Entrainment and chaos in a pulse-driven Hodgkin-Huxley oscillator Kevin K. Lin [email protected] 6 0 January 18,2006 0 2 n Abstract a J The Hodgkin-Huxley model describes action potential generation in certain types of neurons and is a 8 standard model for conductance-based, excitable cells. Following the early work of Winfree and Best, 1 thispaperexplorestheresponseofaspontaneouslyspikingHodgkin-Huxleyneuronmodeltoaperiodic pulsatiledrive. Theresponseasafunctionofdriveperiodandamplitudeissystematicallycharacterized. ] S Awiderangeofqualitativelydistinctresponsesarefound,includingentrainmenttotheinputpulsetrain D andpersistentchaos. Theseobservationsareconsistentwithatheoryofkickedoscillatorsdevelopedby . QiudongWangandLai-SangYoung. InadditiontogeneralfeaturespredictedbyWang-Youngtheory,it h t isfoundthatmostcombinationsofdriveperiodandamplitudeleadtoentrainmentinsteadofchaos.This a preferenceforentrainmentoverchaosisexplainedbythestructureoftheHodgkin-Huxleyphaseresetting m curve. [ 4 v 1 Introduction 1 6 TheHodgkin-Huxleymodeldescribesactionpotentialgenerationincertaintypesofneuronsandisastan- 1 5 dardmodelforconductance-based,excitablecells[5,18,20].Thereisanextensiveliteratureontheresponse 0 oftheHodgkin-Huxleymodeltodifferenttypesofinputs[1,2,11,14,15,16,17,19,24,25],andunderstand- 5 0 ing how single neurons respond to external forcing continues to be relevantfor the study of information / transmission in neural systems [21, 23]. Because neurons typically communicate via pulsatile synaptic h t events, it is natural to investigate the response of the Hodgkin-Huxley model to pulsatile inputs. Early a m studiesbyBestandWinfree[3,38]examinetheresponseofaHodgkin-Huxleymodeltoperiodicimpulse trains,chracterizingindetailthestructureofphasesingularitiesandthetransitionfromdegree1todegree : v 0phaseresetting.However,theirworkdoesnotsystematicallyaddresstheasymptoticdynamicalbehavior i X asafunctionofdriveperiodandamplitude.1 r This paper studies a spontaneously spiking (i.e. oscillatory) Hodgkin-Huxley neuron model driven a byperiodic, pulsatile input of fixed amplitude and period, and systematically classifies the response asa functionofdriveperiodandamplitude. Itisfoundthat: 1. Inresponsetoperiodicpulsatileforcingoffixedamplitude AandperiodT,aspontaneouslyspiking Hodgkin-Huxleysystemcanexhibitawiderangeofdistinctbehaviorsdependingon AandT: (a) Entrainment: The driven system is stably periodic and its period is a rational multiple of the driveperiodT. (b) Transientchaos: Thesystemexperiencesatransientperiodofexponentialinstabilitybeforeen- trainingtotheinput. ThistransientchaosiscausedbyaSmalehorseshoe[13]. 1Takabe,Aihara,andMatsumoto[32]appeartohavecarriedoutsuchasystematicstudy.But,Iwasonlyabletolocateanabstract. 1 (c) Chaos: The system becomes fully chaotic: it possesses a positive Lyapunov exponent and a mixingattractor(see[39]forareviewoftheseconcepts). The response of the pulse-drivenneuron is approximately T -periodic in the drive period T, where 0 T istheintrinsicperiodoftheunforcedHodgkin-Huxleyoscillator. Forexample,ifthepulse-driven 0 oscillatorischaoticforsomedriveamplitudeAanddriveperiodT,thenitislikelytobechaoticwhen drivenbyapulsetrainofamplitude AwithperiodnearT+T . 