Renormalization
Group Method and
its
Application
to
Coupled
Oscillators
Department of Applied MathematicsandPhysics
Kyoto University,Kyoto, 606-8501, Japan
HayatoCHIBA *1
Apri1 8,2008
Abstract: The renormalizationgroup(RG)method fordifferentialequations is
one
oftheperturba-tionmethods whichprovidesnotonlyapproximate solutionsbut alsoapproximatevectorfields. Some
topologicalproperties of
an
original equation,suchas
theexistenceofa
normallyhyperbolicinvariant
manifoldandits stability
are
showntobeinherited fromthoseof theRG equation. This factis appliedtothe Kuramoto model and thestability oftheinvarianttoruswill bedetermined.
1
lntroduction
Therenormalizationgroup(RG)method fordifferentialequations is
one
ofthe perturbation methodsfor obtaining solutions which approximate exact solutions for a long time interval. In their papers
[1,2], Chen, Goldenfeld, Oonohaveestablished the RGmethod forordinarydifferential equationsof
theform
$\dot{x}=\frac{dx}{dt}=f(t,x)+\epsilon g(t, x),$ $x\in R^{n}$, (1)
where$\epsilon^{\backslash }>0$is
a
small parameter. For thisequation,themethodforderivingapproximate solutionsoftheform
$x(t)=x_{0}(t)+\epsilon x_{1}(t)+\epsilon^{2}x_{2}(t)+\cdots$ (2)
is called the naive expansion or the regularperturbation method, where $x_{i}(t)$’s aregovemed by
in-homogeneous linear ODEs obtained by puttingEq.(2) into Eq.(l) andequating the coefficients of$\epsilon^{i}$
of theboth sides ofBq.(l). Itis well knownthat approximate solutions constructed by the naive
ex-pansion
are
valid only in atime interval of$O(1)$ in general, since secularterms divergeas
$tarrow\infty$.
Many techniques for obtainingapproximatesolutions whicharevalidinalong time interval have been
developeduntil now, which
are
collectivelycalledsingularperturbation methods.TheRG method proposed by Chen etal. is
one
of the singular perturbation methods looking likethe variation-of-constantmethod, inwhich the secular terms included in$x_{1}(t),x_{2}(t),$$\cdots$ ofEq.(2)
are
renormalized intotheintegral constant of$x_{0}(t)$
.
TheODEto be satisfied by therenormalized integralconstantiscalledthe$RG$equation.
Intheir
papers
[1,2], it is notclearwhy the RGmethod works well. Kunihiro [9,10] revealed thereason
bycharacterizingtheRGequationas
an
equationforobtainingan
envelopeofafamilyofcurves
constructedby thenaive expansion. Hisidea
gave an
intuitiveconcept ofthe RGmethod, however,theproblems below remainedtobe solved.
(i)An explicitdefinitionof theRGequation
was
notgiven.(ii)It
was
notclearwhetherwe can
detect theexistenceofan
invariantmanifold andits stability.The problem(i)
was
solvedby Ziane [11], DeVilleetal. F7], and Chiba [3]. Ziane and DeVilleetal.gavethedefinition of the firstorderRGequation byusingtheaveragingoperator. Further,Chibagave
the formula for them-th orderRG equationby calculating the envelope ofnaive expansion solutions
upto arbitraryorder of$\epsilon$. Withtheseformulas,
one can
obtaina
higherorderRG equation by usinga
computersoftwarelikeMathematica.
While the problem (i) is
a
computational issue, the problem (ii) ismore
essential. Itwas
shownby Ziane [11], DeVille etal. [7], and Chiba [3] that
an
approximate solution constructed by the RGmethod is close to
an
exact solution ina
time interval $-T/\epsilon<t<T/\epsilon$, where $T$ issome
positiveconstant(seeThm.10for the exactstatement). Now consider thefollowing situation: Supposethat
an
approximatesolutionobtained bytheRGmethodis
a
periodicsolution. However,theoriginalequationmay nothave aperiodic orbit becauseofthe
error
oftheapproximate solution(seethefigure below).In general, since approximate solutions have small errors, it is not obvious that the RG method
can
show theexistenceof
an
invariantmanifold. Further,we
can
notunderstandan
asymptotic behavior ofanexact solution becausethetimeinterval$-T/\epsilon<t<T/\epsilon$is finitein general. Inparticular, we may
notunderstand thestability of
an
invariantmanifold.approximate solution exact solution
Thisproblem
was
solved by Chiba [3]. He considered constructingan
approximatevectorfieldin-stead ofanapproximate solution. It
was
considered that whether there exists anapproximatedifferen-tialequation satisfied byafamilyofapproximate solutions. Itisnotobviousbecausetwoapproximate
solutions may intersect with each other and the uniqueness ofsolutions is violated. However, under
appropriate assumptions, Chiba[3]provedthat
a
family ofapproximatesolutionsdefinesa
vectorfieldwhichissufficiently close to the originalvectorfield associated with the originalequation (Thm.9).
Once
we
obtain the approximate vectorfield,we can use
powerfultoolsofdynamicalsystemstheory.In particular, by using the invariant manifold theory,
we can prove
that if the RG equation hasa
normally hyperbolic invariant manifold $N$, then the original equation also has
an
invariant manifold$N_{\epsilon}$, whichis diffeomorphicto$N$
.
Further thestability of$N_{\epsilon}$ coincideswiththat of$N$(Thm. 11).fact, itis shown that symmetries of the original equation (group invariance)
are
inheritedto those oftheRGequation. Furthermore,the RGequation is invariantunder the action ofthe l-parameter
group
defined by the flow of the unperturbed vector fieldof the original equation (Thm.12). It
means
thattheRG equationhas
more
symmetries than theoriginal equationhas, and it iseasierto solvethantheoriginalequation.
The fact that the RG method unifiesthe traditional singular perturbationmethods
was
alreadysug-gested by Chen, Goldenfeld, Oono [1,2] (without mathematical proofs). Ei, Fujii, Kunihiro [8]
sug-gestedthat theRG method provides
an
approximatecentermanifold,andthisfactwas
proved by Chiba[5]by using invariantmanifold theory. The equivalence of the normal forms of vector fields and the
RG equations
was
provedby DeVilleetal. [7] for thecase
of the firstorderRGequation. Thisresultwasextendedtothehigher order RG equation by Chiba [4]. Further,byfocusing the nonuniqueness of
thehigherorderRGequation,Chiba[4] showedthat
we can
construct the higherorderRG equationso
thatit isequivalentto the hyper-normal form. Inaddition, it is knownthat the RG method unifies the
averagingmethod, the multi-scalemethod, thegeometric singularperturbation, and theLiesymmetry
2 Definitions
Let $f$ be
a
time independent $C^{r}$ vector fieldon a
$C^{r}$ manifold $M$and $\varphi$:
$R\cross Marrow M$ its flow.We denote by$\varphi_{t}(x_{0})\equiv x(t),$ $t\in R$, a solutionto the ODE$\dot{x}=f(x)$ through$x_{0}\in M$, which satisfies
$\varphi_{t}\circ\varphi_{s}=\varphi_{t+s},$ $\varphi 0=id_{M}$,where $id_{M}$ denotes the identity mapof$M$
.
Fora
fixed $t\in R,$ $\varphi_{t}$ : $Marrow M$definesa diffeomorphism of$M$
.
Weassume
$\varphi_{t}$ isdefined for all$t\in R$.
For
a
time-dependentvectorfield, let$x(t, \tau,\xi)$denotea
solutiontoan
ODE$\dot{x}(t)=f(t, x)$through$\xi$at$t=\tau$, whichdefinesaflow$\varphi$ : $R\cross R\cross Marrow M$by$\varphi_{t.\tau}(\xi)=x(t, \tau, \xi)$
.
