AMERICAN

METEOROLOGICAL
SOCIETY

Journal of Climate

EARLY ONLINE RELEASE

This is a preliminary PDF of the author-produced

manuscript that has been peer-reviewed and
accepted for publication. Since it is being posted
so soon after acceptance, it has not yet been
copyedited, formatted, or processed by AMS
Publications. This preliminary version of the
manuscript may be downloaded, distributed, and
cited, but please be aware that there will be visual
differences and possibly some content differences
between this version and the final published version.

The DOI for this manuscript is doi: 10.1175/JCLI-D-15-0848.1

The final published version of this manuscript will replace the

preliminary version at the above DOI once it is available.
If you would like to cite this EOR in a separate work, please use the following full
citation:

Pierini, S., M. Ghil, and M. Chekroun, 2016: Exploring the Attractors of a Low-
Order Quasigeostrophic Ocean Model: The Deterministic Case. J. Climate.
doi:10.1175/JCLI-D-15-0848.1, in press.

© 2016 American Meteorological Society

.tex source Click here to download LaTeX File (.tex, .sty, .cls, .bst, .bib)
Pierini_etal_Revision.tex

1 Exploring the Attractors of a Low-Order Quasigeostrophic Ocean Model:

2 The Deterministic Case

3 Stefano Pierini∗

4 Universita’ di Napoli Parthenope, Napoli, Italy

5 Michael Ghil

6 Ecole Normale Supérieure, Paris, France, and

7 University of California at Los Angeles, Los Angeles, California, USA

8 Mickael D. Chekroun

9 University of California at Los Angeles, Los Angeles, California, USA

10
∗ Corresponding author address: Dipartimento di Scienze e Tecnologie, Universita’ di Napoli

11 Parthenope. Centro Direzionale, Isola C4, 80143 Napoli, Italy

12 E-mail: [email protected]

Generated using v4.3.2 of the AMS LATEX template 1

 ABSTRACT

13 A low-order quasigeostrophic double-gyre ocean model is subjected to an

14 aperiodic forcing that mimics time dependence dominated by interdecadal

15 variability. This model is used as a prototype of an unstable and nonlinear

16 dynamical system with time-dependent forcing to explore basic features of

17 climate change in the presence of natural variability. The study relies on

18 the theoretical framework of non-autonomous dynamical systems and of their

19 pullback attractors (PBAs), i.e., of the time-dependent invariant sets attracting

20 all trajectories initialized in the remote past. The existence of a global PBA

21 is rigorously demonstrated for this weakly dissipative, nonlinear model. En-

22 semble simulations are carried out and the convergence to PBAs is assessed by

23 computing the probability density function (PDF) of localization of the trajec-

24 tories. A sensitivity analysis with respect to forcing amplitude shows that the

25 PBAs experience large modifications if the underlying autonomous system is

26 dominated by small-amplitude limit cycles, while less dramatic changes oc-

27 cur in a regime characterized by large-amplitude relaxation oscillations. The

28 dependence of the attracting sets on the choice of the ensemble of initial states

29 is then analyzed. Two types of basins of attraction coexist for certain param-

30 eter ranges; they contain chaotic and non-chaotic trajectories, respectively.

31 The statistics of the former does not depend on the initial states whereas the

32 trajectories in the latter converge to small portions of the global PBA. This

33 complex scenario requires separate PDFs for chaotic and non-chaotic trajec-

34 tories. General implications for climate predictability are finally discussed.

2
35 1. Introduction and motivation

36 Weather and climate predictability have been investigated for several decades, starting from

37 the pioneering work of Lorenz (1963), by relying on the concepts and methods of autonomous

38 dissipative nonlinear dynamical systems (Ghil and Childress 1987, and references therein). For

39 such systems, any initial volume in phase space contracts in time, on the average, thus eventually

40 reducing to a time-invariant set of zero volume called an attractor. In the autonomous case, an

41 attractor can be a fixed point, a limit cycle, an invariant torus or a strange attractor (Eckmann and

42 Ruelle 1985; Ott 2002; Tél and Gruiz 2006).

43 Investigating a system’s attractors, in particular the strange attractors that describe the statistical

44 properties of deterministic chaos, is fundamental in this context (Eckmann and Ruelle 1985) and

45 has helped considerably in the study of multiple weather regimes and of their predictability on

46 time scales of weeks to months (Ghil and Robertson 2002, and references therein). On the other

47 hand, when studying the longer time scales — interannual, interdecadal and longer — that are as-

48 sociated with climate change, it becomes necessary to take into consideration the time dependence

49 of both anthropogenic and natural forcing (Martinson et al. 1995; Mann et al. 1998; Chang et al.

50 2015). The proper framework for studying changes in the system’s internal variability, and not

51 just in its mean properties, when subject to variable forcing is that of non-autonomous and random

52 dynamical systems (Drótos et al. 2015; Ghil 2015, 2016).

53 In the latter framework, the natural generalization of the attractors is provided by pullback at-

54 tractors (PBAs). A global PBA is defined in the mathematical literature as a time-dependent set

55 A (t) in the system’s phase space X that is invariant under its governing equations — together

56 with the equally time-dependent, invariant measure µ(t) supported on this set — and to which all

57 trajectories in X starting in the distant past converge (Arnold 1998; Rasmussen 2007; Kloeden and

3
58 Rasmussen 2011; Carvalho et al. 2012). As we shall see later, such a global PBA might include

59 two or more local attractors, which only attract trajectories from certain subsets of X. These con-

60 cepts are clarified and discussed further in the present paper’s appendix. When the equations or the

61 forcing include stochastic processes, a PBA is called a random attractor. In the physical literature,

62 Romeiras et al. (1990) introduced the somewhat vaguer concept of a snapshot attractor.

63 The concepts and methods of non-autonomous and random dynamical systems, including pull-

64 back attraction, were introduced into climate dynamics by Ghil et al. (2008) and Chekroun et al.

65 (2011), and they were also pursued vigorously by T. Tél and his group (Bódai et al. 2011; Bódai

66 and Tél 2012; Bódai et al. 2013). Several recent studies have explored the properties of PBAs

67 for low-order climate models using ensemble simulations (Bódai et al. 2011, 2013; Bódai and Tél

68 2012; De Saedeleer et al. 2013; Pierini 2014a; Drótos et al. 2015).

69 Drótos et al. (2015) used the simple conceptual climate model of Lorenz (1984), forced by

70 a periodic seasonal component and a linear decrease in the imposed equator-to-pole temperature

71 difference.These authors argued that the PBAs — or snapshot attractors, using the simpler concept

72 introduced by Romeiras et al. (1990) — and their natural probability distributions are the only

73 tools with which mathematically sound statements can be made at a given point in time within a

74 changing climate. Their claim was based on the finding that the time-dependent, chaotic attractor

75 of the Lorenz (1984) model was independent of the initial states, or initial data (IDs) for short,

76 within the parameter ranges and for the ensembles of IDs they considered. This claim appears to

77 be fully justified for the Lorenz (1984) model, which seems to possess a unique global PBA in

78 those parameter ranges.

79 The ensemble simulations of a periodically forced quasigeostrophic ocean model by Pierini

80 (2014a) also provided evidence of ID-independence in chaotic regimes, but dependence on the

81 IDs did emerge for particular parameter ranges, for which chaotic basins of attraction coexist

4
82 with non-chaotic basins; trajectories starting in the the latter converge onto small subsets of the

83 PBA. Thus, the fundamental problem of the dependence of the natural probability distribution of

84 a changing climate on initial states needs further investigation.

85 In this paper, we carry out such an investigation. The four-variable quasigeostrophic model

86 of the ocean’s wind-driven, double-gyre circulation formulated by Pierini (2011) and used in the

87 periodically forced simulations of Pierini (2014a) is studied here via ensemble simulations sub-

88 ject to a deterministic aperiodic forcing that mimics time dependence dominated by interdecadal

89 climate variability. Two reference cases separated by a tipping point are considered: above the

90 tipping point, large-amplitude relaxation oscillations shape the behavior of the autonomous sys-

91 tem, whereas below it, small-amplitude oscillations are dominant. In this second range the system

92 is excitable, because an appropriate time-dependent forcing can excite the relaxation oscillations

93 (e.g., Pikovsky and Kurths 1997; Pierini 2011, 2012). Various properties of these two reference

94 cases are analyzed, such as the time of convergence from arbitrary IDs to PBAs and the sensitivity

95 of the latter to forcing amplitude, as well as the dependence of the local attractors on the choice of

96 the initial ensemble of states.

