betaFeedback

WorksPhysics and Astronomy

Classification of Chimera States via Fourier Analysis and Unsupervised Learning

Abstract

Chimera states are among the most intriguing phenomena in nonlinear dynamics, characterized by the coexistence of coherent and incoherent behavior in systems of coupled identical oscillators. Many methods have been proposed to detect chimera states and to distinguish their different types. However, such methods often suffer from important limitations that prevent sufficiently precise classification. In this work, we overcome the issue by considering a method based on Fourier analysis to determine key signal characteristics such as amplitude, phase, and frequency, jointly with an unsupervised clustering step acting on normalized total variations, measures of local spatial changes of the above-mentioned dynamical features. The proposed method allows us to identify regions in parameter space returning chimera states, but also to further distinguish between the different types. The method is applied to a network of Rayleigh oscillators, which has been shown to exhibit a rich variety of dynamical patterns.

Introduction

Research on networked nonlinear dynamical systems has revealed the emergence of complex spatiotemporal behaviors such as synchronization, consensus, and chimera states. The latter have attracted particular interest because of their intriguing and counterintuitive collective dynamics. A chimera state is characterized by the spontaneous splitting of a network of identical oscillators into coexisting coherent and incoherent domains. This phenomenon was first observed by Kaneko in the context of coupled chaotic maps [1, 2]. Later, similar behaviors were identified in various numerical studies involving global coupling schemes [3, 4, 5] and nonlocal coupling [6, 7, 8, 9, 10]. Despite these earlier observations, it was only in 2002 that the work of Kuramoto and Battogtokh [11], which is historically regarded as the first study to explicitly characterize the emergence of chimera states, was published. The popularity of this phenomenon further increased with the work of Abrams and Strogatz [12]. In an original and influential interpretation, they compared the coexistence of distinct dynamical behaviors to the chimera, a mythological creature composed of parts from different animals. Since then, chimera states have been reported in a wide variety of systems, including periodic oscillators [13], chaotic oscillators [14], time-delay systems [15, 16, 17, 18], modular networks [19], non-normal networks [20], neural systems [21], and many others, with a growing body of work also devoted to their control [22, 23, 24, 25, 26]. Laboratory experiments have also confirmed the existence of chimera states in different physical settings, by including electro-optical systems [27, 28], mechanical systems [29], electrochemical systems [30, 31], electronic circuits [32, 33], optical frequency combs [34], and chemical oscillators [35]. Depending on the initial conditions, the network topology, and the nature of the interactions within the system, different types of chimera states may emerge, including amplitude-mediated chimeras [36], amplitude chimeras [37, 38, 39, 40, 41], chimera death [37, 38], and phase chimeras [42], to name a few.

Despite the increasing body of research on this topic, determining whether a system exhibits a chimera state of a given kind remains a challenging problem. Although several tools have been introduced for this purpose, we hereby observe that many of them face important limitations when it comes to identifying the precise type of chimera state. In addition, the conclusions one can obtain by using those metrics, are sometimes strongly influenced by user-defined thresholds. It is natural to wonder whether it is possible to collect relevant information from the temporal evolution of the oscillators, particularly their amplitudes, phases, and frequencies, and then develop statistical classification algorithms capable of learning the underlying structure of the data and thus the dynamical behavior. Developing this idea, we hereby propose a clustering algorithm acting on the normalized total variation [43, 44] obtained from time series, that can identify classes in an unsupervised manner, each one associated to a clear dynamical behavior without the need of introducing any threshold. The proposed method is robust and returns reliable results allowing to identify chimera states, as well as other dynamical behaviors, as a function of the model parameters.

A similar strategy has been recently proposed to develop a data-clustering approach based on the normalized total variation and the use of an agglomerative hierarchical clustering method [45]. In the framework of topological signals, the method allowed to clearly identify three main dynamical regimes, namely ordered, chimera, and disordered states. The authors have shown that the classification depends on the depth threshold chosen in the dendrogram [45]. In the present work, the use of alternative clustering algorithms such as kk-means [46] appears promising, since it provides a direct partition once the optimal number of clusters has been determined by using suitable validation metrics. Moreover, we aim to go one step further by investigating the internal structure of the chimera cluster itself in order to identify the different types of chimera states it may contain, which constitutes our main objective.