Forfixed$t,$$\tau\in R,$ $\varphi_{t.\tau}$ : $Marrow M$is
a
diffeomorphism of$M$satisfying$\varphi_{t,l}\circ\varphi_{t’,r}=\varphi_{t\tau},$ $\varphi_{t,t}=id_{M}$
.
(3)Conversely,
a
familyofdiffeomorphisms $\varphi_{t,\tau}$of$M$, whichare
$C^{1}$ with respect to$t$and$\tau$, satisfying
the aboveequalityforany$t,$$\tau\in R$defines
a
time-dependentvectorfieldon
$M$through$f(t,x)= \frac{d}{d\tau}|_{\tau=t}\varphi_{\tau,r}(x)$. (4)
Definition 1. Let $f$ be
a
vector fieldon
$M$, and $\varphi_{t}$ its flow. A submanifold $N$ of $M$ is calledf-invariant if$\varphi_{t}(N)=N$ for$\forall t\in R$
.
An$f$-invariantmanifold$N$iscalledhyperbolic, if thereare
vectorbundles $E^{s},$ $E^{u}$
over
$N$s.t.(i) $TM|_{N}=E^{s}\oplus E^{u}\oplus TN$,
(ii)both $E^{s}\oplus TN$and$E^{u}\oplus TN$
are
$D\varphi_{t}$-invariant,(iii)thereexistconstants$C\geq 1,$ $\alpha,\beta>0$s.t. for$\forall p\in N$,
$\nu\in E_{p}^{s}$ $\Rightarrow$ $||\pi^{s}o(D\varphi_{t})_{p}\nu||\leq Ce^{-\alpha}$‘, $t\geq 0$, (5) $v\in E_{p}^{u}$ $\Rightarrow$ $||\pi^{u}\circ(D\varphi_{-t})_{p}\nu||\leq Ce^{-\beta t},$ $t\geq 0$, (6)
where$\pi^{s},$$\pi^{u}$
are
projectionsfrom$TM|_{N}$ to $E^{s},$ $E^{u}$,respectively.Definition 2. A hyperbolic invariant manifold $N$ is called r-normally hyperbolic, ifthere exist an
integer $r\geq 1$ and constants $C\geq 1,$ $\gamma>0$, such that for $\forall p\in N,$$\nu\in E_{p}^{s},$ $w\in E_{p}^{u},$ $u\in T_{p}N$, the
following inequalities hold.
$||(D\varphi_{t})_{p}u||^{k}||\pi^{s}o(D\varphi_{t})_{p}\nu||\leq Ce^{-\gamma t}||u||^{k}||v||,$ $k=0,1,$
$\cdots,$ $r,$ $t\geq 0$, (7)
$||(D\varphi_{-t})_{p}u||^{k}||\pi^{u}\circ(D\varphi_{-t})_{\rho}w||\leq Ce^{-\gamma t}||u||^{k}||w||,$ $k=0,1,$$\cdots$ ,$r,$ $t\geq 0$
.
(8)Next theoremis
one
of the fundamental theorem of theinvariantmanifoldtheory.Theorem3. (Fenichel, 1971)
Let$M$be
a
$C^{r}$manifold $(r\geq 1)$, and$X^{r}(M)$the setof$C^{r}$vectorfieldson
$M$with $C^{1}$ topology. Let$f$-invariantmanifold. Then, the following holds:
(i) There isaneighborhood$u\subset X^{r}(M)$of$f$ s.t. there exists
an
r-normallyhyperbolic g-invariant $C^{r}$manifold$N_{g}\subset M$for$\forall g\in u$
.
(ii)$N_{g}$ is diffeomorphicto$N$and the diffeomorphism $h$
:
$N_{g}arrow N$is closetothe identity$id$ : $Narrow N$in the$C^{1}$ topology. In particular, $N_{g}$lies within
an
$O(\epsilon)$neighborhoodof$N$if$||f-g||\sim O(\epsilon)$.
3 RG
method
In this section,
we
give the definition ofthe RG equation and fundamental theorems of the RGmethod. Allproofs
are
given inChiba [3].Consider
an
ODEon
$R^{n}$ of the formab$=Fx+\epsilon g(t, x,\epsilon)$
$=Fx+\epsilon g1(t,x)+\epsilon^{2}g_{2}(t, x)+\cdots,$ $x\in R^{n}$, (9)
where$\epsilon\in R$is
a
smallparameter. For this system,we suppose
that(Al) thematrix$F$is
a
diagonalizable$nxn$constantmatrixallofwhoseeigenvalues lieon
theimaginaryaxis.
(A2) thefunction$g(t, x, \epsilon)$is of$C^{\infty}$ class with respectto$t,$$x$and$\epsilon$
.
The formalpowerseriesexpansionof$g(t, x, \epsilon)$in$\epsilon$is given
as
above.(A3) each$g_{i}(t, x)$is periodicin $t\in R$andpolynomial in$x$
.
Remark4. These assumptionscanbeweakenedinvarious ways. For example, if$F$has eigenvalues
on
the left half plane,our
method provides the centermanifoldreduction (see Sec.5.3). Thecase
thatthe unperturbedtermis nonlinearistreated in Sec.5.4. Theassumption(A3)
can
be replacedas :
(A3’) each$g_{j}(t, x)$is almostperiodic functions suchthat the set of whoseFourierexponents does not
haveaccumulationpoints.
If the setof Fourierexponentsof$g_{i}(t, x)$hasaccumulationpoints,theRGtransformationdefinedbelow
may divergeas$tarrow\infty$ (seeChiba [3]).
At first, let
us
attemptthe naive expansion. Replacing $x$ in (9) by $x=x_{0}+\epsilon x_{1}+\epsilon^{2}x_{2}+\cdots$, werewrite(9)
as
$\dot{x}_{0}+\epsilon\dot{x}_{1}+\epsilon^{2}\dot{x}_{2}+\cdots=F(x_{0}+\epsilon x_{1}+\epsilon^{2}x_{2}+\cdots)+\sum_{i=1}^{\infty}\epsilon^{i}g_{i}(t, x_{0}+\epsilon x_{1}+\epsilon^{2}x_{2}+\cdots )$
.
(10)each$\epsilon^{j}$ ofthe both sides, weobtain ODEsof
$x_{0},$$x_{1},$ $x_{2},$$\cdots$ as
$\dot{x}_{0}=Fx_{0}$, (11)
$\dot{x}_{1}=Fx_{1}+G_{1}(t, x_{0})$, (12)
.
$\dot{x}_{i}=Fx_{i}+G_{i}(t,x_{0}, x_{1}, \cdots, x_{i-1})$, (13)
.
wherethe inhomogeneousterm $G_{l}$ is
a
smooth function of$t,$ $x_{0},$ $x_{1},$$\cdots$ ,$x_{i-1}$.
Forinstance, $G_{1},$ $G_{2},$ $G_{3}$and$G_{4}$
are
given by$G_{1}(t, x_{0})=g_{1}(t, x_{0})$, (14)
$G_{2}(t, x_{0}, x_{1})= \frac{\partial g1}{\partial x}(t, x_{0})x_{1}+g_{2}(t, x_{0})$, (15)
$G_{3}(t, x_{0}, x_{1}, x_{2})= \frac{1}{2}\frac{\partial^{2}g_{1}}{\partial x^{2}}(t, x_{0})x_{1}^{2}+\frac{\acute{c}ig_{1}}{\partial x}(t, x_{0})x_{2}+\frac{\partial g_{2}}{\partial x}(t, x_{0})x_{1}+g3(t,x_{0})$, (16)
$G_{4}(t, x_{0}, x_{1}, x_{2}, x_{3})= \frac{1}{6}\frac{\partial^{3}g_{1}}{\partial x^{3}}(t, x_{0})_{X_{1}^{\sim}}^{3}+\frac{\partial^{2}g_{1}}{\partial x^{2}}(t, x_{0})x_{1}x_{2}+\frac{\partial g_{1}}{\partial x}(t, x_{0})x_{3}$
$+ \frac{1}{2}\frac{\partial^{2}g_{2}}{\partial x^{2}}(t, x_{0})x_{1}^{2}+\frac{\partial g2}{\partial x}(t, x_{0})x_{2}+\frac{\partial g3}{\partial x}(t, x_{0})x_{1}+g_{4}(t, x_{0})$, (17)
respectively. Wecanverify theequality (seeLemmaA.2ofChiba [3] forthe$proof\gamma$
$\frac{\partial G_{i}}{\partial x_{j}}=\frac{\partial G_{i-1}}{\partial x_{j-1}}=\cdots=\frac{\partial G_{i-j}}{\partial x_{0}},$ $i>j\geq 0$, (18)
anditmayhelp in deriving$G_{i}$andprovingProp.5 below.