97 The paper is organized as follows. In section 2, the model is described and basic aspects of

98 the autonomous system’s solutions and of the two reference ensemble simulations under aperiodic

99 forcing are discussed. In section 3, the time of convergence from arbitrary IDs to the PBAs is

100 estimated. In section 4, the sensitivity of PBAs with respect to the forcing amplitude is inves-

101 tigated. Section 5 is devoted to the dependence of PBAs on the choice of the ensemble of IDs:

102 chaotic and non-chaotic basins of attraction are analyzed in detail. In section 6, the main results

103 are summarized and discussed, and an appendix provides the rigorous mathematical background.

5
104 2. Model description

105 a. Model formulation

106 The mathematical model used in this study was developed by Pierini (2011) to analyze funda-

107 mental dynamical features of the decadal variability in the Kuroshio Extension (Dijkstra and Ghil

108 2005, and references therein). This model is derived through projection of the governing equations

109 onto an orthonormal basis and subsequent low-order truncation, following the approach of Platz-

110 man (1960), Saltzman (1962), Lorenz (1963) and many others since, e.g. Lorenz (1982), Legras

111 and Ghil (1985), Ghil and Childress (1987), Olbers (2001) and Crucifix (2012). S. Pierini and

112 colleagues had already investigated this decadal variability with a relatively realistic primitive-

113 equation ocean model (Pierini 2006; Pierini et al. 2009; Pierini 2010, 2014b).

114 The model is derived via a severe truncation of the spectral representation of the evolution equa-

115 tion of potential vorticity in the quasigeostrophic, reduced-gravity approximation, in a rectangular

116 domain. The cartesian orthonormal basis onto which the streamfunction ψ(x,t) is projected is

117 composed of two trigonometric functions in the zonal and two in the meridional direction; the

118 zonal ones incorporate, moreover, the exponential factor proposed by Jiang et al. (1995) to account

119 for westward intensification of midlatitude surface currents. Here x is the vector of horizontal co-

120 ordinates and t is time.

121 The nonlinear coupled ordinary differential equations for the variables Ψi (t), i = 1, · · · , 4 are

122 given by:

Ψ̇1 + L11 Ψ1 + L13 Ψ3 + B1 (Ψ, Ψ) = W1 (t),

Ψ̇2 + L22 Ψ2 + L24 Ψ4 + B2 (Ψ, Ψ) = W2 (t),

(1)
Ψ̇3 + L33 Ψ3 + L31 Ψ1 + B3 (Ψ, Ψ) = W3 (t),

Ψ̇4 + L44 Ψ4 + L42 Ψ2 + B4 (Ψ, Ψ) = W4 (t).

6
123 Here Ψ denotes the vector of expansion coefficients (Ψ1 , Ψ2 , Ψ3 , Ψ4 ) of the streamfunction, W

124 will be used as the vector of forcing terms (W1 ,W2 ,W3 ,W4 ), and the bilinear terms Bi are given by

B1 (Ψ, Ψ) = 2J112 Ψ1 Ψ2 + 2J114 Ψ1 Ψ4 + 2J123 Ψ2 Ψ3 + 2J134 Ψ3 Ψ4 ,

B2 (Ψ, Ψ) = J211 Ψ21 + J222 Ψ22 + J233 Ψ23 + J244 Ψ24

+ 2J213 Ψ1 Ψ3 + 2J224 Ψ2 Ψ4 ,

B3 (Ψ, Ψ) = 2J312 Ψ1 Ψ2 + 2J314 Ψ1 Ψ4 + 2J323 Ψ2 Ψ3 + 2J334 Ψ3 Ψ4 ,

B4 (Ψ, Ψ) = J411 Ψ21 + J422 Ψ22 + J433 Ψ23 + J444 Ψ24 (2)

+ 2J413 Ψ1 Ψ3 + 2J424 Ψ2 Ψ4 .

125 The analytical expressions of the Jacobian coefficients Ji jk , the linear ones Li j , and the forcing

126 terms Wi were derived by Pierini (2011); moreover, all the parameter values used here are the

127 same as adopted in that study and in Pierini (2014a). Pierini (2012) used the model governed

128 by Eqs. (1) and (2) to systematically analyze abrupt climate changes seen as noise-induced tran-

129 sitions in excitable systems, while S. Vannitsem and colleagues (Vannitsem 2014a,b; Vannitsem

130 and De Cruz 2014) used it to represent the ocean component in their low-order ocean-atmosphere

131 coupled models. Pierini (2014a) applied periodic forcing to model (1) and explored the PBAs of

132 the forced system. The present study can thus be seen as the extension of the latter study to an

133 aperiodic forcing case.

134 In this study, the forcing W(x,t) is defined as

 
W(x,t) = γ 1 + ε f (t) w(x), f (t) = FT f [ζ (t)]. (3)

135 Here w(x) is the double-gyre wind stress curl used in Pierini (2011), with γ its dimensionless

136 intensity, while ζ (t) is a fixed realization of an Ornstein-Uhlenbeck process, normalized to have

137 unit variance and an autocorrelation time Ta . This fixed realization is smoothed with the sliding

7
138 window FT f of width T f , and the resulting aperiodic forcing f (t) is weighted by the dimensionless

139 parameter ε.

140 b. Model behavior

141 In the autonomous case, there is no t-dependence and we let ε = 0. The bifurcation diagram in

142 this case is plotted in Fig. 1a. This diagram is obtained by forward integration of the model, starting

143 from vanishing initial data and dropping a sufficiently long initial segment of the trajectories. The

144 time needed for the trajectories to land on the model’s attractor is discussed in the appendix.

145 The value γ = 0.96 corresponds to the blue limit cycle shown in Figs. 1b,c, in the planes (Ψ1 , Ψ3 )

146 and (Ψ2 , Ψ4 ), respectively. This limit cycle lies above the supercritical Hopf bifurcation that

147 occurs at γ ' 0.348, cf. Pierini (2011), and resembles a linear, harmonic oscillation.

148 The value γ = 1.1, though, corresponds to a much larger and more complex limit cycle, plotted

149 in red in Figs. 1b,c. This limit cycle is associated with a relaxation oscillation that can be excited

150 by an appropriate noise or deterministic time-dependent forcing when the system lies below a ho-

151 moclinic bifurcation (e.g., Simonnet et al. 2005, Pierini 2011). The sudden growth in the size and

152 complexity of the closed, periodic trajectory at γ = 1.0 corresponds to such a global bifurcation.

153 Pierini (2011, 2012, 2014a) did, in fact, study already the present model in the presence of both

154 stochastic and deterministically periodic forcing, and found that such a homoclinic bifurcation

155 does occur in it and that this bifurcation gives rise to what is currently termed a tipping point by the

156 climate community (e.g., Lenton et al. 2008). Such a dynamical mechanism is a typical paradigm

157 of abrupt climate change associated with an intrinsically nonlinear mode of variability (Lorenz

158 1963; Ghil and Childress 1987; Crucifix 2012), and it is sometimes called coherence resonance

159 (Pikovsky and Kurths 1997). When coherence resonance does occur, the system displays strong

160 dependence on forcing amplitude, as we will see in section 4.

8
161 In the non-autonomous case, ε 6= 0 in Eq. (3), and we use the fixed aperiodic forcing f (t)

162 obtained with autocorrelation time Ta and a sliding window FT f of width T f . Figure 2a shows a

163 realization with Ta = T f = 15 yr and ε = 0.2. This choice provides an idealized aperiodic forcing

164 that mimics a North Pacific multidecadal variability (Chao et al. 2000; Chang et al. 2015) for

165 the minimal representation of the Kuroshio Extension provided by (1). On the other hand, the

166 relaxation oscillations of the autonomous model display decadal variability as well, in agreement

167 with the oceanic phenomenon that inspired this simple model. The dimensional time variable

168 used throughout this paper refers to this multidecadal interpretation, although the present model’s

169 properties are likely to be shared by a wide class of systems of climate relevance.

170 Apart from this oceanographic interpretation, the forcing in Eq. (3) extends the analysis of

171 Pierini (2014a) from purely periodic to deterministically aperiodic. This choice disrupts the cyclo-

172 stationarity and cycloergodicity properties of PBAs and makes the forced model’s behavior more

173 realistic.

174 The evolution of the model solutions subject to the forcing of Fig. 2a is shown in Figs. 2b and

175 2c for two reference cases, γ = 0.96 and γ = 1.1, respectively; the solutions of the corresponding

176 autonomous system, i.e. with ε = 0, are plotted in Figs. 1(b,c). In the present figure, N trajectories

177 Ψk (t) emanate from N IDs uniformly distributed at time t0 = 0 on a reference subset Γ of the

178 (Ψ1 , Ψ3 )-plane, which is shown in panels (b,c) here, while the (Ψ2 , Ψ4 ) coordinates of the initial

179 states (not shown) are chosen the same way as in Pierini (2014a).