In this work, we focus on the Rayleigh model because previous studies have shown that this system can support a wide variety of chimera patterns, including amplitude chimeras, amplitude-mediated chimeras, chimera death, and others [47, 48]. It therefore provides a natural benchmark for testing the proposed classification method. More specifically, we design a method for identifying dynamical regimes in parameter space, with particular emphasis on the coupling strength and the coupling range, two parameters that have been shown in several studies to strongly influence the emergence of chimera states. Our approach is based on unsupervised learning techniques, namely kk-means clustering and Gaussian Mixture Models (GMM). We apply it to a Rayleigh system, but let us remark that the method is general and can be applied beyond the proposed study case. The clustering phase allows us to conclude about the existence of two main parameter regions whose associated behaviors are coherent dynamics and chimera states, namely, phase chimera and amplitude-mediated chimera states. Let us stress the relevant role of Fourier theory in the proposed method, which is used to extract features from the signals, such as amplitude, phase, and frequency.

In particular, for a Rayleigh model with rotational coupling matrix, both in the linear case [47] and in the nonlinear case, we show that conventional metrics commonly used to characterize chimera states do not always allow a clear identification of the observed regimes, even when they are combined.

By contrast, the Fourier-based method together with the clustering approach makes this identification simpler and more reliable. Finally, in the Appendix, we present an additional investigation of several dynamical regimes exhibited by the Rayleigh model with nonlinear coupling and a rotational coupling matrix, which, to the best of our knowledge, has not yet been explored in detail. We show that this model can display interesting states, in particular, chimera death and coherent clusters, which are absent in the case of linear coupling with rotational coupling matrix [47].

The rest of the paper is structured as follows. Section II provides a description of the model under consideration. In Section III, we present the approach developed in this work for the identification of chimera states. Section IV is devoted to an analysis of the limitations of several existing metrics commonly used for chimera characterization. Finally, Section V concludes the paper by summarizing the main results.

The model

Let us consider NN identical Rayleigh oscillators anchored to the nodes of a (2p)(2p)-regular ring and experiencing a diffusive non-linear coupling. The time evolution of the ii-th oscillator, i=1,…,Ni=1,\ldots,N, is given by

x˙i=ωyi+ε2p∑j=i−pi+p[φ11(xjα−xiα)+φ12(yjα−yiα)],y˙i=−ωxi+Δ(1−yi2)yi+ε2p∑j=i−pi+p[φ21(xjα−xiα)+φ22(yjα−yiα)],(1)\begin{aligned} \dot{x}_i &= \omega y_i+\frac{\varepsilon}{2p}\sum_{j=i-p}^{i+p}\left[\varphi_{11}(x_j^\alpha-x_i^\alpha)+\varphi_{12}(y_j^\alpha-y_i^\alpha)\right],\\ \dot{y}_i &= -\omega x_i+\Delta(1-y_i^2)y_i+\frac{\varepsilon}{2p}\sum_{j=i-p}^{i+p}\left[\varphi_{21}(x_j^\alpha-x_i^\alpha)+\varphi_{22}(y_j^\alpha-y_i^\alpha)\right], \tag*{(1)} \end{aligned}

where xix_i and yiy_i denote the state of the ii-th oscillator, ω\omega is the natural frequency of the system, and Δ>0\Delta>0 governs the nonlinear damping. The coupling term is modeled by the sum of interactions between neighboring oscillators, i.e., pp nodes on both “sides” of node ii. Moreover, nodes indexes are considered modulo NN to encode the periodicity of the ring structure. The interaction is realized with a power α>0\alpha>0 of the state variable, while ε>0\varepsilon>0 represents the coupling strength. Finally, the coupling matrix is defined as:

Φ=(φ11φ12φ21φ22).\Phi=\begin{pmatrix}\varphi_{11}&\varphi_{12}\\\varphi_{21}&\varphi_{22}\end{pmatrix}.

For clarity, in this work the term rotational coupling matrix refers to a coupling matrix of the form

Φrot=(cos⁡ϕsin⁡ϕ−sin⁡ϕcos⁡ϕ).\Phi_{\mathrm{rot}}=\begin{pmatrix}\cos\phi&\sin\phi\\-\sin\phi&\cos\phi\end{pmatrix}.

where ϕ\phi denotes the coupling phase. Since this form imposes specific relations between the entries of the coupling matrix

Φ\Phi, namely

φ11=φ22=cos⁡ϕ,φ12=−φ21=sin⁡ϕ,\varphi_{11}=\varphi_{22}=\cos\phi,\qquad\varphi_{12}=-\varphi_{21}=\sin\phi,

the family of rotational coupling matrices Φrot\Phi_{\mathrm{rot}} represents only a particular subset of all possible coupling matrices Φ\Phi.