Solving the system above and constructing the
curve
$x(t)=x_{0}(t)+\epsilon x_{1}(t)+\epsilon^{2}x_{2}(t)+\cdots$ is calledthe naive expansion
as
is mentioned in Sec.1. According to the Kunihiro’s idea, the RG equation isgiven
as an
equationfor obtainingan
envelope ofa
family of the naiveexpansion solutions. Now letus
derive thenaive expansionsolutions.In what follows,
we
denote the fundamental matrix $e^{Ft}$as
$X(t)$. Define the functions $R_{j},$$h_{t}^{(i)},$ $i=$$1,2,$$\cdots$
on
$R^{n}$ by$R_{1}\circ)$ $:= \lim_{tarrow\infty}\frac{1}{t}\int^{t}X(s)^{-1}G_{1}(s, X(s)y)ds$, (19)
$h_{t}^{(1)}(y)$ $:=X(t) \int^{t}(X(s)^{-1}G_{1}(s, X(s)y)-R_{1}(y))ds$, (20)
$R_{i}(y):= \lim_{tarrow\infty}\frac{1}{t}\int^{t}(X(s)^{-1}G_{i}(s,X(s)y, h_{s}^{(1)}(y), \cdots, h_{s}^{(i-1)}(v))$
$-X(s)^{-1} \sum_{k=1}^{i-1}(Dh_{s}^{(k)})_{y}R|-k(y))ds,$ $i=2,3,$$\cdots$ , (21)
$h_{t}^{(i)}(y):=X(t) \int^{t}(X(s)^{-1}G_{i}(s,X(s)y, h_{s}^{(1)}(y), \cdots, h_{s}^{(i-1)}(y))$
respectively. Then,the following statement holds.
Proposition5. (Chiba[3]) Let$x_{0}(t)=X(t)y$bethe solutionto Eq.(ll)whoseinitial value is$y\in R^{n}$
.
Then, for arbitrary$\tau\in R$and$i=1,2,$$\cdots$, the
curve
$x_{i}:=x_{i}(t, \tau;y)=h_{t}^{(i)}O’)+p_{1}^{(i)}(t,y)(t-\tau)+p_{2}^{(i)}(t,y)(t-\tau)^{2}+\cdots+p_{i}^{(i)}(t,y)(t-\tau)^{j}$ (23)
givesasolution toEq.(13), where the functions$p_{1}^{(i)},$ $\cdots$ ,$p_{i}^{(i)}$
are
givenby$p_{1}^{(i)}(t,y)=X(t)R_{i}( J)+\sum_{k=1}^{i-1}(Dh_{t}^{(k)})_{y}R_{i}k(y)$, (24)
$p_{i}^{(i)}(t,y)= \frac{1}{j}\sum_{-k-1}^{i-1}\frac{\partial p_{i-1}^{(k)}}{\partial v}(t,y)R_{i-k}(y),$ $(j=2,3, \cdots, i-1)$, (25)
$p_{\dot{f}}^{(i)}(t,y)= \frac{1}{i}\sum_{k=1}^{i-1}\frac{\partial p_{i-1}^{(k)}}{\partial y}(t,y)R_{i-k}(y)=\frac{1}{i}\frac{\partial p_{i-1}^{(i-1)}}{\partial y}(t,y)R_{1}(\gamma)$, (26)
$p_{j}^{(i)}(t,\}’)=0,$ $(j>i)$
.
(27)Further,the functions $h_{i}^{(i)}(y)$
are
boundeduniformlyin$t$.
Remark 6. Note that
we
gave
the solution to Eq.(13)so
thatit is split into the bounded term $h_{t}^{(i)}$andthe divergence terms. Inparticular,the linearly increasingterm$p_{1}^{(i)}(t,y)(t-\tau)$is called the secular
$tem$
.
To
see
whatwe
did,letus
derivethe functions$R_{1}$ and$h_{t}^{(1)}$.
Withthe0-th ordersolution$x_{0}(t)=X(t)y$,the first orderequation (12)is written as
$\dot{x}_{1}=Fx_{1}+G_{1}(t,X(t)y)$
.
(28)The solutiontothisequationis givenby
$x_{1}=X(t)X(\tau)^{-1}h+x(t)l^{r_{X(S)^{-1}G_{1}(s,X(t)y)ds}}$, (29)
where $h$ is
an
initial value and $\tau$ isan
initial time. The integrand in the right hand side is a almostperiodic function because of the assumptions (Al) to (A3). In particular, it is written
as
“constantterm” $+$ “almost periodic term” by virtue of the Fourier expansion. The linearly increasing term,
namely secularterm, arisesfrom the integral of“theconstant“. On the otherhand, theintegral ofthe
“almostperiodic term“ isalmostperiodic yet. OurpurposeistorewriteEq.(29) sothat the right hand
side is split explicitly intothe secular term and theboundedterm. Since theintegral intherighthand
side is’‘secular term” $+$“almost periodicterm“,if
we
dividetheintegralby$t$andtakethelimit$tarrow\infty$,the almost periodic part vanishes and the coefficient ofthe secularterm remains. This coefficient of
the secularterm is justthe function$R_{1}$ defined by Eq.(19). To obtain theboundedterm,
we
subtractthe secular part$R_{1}$ fromthe integrandofEq.(29)
as
Now wedefine$h_{t}^{(1)}$byEq.(20) and put$h=h_{\tau}^{(1)}(y)$inthe above. Then
we
obtain$x_{1}=h_{t}^{(1)}(y)+X(t)R_{1}Cv)(t-\tau)$ (31)
and thisproves Eq.(23)for$i=1$.
Nowthat
we
know thenaive expansion solutionsupto the firstorder$x(t, \tau,y)=X(t)y+\epsilon(h_{t}^{(1)}(y)+$$X(t)R_{}(y)(t-\tau))$, let
us
calculate the envelope ofthefamily of thesecurves
parameterized by $\tau$.
Wevarythe parameter$y$, whichis
an
initial value of the 0-th orderequation, alongan
exactsolution whenwe
vary
the initialtime$\tau$(seethe figurebelow).Then, it
seems
thatthe envelopegivesa
goodapproximate solution. Todo this, put$y=y(\tau)$,andtheenvelope is given
as
follows: Atfirst,we
differentiatethe family by $\tau$ at $t$ and determine$y(\tau)$so
thatthederivative is equalto
zero.
$\frac{d}{d\tau}|_{\tau=t}x(t,\tau,y(\tau))=X(t)\frac{dy}{dt}(t)+\epsilon(\frac{\partial h_{t}^{(1)}}{\partial y}\frac{dy}{dt}(t)-X(t)R_{1}(y))=0$ . (32)
Itiseasytoverify that if
we
put$\frac{dy}{dt}=\epsilon R_{1}(y)+O(\epsilon^{2})$, (33)
then the equality (32) is satisfied. Let$y(t)$be
an
solution tothis equation. Then, the envelope for thefamilyof the naive expansion solutions is givenby
$x(t, t,y(t))=X(t)y(t)+\epsilon h_{t}^{(1)}(y(t))$
.