180 Figs. 2(b,c) provide a first representation of the sets that approximate the corresponding PBAs;

181 the small number of trajectories is in fact chosen for the sake of graphical clarity. The correct

182 identification and characterization of the PBAs requires, however, an analysis of the PDFs that

183 evolve along the trajectories and of the convergence to the appropriate invariant time-dependent

184 sets. This convergence is investigated in the next section.

9
185 3. Convergence to PBAs

186 As mentioned in the introduction, PBAs are the extension of the attractors of dissipative au-

187 tonomous dynamical systems to the non-autonomous case. When time-dependent forcing is

188 present, the trajectories originating from the set Γ = (−70 ≤ Ψ1 ≤ 150, −150 ≤ Ψ3 ≤ 120) of

189 IDs at time τ tend to a subset A (τ,t) of phase space at time t > τ, but while A (t) depends on τ

190 in general, in the limit of τ → −∞ this dependence vanishes.

191 The limit τ → −∞ is the pullback limit and it provides the natural generalization of the forward

192 asymptotics associated with autonomous systems (Arnold 1998; Rasmussen 2007; Kloeden and

193 Rasmussen 2011; Carvalho et al. 2012; Ghil et al. 2008; Chekroun et al. 2011); see appendix

194 A here. In practical terms, though, a question arises: how large should |τ| be for A (τ,t) to be

195 virtually independent of τ.

196 An effective way to answer this question is to compute the PDF of the localization of the trajec-

197 tories that emanate from Γ. To obtain a function that provides a clear graphical representation of

198 the invariant measure on the attracting set, the following parameter is computed:

PΨ3 = log10 (1000pΨ3 ) (4)

199 where pΨ3 (t) is the PDF of localization of the Ψ3 -variable using the N = 644 trajectories shown

200 in Figs. 2(b,c). In order to make pΨ3 independent of the length Ttot = tfin − τ of the signal ( this is

201 important to make the panels of Fig. 3 comparable, see below), Ttot is decomposed into n cells and

202 pΨ3 is scaled in such a way that its integral over each cell equals 1. In our case Ttot = 400 yr and

203 n = 300.

204 In Figs. 2(b’,c’), we plot PΨ3 (t) for the trajectories shown in Figs. 2(b, c), respectively. The vari-

205 ability in PΨ3 induced by the forcing of panel (a) is impressive when compared with Figs. 4(a,f)

206 that are discussed in section 4 and were obtained with ε = 0. Of particular interest is the remark-

10
207 able reduction of the range of variability of PΨ3 found at t ' 170 − 185 yr and at t ' 295 − 310 yr.

208 This variability reduction is most evident in the case with γ < 1 (panel (b’)) and it corresponds

209 to a partial synchronization of the trajectories. As we will see in Sect. 4, both synchronization

210 episodes occur when the total forcing γ[1 + ε f (t)] decreases well below unity.

211 To estimate the rate of converge to the PBA, five cases are considered in Fig. 3, for both γ = 0.96

212 and γ = 1.1. The trajectories emanate from the same distribution of IDs but at the different initial

213 times τ = t01 , · · · ,t05 = 200, 150, 100, 50, 0 yr. Comparison of PΨ3 (t;t0i ) with PΨ3 (t;t0i+1 ) shows

214 that the two distributions coincide when t in the former is greater than t ' t0i + 15 yr. We can

215 therefore conclude that the set to which the trajectories converge approximates well the PBA after

216 roughly 15 yr from the beginning of the integration, and that this important result holds at any time

217 during the nonstationary process.

218 4. Sensitivity of PBAs to the amplitude of the forcing

219 Model robustness is important in making any inferences about the role of time-dependent forcing

220 (Ghil 2015, and references therein). We thus proceed to analyze in this section the sensitivity of

221 the present model’s PBAs with respect to the amplitude ε of the time-dependent component of the

222 forcing, while in the next section dependence on IDs will be analyzed.

223 Figure 4 shows the dependence of PΨ3 (t) on ε, with this amplitude increasing from ε = 0 —

224 i.e., the autonomous case — in the top two panels, to ε = 0.2 — the value used in Fig. 2 — in

225 the bottom two panels, while the overall intensity γ of the total wind stress equals 0.96 in the left

226 column and 1.1 in the right column, as in Figs. 2 and 3. The PBAs corresponding to the blue and

227 red limit cycles visible in Fig. 1b are shown here in panels (a,f). It is interesting to notice that,

228 for ε = 0, any point of Γ evolves toward the same limit cycle, while the phase of each orbit on its

11
229 limit cycle does depend on the IDs. Note that the limit cycles in Figs. 1(b,c) were obtained, like

230 the bifurcation diagram in Fig. 1a, from vanishing initial data, {Ψi (0) = 0 : i = 1, 2, 3, 4}.

231 Inspecting the panels of the left column, which correspond to (γ < 1), from top to bottom,

232 shows a dramatic transition from the small limit cycles, in the top two rows, to the highly variable

233 ones of the bottom two rows. To better understand this behavior, the total time-dependent forcing

234 amplitude γ̄ = γ[1 + ε f (t)] is plotted at the bottom of each panel. Each transition from small- to

235 large-amplitude oscillations is preceded by an increase of γ̄ above the critical level γ̄ = 1; consider,

236 for instance, the one already noticed in Figs. 2(b’) and 4(e). More generally, coherence resonance

237 is activated when γ̄ > 1, provided this has occurred at least 10–30 yr in advance; note that even

238 very small positive excursions above γ̄ = 1 yield the same behavior, as shown for example in

239 Fig. 4(c).

240 This result agrees with Pierini (2012), who noticed that model (1), when subjected to stochastic

241 forcing, behaves approximately like an adiabatically changing autonomous dynamical system, as

242 far as coherence resonance is concerned. Likewise, when γ̄ decreases well below unity, the system

243 tends to produce small-amplitude limit cycles, according to the bifurcation diagram of Fig. 1(a):

244 this corresponds to the synchronization events already discussed.

245 The same behavior is found for γ > 1, in the right column of Fig. 4, but now the relaxation

246 oscillations are typically present, since they are self-sustained and since γ̄ > 1 is the norm. The

247 consequence of the fluctuations of γ̄ on PΨ3 (t) can again be easily interpreted by taking into ac-

248 count the diagram of Fig. 1(a). Three synchronization events are present for ε = 0.2, panel (j),

249 after intervals during which γ̄ < 1. In general, if γ > 1, the changes in the PBA are less dramatic.

12
250 5. PBA dependence on the choice of initial ensemble

251 In Section 3, we have investigated the evolution of the same set of IDs starting from different

252 initial times τ; here we study the evolution of different initial states starting from the same ini-

253 tial time τ = 0. This second analysis is fundamental in a climate perspective, as it concerns the

254 dynamical origin of the PBAs and of the invariant measures supported on them, as well as the

255 dependence of predictability on IDs.

256 In the periodic-forcing case, Pierini (2014a) provided numerical evidence that the present ocean

257 model’s PBA can be obtained from any set of IDs, provided its evolution is chaotic; see Figures

258 3(e,f) therein. Drótos et al. (2015) have recently stressed that the chaotic PBAs of the Lorenz

259 (1984) atmospheric model are independent of IDs and have suggested that this is likely to be the

260 typical situation in climate models.

261 PBA dependence on IDs may, however, be more complex. In his systematic study of the present

262 model’s PBAs as a function of the forcing period, Pierini (2014a) identified periods for which a

263 chaotic basin of attraction — which converges onto the global PBA — coexists with non-chaotic

264 basins. Thus, for example, a chaotic basin can coexist with another basin whose phase-space

265 flow converges onto a single periodic trajectory lying on the global PBA; see, for example Figs.

266 6(a–c) of Pierini (2014a). It is therefore natural to investigate whether a similar behavior may be

267 present also in the aperiodic-forcing case studied in the present paper. In the following, we use

268 two complementary mathematical tools to explore this possibility.

269 a. Finite-time Lyapunov exponents

270 The typical approach to the investigation of chaotic vs. non-chaotic behavior is to evaluate the

271 finite-time Lyapunov exponents (FTLEs) of the system (Nese 1989). FTLEs have been widely

13
272 used in physical oceanography to help identify localized coherent structures and material surfaces

273 (Haller 2001; D’Ovidio et al. 2009).

274 If two nearby points — whose distance according to Eulerian metrics in phase space is δ0 at

275 time t = 0 and δt at time t — diverge exponentially in time, the Lyapunov exponent

1 δt
λt = ln , (5)
t δ0

276 is independent of t and it is positive; if, on the contrary, the trajectories converge exponentially,

277 then λt is still independent of t and it is negative. In the presence of a bound on the region of

278 phase space occupied by all possible model trajectories, divergence of trajectories indicates that

279 the behavior is chaotic (e.g., Ghil and Childress 1987).