Our analysis will rely on a numerical study of the solutions of system (1) initialized with clustered initial conditions, (xi,yi)=(1,−1)(x_i,y_i)=(1,-1) for i∈[1,N/2]i\in[1,N/2] and (xi,yi)=(−1,1)(x_i,y_i)=(-1,1) for i∈[N/2+1,N]i\in[N/2+1,N], obtained by using a fourth-order Runge–Kutta method.

A method to detect chimera states

Characterizing chimera states requires the use of metrics that are capable of capturing their essential nature, i.e., being sensitive enough to determine the coexistence of coherent and incoherent dynamics, and at the same time, to be general enough to be applied across different types of chimera states. Indeed, many existing indicators have been designed to target specific chimera states, which limits their broader applicability, as we explain in Sec. IV. To overcome this limitation, we propose a method resulting from the use of Fourier analysis combined with a statistical clustering model. The latter allows us to avoid ad hoc thresholds when identifying chimera states, thereby providing a self-consistent definition. Note that the Fourier method has already been applied to detect chimera states in different settings [42, 43, 44]. Let us also observe that, in [45], a method grounded on Fourier analysis and statistical classification has been used to determine chimeras for systems of oscillatory topological signals.

More specifically, the Fast Fourier transform can be used to extract information from temporal signals. In particular, it enables the computation of quantities that are effectively “local” in time, such as amplitudes, frequencies, and phases. By restricting attention to the dominant amplitude within a given time window, any sufficiently regular signal y(t)y(t) can be approximated as

y(t)∼a0(w)+a(w)eι(2πΩ(w)t+θ(w))withι=−1,y(t) \sim a_{0}^{(w)} + a^{(w)}e^{\iota(2\pi\Omega^{(w)}t+\theta^{(w)})} \quad\text{with} \quad\iota=\sqrt{-1},

where a0(w)∈Ra_{0}^{(w)} \in\mathbb{R} represents the baseline level of the oscillation, a(w)∈R+a^{(w)} \in\mathbb{R}_{+} is the positive amplitude, Ω(w)∈R+\Omega^{(w)} \in\mathbb{R}_{+} is the frequency, and θ(w)∈[−π,π)\theta^{(w)} \in[-\pi,\pi) is the phase. The superscript ww indicates that these quantities are defined within the time window ww, over which the approximation is assumed to hold true. Notably, if the signal is strictly periodic, these quantities are independent of the chosen window; therefore, by examining their variation across adjacent windows, one can infer the degree of regularity of the signal.

Given a temporal signal, the accuracy of the reconstructed amplitude, frequency and phase, by means of Fourier analysis, strongly depends on the length of the signal, the longer the signal the better the accuracy. However, we hereby consider time series that are not necessarily periodic, and thus we must resort to relatively short time windows to capture local, in time, amplitude, frequency and phase. To tackle this issue we propose a modified Fourier method divided into three steps.

We start by computing the Fast Fourier Transform (FFT) of the signal y(t)−⟨y⟩y(t)-\langle y\rangle on a given time window, w=[t0,t1]w=[t_{0},t_{1}], where ⟨y⟩\langle y\rangle is the time average of y(t)y(t) in the given window. We assume the latter to be large enough to contain sufficiently many oscillations. This allows us to determine a preliminary approximation of amplitude, frequency and phase, say a~(w)\widetilde{a}^{(w)}, Ω~(w)\widetilde{\Omega}^{(w)}, θ~(w)\widetilde{\theta}^{(w)}, and the signal average a~0(w)=⟨y⟩\widetilde{a}_{0}^{(w)}=\langle y\rangle.

To increase the accuracy we look for a better approximation of Ω(w)\Omega^{(w)}. The basic observation is the amplitude of the FFT power spectrum is roughly quadratic close to its maximum. Hence, via a parabolic fit to upon the power spectrum, we obtain the sought better approximation for amplitude and frequency, namely, a^(w)\widehat{a}^{(w)} and Ω^(w)\widehat{\Omega}^{(w)}.

We can eventually overall improve the computed quantities by performing a nonlinear fit of the signal y(t)y(t) in the form

y~(t)=p1cos⁡(2πΩ^(w)t+p2)+p3,\widetilde{y}(t)=p_{1}\cos\left(2\pi\widehat{\Omega}^{(w)}t+p_{2}\right)+p_{3},

where we want to determine the unknown amplitude, p1p_{1}, phase, p2p_{2}, and baseline oscillation, p3p_{3}, by assuming Ω^(w)\widehat{\Omega}^{(w)} to be precise enough. The already computed values a^(w)\widehat{a}^{(w)}, θ~(w)\widetilde{\theta}^{(w)}, and a~0(w)\widetilde{a}_{0}^{(w)}, are used as starting point for the optimization process.