(34)Calculating the higher order
case
ina
similar manner,we
find the definition of the RG equation.Deflnition 7. Along with $R_{1}(y),$$\cdots,R_{m}(y)$ defined in Eqs.(19), (21),
we
define the m-th order$RG$equation forEq.(9) tobe
$\dot{y}=\epsilon R_{1}(y)+\epsilon^{2}R_{2}(y)+\cdots+\epsilon^{m}R_{m}(y),$ $y\in R^{n}$
.
(35)Using $h_{t}^{(1)}(y),$$\cdots,h_{t}^{(m)}(y)$defined in Eqs.(20), (22),
we
definethe m-th order$RG$transformation
$\alpha_{t}$:
$R^{n}arrow R^{n}$ tobe
$\alpha_{t}(y)=X(t)y+\epsilon h_{t}^{(1)}(y)+\cdots+\epsilon^{m}h_{t}^{(m)}(y)$. (36)
Remark 8. Since $X(t)$ is nonsingular and $h_{t}^{(1)}O’$),
sufficientlysmall$|\epsilon|$,there exists
an
open set$U=U(\epsilon)$ such that$\overline{U}$is compact and theresmictionof $\alpha_{t}$
to $U$is diffeomorphismfrom $U$into$R^{n}$.
An approximate solution is constructed
as
above, however,we
want to constructan
approximatevector field toinvestigate topologicalpropertiesof the original equation
as
discussed in Sec. 1.Funda-mentaltheorems oftheRGmethod
are
listed below.Theorem9. (Approximationof VectorFields)
Let$\varphi_{t}^{RG}$ be the fl$ow$ ofthem-thorder RGequationforBq.(9)and $\alpha_{t}$them-th orderRG
transforma-tion. Then,thereexists
a
positive constant$\epsilon_{0}$ such thatthefollowing holds for$\forall|\epsilon|<\epsilon_{0}$:
(i)The
map
$\Phi_{t.f_{0}}:=\alpha_{t}\circ\varphi_{t-t_{0}}^{RG}0\alpha_{t_{0}}^{-1};\alpha_{t_{0}}(U)arrow R^{n}$ (37)
defines
a
local flowon
$\alpha_{t_{0}}(U)$ for each $t_{0}\in R$, where $U=U(\epsilon)$ isan open
seton
which $\alpha_{t_{0}}$ is adiffeomorphism(see Rem.8). This$\Phi_{t,t_{0}}$ induces
a
time-dependentvectorfield$F_{b}$ through$F_{\epsilon}(t,x):= \frac{d}{da}|_{a=t}\Phi_{a.t}(x),$ $x\in\alpha_{t}(U)$
.
(38)(ii)Thereexists atime-dependentvectorfield $\tilde{F}_{F}(t, x)$such that
$F_{\epsilon}(t,x)=Fx+\epsilon g_{1}(t, x)+\cdots+\epsilon^{m}g_{m}(t, x)+\epsilon^{m+1}\overline{F}_{\epsilon}(t, x)$, (39)
where$\overline{F}_{\epsilon}(t, x)$ is
a
$C^{\infty}$ function with respect to$\epsilon,$$x,$$t$andbounded uniformly in$t\in R$withits
deriva-tives. Inparticular,the vector field$F_{\epsilon}(t, x)$isclosetotheoriginalvector field$Fx+\epsilon g_{1}(t, x)+\cdots$ within
of$O(\epsilon^{m+1})$
.
Theorem 10. (Error Estimate)
(i)Let$y(t)$be
a
solutiontothe m-thorderRG equationforBq.(9) and$\alpha_{t}$ the m-th orderRGtransfor-mation. Then, integralcurvesof the approximatevector field$F_{\epsilon}(t, x)$aregivenby
$\overline{x}(t)=\alpha_{t}(y(t))=X(t)y(t)+\epsilon h_{t}^{(1)}(y(t))+\cdots+\epsilon^{m}h_{t}^{(m)}(y(t))$. (40)
(ii) There existpositive constants $\epsilon_{0},$$C,$$T$, and acompact subset $V=V(\epsilon)\subset R^{n}$includingthe origin
such that for$\forall|\epsilon|<\epsilon_{0}$, every solution$x(t)$ ofEq.(9) and$\overline{x}(t)$definedbyBq.(40)with $x(O)=\overline{x}(0)\in V$
satisfy the inequality
$||x(t)-\tilde{x}(t)||<C\epsilon^{m}$, for $0\leq t\leq T/\epsilon$
.
(41)The followingtwotheoremsareconcemedwith
an
autonomousequation$\dot{x}=Fx+\epsilon g_{1}(x)+\epsilon^{2}g2(X)+\cdots$ , (42)
where$\epsilon\in R$ is
a
small parameter, $F$ is adiagonalizable$nxn$ matrix all of whose eigenvalueslieon
theimaginary axis, and$g_{i}(x)$
are
$C^{\infty}$ vector fields on$R^{n}$.
I.et$\epsilon^{k}R_{k}(y)$beafirst
non
zeroterm intheRGequation(35). If the vectorfield$\epsilon^{k}R_{k}(y)$hasa
normallyhyperbolicinvariantmanifold$N$,then theoriginalequation(9)also has
a
normallyhyperbolicinvariantmanifold$N_{\epsilon}$, which isdiffeomorphic to $N$, for sufficiently small $|\epsilon|$. In particular, the stability of$N_{\epsilon}$
coincides withthatof$N$
.
Theorem 12. (InheritanceoftheSymmetries)
(i)If vector fields$Fx$and$g_{1}(x),$ $g_{2}(x),$$\cdots$
are
invariantunder theaction ofaLiegroup$G$,then the m-thorderRGequation is alsoinvariantunder theaction of$G$
.
(ii)The m-thorderRG equation commutes with thelinearvectorfieldFXwith respect toLie bracket
product. Equivalently, each$R_{i}(y),$ $i=1,2,$$\cdots$ , satisfies
$X(t)R_{i}(y)=R_{i}(X(t)y),$ $y\in R^{n}$
.
(43)Theorem9
means
that thefamilyof approximatesolutions (40)defines the vector field $F_{\epsilon}(t, x)$ andit approximates to the original vector field well. In other words, the
curves
(40)are
solutionsto the‘’approximatedifferentialequation”
$\frac{d\overline{x}}{dt}=F_{\epsilon}(t,X\gamma.$ (44)
Onceweobtaintheapproximate differential equation, subtractingthe aboveequationfrom theoriginal
equation(9)andusingtheGronwall’s inequality,
we
can prove
theTheorem10. Sincetheapproximatevectorfield$F_{\epsilon}(t, x)$isclosetothe originalvectorfield,byvirtueof theFenichel’s theorem(Thm.3),itis
expectedthat
an
invariantmanifold oftheoriginal vector fieldisinheritedfromthat of theapproximatevector field $F_{\epsilon}(t, x)$
.
In fact, we can show that it is inherited froman
invariant manifold ofthe RGequation (Thm.11). It is becauseEqs.(37,38) show that the flow of the RG equation is topological
conjugate to that of the approximate vector field. Thus,
we
hope that the RG equation is easier toanalyze than the original equation. Theorem 12
assures
it. Actually, itmeans
that if the originalequation is autonomous and invariant under the action of
a
$k$ dimensional Lie group, then its RGequation is invariant under the action ofa$k+1$ dimensional Lie group. It is worth pointing outthat
Thm.12 (i)holds
even
if for non-autonomousequations,while Thm.12 (ii)holds only forautonomousequations. However,sincetheRGequation is anautonomousequation evenif the original equation is
non-autonomous,the RGequationhas simpler structure than the originalequationyet.
Remark13. Theinfinite orderRGequationand the infinite orderRG transformationdo not
converge
as
powerseriesof$\epsilon$in general. However,thenecessary
and sufficient conditionfor theconvergenceisobtainedby Chiba [6]. Roughly speaking, the infinite orderRGequationconvergesif and only if the
original equation is invariantunder theaction of
some
Liegroup
whichis diffeomorphicto $S^{1}$.