280 To test whether chaotic behavior may coexist with non-chaotic trajectories, we computed maps

281 of λt = λ (Ψ1 , Ψ3 ,t) for trajectories that originate from a homogeneous distribution of δ0 at times

282 t = 20, 40 and 60 yr. These maps are plotted in Fig. 5 for both γ = 0.96 and γ = 1.1.

283 The first thing to note in all six panels is the existence of extensive regions in which λt > 0 that

284 coexist with regions where λt < 0, so the chaotic nature of the flow is far from being independent

285 on IDs for the model (1), subject to the aperiodic forcing of Fig. 2a. Another notable feature is

286 that λt exhibits strong time dependence, for both γ < 1 (upper panels) and γ > 1 (lower panels).

287 The remarkable decrease of the range of λt as t increases from t = 20 yr to t = 60 yr, both

288 in Figs. 5(a–c) and in Figs. 5(d–f), is to be expected, due to the saturation of the exponential

289 behavior, but the spatial distribution depends strongly on time as well. This dependence clearly

290 indicates that complex temporal behavior lies behind these maps, which cannot, therefore, provide

291 an unambiguous indication of the flow’s chaoticity depending on (Ψ1 , Ψ3 ). Such complexities are

292 further discussed in connection with Fig. 8 below.

14
293 b. Mean normalized distance

294 A more robust identification of chaotic and non-chaotic regions within the set Γ of initial states

295 can be obtained by introducing and evaluating the metric σ , defined by

σ (Ψ1 , Ψ3 ) = hδn iTtot , (6)

296 where the normalized distance δn , given by δn = δt /δ0 , is averaged over the whole forward time

297 integration Ttot = tfin −tinit of the available trajectories. The maps of σ in Fig. 6 reveal large chaotic

298 regions where δn  1 on average (warm colors) but also non-chaotic regions where σ ≤ 1 (blue)

299 and thus initially close trajectories do remain close on average.

300 To check the character of the two types of evolution, we plot in Fig. 7 many trajectories that

301 leave two small subsets of Γ for which σ  1 — namely the sets A and C in Fig. 6, for γ = 0.96

302 and γ = 1.1, respectively — to compare with those that leave the sets B and D of Fig. 6, for which

303 σ ≤ 1.

304 Figures 7(a,c) confirm the extreme sensitivity to IDs that lie in boxes A and C: a comparison of

305 these two panels with Figs. 2(b, c), respectively, shows that the attracting sets to which trajectories

306 started in A and B converge are dense in the PBAs obtained by starting from the whole Γ. The

307 corresponding PDFs, as represented by PΨ3 (not shown) coincide in fact with those of Figs. 2(b’,c’)

308 and the same result is obtained by starting from any small region of Γ for which σ  1. Panels

309 (b,d) of Fig. 7 also confirm that the trajectories leaving from boxes B and D converge to time-

310 dependent sets in the (Ψ1 , Ψ3 )-plane whose measure is much smaller than that of either B or D.

311 c. Discussion

312 It is worth exploring in greater detail the evolution of some of the trajectories plotted in Fig. 7.

313 We display in Figs. 8(a–d) the Ψ3 -time series of two initially nearby trajectories selected among

15
314 those of Figs. 7(a,b,c,d), respectively; each pair of trajectories has lines one in blue and one in

315 red, both drawn from the same panel of Fig. 7. The corresponding normalized distances δn (t) are

316 shown in Figs. 8(a’,b’,c’,d’).

317 In the chaotic cases, represented by panels (a,a’) and (c,c’), the values of δn (t) range from

318 δn (t) ' 1 to δn (t)  1, up to order 102 ; the former very small values occur over the same time

319 intervals — already apparent in Figs. 2(b’,c’) — in which synchronization sets in. In the non-

320 chaotic cases, like those of panels (b,b’) and (d,d’), the values of δn lie typically below unity,

321 although an initial divergence is noticeable in panel (d’); this slow divergence may be due to a

322 transient growth of polynomial, rather than exponential nature.

323 Inspection of Figs. 8(a’–d’) confirms the validity of the mean normalized distance σ of Eq. (6)

324 as a sharper diagnostic tool than the FTLEs. Longer integrations (not shown) simply extend in

325 time the ensemble of trajectories shown in Figs. 7(a,b), but the chaotic vs. non-chaotic character

326 of the trajectories —as diagnosed by σ — is preserved.

327 To further analyze the structure of the model’s PBAs, the intersection of many trajectories with

328 the (Ψ1 , Ψ3 )-plane at a reference time t = 330 yr is shown in Figs. 9(a,b) for γ = 0.96 and γ = 1.1,

329 respectively. The grey dots in both panels arise from 15 000 IDs evenly distributed in Γ and are

330 associated with the filamentary patterns that characterize chaotic attractors (e.g., Chekroun et al.

331 2011; Drótos et al. 2015).

332 The red dots arise from the small boxes A and C that are located in the chaotic regions of

333 Figs. 6(a,b), respectively. These red dots, too, are distributed all over the corresponding attractors

334 in the two panels, as expected from the previous analysis; in fact, only the 420 red dots that

335 correspond to the trajectories of Figs. 7(a,c) have been plotted here, to avoid complete overlapping

336 with the grey dots.

16
337 On the contrary, the cyan dots that arise from the large non-chaotic boxes 1,2,3, and 4

338 of Figs. 6(a,b) converge onto small subsets of the PBAs. This convergence onto a smaller-

339 dimensional set is very similar to what Pierini (2014a) found for some forcing periods that yielded

340 periodic PBAs; for example, see Figs. 6(a–c) therein.

341 For γ > 1 in Fig. 9(b), the similarity with the result of Pierini (2014a) is striking because all

342 the trajectories originating from boxes 3 and 4 of Fig. 6(b) converge toward two distinct points.

343 It is worth noting that the blue boxes 3 and 4 of Fig. 6(b) are separated by a chaotic ridge which

344 apparently acts as a barrier separating two distinct basins of attraction, each converging to a single

345 point. I this case the asymptotic evolution is independent of initial data. On the other hand, for

346 γ < 1 (Fig. 9(a)) the attracting cyan sets are small segments, so the evolution of the non-chaotic

347 trajectories does depend, though very weakly, on IDs.

348 Graphs like those of Fig. 9 cannot provide quantitative information on the probability of local-

349 ization of trajectories on the model’s global PBA. Such information can be obtained, though, from

350 the three-dimensional PDFs plotted in Fig. 10.

351 For the subcritical case γ < 1 in the left column of the figure, the distribution is highly inhomo-

352 geneous and concentrates in two areas labeled 1 and 2 in panel (a); these concentrations extend

353 over very small portions of the filament and result from large contributions of both chaotic (panel

354 (c)) and non-chaotic (panel (e)) trajectories, with a larger contribution from the latter. In the super-

355 critical case γ > 1 (right column), the same situation holds, with two peaks contained in the small

356 area labeled 3.

357 In summary, in the small regions of phase space in which the invariant sample measure on

358 our model’s global PBA peaks in a filamentary pattern, both chaotic and non-chaotic trajectories

359 coexist. The PDF peaks in Figs. 10(a,b) are clearly an indication of enhanced overall predictability,

17
360 and recall those observed in Figs. 6 and 7 of Chekroun et al. (2011) for a highly idealized model

361 of the El Niño–Southern Oscillation with seasonal forcing.

362 6. Summary and conclusions

363 In this study, a low-order quasigeostrophic model of the wind-driven ocean circulation has been

364 used as a prototype of an unstable, nonlinear, non-autonomous dynamical system of climatic rel-

365 evance. We have studied this double-gyre model’s PBAs — i.e., those time-dependent invariant

366 sets that attract all trajectories initialized in the remote past — along with the invariant sample

367 measures that live on these sets. PBAs (Arnold 1998; Ghil et al. 2008; Chekroun et al. 2011) and

368 their close cousins, snapshot attractors (Romeiras et al. 1990; Bódai et al. 2011, 2013; Bódai and

369 Tél 2012; Drótos et al. 2015) have only recently been recognized as the natural tools for investi-

370 gating basic features of a changing non-equilibrium climate, and the theory of non-autonomous

371 dynamical systems as the proper mathematical framework for such investigations.

372 We chose here an aperiodic forcing dominated by interdecadal variability to mimic the effects

373 of multidecadal climate changes (Chang et al. 2015) on the midlatitude double-gyre circulation;

374 see section 2. The present study generalizes therewith a set of results previously obtained in the

375 presence of periodic forcing (Pierini 2014a). We recall that, in the absence of time-dependent

376 forcing, i.e. when ε = 0 in Eq. (3), a tipping point at γ = 1 separates small-amplitude, nearly

377 harmonic oscillations from large-amplitude relaxation oscillations.