The eventually obtained quantities, a0(w)a_{0}^{(w)}, a(w)a^{(w)}, and θ(w)\theta^{(w)} depend on the used time window w=[t0,t1]w=[t_{0},t_{1}], while Ω(w)=Ω^(w)\Omega^{(w)}=\widehat{\Omega}^{(w)}. We can thus consider another time window, w′=[t0′,t1′]w'=[t'_{0},t'_{1}], and repeat the same construction to get a0(w′)a_{0}^{(w')}, a(w′)a^{(w')}, θ(w′)\theta^{(w')}, and Ω(w′)\Omega^{(w')}. By imposing the two time windows to overlap, [t0,t1]∩[t0′,t1′]≠∅[t_{0},t_{1}]\cap[t'_{0},t'_{1}]\ne\emptyset, we can ensure “some kind of continuity” of the computed values.

By eventually considering QQ time windows, wqw_{q}, q=1,…,Qq=1,\ldots,Q, the above presented procedure allows to obtain accurate estimates of baseline, amplitude, phase and frequency, a0(wq)a_{0}^{(w_{q})}, a(wq)a^{(w_{q})}, θ(wq)\theta^{(w_{q})}, and Ω(wq)\Omega^{(w_{q})}. The reconstructed phase, amplitude, and frequency are independent of the observation window only for strictly periodic signals. Otherwise, each time window yields distinct values aia_{i}, Ωi\Omega_{i}, and θi\theta_{i} for each oscillator ii. To account for this variability, we compute averages and variations over multiple windows (see [44]). Specifically, we perform the Fourier analysis on the interval [700,1000][700,1000], which is divided into QQ time windows wqw_{q}, q=1,…,Qq=1,\ldots,Q. For each oscillator i=1,…,Ni=1,\ldots,N and each time window wqw_{q}, we extract the quantity Zi(q)Z_{i}^{(q)}, with Z∈{a,Ω,θ}Z\in\{a,\Omega,\theta\}, and define

⟨Zi⟩=1Q∑q=1QZi(q),σ2(Zi)=1Q∑q=1Q(Zi(q)−⟨Zi⟩)2.(2)\langle Z_{i}\rangle=\frac{1}{Q}\sum_{q=1}^{Q}Z_{i}^{(q)}, \qquad\sigma^{2}(Z_{i})=\frac{1}{Q}\sum_{q=1}^{Q}\left(Z_{i}^{(q)}-\langle Z_{i}\rangle\right)^{2}. \tag*{(2)}

In the following figures, particularly in the figure showing the amplitude, frequency, and phase of each oscillator, this temporal variability measure with σ2(Zi)\sigma^{2}(Z_{i}) is indicated by shaded blue regions, while the average values, ⟨Zi⟩\langle Z_{i}\rangle, are denoted with blue dots.

The method is thus applied to the signal yi(t)y_{i}(t) of each oscillator i=1,…,Ni=1,\ldots,N, yielding the spatial profiles ⟨ai⟩\langle a_{i}\rangle, ⟨Ωi⟩\langle\Omega_{i}\rangle, and ⟨θi⟩\langle\theta_{i}\rangle. By studying their spatial dependence, namely how they vary with the oscillator index ii, one can distinguish different types of chimera states. In the case of a phase chimera, both ⟨ai⟩\langle a_{i}\rangle and ⟨Ωi⟩\langle\Omega_{i}\rangle remain approximately constant across the oscillators, whereas ⟨θi⟩\langle\theta_{i}\rangle exhibits the coexistence of coherent and incoherent spatial domains. Furthermore, amplitude-mediated chimera (AMC) states are characterized by the coexistence of coherent and incoherent domains in ⟨θi⟩\langle\theta_{i}\rangle, ⟨ai⟩\langle a_{i}\rangle, and ⟨Ωi⟩\langle\Omega_{i}\rangle across groups of indices ii. Finally, when ⟨θi⟩\langle\theta_{i}\rangle and ⟨Ωi⟩\langle\Omega_{i}\rangle are approximately constant across the nodes, while ⟨ai⟩\langle a_{i}\rangle exhibits the coexistence of coherent and incoherent domains, the system exhibits an amplitude chimera (AC).