0 2. Thescenariosenumeratedaboveareprevalentinthesensethattheycorrespondtopositive-areasub- setsofthedriveperiod-driveamplitudespace.Prevalence,togetherwiththeapproximateperiodicity statedabove,implythateachscenariooccurswithpositive“probability.”(SeethediscussionofFig.3 in§3fortheprecisemeaningofprobabilityinthiscontext.) Therangeofresponsesandtheirpreva- lenceareconsistentwithatheoryofnonlinearoscillatorsdevelopedrecentlybyQiudongWangand Lai-SangYoung[33,34,35,36]. 3. While chaotic behavior is readily observable, most combinations of drive period and drive ampli- tude leadto entrainment instead of chaos. This preferencefor entrainment canbe explained by the structureofthephaseresettingcurve(see§4)oftheHodgkin-Huxleysystem. This paperreliesheavilyon numerical computation and the conceptual frameworkprovidedbyWin- free’stheoryofbiologicalrhythms[38]andtheworkofQ.WangandL.-S.Youngonnonlinearoscillators [34,35]. Phaseresettingcurves,introducedbyWinfree,playaparticularlyimportantrolehere. Thephase resettingcurveofanonlinearoscillatorisanintervalmapwhichcapturestheasymptoticresponseofanon- linearoscillatortoasingle,pulsatileperturbation. Becausetheyare1-dimensionalobjects,phaseresetting curvesareofteneasiertounderstandthanthenonlinearoscillatorstheyrepresent.Theyarefrequentlyused toinferstabledynamicalbehaviorlikephaselocking. Wang-Youngtheoryprovidesamathematicalframe- workforusingphaseresettingcurvestoinfertheexistenceandprevalenceofchaoticbehavior.Ratherthan numerically verify the hypotheses of their theorems, we have opted to examine the consequences of the theorydirectly,relyingonacombinationofnumericalsimulationandgeometricreasoningtocharacterize thespecificresponseoftheHodgkin-Huxleymodeltoaperiodicpulsatiledrive. For the sake of clarity, parametersareselected toensure that the Hodgkin-Huxley system possesses a unique limit cycle and no other attractinginvariantset. This correspondsto a repeatedlyspiking neuron withanunstablereststate. Whilethescenariosstatedaboveshouldstillholdwhenthelimitcyclecoexists with other stable invariant sets, this choice simplifies the interpretation of numerical simulations. Other- wise,atrajectorymayjumpoutofthebasinofthelimitcycle,whichobscuresthemechanismdescribedby Wang-YoungtheoryandwhichWinfreeandBesthavealreadyinvestigatedthoroughly[3,38]. Therestof thispaperisorganizedasfollows: Section2brieflyreviewsthe unforcedHodgkin-Huxley equationsanditsproperties. Mainnumericalresultsaresummarizedin§3anddiscussedin§4. Section5 discussesfurthernumericalresults,addressingsomeissuesraisedinSections3and4. Section6discusses possibleextensionsandgeneralizations. 2 Brief review of the Hodgkin-Huxley model TheHodgkin-Huxleyequationsareasystemofnonlinear ordinarydifferentialequations2 whichdescribe the way neurons generate spatially and temporally localized electrical pulses [5, 18, 20]. These electrical 2ThispaperdoesnottreattheHodgkin-HuxleyPDEs:spatialdependenceisnotrelevanthere. 