378 In section 3, the convergence of ensembles of trajectories to PBAs was assessed in order to esti-

379 mate the time required to identify the PBAs. To do so, we considered — in addition to projections

380 of the trajectories from the full, four-dimensional phase space onto a plane spanned by the Ψ1 and

381 Ψ3 variables — the suitably scaled, time-dependent statistic PΨ3 (t); see Eq. (4) and Fig. 2. A key

382 result of this section is that about 15 yr suffice for trajectories to converge to the PBA.

18
383 The sensitivity of PBAs with respect to the amplitude ε of the time-dependent portion of the

384 forcing was studied in section 4. Below the tipping point γ = 1, the range of variability within

385 the PBA increases dramatically as ε increases, and so does the intermittency; see left column of

386 Fig. 4. For γ > 1, the changes in both range and intermittency are less striking, cf. right column of

387 this figure. The difference in behavior is clearly associated with the solutions of the autonomous

388 system: a small-amplitude limit cycle for γ < 1 vs. a higher-amplitude, more complex and self-

389 sustained relaxation oscillation for γ > 1.

390 Finally, the dependence of the attracting sets on the choice of the ensemble of initial states is

391 investigated in section 5. Our investigation relies, on the one hand, on the classical FTLEs to

392 measure the rate of divergence of trajectories, but introduces also a novel metric, namely the time

393 mean σ of the normalized distance δn between two initially nearby trajectories. The use of δn

394 and σ , cf. Eq. (6), is required by the extreme complexity of the evolution of the distance between

395 the trajectories: this evolution is far from purely exponential and it involves large variations over

396 different time scales and episodes of synchronization induced by the forcing, as seen in Fig. 8.

397 In our model, the statistics of an ensemble of trajectories in a chaotic regime does depend on the

398 ensemble of states chosen in the remote past. Figures 6, 7 and 9, in particular, illustrate very well

399 that the model’s global PBA — whose existence is rigorously proven in the appendix — contains

400 both chaotic and non-chaotic trajectories.

401 Both of these types of trajectories have attractor basins that seem to be separated by fractal

402 boundaries: the former fill the global PBA, while the latter form small and smooth subsets of the

403 PBA. This complex, interwoven pattern of rapidly divergent and non-divergent trajectories shows

404 that a time-independent PDF, as often used in predictability studies, is not sufficient to describe

405 the model’s long-term behavior nor to characterize its predictability.

19
406 These intriguing results have been obtained with a highly idealized nonlinear model subject to

407 deterministic aperiodic forcing. As usual in this kind of approach, they raise as many questions

408 as they answer about the broader issues of climate modeling and climate predictability. The two

409 most urgent questions for the climate sciences are:

410 • To which extent is this simple model’s behavior characteristic of a wide class of non-

411 autonomous dynamical systems, whether deterministic or stochastic?

412 • How relevant is this class to climate modeling and climate predictability studies?

413 Again as usual, these questions require one to climb the rungs of a hierarchy of successively more

414 complex and detailed climate models (Ghil 2001; Dijkstra and Ghil 2005; Ghil 2015, and refer-

415 ences therein). Naturally, introducing noise in the forcing may substantially modify the system’s

416 behavior: a random dynamical system extension of the present study is in progress.

20
417 Acknowledgments. SP and MG would like to thank the TEMASAV Project of the Regione

418 Campania of Italy (POR Campania FSE 2007/2013) and the coordinator, Giancarlo Spezie, for

419 having supported exchange visits between the Environmental Research and Teaching Institute

420 (CERES-ERTI) in Paris and the Department of Science and Technology (DiST) of the University

421 of Naples Parthenope, thus promoting this collaboration. The research herein was supported by

422 grants N00014-12-1-0911 and N00014-16-1-2073 from the Multidisciplinary University Research

423 Initiative (MURI) of the Office of Naval Research and by the US National Science Foundation

424 grant OCE 1243175 (MDC and MG). The authors wish to thank two anonymous reviewers, whose

425 detailed and constructive comments helped improve the manuscript.

21
426 APPENDIX

427 Existence of a Global Pullback Attractor

428 We present in this appendix rigorous mathematical results on the existence of pullback attractors

429 (PBAs) in ordinary differential equation (ODE) models — like the one given by Eqs. (1) and (2)

430 here — and, in particular, the conditions on the forcing term W = (W1 , · · · ,W4 ) for such an exis-

431 tence to be guaranteed. General results are known from the specialized literature on the existence

432 of pullback attractors or related invariant manifolds (Kloeden and Rasmussen 2011; Carvalho et al.

433 2012; Chekroun et al. 2015a,b). For the sake of concision and clarity, however, we provide below

434 the main elements of such an existence theory, while emphasizing the energy estimates involved;

435 see also Kondrashov et al. (2015, Thm. 3.1 and Cor. 3.2).

436 In particular, these existence results are shown to apply to Eqs. (1) and (2) considered in the main

437 text. We mention also that the approach presented below can be adapted to the infinite-dimensional

438 setting and thus to the original two-dimensional quasi-geostrophic partial differential equation

439 model, by working in the appropriate function spaces to define the corresponding solutions.

440 The less mathematically inclined reader can skip section A.a and proceed directly to section A.b.

441 a. Rigorous mathematical background

442 In what follows X denotes the Euclidean space Rd (d ≥ 1), endowed with its natural inner

443 product h·, ·i, and B(0, r(t)) denotes the closed ball in X centered at zero, with time-dependent

444 radius r(t) > 0.

445 Theorem 1. Consider the following evolution equation for Ψ ∈ X,

dΨ
+ LΨ + B(Ψ, Ψ) = W (t), (A1)
dt

22
446 for which the initial value problem with Ψ(0) = Ψ0 possesses a unique global solution for any

447 Ψ0 ∈ X.

448 Assume that the bilinear term B is energy preserving, i.e.

hB(Ψ, Ψ), Ψi = 0, for all Ψ ∈ X, (A2)

449 and that there exists α > 0 such that, for all Ψ ∈ X,

hLΨ, Ψi ≥ αkΨk2 , (A3)

450
2 (R, X) — the space of X-valued, locally square-
Suppose furthermore that W belongs to Lloc

451 integrable functions — and that

Z t
eσ s kW(s)k2 ds < ∞, for all t ∈ R, (A4)
−∞

452 with

σ = 2(α − ε), (A5)

453 for some ε > 0 such that α > ε.

454 Then the propagator U associated with (A1) possesses a unique pullback attractor that pullback

455 attracts any time-dependent set D(t) such that D(t) ⊂ B(0, rσ (t)), with rσ (t) that satisfies the

456 growth condition

lim eσt rσ2 (t) = 0. (A6)

t→−∞

457 Remark 1. Inequality (A3) states that the linear terms in Eq. (1) include dissipative effects.

458 Remark 2.

459 (i) We recall that a propagator on X is a two-parameter family of continuous mappings U(t, τ) :

460 X → X, such that U(τ, τ) is the identity operator in X, U(τ, τ)x = x, and that the “multiplica-

461 tion rule”

U(t, τ) = U(t, s)U(s, τ) for all τ ≤ s ≤ t (A7)

23
462 applies. It is trivial to show that when a non-autonomous ODE system, such as Eq. (A1),

463 possesses a unique global solution for any Ψ0 ∈ X, then a propagator U is well defined and

464 is actually given by U(t, τ)Ψ0 := Ψ(t, τ; Ψ0 ), where Ψ(t, τ; Ψ0 ) denotes the solution at time

465 t of Eq. (A1) which emanates from Ψ0 at time τ ≤ t.

466 (ii) A time-dependent set A (t) is said to pullback attract the time-dependent set D(t) if

lim distX (U(t, τ)D(τ), A (t)) = 0, for all t, (A8)

τ→−∞

467 where distX denotes the Hausdorff semi-distance in X (Carvalho et al. 2012; Chekroun et al.

468 2011).