To define an indicator capable of describing the emergence and the type of chimera state, we further use the concept of total variation [44]. Borrowed from mathematical analysis, it allows us to measure the smoothness of a function: small values correspond to (local) regularity, whereas once it is large, the function can exhibit “jumps” or sudden changes in values at nearby indices. More precisely, to any of the above introduced quantities we associate its total (normalized) variation:

V(⟨a⟩)=1N∑i=1N∣⟨ai+1⟩−⟨ai⟩∣,V(⟨ω⟩)=2πN∑i=1N∣⟨Ωi+1⟩−⟨Ωi⟩∣andV(⟨θ⟩)=1πN∑i=1N∥⟨θi+1⟩−⟨θi⟩∥,(3)V(\langle a\rangle) = \frac{1}{N}\sum_{i=1}^{N} | \langle a_{i+1}\rangle- \langle a_i\rangle|,\quad V(\langle\omega\rangle) = \frac{2\pi}{N}\sum_{i=1}^{N} | \langle\Omega_{i+1}\rangle- \langle\Omega_i\rangle| \quad\text{and} \quad V(\langle\theta\rangle) = \frac{1}{\pi N}\sum_{i=1}^{N} \lVert\langle\theta_{i+1}\rangle- \langle\theta_i\rangle\rVert, \tag*{(3)}

where ∥θ∥\lVert\theta\rVert is the distance on the circle, namely ∥θ∥=min⁡{θ,2π−θ}\lVert\theta\rVert= \min\{\theta, 2\pi-\theta\}, for any θ∈[0,2π)\theta\in[0,2\pi). Moreover indexes in the above sums have to be considered modulo-NN, i.e., N+1≡1N+1\equiv1. Let us observe that the normalized total variation has been already used in the framework of chimera states [43, 44, 45].

Based on the normalized total variations of amplitude, V(⟨a⟩)V(\langle a\rangle), phase, V(⟨θ⟩)V(\langle\theta\rangle), and frequency, V(⟨ω⟩)V(\langle\omega\rangle), we can classify the dynamical states of the system, as summarized in Table 1. When the normalized amplitude and frequency variations vanish, and the phase variation is either zero or small, the system is said to be in a coherent state. This category includes both fully synchronized configurations and traveling wave regimes. When the phase variation becomes large, but still smaller than 1 because of the normalization while V(⟨a⟩)≈0V(\langle a\rangle)\approx0 and V(⟨ω⟩)≈0V(\langle\omega\rangle)\approx0, the system exhibits a phase chimera. When the normalized amplitude variation is large while V(⟨θ⟩)≈0V(\langle\theta\rangle)\approx0 and V(⟨ω⟩)≈0V(\langle\omega\rangle)\approx0, the behavior is identified as amplitude chimera (AC). Now, when amplitude, phase, and frequency variations are all non-negligible, the dynamics is classified as amplitude-mediated chimera (AMC). Finally, when V(⟨θ⟩)V(\langle\theta\rangle) is too large (≤1\leq1), this behavior corresponds to incoherent state. Let us observe that the above classification is left, on purpose, vague, indeed it strongly rely on the definition of “too large”, “large” and “small”. The goal of this work is to tackle this problem and define a self-consistent method capable to identify chimera states without explicitly defining a threshold for “too large”, “large” and “small”.

StateV(⟨a⟩)V(\langle a\rangle)V(⟨θ⟩)V(\langle\theta\rangle)V(⟨ω⟩)V(\langle\omega\rangle)
Coherent state≈0\approx0small≈0\approx0
Phase chimera≈0\approx0large (≤1\le1)≈0\approx0
Amplitude chimera (AC)large≈0\approx0≈0\approx0
Amplitude-mediated chimera (AMC)largelarge (≤1\le1)large

Table 1. Classification of dynamical states based on the normalized total variations of amplitude V(⟨a⟩)V(\langle a\rangle), phase V(⟨θ⟩)V(\langle\theta\rangle), and frequency V(⟨ω⟩)V(\langle\omega\rangle).

Clustering Analysis

As mentioned in the previous section, the identification of the different dynamical regimes requires an objective criterion capable of distinguishing between “too large”, “small” and “large” values of the normalized total variations. In order to illustrate the proposed methodology, we focus here on the xx–xx coupling configuration for the considered model (1), i.e., that given by the coupling matrix

Φxx=(1010).\boldsymbol{\Phi}_{xx} = \begin{pmatrix} 1 & 0 \\ 1 & 0 \end{pmatrix}.