2 pulses,calledactionpotentials,aretheprimarywayinwhichneuronstransmitinformation. Actionpoten- tialsaretriggered bysufficiently largemembrane voltages, which can be setup by the influx of ions into thecell. Aneuronissaidtofireorspikewhenitgeneratesanactionpotential(Fig. 1). TheHodgkin-Huxley model describes action potential generation in terms of the membrane voltage and dimensionless gating variables which quantify the effective permeability (or conductance) of the membrane for various types of ions. TheoriginalHodgkin-Huxleyequationsmodelactionpotentialgenerationinthesquidgiantaxon.This giantaxoncontainstwotypesofmembraneionchannels. Onetypeofchannelisspecifictopotassiumions, theothertosodiumions. Thestatevariablesofthemodelarethemembranevoltagev,theactivationnof thepotassiumchannels,andtheactivationmandinactivationhofthesodiumchannels. Theequationsare [18] v˙ = C−1 −I−g¯ n4(v−v )−g¯ m3h(v−v )−g¯ (v−v ) K K Na Na leak leak m˙ = α (v)(1−m)−β (v)m m (cid:2) m (cid:3) (1) n˙ = α(v)(1−n)−β(v)n n n h˙ = α(v)(1−h)−β(v)h h h where α (v)= Ψ v+25 , β (v)= 4exp(v/18), m 10 m α(v)=0.1(cid:16)Ψ v+(cid:17)10 , β(v)=0.125exp(v/80), n 10 n (2) α(v)=0.07e(cid:16)xp(v/(cid:17)20), β(v)= 1 , h h 1+exp(v+30) 10 Ψ(v)= v . exp(v)−1 Each ion channel consists of independent, identical subunits which must all open to allow ions to pass through. Thegatingvariablesm,n,andhtakevaluein(0,1)andrepresentthefractionofsubunitswhich are open. The term n4 enters into the potassium conductance because potassium channels consist of 4 identical subunits; analogous structures account for the m3h term in the sodium conductance [5]. The gatingvariableequationsaremasterequationsforcontinuous-timeMarkovchainswithvoltage-dependent transitionratesαandβ;theMarkovchainsdescribetheopeningandclosingofthecorrespondingchannel subunits. The v˙ equation is Kirchoff’s current law. Action potentials are downward voltage spikes and a positive I corresponds toan inflow of positively-chargedions. The voltage convention hereis thatof [18] andoppositecontemporaryusage: themembranevoltagevisdefinedby v=voltageoutside−voltageinside. Actionpotentialsaregenerallyinitiatedbyperturbationstothemembranevoltage. Suchperturbations maybecaused,forinstance,bytheflowofionsacrossthecellmembrane.Becauseneuronstransmitsignals through spatially and temporally localized pulses, it is natural to model stimuli as impulses [31]. The simplesttypeofrepetitive,pulsatilestimulustoaneuronisaperiodicimpulsetrain. Thismeansreplacing thev˙ equationaboveby v˙ =C−1 −I−g¯ n4(v−v )−g¯ m3h(v−v )−g¯ (v−v ) (3) K K Na Na leak leak h(cid:229) i +A G(t−kT), k∈Z where G is a “bump” function such that G(t)dt = 1. For simplicity, most of this paperuses the choice G(t)=δ(t);Section5.2discussestheresponseoftheHodgkin-Huxleysystemtoapulsatiledrivewith R 1/t , 0 ≤t ≤t G(t)= 0 0 . (4) ( 0, otherwise 3 0.0 ) v ( e g -20.0 a t ol v -40.0 e n a r -60.0 b m e M -80.0 0.0 10.0 20.0 30.0 40.0 Time (t) Figure1:ThetimecoursefortheHodgkin-Huxleyequationsattheparametervalues(5). Therapid“spike” followedbyalong“recovery”periodistypicaloftheHodgkin-Huxleyequations. Mathematically,onecanalsochoosetoperturbthegatingvariables,butsuchperturbationsarenotentirely naturalandarenotconsideredhere. ThispaperusestheoriginalHodgkin-Huxleyparameters[18]: v = −115mV, g¯ =120mΩ−1/cm2, Na Na v =+12mV, g¯ =36mΩ−1/cm2, K K (5) v =−10.613mV, g¯ =0.3mΩ−1/cm2, leak leak C =1µF/cm2. TimeismeasuredinmillisecondsandcurrentdensityinµA/cm2. Figure 