469 Proof. Let us multiply Eq. (A1) by Ψ. Energy conservation by the quadratic term, cf. Eq. (A2),

470 then yields

1 dkΨk2
+ hLΨ, Ψi = hW, Ψi. (A9)
2 dt

471 By using the dissipativity condition of Eq. (A3) on hLΨ, Ψi and Young’s inequality (Brézis 2010,

472 Ch. 2, Eq. (2) and footnote on p. 92) on hW, Ψi, we obtain that, for all ε > 0, there exists a constant

473 Cε > 0 such that

dkΨk2
+ 2(α − ε)kΨk2 ≤ Cε kWk2 . (A10)
dt

474 Now by choosing ε sufficiently small for σ = 2(α − ε) > 0 to hold, we obtain, after integration

475 over [τ,t],

Z t
2 −σ (t−τ) 2 −σt
kΨ(t)k ≤ e kΨ(τ)k +Cε e eσ s kW(s)k2 ds. (A11)
τ

476 If Ψ(τ) ∈ D(τ) ⊂ B(0, rσ (τ)) with rσ that satisfies the slow-growth condition of Eq. (A6), then

477 the first term on the right-hand side converges to zero as τ → −∞, ensuring thus the existence of a

24
478 dissipation time τ0 (t) such that, for all τ ≤ τ0 (t),

Z t
2 −σt
kΨ(t)k ≤ 2Cε e eσ s kW(s)k2 ds ≡ R(t). (A12)
−∞

479 We have thus proved that U(t, τ)D(τ) ⊂ B(0, R(t)) for τ ≤ τ0 (t), ensuring in turn the desired

480 pullback dissipation. The existence and uniqueness proceeds then from this dissipation property

481 and the theory of pullback attractors [e.g., Carvalho et al. (2012)].

482 b. Interpretation

483 We provide here some interpretations of the application of Theorem 1 in the context of our

484 idealized climate model governed by Eqs. (1) and (2). First, the dissipativity parameter α of

485 condition (A3) above is given here by

α = min (Re(λ )), (A13)

λ ∈σ (L)

486 where σ (L) denotes the spectrum of the linear operator L. For the parameter regime presented in

487 the main text, we find α = 1.1487 × 10−5 .

488 The result of Theorem 1 states thus that, for any forcing W for which σ defined in (A5) is

489 strictly positive, the double-gyre ocean model of Eq. (1) possesses a pullback attractor in the sense

490 of Theorem 1. The constant σ in Eq. (A11) is a measure of the rate of pullback dissipation. This

491 constant is very small, of the order of 10−5 , which helps explain, as we shall see, the complexity

492 of the model’s phase portrait.

493 The identity (A8) means that A (t) attracts — with respect to the Hausdorff semi-distance and

494 in a pullback sense — any time-dependent set D(τ) of initial states at time τ, as they evolve with

495 the action of the propagator U(t, τ) forward in the time t and as the pullback time τ is sent further

496 and further away into the past.

25
497 However the Hausdorff semi-distance is not a distance in the classical sense. For instance,

498 distX (E , F ) = 0 implies that the set E is included in the set F , but does not guarantee the con-

499 verse. So if the global attractor A (t) turns out to be the union of a strange part and of some

500 topologically simpler parts, such as fixed points or periodic orbits, it is definitely possible to have

501 pullback attraction toward the global PBA A (t), while single trajectories land in different regions

502 of A (t) that are either chaotic or not. We are then back, in a pullback setting, to the notion of

503 local attractors that compose the global attractor and that can be of topologically very different na-

504 tures. In the autonomous context, such coexistence of topologically distinct local attractors is well

505 known in the climate sciences (Ghil and Childress 1987; Dijkstra and Ghil 2005; Simonnet et al.

506 2005, 2009, and references therein), while it is also observed in the numerical results of Figs. 5

507 and 6 here for the present ocean model.

508 As shown in these figures, the boundaries between the corresponding local basins of attractions,

509 as measured with the metrics of Section 5, are seemingly intricate and strongly suggest that these

510 objects have fractal features. An attractor basin of chaotic solutions being separated by a fractal

511 boundary from the attractor basin of a fixed point has been documented in a quadratic, autonomous

512 model governed by four ODEs that arises in population dynamics; see Fig. 5a in Roques and

513 Chekroun (2011).

514 In that ecological model, however, no abutting of two different attractor basins of chaotic regions

515 on the global attractor is present. A key difference between the ocean model studied herein and

516 the population dynamics model of Roques and Chekroun (2011) lies in the quadratic terms here

517 being energy preserving, cf. Eq. (A2), while this is not the case in population dynamics.

518 There is, therefore, a Hamiltonian skeleton here that may help explain the difference between

519 the two phase portraits: The pullback dissipation being very small argues for the asymptotic trajec-

520 tories — in spite of the dissipative nature of the dynamics — lying quite close to this skeleton. It

26
521 is thus not surprising to recover interleaved chaotic islands, as shown in Fig. 6. In the autonomous

522 context, the effects of small dissipation have been studied by Ghil and Wolansky (1992), Feudel

523 and Grebogi (1997) and Seoane et al. (2007), among others.

27
524 References

525 Arnold, L., 1998: Random Dynamical Systems. Springer.

526 Bódai, T., G. Károlyi, and T. Tél, 2011: A chaotically driven model climate: extreme events and

527 snapshot attractors. Nonl. Proc. Geophys., 18 (5), 573–580.

528 Bódai, T., G. Károlyi, and T. Tél, 2013: Driving a conceptual model climate by different processes:

529 Snapshot attractors and extreme events. Phys. Rev. E, 87 (2), 022 822.

530 Bódai, T., and T. Tél, 2012: Annual variability in a conceptual climate model: Snapshot attrac-

531 tors, hysteresis in extreme events, and climate sensitivity. Chaos: An Int. J. Nonl. Sci., 22 (2),

532 023 110.

533 Brézis, H., 2010: Functional Analysis, Sobolev Spaces and Partial Differential Equations.

534 Springer.

535 Carvalho, A., J. A. Langa, and J. Robinson, 2012: Attractors for Infinite-Dimensional Non-

536 Autonomous Dynamical Systems. Springer.

537 Chang, C. P., M. Ghil, L. M., and J. M. Wallace, Eds., 2015: Climate Change: Multidecadal and

538 Beyond. World Scientific Publ. Co./Imperial College Press.

539 Chao, Y., M. Ghil, and J. C. McWilliams, 2000: Pacific interdecadal variability in this century’s

540 sea surface temperatures. Geophys. Res. Lett., 27, 2261–2264.

541 Chekroun, M. D., H. Liu, and S. Wang, 2015a: Approximation of Stochastic Invariant Manifolds:

542 Stochastic Manifolds for Nonlinear SPDEs I. Springer Briefs in Mathematics, Springer, New

543 York.

28
544 Chekroun, M. D., H. Liu, and S. Wang, 2015b: Parameterizing Manifolds and Non-Markovian

545 Reduced Equations: Stochastic Manifolds for Nonlinear SPDEs II. Springer Briefs in Mathe-

546 matics, Springer, New York.

547 Chekroun, M. D., E. Simonnet, and M. Ghil, 2011: Stochastic climate dynamics: Random attrac-

548 tors and time-dependent invariant measures. Physica D, 240, 1685–1700.

549 Crucifix, M., 2012: Oscillators and relaxation phenomena in Pleistocene climate theory. Phil.

550 Trans. Roy. Soc. A, 370, 1140–1165.

551 De Saedeleer, B., M. Crucifix, and S. Wieczorek, 2013: Is the astronomical forcing a reliable and

552 unique pacemaker for climate? A conceptual model study. Clim. Dyn., 40 (1-2), 273–294.

553 Dijkstra, H. A., and M. Ghil, 2005: Low-frequency variability of the large-scale ocean circulation:

554 A dynamical systems approach. Rev. Geophys., 43, RG3002.

555 D’Ovidio, F., J. Iserne-Fontanet, C. Lopez, E. Hernandez-Garcia, and E. Garcia-Ladon, 2009:

556 Comparison between Eulerian diagnostics and finite-size Lyapunov exponents computed from

557 altimetry in the Algerian basin. Deep-Sea Res. Part I, 56, 15–31.

558 Drótos, G., T. Bódai, and T. Tél, 2015: Probabilistic concepts in a changing climate: A snapshot

559 attractor picture. J. Clim., 28 (8), 3275–3288.

560 Eckmann, J.-P., and D. Ruelle, 1985: Ergodic theory of chaos and strange attractors. Rev. Modern

561 Phys., 57, 617–656.

562 Feudel, U., and C. Grebogi, 1997: Multistability and the control of complexity. Chaos: an Inter-

563 disciplinary Journal of Nonlinear Science, 7 (4), 597–604.

564 Ghil, M., 2001: Hilbert problems for the geosciences in the 21st century. Nonlin. Processes Geo-

565 phys., 8, 211–222.

29
566 Ghil, M., 2015: A mathematical theory of climate sensitivity or, how to deal with both anthro-

567 pogenic forcing and natural variability? Climate Change: Multidecadal and Beyond, C. P.

568 Chang, M. Ghil, L. M., and J. M. Wallace, Eds., World Scientific Publ. Co./Imperial College

569 Press, 31–51.

570 Ghil, M., 2016: The wind-driven ocean circulation: Applying dynamical systems theory to a

571 climate problem. Discr. Cont. Dyn. Syst. A, submitted.