However, the proposed method is not restricted to this particular coupling configuration. Its applicability is more general and additional tests have also been performed for other coupling matrices, such as Φyy\boldsymbol{\Phi}_{yy}, Φyx\boldsymbol{\Phi}_{yx}, and Φxy\boldsymbol{\Phi}_{xy}.

For each parameter pair (p,ε)(p,\varepsilon), we numerically solve (1), use Fourier analysis to extract the signal features for each node and eventually compute the three indicators V(⟨θ⟩)V(\langle\theta\rangle), V(⟨a⟩)V(\langle a\rangle), and V(⟨ω⟩)V(\langle\omega\rangle).

Those quantities depend on the chosen parameters and, by varying the latter in a suitable domain, we can thus obtain a 3D3D dataset containing the information about the system behavior. Before applying the clustering procedure, each feature is rescaled by min–max normalization [49], that is, by shifting its minimum value to 0 and its maximum value to 1, while all intermediate values are proportionally mapped between these two bounds. This normalization step is essential, since it ensures that the three observables contribute on the same footing to the classification and prevents the clustering from being biased by differences in numerical scale. Let us observe that the classification task can be realized by using different tools, we hereby show and compare the results obtained by using kk-means, while in Appendix A we will present an alternative approach based on the use of Gaussian Mixture Model (GMM).

k-means. The kk-means algorithm [46] partitions a set of BB points {x1,…,xB}⊂Rn\{\mathbf{x}_1,\ldots,\mathbf{x}_B\}\subset\mathbb{R}^n into kk clusters {C1,…,Ck}\{C_1,\ldots,C_k\} by minimizing the total intra-cluster sum of squared distances, commonly referred to as the inertia:

J=∑c=1k∑x∈Cc∥x−μc∥2,(4)\mathcal{J} = \sum_{c=1}^{k}\sum_{\mathbf{x}\in C_c}\lVert\mathbf{x}-\boldsymbol{\mu}_c\rVert^2, \tag*{(4)}

where μc=1∣Cc∣∑x∈Ccx\boldsymbol{\mu}_c=\frac{1}{|C_c|}\sum_{\mathbf{x}\in C_c}\mathbf{x} denotes the centroid of cluster CcC_c. The algorithm proceeds iteratively through two successive steps. In the assignment step, each point is assigned to the nearest centroid:

Cc={xb:∥xb−μc∥2≤∥xb−μℓ∥2, ∀ ℓ≠c},(5)C_c=\{\mathbf{x}_b:\lVert\mathbf{x}_b-\boldsymbol{\mu}_c\rVert^2\leq\lVert\mathbf{x}_b-\boldsymbol{\mu}_{\ell}\rVert^2,\ \forall\ \ell\ne c\}, \tag*{(5)}

while in the update step, the centroids are recomputed from the newly assigned clusters. These two steps are repeated until convergence, that is, until the cluster assignments no longer change.

For this algorithm, as well as for the GMM presented in Appendix A, an essential step is to determine the optimal number of clusters. Several metrics can be used for this purpose; here, we consider two of them: the silhouette score [50] and the Davies–Bouldin index [51]. This choice is motivated by the need to identify a number of clusters that are both compact and well separated from each other.

For each point xb\mathbf{x}_b, let ded_e denote the mean intra-cluster distance, i.e., the average distance to all other points in the same cluster, and let dfd_f denote the mean nearest-cluster distance, i.e., the average distance to the points belonging to the closest neighboring cluster. The silhouette value of xb\mathbf{x}_b is defined as

s(xb)=df(xb)−de(xb)max⁡{df(xb),de(xb)}.(6)s(\mathbf{x}_b)=\frac{d_f(\mathbf{x}_b)-d_e(\mathbf{x}_b)}{\max\{d_f(\mathbf{x}_b),d_e(\mathbf{x}_b)\}}. \tag*{(6)}

The overall silhouette score is then obtained by averaging over all points:

S=1B∑b=1Bs(xb),(7)S=\frac{1}{B}\sum_{b=1}^{B}s(\mathbf{x}_b), \tag*{(7)}

and takes values in the interval [−1,1][-1, 1]. Values close to 1 indicate compact and well-separated clusters, whereas values close to −1-1 correspond to poorly separated or strongly overlapping clusters. The Davies–Bouldin index quantifies the clustering quality by jointly evaluating intra-cluster compactness and inter-cluster separation. For each cluster CcC_c, let σc\sigma_c denote the average distance between the points of CcC_c and their centroid μc\mu_c, and let DcrD_{cr} be the distance between the centroids of clusters CcC_c and CrC_r. The index is defined as

DB=1k∑c=1kmax⁡r≠c(σc+σrDcr),(8)\mathrm{DB} = \frac{1}{k}\sum_{c=1}^{k}\max_{r\ne c}\left(\frac{\sigma_c+\sigma_r}{D_{cr}}\right), \tag*{(8)}

where kk is the total number of clusters. Lower values of DB indicate better clustering quality, corresponding to more compact and better separated clusters. Although the index is unbounded from above, larger values indicate poorer clustering quality, reflecting increased intra-cluster dispersion and stronger inter-cluster overlap.