2 shows a bifurcation diagram for the unforced Hodgkin-Huxley equations. When I = 0, the neuron maintains a stable rest state, corresponding to the branch of stable fixed points on the left of the diagram.AsufficientlylargevalueofIcausesaneurontofirerepeatedly,whichcorrespondstothecreation of a limit cycle through a saddle-node bifurcation of periodic orbits. Further increasing I destablizes the reststatethroughasubcriticalHopfbifurcation. Inthispaper,the injectedcurrentisalwayssettoavaluenear I ≈ 14,correspondingtoasteadyionic currentwhichdestabilizesthereststate. Thephenomena studiedhereareinsensitive totheprecisevalue of I aslong asitensuresthe existenceof astable limitcycle andanunstablefixed point. Asexplainedin theIntroduction,thesepropertiessimplifytheinterpretationofnumericalsimulations. Forthischoiceof I, theJacobianoftheHodgkin-Huxleyvectorfield(Eq. 1)attheunstablefixedpointhastworealeigenvalues {−4.97815,−0.146991} in the left half plane, and a complex conjugate pair 0.0763367±0.61866i in the right half plane. The fixed point thus has 2-dimensional stable and unstable manifolds. The Lyapunov exponentsassociatedwiththelimitcycleare0,≈ −0.20,≈ −2.0,and≈ −8.3.ItsperiodisT ≈ 12.944. 0 4 20 0 Stable fixed point Unstable fixed point -20 V -40 Unstable cycle -60 -80 Limit cycle -100 6 7 8 9 10 11 12 13 14 I Figure2:ThebifurcationdiagramfortheHodgkin-HuxleyequationsastheinjectedcurrentIisvaried.The lineinthemiddlemarksthevcoordinateofthereststate. Thesolidbluepartisstablewhilethedashedred partisunstable. Solidblackdotsnearthetopandthebottomofthefigurearethemaximumandminimum v valuesof limit cycles. Empty blackcirclesarethe maximum and minimum v valuesof unstable cycles. The fixed point undergoes a subcritical Hopf bifurcationas I increases. This diagramis computed using XPPAUT[7]. 5 A = 5 A = 20 0 0 t t n n e e n n o o p-2 p-2 x x e e v v o o n n u u p p a a y y L L st-4 st-4 e e g g r r a a L L -6 -6 2 3 4 5 6 7 8 2 3 4 5 6 7 8 Multiple of intrinsic period (T/T0) Multiple of intrinsic period (T/T0) Figure3: Asymptoticpropertiesofthepulse-drivenflowaredescribedbythedynamicsofthetime-Tmap F (see Eq. 6) and its largest Lyapunov exponent λ . Entrainment corresponds to λ < 0, and chaos T max max corresponds toλ > 0. This figure showsλ as a function of the drive period T, with T ranging from max max T ≈ 13 (the intrinsic period of the unforced Hodgkin-Huxleysystem; see §2) to 8·T ≈ 101. Left: Kick 0 0 amplitudeis A = 5. Right: Kickamplitudeis A = 40. Note(i)λ (T+T ) ≈λ (T);(ii)presenceofboth max 0 max positive and negative exponents for strong kicks (right), and only zero and negative exponents for weak kicks (left); and (iii) the presenceof more negative exponents than positive ones. See §4 for a discussion. LyapunovexponentsareestimatedbyiteratingF for1000stepsandtrackingtherateofgrowthofatangent T vector. 