Thisfact is understood as follows : By Thm.12 (ii), theRG equation is invariant under the action of the
fundamental matrix $e^{Ft}$, which is diffeomorphic to $S^{1}$. Since the approximate vector field $F_{\epsilon}(t, x)$is
the$S^{1}$ action,which is obtainedbytransforming the$e^{Ft}$by the RG transformation. If the infinite order
RGequation andthe infinite order RG transformation converge, then the left hand side ofBq.(39) is
well-defined
as
$marrow\infty$ and it is invariantunder theactionof$S^{1}$.
Therefore theright handside, whichcorresponds to the originalequationas$marrow\infty$, is alsoinvariantunder theaction of$S^{1}$
.
4
Examples
Inthissection,
we
givetwo simple examples.Example 14. considerthe systemon$R^{2}$
$\{\begin{array}{l}\dot{x}=y-x^{3}+\epsilon x,\dot{y}=-x.\end{array}$ (45)
To bringthe nonlinear term $x^{3}$ into thefirst orderterm with respect to $\epsilon$, put $(x,y)=(\epsilon^{1/2}X, \epsilon^{1/2}Y)$
.
Then
we
obtain$\{\begin{array}{l}\dot{X}=Y+\epsilon(X-X^{3}),Y=-x.\end{array}$ (46)
Introduce the complex variable $z$by$X=z+\overline{z},$ $Y=i(z-\overline{z})$ todiagonalizetheunperturbedterm
as
$\{\begin{array}{l}\dot{z}=iz+\frac{\epsilon}{2}(z+\overline{z})-\frac{\epsilon}{2}(z+\overline{z})^{3}\epsilon’\overline{z}=-\iota^{arrow}z+\frac{\epsilon}{2}(z+\overline{z})-(z+\overline{z})^{3}\overline{2}.\end{array}$ (47)
For thissystem,the first orderRGequationis given by
$\dot{z}=\frac{\epsilon}{2}(z-3|z|^{2}z)$
.
(48)Putting$z=re^{ir)}$ yields
$\{\begin{array}{l}\dot{r}=\frac{\epsilon r}{2}(1-3r^{2}),\dot{\theta}=0.\end{array}$ (49)
Now Thm.12 (ii)
means
thattheRGequationis invariant under the flow $e^{F}$‘, $F=(\begin{array}{ll}i 00 -i\end{array})$ definedbytheharmonicoscillator(rotation invariance). Thus,written in the polarcoordinate, the RG equationis
split intotheequationof$r$direction and theequationof$\theta$direction. Generally,theRG equation for
a
perturbedharmonic oscillatoriseasily solved in thepolarcoordinate.
It is easyto show that this RGequationhas
a
stableperiodicorbit$r=\sqrt{1/3}$if$\epsilon>0$.
NowThm.11proves
that the system (46) also hasa
stable periodicorbit, whoseradius is of$O(1)$.
By transformingit intotheoriginal $(x,y)$coordinate, itis shown that the system(45)has
a
stableperiodic orbitwhoseradiusisof$O(\epsilon^{1/2})$
.
This resultcoincideswith theclassical Hopftheorem.(50)
Example15. Consider the systemon $R^{2}$
(51)
Changingthe coordinatesby $(x,y)=(\epsilon X, \epsilon Y)$yields
$\{\begin{array}{l}\dot{X}=Y+\epsilon Y^{2}\dot{Y}=-X+\epsilon^{\backslash }(Y^{2}-XY)+\epsilon^{2}Y\end{array}$
We introduceacomplexvariable$z$by$X=z+\overline{l}_{2}Y=i(z-\overline{z})$
.
Then,theabove systemis rewritten as$\{\begin{array}{l}i=iz+\frac{\epsilon}{2}(z-2z^{2}+2z\overline{z})+\frac{\epsilon^{2}}{2}(z-\overline{z}),\overline{z}=-,\overline{z}+\frac{\epsilon}{2}(-i(z-\overline{z})^{2}-2\overline{z}^{2}+2z\overline{z})-\frac{\epsilon^{2}}{2}(z-\overline{z}).\end{array}$ (52)
(54)
Forthis system, the second orderRGequation is given by
$\dot{z}=\frac{1}{2}\epsilon^{2}(z-3|z|^{2}z-\frac{16i}{3}|z|^{2}z)$
.
(53)Note that the first orderterm$R_{1}(y)$ vanishes. Putting$z=re^{i\theta}$resultsin
$\{\dot{\theta}=-\frac{8}{3}\epsilon^{2}r^{2}\dot{r}=\frac{1}{2}\epsilon^{2}r(1-3r^{2})$
Itiseasytoverifythat thisRGequationhasastableperiodic orbit$r=\sqrt{1/3}$
.
Since$R_{1}(y)=0$,Thm.11for$k=2$implies thattheoriginal system(50)also has
a
stable periodicorbit, whose radiusisof$O(\epsilon)$,forsmall$\epsilon$
.
Such a caseisknownas
thedegenerate Hopfbifurcation.5
Relation
to
the
traditional singular perturbation methods
It is shown in [1,2,6] that theRG method unifies the traditional singular perturbation methods. In
this section,
we
givea
briefreviewof this fact.5.1
Multi-scale
method
$[18|$The multi-scale methodis
one
of the mostfamousperturbation methods which is widely used. Thismethodisbasedonideasofintroducing varioustime scales andremoving secular terms. Considerthe
system
on
$R^{n}$$\frac{dx}{dt}=\dot{x}=Fx+\epsilon gI(x)+\epsilon^{2}g2(X)+\cdots$ , (55)
where the matrix $F$ is
a
diagonalizable $n\cross n$ constant matrix all of whose eigenvalues lieon
theimaginaryaxisand wherethe functions$g_{l}(x)$’s
are
polynomial in$x$.
Assume that thereexistmany
timescales$t_{0},$$t_{1},$$\cdots$ ,$t_{m}$ satisfying
$t_{0}=t,$ $t_{1}=\epsilon t,$$\cdots,$$t_{m}=\epsilon^{m}t$
.
(56)Then, $d/dt$isrewritten
as
Furtherwe
suppose
that thedependentvariable$x$is expandedin$\epsilon$as$x(t)=x_{0}(t_{0}, t_{1}, \cdots, t_{m})+\epsilon x_{1}(t_{0}, t_{1}, \cdots, t_{m})+\epsilon^{2}x_{2}(t_{0}, t_{1}, \cdots, t_{m})+\cdots$
.
(58)SubstitutingEqs.(57,58)into Eq.(55)yields
$\frac{\partial x_{0}}{\partial t_{0}}=Fx_{0}$, (59)
$\frac{\partial x_{1}}{\partial t_{0}}+\frac{\partial x_{0}}{\partial t_{1}}=Fx_{1}+g_{1}(x_{0})$
.
(60)The solutionto the formeris givenby$x_{0}=e^{Ft_{0}}y$, where the initial value$y=y(t_{1}, \cdots , t_{m})$depends
on
$t_{1},$$\cdots$ ,$t_{m}$. Substituting the $x_{0}$intoEq.(60),
we
obtain the general solution of$x_{1}$as
$x_{t(i_{0},\cdots,t_{m})=e^{F(t_{0}-\tau)}h+e^{Ft_{0}}\int_{\tau}^{t_{0}}e^{-Fs}}(g_{1}(e^{Fs}y)-e^{Fs} \frac{\partial y}{\partial t_{1}})ds$ .