To quantitatively compare the performance of the kk-mean clustering algorithm, we evaluate the silhouette score SS and the Davies–Bouldin index DB by testing values ranging from k=2k=2 to k=20k=20 and evaluating two complementary validation metrics. Both methods identify k=2k=2 as the optimal partition. kk-means provides better clustering quality with respect to GMM, indeed the former exhibits a higher silhouette score (S=0.925S=0.925) and a lower Davies–Bouldin index (DB = 0.421), as one can appreciate by looking at Figure 1(a) and (b) and compare with Figure A1(a) and (b), returning (S=0.853S=0.853, DB = 0.548) for GMM. These results indicate that kk-means yields clusters that are more compact and better separated than those obtained with GMM. A possible explanation for this superiority could probably be that the data consist of roughly spherical, similarly sized, and well-separated clusters. Under these conditions, kk-means’ simple hard assignment produces more compact and clearly delineated clusters, whereas, in contrast, the added flexibility of GMM’s soft probabilistic assignment introduces unnecessary complexity that could slightly degrade the separation quality [52].

Figure 1. Results of the clustering methods kk-means for the xx–xx coupling. Panel (a) shows the silhouette score as a function of the number of clusters kk, while panel (b) displays the Davies–Bouldin index. Panel (c) presents the corresponding three-dimensional cluster distributions in the feature space (V(⟨θ⟩)V(\langle\theta\rangle), V(⟨a⟩)V(\langle a\rangle), V(⟨ω⟩)V(\langle\omega\rangle)). Panel (d) shows the associated cluster assignments in the (pp, ε\varepsilon) parameter plane. The red circles in panels (a) and (b), indicate the optimal number of clusters selected by the validation metrics.

Results of the clustering methods GMM for the $x$–$x$ coupling

Figure A1. Results of the clustering methods GMM for the xx–xx coupling. Panel (a) shows the silhouette score as a function of the number of clusters kk, while panel (b) displays the Davies–Bouldin index. Panel (c) presents the corresponding three-dimensional cluster distributions in the feature space spanned by V(⟨θ⟩)V(\langle\theta\rangle), V(⟨a⟩)V(\langle a\rangle), and V(⟨ω⟩)V(\langle\omega\rangle). Panel (d) shows the associated cluster assignments in the (p,ε)(p,\varepsilon) parameter plane. The red circles indicate the optimal number of clusters selected by the validation metrics.

Furthermore, we display a three-dimensional representation of kk-means in Figure 1(c) (see Figure A1(c) for the analogous plot for GMM), aimed at visualize the spatial extent of the clusters; one can clearly appreciate the presence of two clusters, the blue and the cyan one. The corresponding partition in the parameter plane (p,ε)(p,\varepsilon), shown in Figure 1(d), exhibits two contiguous and clearly distinguishable regions. Let us observe that the same overall structure is recovered by GMM (see Figure A1(d)).

To analyze the two clusters and emphasize their differences, we represent the data by using the boxplots shown in Figure 2.

Figure 2. Analysis of the normalized total variations for the two clusters shown in Figure 1. Panels display the boxplot distributions respectively for : V(⟨θ⟩)V(\langle\theta\rangle) panel (a), V(⟨a⟩)V(\langle a\rangle) panel (b) and V(⟨ω⟩)V(\langle\omega\rangle) panel (c). The median (shown in red) provides the most robust comparison between clusters by indicating the typical level of variation in each observable. Cluster 1 exhibits medians close to zero for all three quantities, which is characteristic of a coherent regime, whereas Cluster 2 displays markedly higher medians, revealing a chimera regime with stronger phase, amplitude, and frequency variations.