572 Ghil, M., M. D. Chekroun, and E. Simonnet, 2008: Climate dynamics and fluid mechanics: Natu-

573 ral variability and related uncertainties. Physica D, 237, 2111–2126.

574 Ghil, M., and S. Childress, 1987: Topics in Geophysical Fluid Dynamics: Atmospheric Dynamics,

575 Dynamo Theory and Climate Dynamics. Springer-Verlag.

576 Ghil, M., and A. W. Robertson, 2002: “Waves” vs. “particles” in the atmosphere’s phase space: A

577 pathway to long-range forecasting? Proc. Natl. Acad. Sci. USA, 99 (Suppl. 1), 2493–2500.

578 Ghil, M., and G. Wolansky, 1992: Non-hamiltonian perturbations of integrable systems and reso-

579 nance trapping. SIAM J. Appl. Math., 52, 1148–1171.

580 Haller, G., 2001: Distinguished material surfaces and coherent structures in three-dimensional

581 fluid flows. Physica D: Nonlinear Phenomena, 149 (4), 248–277.

582 Jiang, S., F.-F. Jin, and M. Ghil, 1995: Multiple equilibria, periodic, and aperiodic solutions in a

583 wind-driven, double-gyre, shallow-water model. J. of Phys. Ocean., 25 (5), 764–786.

584 Kloeden, P. E., and M. Rasmussen, 2011: Nonautonomous Dynamical Systems. 176, Amer. Math.

585 Soc.

586 Kondrashov, D., M. D. Chekroun, and M. Ghil, 2015: Data-driven non-Markovian closure models.

587 Physica D: Nonlinear Phenomena, 297, 33–55.

30
588 Legras, B., and M. Ghil, 1985: Persistent anomalies, blocking and variations in atmospheric pre-

589 dictability. J. Atmos. Sci., 42, 433–471.

590 Lenton, T. M., H. Held, E. Kriegler, J. Hall, W. Lucht, S. Rahmstorf, and H. Schellnhuber, 2008:

591 Tipping elements in the earth’s climate system. Proc. Natl. Acad. Sci. USA, 105 (6), 1786–1793.

592 Lorenz, E. N., 1963: Deterministic nonperiodic flow. J. Atmos. Sci., 20, 130–141.

593 Lorenz, E. N., 1982: Low-order models of atmospheric circulations. J. Meteor. Soc. Japan, 60,

594 255–267.

595 Lorenz, E. N., 1984: Irregularity: a fundamental property of the atmosphere. Tellus A, 36 (2),

596 98–110.

597 Mann, M. E., R. S. Bradley, and M. K. Hughes, 1998: Global-scale temperature patterns and

598 climate forcing over the past six centuries. Nature, 392, 779–787.

599 Martinson, D. G., K. Bryan, M. Ghil, M. Hall, T. R. Karl, E. S. Sarachik, S. Sorooshian, and L. D.

600 Talley, Eds., 1995: Natural Climate Variability on Decade-to-Century Time Scales. National

601 Research Council/National Academy Press.

602 Nese, J., 1989: Quantifying local predictability in phase space. Physica D: Nonlinear Phenomena,

603 35 (1), 237–250.

604 Olbers, D., 2001: A gallery of simple models from climate physics. Progress in Probability,

605 P. Imkeller, and J. von Storch, Eds., Vol. 49, Birkhäuser Verlag, 3–63.

606 Ott, E., 2002: Chaos in Dynamical Systems. Cambridge University Press.

607 Pierini, S., 2006: A Kuroshio Extension system model study: Decadal chaotic self-sustained os-

608 cillations. J. Phys. Ocean., 36 (8), 1605–1625.

31
609 Pierini, S., 2010: Coherence resonance in a double-gyre model of the Kuroshio Extension. J. Phys.

610 Ocean., 40 (1), 238–248.

611 Pierini, S., 2011: Low-frequency variability, coherence resonance, and phase selection in a low-

612 order model of the wind-driven ocean circulation. J. Phys. Ocean., 41 (9), 1585–1604.

613 Pierini, S., 2012: Stochastic tipping points in climate dynamics. Phys. Rev. E, 85 (2), 027 101.

614 Pierini, S., 2014a: Ensemble simulations and pullback attractors of a periodically forced double-

615 gyre system. J. Phys. Ocean., 44 (12), 3245–3254.

616 Pierini, S., 2014b: Kuroshio Extension bimodality and the North Pacific Oscillation: A case of

617 intrinsic variability paced by external forcing. J. Clim., 27 (1), 448–454.

618 Pierini, S., H. A. Dijkstra, and A. Riccio, 2009: A nonlinear theory of the Kuroshio Extension

619 bimodality. J. Phys. Ocean., 39 (9), 2212–2229.

620 Pikovsky, A., and J. Kurths, 1997: Coherence resonance in a noise-driven excitable system. Phys.

621 Rev. Lett., 78 (5), 775.

622 Platzman, G., 1960: The spectral form of the vorticity equation. J. Meteor., 17, 635–644.

623 Rasmussen, M., 2007: Attractivity and Bifurcation for Nonautonomous Dynamical Systems.

624 Springer.

625 Romeiras, F. J., C. Grebogi, and E. Ott, 1990: Multifractal properties of snapshot attractors of

626 random maps. Phys. Rev. A, 41 (2), 784.

627 Roques, L., and M. D. Chekroun, 2011: Probing chaos and biodiversity in a simple competition

628 model. Ecological Complexity, 8 (1), 98–104.

32
629 Saltzman, B., 1962: Finite amplitude free convection as an initial value problem. J. Atmos. Sci.,

630 19, 329–341.

631 Seoane, J. M., M. A. Sanjuán, and Y.-C. Lai, 2007: Fractal dimension in dissipative chaotic scat-

632 tering. Physical Review E, 76 (1), 016 208.

633 Simonnet, E., H. A. Dijkstra, and M. Ghil, 2009: Bifurcation analysis of ocean, atmosphere and

634 climate models. Computational Methods for the Ocean and the Atmosphere, R. Temam, and

635 J. J. Tribbia, Eds., North-Holland, 187–229.

636 Simonnet, E., M. Ghil, and H. Dijkstra, 2005: Homoclinic bifurcations in the quasi-geostrophic

637 double-gyre circulation. J. of Marine Res., 63 (5), 931–956.

638 Tél, T., and M. Gruiz, 2006: Chaotic Dynamics: An Introduction Based on Classical Mechanics.

639 Cambridge University Press.

640 Vannitsem, S., 2014a: Dynamics and predictability of a low-order wind-driven ocean–atmosphere

641 coupled model. Clim. Dyn., 42 (7-8), 1981–1998.

642 Vannitsem, S., 2014b: Stochastic modelling and predictability: analysis of a low-order coupled

643 ocean–atmosphere model. Phil. Trans. Roy. Soc. A, 372 (2018), 20130 282.

644 Vannitsem, S., and L. De Cruz, 2014: A 24-variable low-order coupled ocean–atmosphere model:

645 Oa-qg-ws v2. Geos. Model Dev., 7 (2), 649–662.

33
646 LIST OF FIGURES
647 Fig. 1. Bifurcation diagram of our idealized ocean model (1) in the autonomous case, i.e. with ε = 0
648 in Eq. (3). (a) The range of the variable Ψ1 is plotted vs. the wind stress intensity γ. (b, c)
649 The limit cycles corresponding to the blue and red vertical segments in panel (a) are shown
650 here projected onto the planes (Ψ1 , Ψ3 ) and (Ψ2 , Ψ4 ), respectively. . . . . . . . . 34

651 Fig. 2. Ensemble behavior of forced solutions of the double-gyre ocean model. (a) Time depen-
652 dence of the total forcing 1 + ε f (t), for ε = 0.2 . (b,c) Evolution of 644 initial states em-
653 anating from the subset Γ in the (Ψ1 , Ψ3 )-plane for (b) γ = 0.96 and (c) γ = 1.1 . (b’,c’)
654 Corresponding time series of PΨ3 . The set Γ is given by {−70 ≤ Ψ1 ≤ 150} × {−150 ≤
655 Ψ3 ≤ 120}. . . . . . . . . . . . . . . . . . . . . . . . 35

656 Fig. 3. Evolution in time of PΨ3 as a function of the initial time tinit . (Left column, a–e) γ = 0.96,
657 and (right column, f–j) γ = 1.1, for different initial times but with the same distribution of
658 IDs. . . . . . . . . . . . . . . . . . . . . . . . . . 36