Sincethe secular term arising from the integral intherighthandside is givenby
$\lim_{t_{0}arrow\infty}\frac{1}{t_{0}}\int_{\tau}^{l_{0}}e^{-Fs}(g1(e^{Fs}y)-e^{Fs}\frac{\partial y}{\partial t_{1}})ds=\lim_{t_{0}arrow\infty}\frac{1}{t_{0}}\int_{\tau}^{t_{0}}e^{-Fs}g_{1}(e^{Fs}y)ds-\frac{r?y}{\partial t_{1}}$, (61)
if
we
determine $y$ so thatthe secular term vanishes,we
obtain the first order RG equation $\partial y/\partial t_{1}=$$R_{1}(y)\Rightarrow\partial y/\partial t=\epsilon R_{1}(y)$. We
can
obtain thehigherorderRG equation inasimilarmanner.
5.2
Normal forms
[15],[16]For Bq.(55), ifthere exists
a
time independent (local) coordinate transformation $x\mapsto z$ such thatEq.(55)is brought into
$\{\begin{array}{l}\dot{z}=Fz+\epsilon\tilde{g}_{1}(z)+\epsilon^{2}\tilde{g}2(Z)+\cdots(62)s.t. \tilde{g}_{i}(e^{Ft}z)=e^{Ft}\tilde{g}_{i}(z), for i=1,2, \cdots ,\end{array}$
thenEq.(62) is calledthe normal
form
ofBq.(55). Sincea purpose
of normal forms is not obtainingapproximate solutions but transforming vector fields by coordinate transformations, it
answers
our
purpose
ofinvestigatingtopologicalproperties ofa
givenvector field.ByThm.9, if
we
apply the RG transformation$x=\alpha_{t}(y)$toEq.(55),we
obtainthe RGequation(35).Further,because of Thm.12(ii), changing thecoordinatesby$y=e^{-Ft}z$yieldsthe system
$\dot{z}=Fz+\epsilon R_{1}(z)+\epsilon^{2}R_{2}(z)+\cdots$
.
(63)Since$R_{i}$ commutes with $e^{Ft}$,this systemseems tobe the normal form ofEq.(55). However, wehave
toshowthatthe coordinate transformation$x=\alpha_{t}(e^{-Ft}z)$ fromEq.(55) to Bq.(63) is independent of$t$
.
Infact, it immediately follows fromthe next lemma.
Lemma
16.
TheRG transformation$\alpha_{t}$ satisfiestheequality$\alpha_{t}(e^{Ft’}y)=\alpha_{t+t’}(y)$.
Sincethis lemma shows that$\alpha_{t}(e^{-Ft}z)=\alpha_{0}(z)$is independentof$t$,Bq(63) is provedtobethe normal
is applicableto non-autonomous equations while the normal forms are defined only forautonomous equations.
Remark 17. Itis known that for agiven equation (55),its normal forms
are
not unique in general.Thus, the RGequations
are
also notunique. The non-uniqueness resultsfrom undetermined integralconstantsinEqs.(20,22) (notethatintegralconstants inEqs.(19,21)vanish
as
thelimit$tarrow\infty$). Sincethe integrating variable is $s$,
we can
choose arbitrary functions of$y$ as integral constants in Eqs.(20,22). An integralconstant in the definition of$h_{t}^{(i)}$ affectsthedefinitions of$R_{i+1},R_{i+2},$$\cdots$ . Ifwechoose
integral constants appropriatelysothat$R_{2},$ $R_{3},$$\cdots$ take the simplestformsin
some
sense, the resultantRGequationisprovedtobeequivalenttothehyper-normal forms [4].
5.3 Center
manifold reduction
[14]ForEq.(55), we
suppose
that(Cl)all eigenvalues of the matrix$F$
are
on
the imaginary axisor
the left half plane. The Jordan blockcorrespondingtoeigenvalues
on
the imaginaryaxis is diagonalizable.(C2) each$g_{i}(x)$ is polynomial in$x$.
Ifalleigenvaluesof$F$are onthelefthalf plane, the origin is stableandtheflow
near
theorigin is trivial.In what follows,
we
supposethat atleastone
eigenvalue ison
theimaginary axis. Inthiscase,Bq.(55)hasacentermanifold whichistangent to thecentersubspaceattheorigin. Sincenontrivialphenomena
such
as
bifurcationsoccur on a
centermanifold and orbits outof thecentermanifold approachtothecentermanifold
as
$tarrow\infty$, it isimportanttoinvestigatethe flow onthecentermanifold. Thepurposeofthissectionistoconstruct
a
centermanifold andaflowon itby the RG method.Let$N_{0}$bethe center subspaceof$F$, whichis spanned bytheeigenvectorsassociated with the
eigen-values
on
theimaginary axis. For Eq.(55),we
define the functions $R_{i}$:
$N_{0}arrow R^{n}$ and $h_{t}^{(i)}$:
$N_{0}arrow$$R^{n},$ $i=1,2,$$\cdots$ tobe
$R_{1}(y)$ $:= \lim_{tarrow-\infty}\frac{1}{t}\int^{t}X(s)^{-1}g_{1}(s,X(s)y)ds$, (64)
$h_{t}^{(1)}Cv)$ $:=X(t) \int^{t}(X(s)^{-}g1(s,X(s)y)-R_{1}(y))ds$, (65)
and
$R_{i}(y):= \lim_{tarrow-\infty}\frac{1}{t}\int(X(s)^{-1}G_{i}(s,X(s)y, h_{s}^{(1)}(y), \cdots, h_{s}^{(i-1)}(y))$
$-X(s)^{-1} \sum_{k=1}^{i-1}(Dh_{s}^{(k)})_{y}R_{i-k}(y))ds$, (66)
$h_{t}^{(i)}(y):=X(t) \int^{t}(X(s)^{-1}G_{i}(s,X(s)y, h_{s}^{(1)}(y), \cdots, h_{s}^{(i-1)}(y))$
for $i=2,3,$$\cdots$, respectively. They
are
the sameas
Eqs.(19,20,21,22), except that the domain isrestricted to the center subspace $N_{0}$
.
Inthe abovedefinitions, theintegral constantsare
assumedtobezero.
Wecan
prove
the nextlemma.Lemma 18. The functions$R_{1}(y),R_{2}(y),$ $\cdots$
are
well-defined (namely, the limits converge) and thefollowingholds.
(i)$R_{i}(y)\in N_{0}$ for all$y\in N_{0},$ $i=1,2,$ $\cdots$
.
(ii)$h_{t}^{(i)}(y)$is bounded uniformly in $t\in R$forall$y\in N_{0},$ $i=1,2,$$\cdots$
.
Withthese$R_{i},h_{t}^{(i)}$,
we
define them-th order$RG$equation on$N_{0}$ tobe$\dot{y}=\epsilon R_{1}(y)+\epsilon^{2}R_{2}(y)+\cdots+\epsilon^{m}R_{m}(y),$ $y\in N_{0}$, (68)
and define the m-th order$RG$
tranfformation
on$N_{0}\alpha_{f}$ : $N_{0}arrow R^{n}$ by$\alpha_{t}(y)=X(t)y+\epsilon h_{t}^{(1)}(y)+\cdots+\epsilon^{m}h_{t}^{(m)}(J’)$
.
(69)They
are
thesame
as
Bqs.(35,36),exceptthatthedomainisrestrictedtothe center subspace$N_{0}$as
be-fore. SinceLem.18shows that thedomainand therangeof$R_{i}(y)$
are
$N_{0}$,Eq(68)definesthe differentialequationson$N_{0}$
.
Itmeans
that the system(68)has $\dim N_{0}$ linearly independent equations.In thissituation, Thm.9toThm.12hold
on
$N_{0}$.
Further, wecan prove
the next theorem.Theorem 19. (Approximationof CenterManifolds, [5])
Let$\alpha_{t}$bethe m-th orderRG transformationon$N_{0}$ and$W$acompactneighborhoodoftheoriginsuch
that$\alpha_{f}$ is diffeomorphism
on
$W\cap N_{0}$ (see Rem.8). Then, the set$\alpha_{t}(W\cap N_{0})$ lies withinan
$O(\epsilon^{m+1})$
neighborhoodofthecentermanifoldofBq.(55).