6 1.0 Rotations 0.8 Sinks y 0.6 t bili a b o Pr 0.4 Chaos 0.2 Unknown 0.0 10.0 20.0 30.0 40.0 Drive amplitude Figure 4: The probability of different response types, as a function of the drive amplitude A. For each driveamplitude A,thefractionofdriveperiodsT ∈ [T ,8T ]forwhichλ (F ) > 0,etc.,iscomputedby 0 0 max T samplingfromauniformgridin[T ,8T ]. Itisnaturaltoequatethesefractionswithprobabilitiesbecause 0 0 the Lyapunov exponentsareroughlyperiodicfunctions of T (andbecome moresoas T → ∞), asshown inFigure3andexplainedin§4. Empiricaldefinitions: Letλˆ denotetheestimatedLyapunovexponentandǫ theestimatedstandarderror. Then“chaos”isdefinedasλˆ > 3ǫ,“entrainment”λˆ < −3ǫ,and“rotation” λˆ <ǫ/3. (cid:12) (cid:12) (cid:12) (cid:12) 7 3 Main numerical results Lyapunov exponents provide a convenient way to characterize the asymptotic dynamics of Eq. 3. Let φ : R4 → R4 denote the flow map generated by the unforced Hodgkin-Huxley equations, T the drive t period,and Athedriveamplitude. ThePoincare´ map F (v,m,n,h)=φ (v+A,m,n,h) (6) T T takesaHodgkin-Huxleystatevector(v,m,n,h),appliesapulseofamplitude Atothemembranevoltage, thenevolvesitfor time T. Iteratingthe map F thusgivesastroboscopic recordofthe stateof ourpulse- T drivenHodgkin-Huxleysystembeforethearrivalofeachpulse. Thelong-termdynamicalbehaviorofthe pulse-driven Hodgkin-Huxley oscillator can be deduced from the asymptotic dynamics of F , which is T characterizedbyits(largest)Lyapunovexponentλ [13]: max λ <0 ⇔ F hassinks ⇔ kickedflowisentrainedtoinput max T λ =0 ⇔ F isquasiperiodic ⇔ kickedflowdriftsrelativetoinput max T λ >0 ⇔ F chaotic ⇔ kickedflowischaotic max T (Sinksrefertostablefixedpointsandstableperiodicorbits.) NotethatofthescenariosgivenintheIntro- duction, transientchaosalonedoesnotappearinthislist: Lyapunovexponents, beinglong-time, average quantities,cannotdetecttransientchaos. Figure 3 shows the Lyapunov exponents of F as a function of T/T , where T is the period of the T 0 0 Hodgkin-Huxley limit cycle. Different colors correspond to differentvalues of A. The periodicity of the response as a function of T is apparent. Because the response type as a function of T is approximately identicalovereachperiod[nT ,(n+1)T ],itmakessensetospeakoftheprobabilitythatarandomlychosen 0 0 driveperiod T willelicita particularasymptotic behavior, forexamplechaos. Moreprecisely, periodicity ensures that the fraction p of drive periods T in [0,nT ] for which λ > 0 converges to a well-defined n 0 max limitasn→ ∞. Similarstatementsholdforλ <0andλ = 0. max max Figure4showstheseprobabilitiesasfunctionsof A. At A = 10,theprobabilityofobtainingapositive exponent is roughly 20% and the probability of obtaining a negative exponent is roughly 70%. Thus, if oneweretopickT randomlyoutofaninterval[NT ,(N+1)T ]forlarge,fixedinteger N,theprobability 0 0 thatλ (F ) > 0 is about 20%. Figure 4 shows thatwhen A is small, the most likely type of behavior is max T rotation-likebehavior.ThispossibilitybecomeslesslikelyasAincreases.Atthesametime,sinksandchaos both become more likely, with sinks dominating the scene. One feature of Figure 4 specific to the pulse- drivenHodgkin-HuxleyflowisthatwhenAislarge,thesystemprefersentrainmentoverchaosinthesense thatentrainmenthashigher probability. Thispreferenceismorepronounced as A increases. Note thatin computingLyapunovexponentsnumerically,weonlyhaveaccesstofinitetimeinformation. Inprinciple, thismeansitisvirtuallyimpossibletodistinguishpersistentchaoticbehaviorfromtransientchaoscaused