretestructurepreservingschemes
Inspired by [17,18,22,44,51], we construct a discrete numerical scheme in the variable
ϑ
for the above equation as follows.Fori=
1,
· · · ,
N,wefirstintroduceauniform spatialgridϑ
i∈ T
suchthatϑ
i+1− ϑ
i= ϑ
andϑ
N+k= ϑ
kfork∈ R
. Withoutlossofgenerality,wesetϑ
1/2= ϑ
N+1/2=
0≡
2π
,andwethendefineρ
i(
ω
,
t)
:=
1ϑ
ϑi+1/2 ϑi−1/2ρ
(ϑ,
ω
,
t)
dϑ.
Weconsiderthefollowingapproximationsfor(3.1)
d
dt
ρ
i(
ω
,
t)
=
Fi+1/2
[
ρ
](
ω
,
t)
−
Fi−1/2[
ρ
](
ω
,
t)
ϑ
for i=
1,
· · · ,
N,
(3.3)wherethenumericalfluxfunction Fi±1/2
[
ρ
](
ω
,
t)
isgivenbyFi+1/2
[
ρ
](
ω
,
t)
:=
Dρ
i+1−
ρ
iϑ
−
ui+1/2ρ
˜
i+1/2,
(3.4) with ui+1/2(
ω
,
t)
:=
1ϑ
ϑi+1 ϑi u(ϑ,
ω
,
t)
dϑ,
ρ
˜
i+1/2:= (
1− δ
i+1/2)
ρ
i+1+ δ
i+1/2ρi,
and u(ϑ,
ω
,
t)
=
ω
+
Kϑ
N j=1 sin(ϑ
j− ϑ)
Rρ
j(
ω
∗,
t)
g(
ω
∗)
dω
∗.
(3.5)As in[22,44] asuitablechoiceoftheweightfunctions
δ
i+1/2 yieldsamethodthatmaintainsnonnegativityofthesolution (withoutrestrictionsonϑ
)andpreservesthesteadystateofthesystemwitharbitraryorderofaccuracy.Wewillreferto theschemesobtainedinthiswayasChang–Coopertypeschemes[22].First,observethatwhenthenumericalflux(3.4) vanishesweget
ρ
i+1ρ
i=
Dϑ
+ δ
i+1/2ui+1/2 Dϑ
− (
1− δ
i+1/2)
ui+1/2.
(3.6)Similarly,ifweconsidertheexactflux(3.2),byimposing F
[
ρ
]
≡
0,wehaveD
∂
ϑρ
=
uρ
.
Integratingtheaboveequationonthecell
[ϑ
i,
ϑ
i+1]
wegetϑi+1
ϑi 1ρ
∂
ϑρ
ϑ
=
ϑi+1 ϑi u Ddϑ,
whichgivesρ
i+1ρ
i=
expϑ
D ui+1/2.
(3.7)Therefore,byequating(3.6) and(3.7) werecover
δ
i+1/2=
1ξ
i+1/2+
1 1−
exp(ξ
i+1/2)
withξ
i+1/2= −
ϑ
D ui+1/2.
(3.8)Wecanstatethefollowing.
Proposition 3.1. Thenumericalfluxfunction(3.4) with
δ
i+1/2definedby(3.8) vanisheswhenthecorrespondingflux(3.2) isequaltoProof. The latterresultfollowsfromthesimpleinequalityexp
(
x)
≥
1+
x.2
Remark 3.1. Since r cos(
ϕ
− ϑ
i)
= ϑ
N j=1 cos(ϑ
j− ϑ
i)
Rρ
j(
ω
,
t)
g(
ω
)
dω
,
weget ui+1/2(
ω
,
t)
=
ω
+
K N j=1 cos(ϑ
j− ϑ
i+1)
−
cos(ϑ
j− ϑ
i)
R
ρ
j(
ω
,
t)
g(
ω
)
dω
=
ω
+
K rϑ
(
cos(
ϕ
− ϑ
i+1)
−
cos(
ϕ
− ϑ
i)) .
Ontheother hand,the aboveexpressionforui+1/2 isnot thatusefulinpracticesinceitcontains r whichdependson
ρ
. Thus,itwouldbemoretechnicallyusefultowriteui+1/2
(
ω
,
t)
=
ω
+
2K sin( ϑ/
2)
N j=1 sin(ϑ
j− ϑ
i+1/2)
Rρ
j(
ω
,
t)
g(
ω
)
dω
=
ω
+
2K sin( ϑ/
2)
⎛
⎝
cos(ϑ
i+1/2)
N j=1ρ
sj(
t)
−
sin(ϑ
i+1/2)
N j=1ρ
cj(
t)
⎞
⎠ ,
(3.9) wherewesetρ
sj(
t)
:=
sin(ϑ
j)
Rρ
j(
ω
,
t)
g(
ω
)
dω
andρ
cj(
t)
:=
cos(ϑ
j)
Rρ
j(
ω
,
t)
g(
ω
)
dω
.