5.4
Averaging method
[17]Considerthe system
on
amanifold $M$$\dot{x}=\epsilon g1(t,x)+\epsilon^{2}g_{2}(t, x)+\cdots$ , (70)
where each $g_{i}$ is
a
time-dependent smooth vector fieldon
$M$, which is almost periodic in$t$, the set
of whose Fourierexponents has
no
accumulation points on R. For this system, we define themaps
$R;,$$u_{t}^{(i)}$ : $Marrow M$tobe
$R_{1}(y)= \lim_{farrow\infty}\frac{1}{t}\int^{t}g_{1}(s,y)ds$, (71)
$u_{t}^{(1)}(y)= \int(g\iota(s,y)-R_{1}(y))ds$, (72)
and
$R;(y)= \lim_{tarrow\infty}\frac{1}{t}\int^{t}(G_{i}(s,y, u_{s}^{(1)}(y), \cdots, u_{s}^{(i-1)}(y))-\sum_{k=1}^{i-1}(Du_{s}^{(k)})_{y}R_{i-k}(y))ds$, (73)
for $i=2,3,$$\cdots$, respectively. Define the m-th order RGequation by Eq.(35) and the m-th order RG
transformationby
$\alpha_{t}(y)=y+\epsilon u_{t}^{(1)}(y)+\cdots+\epsilon^{m}u_{t}^{(m)}(y)$. (75)
In thissituation,
we
can
prove
Thm.9toThm.12(i).Remark
20.
Consider thesystem ofthe form$\dot{x}=f(x)+\epsilon g_{1}(t,x)+\epsilon^{2}g2(t, x)+\cdots$
.
(76)Let$\varphi_{t}$ be the flow of the vectorfield $f$and
suppose
thatitis almost periodic in$t$.
Then, ifwe
changethecoordinates
as
$x=\varphi_{t}(X)$, the above systemis broughtintothe system$\dot{X}=\epsilon(D\varphi_{t})_{X}^{-1}g1(t, \varphi_{t}(X))+\epsilon^{2}(D\varphi_{t})_{X}^{-1}g_{2}(t, \varphi_{t}(X))+\cdots$ , (77)
which isof the form ofBq.(70). Thus, theRGmethodin the presentsectiongives extension of those
in the
cases
of theprevious sections. TheRGequation in this sectionisprovedtobeequivalent to theaveraging equation in the averaging method.
6
Analysis
of the Kuramoto model
TheKuramotomodelofcoupled phase oscillators
$\dot{\theta}_{i}=\omega_{i}+\frac{\epsilon}{N}\sum_{i=1}^{N}\sin(\theta_{j}-\theta_{i}),$ $i=1,$
$\cdots,$$N$ (78)
is
one
of the moststudied modelsofnonlinear phenomenaofglobally coupled limit-cycle oscillators[19], where $\theta_{j}\in S^{1}$ is
a
particleon
a
circle, $\omega$; isa
constant called the natural frequency, $N$ isa
number of particles, and$\epsilon$is thecoupling strength. It isknownnumerically that if$\epsilon$is largerthan the
threshold $\epsilon_{0}$, then $<\dot{\theta}_{i}>-<\dot{\theta}_{j}>$ tends to
zero as
$tarrow\infty$ for all $i,$$j$, where $<>$ denotes averagingover
time. Such aphenomenon is called the synchronization. However, mechanism of the transitionfrom thecoupledharmonicoscillators, namely$\epsilon=0$(uncoupledsystem),tothesynchronization state
is not well understood [20]. Most recently, bifurcation diagrams ofthe Kuramoto model for small$N$
andsmall$\epsilon$
were
investigated by Maistrenkoetal. [21,22] andPopovychetal. [23],in which naturalfrequenciesareassumed tobe distributed symmetrically arounda
mean
frequency$\Omega$ :$\omega_{i}-\Omega=-(\omega_{N-i+1}-\Omega),$ $;=1,2,$$\cdots$ ,$N$
.
(79)Note that
we
can
put$\Omega=0$without loss ofgenerality because the Kuramotomodelis invariantundertherotation$\theta_{i}\mapsto\theta_{i}+\Omega$
.
In thiscase,itiseasy
to showthat the Kuramoto model has theinvarianttorus$M$defined by
$M=(\theta_{i}=-\theta_{N-i+1},$ $i=1$ ,–,$N\}$
.
(80)It is important to determine the stability of $M$ because the synchronization solutions
are
alwayswhen$\epsilon$is small andin this
case
it is difficulttodeterminethe stability by numerical simulation. Inthisarticle,
we
apply the RG methodto the Kuramoto modelto determine thestability of$M$for small $\epsilon$.
Wecan provethefollowing results.
Theorem21. Supposethat$N=2M-1$ is
an
odd number. Suppose thenaturalfrequencies satisfythesymmetric condition(79) (weput$\Omega=0$) andthe followingnonresonancecondition :
$\omega;\neq\omega_{j}$ forall $i,$$j$,
$\omega_{k}+\omega_{j}=2\omega_{i}$ ifand only if $i=k=j$
or
$j=2M-k,$$i=M$,$\omega;+\omega_{j}=\omega_{k}+\omega_{l}$ if andonlyif $i=j=k=l$
or
$j=2M-i,$$l=2M-k$, $3\omega;=\omega_{j}+\omega_{k}+\omega_{l}$ ifand only if $i=j=k=l$,$\omega_{i}+2\omega_{k}=\omega_{j}+2\omega_{l}$ ifandonlyif $i=j,$ $k=l$
or
$j=2M-i,$$k=M,$ $l=i$.
Then thereexistspositive constant$F_{0},$, which depends
on
thenatural frequencies, such that if$0<\epsilon<$$\epsilon_{0}$, theinvarianttoms $M$isstable and thetransverseLyapunov exponents of$M$isof
$O(\epsilon^{3})$
.
When$N=3$, the
nonresonance
condition isviolated if andonlyif$\omega_{1}=\omega_{2}=\omega_{3}=0$.
However, inthiscase, the phase portraitof theKuramotomodelis independentof$\epsilon$because
we
can
divide therighthand side ofEq.(78)by$\epsilon$by changingthetimescale
so
that thesystemisindependentof$\epsilon$. Similarly,if$\omega_{i}=0$forall$i$,the phaseportrait oftheKuramotomodelis independent of$\epsilon$
.
Otherwise,for$N=5$,thenonresonanceconditionis violatedif and only if
$\omega_{1}=0,$ $\omega_{2},3\omega_{2}/2,2\omega_{2},3\omega_{2},4\omega_{2},5\omega_{2}$
.
(81)In these cases, theRG equationstakedifferent forms fromBq.(83), which is the RG equationfor the
nonresonance case.
To determine the stability of$M$for the abovecases, deriving theRG equationsforindividual
resonance cases
and investigatingthem,we can
provethenexttheorem.Theorem 22. Suppose that $N=5$ and the natural frequencies satisfy the symmetric condition
(79). Then the invariant torus $M$is stable for sufficiently small $\epsilon$
.
In particular, ifall of the naturalfrequencies
are
notidentical, the transverseLyapunovexponents of$M$isof$O(\epsilon^{3})$.In thisarticle, wegivetheproofof Thm.21. Theproofof Thm.22 needs
more
hardanalysis anditisomitted here.
Proof of Thm.21 To write downthe system (78) in the Cartesian coordinate, put $x_{i}=\cos\theta;,$ $y_{i}=$
$\sin\theta_{j}$
.