bya“large”horseshoe(butsee§5.1). In all numerical simulations shown in this paper, Eq. 1 is integrated using an adaptive integrator of Runge-Kutta-Fehlbergtype,withanerrortoleranceof 10−6 (inthesup norm)[30]. ThelargestLyapunov exponentλ of F is computed in a straightforward manner, by choosing a nonzero unit vector w ∈ R4 max T andestimatingtherateofgrowthof ||(DF )nw||. Thematrix-vectorproduct(DF )nw iseasilycomputed T T viathevariationalequations x˙ = H(x),w˙ = DH(x)·wfortheHodgkin-Huxleyvectorfield H (DHisthe Jacobianmatrixof H;see[10]fordetails). 8 4 Discussion 4.1 Response to asingle pulse: phaseresetting curves Thissectionreviewsphaseresettingcurves. SeeWinfree[38],GlassandMackey[11],andBrown,Moehlis, and Holmes [4] for more details and applications, and Guckenheimer and Holmes [13] for background informationondynamicalsystemstheory. See[7,8,9,37]forfurtherdiscussionsofphaseresettingcurves. Letφ : Rn → Rn be a flow generated bya smooth vector field with a hyperbolic limit cycleγ. Such t a limit cycle represents a stable nonlinear oscillator. The basin of attraction of γis denoted B(γ). The hyperbolicityofγguaranteesthatpointsinB(γ)convergetoγexponentiallyfast.(Itisconvenienttouseγ torefertoboththetrajectoryγ:R→Rn andtheinvariantpointsetitdefines.) Animpulsiveperturbation (“kick”) to the nonlinear oscillator can be defined by specifying a kick amplitude A and a kick direction Kˆ :Rn →Rn anddefiningafamilyofkickmaps K (x)= x+A·Kˆ(x), (7) A so that kicks send each point x ∈ Rn to K (x). For what follows, K should be smooth and satisfy A A K (B(γ))⊂ B(γ). A The Hodgkin-Huxley system with the value of I given in §2 is a nonlinear oscillator whose basin B(γ) is an open subset of R4. The kick map corresponding to an instantaneous voltage spike is simply K (v,m,n,h)=(v+A,m,n,h).Asin§3,itisconvenienttointroducethetime-Tmap A F =φ ◦K , (8) T T A where◦denotesfunctioncomposition. IteratingF givesastroboscopicrecordofthesystemstatebeforethe T arrivalofeachkick, andthusdescribesthelong-time dynamicsofthe flowφ underrepeated, T-periodic t kicks. Becausethephasedimensionnmaybelarge,thedynamicsofF :Rn →Rnmaybedifficulttoanalyze. T Winfreeobservedthateverypointnearthelimitcycleγmustconvergetoγast →∞,sotheflownearγis dominatedbytherotationalmotionalongγ. Thus,onecanreducethedimensionofthephasespacefrom nto1,atleastheuristically. Todothis, firstdefinethephasefunctionθ:γ→ [0,T ) byfixingareference 0 pointx ∈γandrequiringthatφ (x ) = xforallx ∈γ. Byconstruction,θsatisfies d (θ(γ(t))) =1,0≤ 0 θ(x) 0 dt t < T . The functionθcanbe extendedtoafunctionθ: B(γ) → [0,T ) byprojectingalongstrong-stable 0 0 manifolds3: if yisapointinthebasinofγthenθ(y)isdefinedtobeθ(x),wherexistheuniquepointsuch that y∈W (x).Thisdefinitionofphasepreservesthepropertythat d (θ(φ(x))) =1. ss dt t Considerthelimit[12] F¯ = lim F . (10) T n→+∞ T+nT0 Themap F¯ iswell-definedonthebasinofγandretractsthebasinontoγ,i.e. F¯ (x) ∈γforallx ∈ B(γ). T T Thus, F¯ induces an interval map f : [0,T ) → [0,T ) which, given the current phase of the system, T T 0 0 yieldsthenewphaseafterkickingandevolvingthesystemfortimeT. Thatis, f (θ(x)) =θ(F (x))forall T T x ∈ B(γ). 