Remark 3.2. Theresultingschemeissecondorderaccuratein
ϑ
forD>
0,anddegeneratetosimplefirstorderupwinding inthelimitcaseD=
0.Infact,itisimmediatetoshowthatasD→
0 weobtaintheweightsδ
i+1/2=
0
,
ui+1/2<
0,
1
,
ui+1/2>
0.
Inthelemmabelow,weshowthatthenumericalschemeconservesthemass.
Lemma 3.1. Considerthenumericalscheme(3.3).Thenwehave
d dt N
i=1ρ
i(
ω
,
t)
=
0 for(
ω
,
t)
∈ R × R
+.
Proof. It followsfromtheperiodicityofdomainthat
ρ
N+1=
ρ1
andρ
N=
ρ0
.
(3.10)Similarly, we getuN+1/2
=
u1/2 andsubsequentlythisimpliesξ
N+1/2= ξ
1/2,δ
N+1/2= δ
1/2,andρ
˜
N+1/2= ˜
ρ
1/2.From the aboveproperties,wecaneasilyobtainFN+1/2
[
ρ
](
ω
,
t)
=
F1/2[
ρ
](
ω
,
t)
for(
ω
,
t)
∈ R × R
+.
This,togetherwiththefollowingstraightforwardcomputationd dt N
i=1ρ
i(
ω
,
t)
=
1ϑ
N i=1 Fi+1/2[
ρ
](
ω
,
t)
−
Fi−1/2[
ρ
](
ω
,
t)
=
1ϑ
FN+1/2[
ρ
](
ω
,
t)
−
F1/2[
ρ
](
ω
,
t)
=
0,
We next providethe positivity preservation whose proof can be obtained by usingalmost sameargument asin [44, Proposition 1]. However, for the completeness of thiswork, we sketch the proof in the propositionbelow. For this, we introducethetimediscretizationtn
=
nt with
t
>
0 andn∈ N
0andconsiderthefollowingforwardEulermethod
ρ
ni+1(
ω
,
t)
=
ρ
ni(
ω
,
t)
+
tF n i+1/2[
ρ
](
ω
,
t)
−
Fni−1/2[
ρ
](
ω
,
t)
ϑ
for i=
1,
· · · ,
N,
(3.11) where Fni+1/2[
ρ
](
ω
,
t)
:=
Dρ
n i+1−
ρ
n iϑ
−
u n i+1/2ρ
˜
in+1/2.
Proposition 3.2. Supposethatg iscompactlysupportedandthetimestep
t satisfies
t
≤
( ϑ)
2
2
(
C0ϑ
+
D)
where C0
= |
supp(
g)
| +
K.
Thentheexplicitscheme(3.11) preservesnonnegativity,i.e.,
ρ
n+1i
≥
0 ifρ
in≥
0 fori=
1,
· · · ,
N.Proof. It followsfrom(3.11) that
Fni+1/2
[
ρ
](
ω
,
t)
−
Fin−1/2[
ρ
](
ω
,
t)
=
ρ
in+1 Dϑ
−
u n i+1/2(
1− δ
ni+1/2)
+
ρ
in uni−1/2(
1− δ
in−1/2)
−
uni+1/2δ
ni+1/2−
2Dϑ
+
ρ
ni−1 Dϑ
+
u n i−1/2δ
in−1/2.
Ontheotherhand,weeasilyfind
D
ϑ
−
u n i+1/2(
1− δ
ni+1/2)
=
Dξ
i+1/2ϑ
1−
1 1−
expξ
i+1/2≥
0,
duetox
(
1− (
1−
ex)
−1)
≥
0 forx∈ R
.Similarly,wegetD
ϑ
+
u n i−1/2δ
ni−1/2=
Dϑ
ξ
i−1/2 expξ
i−1/2−
1≥
0.
Wealsonoticethat
max 0≤i≤N
uni−1/2(
1− δ
in−1/2)
−
uni+1/2δ
ni+1/2−
2Dϑ
≤
max 0≤i≤N|
uni+1/2| + |
uni−1/2|
+
2Dϑ
≤
2|
supp(
g)
| +
K+
Dϑ
.
This, together with the property of convex combination, yields that the nonnegativity is preserved ifthe time step
t
satisfies
t
≤
ϑ
2
(
|
supp(
g)
| +
K+
D/ ϑ)
.
2
Remark 3.3. The parabolicstability restriction