Furtherputting $x_{i}=z_{i}+\overline{z}_{i},$ $y_{i}=i(z_{t}-\overline{z}_{i})$,we
obtainthe systemoftheform$\{\begin{array}{l}\dot{z}_{i}=-i\omega_{i}z_{i}+\frac{2\epsilon}{N}\sum_{j=1}^{N}z;(\overline{z}_{i}z_{j}-z\mathfrak{s}\overline{z}_{j}),\overline{z}_{i}=i\omega_{i}\overline{z}_{i}+\frac{2\epsilon}{N}\sum_{j=1}^{N}\overline{z}_{i}(z_{i}\overline{z}_{j}-\overline{z}_{i}z_{j}).\end{array}$ (82)
Since this system is the perturbed harmonic oscillators whose unperturbed term has eigenvalues
equation forEq.(82),
we
put$z_{i}=e^{i\theta_{i}}$ to change tothe polar coordinate(FortheRGequation, weuse
the
same
notations $z;^{\theta_{j}}$ with those of the original system). Thenwe
obtain the RG equation of theform
$\{\begin{array}{l}\dot{\theta}_{M}=-\frac{16\epsilon^{3}}{N^{3}}\sum_{k\neq M}\frac{1}{\omega_{k}^{2}}\sin(2\theta_{M}-\theta_{k}-\theta_{2M-k}),\dot{\theta}_{i}=\frac{8\epsilon^{2}}{N^{2}}[2\sum_{k\neq i}\frac{1}{\omega_{i}-\omega_{k}}-\frac{1}{\omega_{l}}\cos(\theta_{i}-2\theta_{M}+\theta_{2M-i}))+\frac{16\epsilon^{3}}{N^{3}}(\sum_{\neq i.2M-i}\frac{1}{\omega_{i}^{2}-\omega_{k}^{2}}\sin(\theta_{i}-\theta_{k}-\theta_{2M-k}+\theta_{2M-;})-2\sum_{k\neq i.2M-i}\frac{1}{\omega_{i}(\omega_{i}-\omega_{k})}\sin(\theta_{i}-\theta_{k}-\theta_{2M-k}+\theta_{2M-;})+2\sum_{k\neq i.M}\frac{1}{\omega_{k}(\omega_{i}-\omega_{k})}\sin(2\theta_{M}-\theta_{k}-\theta_{2M-k})-\sum_{k\neq M,i,2M-i}\frac{1}{\omega_{k}(\omega_{i}+\omega_{k})}\sin(\theta_{i}-\theta_{k}-\theta_{2M-k}+\theta_{2M-i})-2\sum_{k\neq M.2M-i}\frac{1}{\omega_{i}(\omega_{i}+\omega_{k})}\sin(\theta_{i}-2\theta_{M}+\theta_{2M-i})), (i\neq M).\end{array}$
(83)
Note that the first order term vanishes. Since the invarianttorus $M$ corresponds to the solution$\theta_{i}+$
$\theta_{2M-;}=0$,
we
put$\phi_{i}=\theta_{i}+\theta_{2M-i}$and$\phi_{M}=2\theta_{M}$. Thenwe obtain the systemof$\phi_{i}$$\{\begin{array}{l}\phi_{M}=-\frac{64\epsilon^{3}}{N^{3}}\sum_{k=1}^{M-1}\frac{1}{\omega_{k}^{2}}\sin(\phi_{M}-\phi_{k}),\phi_{j}=\frac{32_{6^{\backslash }}^{3}}{N^{3}}(-\frac{1}{\omega_{i}^{2}}\sin(\phi_{i}-\phi_{M})-4\sum_{k\neq i}^{M-1}\frac{1}{\omega_{i}^{2}-\omega_{k}^{2}}\sin(\phi_{i}-\phi_{M})+4\sum_{k\neq i}^{M-1}\frac{1}{\omega_{i}^{2}-\omega_{k}^{2}}\sin(\phi_{M}-\phi_{k})), (i=l, \cdots, M-1).\end{array}$ (84)
Now that the second order termvanishes, Thm.11 for$k=3$ is applicabletothis system.Wecan prove
thatthe eigenvalues of the Jacobian matrix atthe fixed point$\phi_{i}=0$ $(i=1, \cdots , M)$ of the righthand
side ofEq.(84) have negative real parts except to a
zero
eigenvalue, which results from the rotationinvariance ofEq.(78). Thus, the solution $\phi_{i}=\theta_{i}+\theta_{2M-i}=0(i=1, \cdots, M)$ of the RG equation is
stable andthisprovesthat theinvarianttorus $M$is stablefor small$\epsilon>0$
.
Reference
[1] L. Y.Chen,N. Goldenfeld, Y.Oono,Renormalizationgrouptheoryforglobalasymptotic
anal-ysis, Phys. Rev. Lett. 73 (1994),no. 10, 1311-15
[2] L. Y. Chen, N.Goldenfeld, Y. Oono, Renormalizationgroup and singular perturbations:
Mul-tiple scales, boundary layers. and reductive perturbation theory, Phys. Rev. E54, (1996),
376-394
[3] H.Chiba,$C^{1}$ approximation ofvectorfields based
on
therenormalizationgroup
method, SIAMJ. Appl. Dym. Syst.,(2008, preprint)
[4] H. Chiba, Simplified Renormalization Group Equations for Ordinary Differential Equations,
(submitted)
[5] H.Chiba,Approximation of Center Manifolds ontheRenormalization GroupMethod,
(submit-ted)
[6] H.Chiba,ExtensionandUnificationofTraditional Singular PerturbationMethods, (in
[7] R. E. Lee DeVille, A. Harkin, M. Holzer, K. Josi\v{c}, T. Kaper, Analysis of
a
renormalizationgroupmethod and normal form theory for perturbedordinarydifferential equations, PhysicaD,
(2008)
[8] S. Bi, K. Fujii, T. Kunihiro, Renormalization-group method for reduction of evolution
equa-tions; invariant manifolds and envelopes, Ann. Physics280(2000),
no.
2,$236arrow 298$[9] T. Kunihiro, A geometrical formulation of therenormalization groupmethod for global
analy-sis,Progr. Theoret. Phys. 94(1995),
no.
4,503-514[10] T. Kunihiro,Therenormalization-groupmethodappliedtoasymptotic analysisof vectorfields,
Progr. Theoret. Phys. 97 (1997),no. 2,
179-200
[11] M. Ziane,On
a
certain renormalization groupmethod, J. Math. Phys. 41 (2000),no.
5,3290-3299
[12] Fenichel, Geometric singular perturbation theoryfor ordinary differential equations, J.
Differ-entialEquations 31 (1979),
no.
1,pp.53-98[13] L. Amoldet al,Dynamicalsystems,Lecture notesinmathematics 1609,SpringerVerlag,
1995
[14] J. Carr, Applicationsof Centre Manifold Theory, Springer-Verlag,
1981
[15] S. N. Chow,C. Li, D.Wang, Normal formsandbifurcation of planarvector fields, Cambridge
University Press, 1994
[16] J. Murdock,Normalformsandunfoldingsforlocaldynamical systems,Springer, 2003
[17] J.A.Sanders,F.Verhulst,Averagingmethodsinnonlineardynamical systems, Springer-Verlag,
1985
[18] A.H. Nayfeh,Methodof NormalForms,JohnWiley&Sons, INC.,
1993
[19] A. Pikovsky, M. Rosenblum, J. Kurths, “Synchronization:
a
universal concept in nonlinearsciences”,Cambridge UniversityPress, (2001)
[20] S. H.Strogatz, From Kuramoto to Crawford: exploring the onsetof synchronization in
popula-tionsof coupledoscillators, Phys. D143 (2000),
no.
1-4, 1-20[21] Y. Maistrenko, O.Popovych, O. Burylko,P. A. Tass, Mechanism ofDesynchronization in the
Finite-Dimensional KuramotoModel,Phys. Rev. Lett. 93, 084102(2004)
[22] Y. L.Maistrenko, O.V. Popovych, P.A. Tass, Chaotic attractorintheKuramotomodel,Intemat.
J. Bifur. Chaos, 15 (2005),
no.
11, 3457-3466[23] O. V. Popovych, Y. L.Maistrenko, P. A.Tass, Phase chaos in coupledoscillators, Phys. Rev. $E$