3Thestrong-stablemanifoldW (x)ofx∈γistheset ss Wss(x)=(cid:26)y∈Rn:n∈Zl,inm→+∞φnT0(y)→x(cid:27). (9) Whenthevectorfieldgeneratingφtissmooth,thestrong-stablemanifoldsare(locally)smoothsubmanifoldsofRn.Thestrong-stable linearsubspaceE (x)isthetangentspaceofW (x)atx.See[12,13]. ss ss 9 Driveamplitude A Prob. ofsinknearplateau Prob. ofλ <0 max 5 41% 48% 10 58% 62% 20 68% 70% 30 76% 78% Table 1: Estimates of the probabilityof obtaining sinks near the plateau, as a function of A. The datafor thistableiscomputedbytryingabout40valuesof T foreachchoiceof A andexaminingthegraphofthe first return map to the interval [4,10] (chosen to coincide with the “plateau”)and its intersection(s) with thediagonal. The map f isthe phaseresetting curve4, or more preciselythe finite phase resettingcurve(infinitesimal T phaseresettingcurves[4,7]arenotneededhere).Byconstruction,ithasthepropertythat fT+δ(t)= fT(t)+δ(modT0). (11) Thus,thefamilyofmaps{f }isperiodicinT. T PeriodicityindriveperiodT. Theapproximateperiodicityofλ (F )seeninFigure3iseasytounder- max T standheuristically: kickingtheoscillatoreveryTsecondsandkickingiteveryT+T secondsshouldyield 0 the same asymptotic response because the oscillator simply traversesγatfrequency 1/T between kicks. 0 Onecanrestatethisusingphaseresettingcurves: ifthedriveperiodT issufficientlylargeandθ(x ) = t , 0 0 thenthe f -orbit t , f (t ), f2(t ),... shouldcloselyfollowthephases θ(x ),θ(F (x )),θ(F2(x )),... of T 0 T 0 T 0 0 T 0 T 0 thecorrespondingF -orbit. Since f = f ,thissuggeststhatλ (F )≈λ (F ). (cid:0) T T+(cid:1)T0 T max T+(cid:0)T0 max T (cid:1) Preference for entrainment. Figure 5 shows phase resetting curves for the Hodgkin-Huxley equations forvariousvaluesofdriveperiod T anddriveamplitude A. Forsufficientllysmallvaluesof A,thephase resettingcurvesarecirclediffeomorphisms: eithertherearesinks(i.e. stablefixedpiontsorstableperiodic orbits), or the map is conjugate to a rotation on a circle and the response of the kicked oscillator drifts relativetotheperiodicdrive. AsAincreases,thegraphof f ratherquicklyfoldsoverandacquirescritical T points. AstrikingfeatureofthegraphsinFigure5isthe“plateau,”aphaseintervaloverwhich f varies T veryslowly. Another striking featureisthe “kink” aroundθ≈ 9.8. These featuresarediscussed inmore depthin§5. Fornow,noticethattheplateauprovidesasimplemechanismforcreatingsinks: changingthe kickperiodTshiftsthegraphof f vertically. Wheneverthegraphintersectsthediagonalwithaderivative T f′ <1,thenastablefixedpointiscreated. T ThismechanismcanbeusedtoverifytheresultsofFigure4: computethegraphofthefirstreturnmap (cid:12) (cid:12) o(cid:12)f f(cid:12) toanintervalaround the plateau, then shiftthe graphverticallyusing a number of differentvalues T of T andestimatethefractionofT’sforwhich f hasastablefixedpoint(seeFigure6). Table1showsthe T results. For A = 10, the 58% probability of sinks corresponds fairly closely with Figure 4. It is unclear whether the ambiguous exponents in Figure 4 reallyrepresentpositive or negative Lyapunov exponents. If a significant fraction of the ambiguous exponents are really negative, then they must come from small sinks. 4WangandYoungrefertophaseresettingcurvesassingularlimits.Phaseresettingcurvesarealsosometimescalledphasetransition curves[11]. 10