659 Fig. 4. Evolution in time of PΨ3 as a function of the amplitude ε of the time-dependent part of
660 the forcing, ε f (t). (Left column) γ = 0.96, and (right column) γ = 1.1. The amplitudes
661 are (a,f) ε = 0, (b,g) ε = 0.01, (c,h) ε = 0.05, (d,i) ε = 0.1, and (e,j) ε = 0.2. The total
662 forcing amplitude γ̄(t) = γ[1 + ε f (t)] is reported at the bottom of each panel (the black
663 areas correspond to γ̄ > 1). . . . . . . . . . . . . . . . . . . . 37

664 Fig. 5. Maps of FTLEs for the present model. Upper panels: Maps of λ (Ψ1 , Ψ3 ,t) at (a) t = 20 yr;
665 (b) t = 40 yr, and (c) t = 60 yr, for γ = 0.96; lower panels (d–f): same for γ = 1.1. In
666 the color bars to the right, warm colors indicate instability, i.e. λt > 0, and cool colors the
667 opposite, while heavy black lines in the panels correspond to λ = 0. . . . . . . . . 38

668 Fig. 6. Mean normalized distance σ (Ψ1 , Ψ3 ) for 15 000 trajectories starting in the initial set Γ: (a)
669 γ = 0.96, and (b) γ = 1.1. The boxes A-D and 1-4 indicate subdomains of Γ from which
670 IDs are taken for our analysis (see text). . . . . . . . . . . . . . . . 39

671 Fig. 7. Evolution of ensembles of 420 trajectories each that emanate from small subsets of Γ. (a,
672 b) γ = 0.96 and IDs in boxes A and B of Fig. 6a, respectively; and (c,d) γ = 1.1 and IDs in
673 boxes C and D of Fig. 6b, respectively. . . . . . . . . . . . . . . . . 40

674 Fig. 8. Superposition of two time series of Ψ3 (t) starting from nearby IDs in Γ. The initial points
675 for γ = 0.96 are (a) (16, 36) and (b) (112, −132), while for γ = 1.1 they are (c) (−9, 51)
676 and (d) (95, −55); (a’,b’,c’,d’) are the corresponding normalized distances δn (t). The two
677 nearby trajectories of panels (a–d) are included in those of Figs. 7a–d, respectively. The
678 horizontal black lines mark the value δn = 1. . . . . . . . . . . . . . . 41

679 Fig. 9. (a) Intersection with the (Ψ1 , Ψ3 )-plane at t = 330 yr of 15 000 trajectories (grey dots)
680 emanating from the whole set Γ, and from boxes 1 and 2 (cyan dots, insets) and A (red dots)
681 of Fig. 6a for γ = 0.96; (b) same but from boxes 3 and 4 (cyan) and C (red) of Fig. 6b for
682 γ = 1.1. The insets in both panels zoom in on the way that the red dots associated with the
683 chaotic orbits and the cyan dots associated with the non-chaotic orbits are embedded into
684 the global PBA (grey dots). . . . . . . . . . . . . . . . . . . . 42

685 Fig. 10. PDF of localization of the 15 000 trajectories of Figs. 6 and 9 projected onto the (Ψ1 , Ψ3 )-
686 plane at t = 330 yr. (a,c,e) γ = 0.96 and (b,d,f) γ = 1.1. The PDF is computed in panels (a,b)
687 from all 15 000 trajectories emanating from Γ; in panels (c,d) only the chaotic trajectories
688 emanating from boxes A and C, respectively, are used; and in panels (e,f) only the non-

34
689 chaotic trajectories emanating from boxes 1-2 and 3-4, respectively, are used (compare with
690 Fig. 9). . . . . . . . . . . . . . . . . . . . . . . . . 43

35
 (a) (b) (c)

691 F IG . 1. Bifurcation diagram of our idealized ocean model (1) in the autonomous case, i.e. with ε = 0 in Eq. (3).
692 (a) The range of the variable Ψ1 is plotted vs. the wind stress intensity γ. (b, c) The limit cycles corresponding to
693 the blue and red vertical segments in panel (a) are shown here projected onto the planes (Ψ1 , Ψ3 ) and (Ψ2 , Ψ4 ),
694 respectively.

36
 (a) (b) (c)

(b’) (c’)

695 F IG . 2. Ensemble behavior of forced solutions of the double-gyre ocean model. (a) Time dependence of the
696 total forcing 1 + ε f (t), for ε = 0.2 . (b,c) Evolution of 644 initial states emanating from the subset Γ in the
697 (Ψ1 , Ψ3 )-plane for (b) γ = 0.96 and (c) γ = 1.1 . (b’,c’) Corresponding time series of PΨ3 . The set Γ is given by
698 {−70 ≤ Ψ1 ≤ 150} × {−150 ≤ Ψ3 ≤ 120}.

37
 t (yr)t (yr)

699 F IG . 3. Evolution in time of PΨ3 as a function of the initial time tinit . (Left column, a–e) γ = 0.96, and (right
700 column, f–j) γ = 1.1, for different initial times but with the same distribution of IDs.

38
 (a) (a) (f) (f)

(b) (b) (g) (g)

(c) (c) (h) (h)

(d) (d) (i) (i)

(e) (e) (j) (j)

t (yr)t (yr)

701 F IG . 4. Evolution in time of PΨ3 as a function of the amplitude ε of the time-dependent part of the forcing,
702 ε f (t). (Left column) γ = 0.96, and (right column) γ = 1.1. The amplitudes are (a,f) ε = 0, (b,g) ε = 0.01,
703 (c,h) ε = 0.05, (d,i) ε = 0.1, and (e,j) ε = 0.2. The total forcing amplitude γ̄(t) = γ[1 + ε f (t)] is reported at the
704 bottom of each panel (the black areas correspond to γ̄ > 1).

39
 (a) (b) (c)

(d) (e) (f)

705 F IG . 5. Maps of FTLEs for the present model. Upper panels: Maps of λ (Ψ1 , Ψ3 ,t) at (a) t = 20 yr; (b)
706 t = 40 yr, and (c) t = 60 yr, for γ = 0.96; lower panels (d–f): same for γ = 1.1. In the color bars to the right,
707 warm colors indicate instability, i.e. λt > 0, and cool colors the opposite, while heavy black lines in the panels
708 correspond to λ = 0.

40
 3 4
1
C
A

D
2

B (b)
(a)

709 F IG . 6. Mean normalized distance σ (Ψ1 , Ψ3 ) for 15 000 trajectories starting in the initial set Γ: (a) γ = 0.96,
710 and (b) γ = 1.1. The boxes A-D and 1-4 indicate subdomains of Γ from which IDs are taken for our analysis
711 (see text).

41
 (a) (c)

from A from C

(b) (d)

from B from D

712 F IG . 7. Evolution of ensembles of 420 trajectories each that emanate from small subsets of Γ. (a, b) γ = 0.96
713 and IDs in boxes A and B of Fig. 6a, respectively; and (c,d) γ = 1.1 and IDs in boxes C and D of Fig. 6b,
714 respectively.

42
 (c)

(a)
(a’) (c’)

(d)

(b)

(b’) (d’)

715 F IG . 8. Superposition of two time series of Ψ3 (t) starting from nearby IDs in Γ. The initial points for γ = 0.96
716 are (a) (16, 36) and (b) (112, −132), while for γ = 1.1 they are (c) (−9, 51) and (d) (95, −55); (a’,b’,c’,d’) are
717 the corresponding normalized distances δn (t). The two nearby trajectories of panels (a–d) are included in those
718 of Figs. 7a–d, respectively. The horizontal black lines mark the value δn = 1.

43
719 F IG . 9. (a) Intersection with the (Ψ1 , Ψ3 )-plane at t = 330 yr of 15 000 trajectories (grey dots) emanating
720 from the whole set Γ, and from boxes 1 and 2 (cyan dots, insets) and A (red dots) of Fig. 6a for γ = 0.96; (b)
721 same but from boxes 3 and 4 (cyan) and C (red) of Fig. 6b for γ = 1.1. The insets in both panels zoom in on the
722 way that the red dots associated with the chaotic orbits and the cyan dots associated with the non-chaotic orbits
723 are embedded into the global PBA (grey dots).

44
 1
2 3

1
2 3

1
2 3

724 F IG . 10. PDF of localization of the 15 000 trajectories of Figs. 6 and 9 projected onto the (Ψ1 , Ψ3 )-plane at
725 t = 330 yr. (a,c,e) γ = 0.96 and (b,d,f) γ = 1.1. The PDF is computed in panels (a,b) from all 15 000 trajectories
726 emanating from Γ; in panels (c,d) only the chaotic trajectories emanating from boxes A and C, respectively, are
727 used; and in panels (e,f) only the non-chaotic trajectories emanating from boxes 1-2 and 3-4, respectively, are
728 used (compare with Fig. 9).

45