Let us observe that the latter, not only illustrate the spread of the data, but they also provide a direct statistical comparison of the typical values of the variations in each cluster. Therefore, when the median of one cluster is systematically lower than the one of another, we can conclude that the corresponding observable is globally small for the majority of parameter pairs in that cluster. By closely observing Figure 2, we can see that the medians of V(⟨θ⟩)V(\langle\theta\rangle), V(⟨a⟩)V(\langle a\rangle), and V(⟨ω⟩)V(\langle\omega\rangle) for Cluster 1 are all located very close to zero, and the associated interquartile ranges remain narrow. This demonstrates that most points belonging to Cluster 1 are characterized by very small variations in phase, amplitude, and frequency. By contrast, Cluster 2 exhibits significantly higher medians for the three observables compared to Cluster 1, together with a broader dispersion. The upward shift of the medians is particularly important here, because it shows that the larger variations are not caused by a few isolated outliers, but instead represent the typical behavior of the cluster. Consequently, parameter values (p,ε)(p,\varepsilon) associated to points belonging to Cluster 1, can be identified as coherent states, whereas those referring to Cluster 2 to chimera behavior.

Having established that the first clustering stage separates coherent states from chimera states, we next refine the analysis by focusing exclusively on data points belonging to the chimera cluster. We apply again kk-means to this subset in order to determine whether several types of chimera states coexist inside this cluster. The results of the finer clustering are reported in Figure 3. The validation metrics (data not shown) again indicate that the optimal number of sub-clusters is k=2k=2, implying that the chimera region itself can be decomposed into two distinct subclasses. Figure 3(a) shows the corresponding separation in the 3D3D features space, while Figure 3(b) displays the induced subdivision in the parameter plane.

Figure 3. Results of the refined clustering methods kk-means for the xx–xx coupling. Panel (a) shows a three-dimensional representation of the two subclusters in the feature space, previously being identified with a single cluster 2. Panel (b) displays the corresponding subdivision in cluster into the (p,ε)(p,\varepsilon) parameter plane.

The use of boxplots allows again to differentiate the dynamics associated to those two new classes. Indeed, in Figure 4(a) we report the distribution of V(⟨θ⟩)V(\langle\theta\rangle), as already observed Cluster 1 has a very low median, confirming the weak phase variation of the coherent regime. On the other hand the refined classification allowed to split the second group into two new ones, Clusters 2 and 3. They both display large phase variations, but Cluster 2 shows larger values with respect to Cluster 3. In Figure 4(b), associated with V(⟨a⟩)V(\langle a\rangle), Cluster 2 is clearly distinguished by a much higher median than the other two clusters, indicating that strong amplitude variation is a typical and persistent property of this group. Finally, Figure 4(c), corresponding to V(⟨ω⟩)V(\langle\omega\rangle), shows that Cluster 2 also possesses the highest median frequency variation, whereas Cluster 3 remains at a lower value. This analysis leads to a clear dynamical interpretation of the three clusters. Cluster 1 corresponds to the coherent state, since the three observables remain globally close to zero. Cluster 2 is characterized by simultaneously large phase, amplitude, and frequency variations, and is therefore identified as an amplitude-mediated chimera (AMC). Cluster 3, on the other hand, exhibits a pronounced phase variation while keeping comparatively weak amplitude and frequency variations. This signature is consistent with a phase chimera. Note that further validation of this conclusion can be obtained by observing the typical time series associated to each cluster.

Figure 4. Analysis of the normalized total variations for the three clusters obtained after the second clustering step. Panel (a) shows the case V(⟨θ⟩)V(\langle\theta\rangle), panel (b) V(⟨a⟩)V(\langle a\rangle), and panel (c) V(⟨ω⟩)V(\langle\omega\rangle). The relative positions of the medians allow a direct identification of the dynamical regimes. Cluster 1 is characterized by low medians for all three observables and corresponds to the coherent state. Cluster 2 displays the highest medians in phase, amplitude, and frequency variations and is therefore associated with an amplitude-mediated chimera (AMC). Cluster 3 is marked by a strong phase variation together with comparatively weak amplitude and frequency variations, which identifies it as a phase chimera.

Overall, the proposed clustering procedure based on features extracted by using Fourier analysis, provides a fully data-driven and threshold-free classification of the observed dynamical states. The first stage separates coherent dynamics from chimera dynamics, while the second stage resolves the chimera region into two distinct subclasses, namely amplitude-mediated chimera and phase chimera. In Figure 5 we report the total normalized variations for amplitude, phase and frequency, as a function of the coupling strength, ε\varepsilon, for some chosen values of the coupling range, pp. To better appreciate the results obtained with the clustering method, we add the information about the classes (background colors in the panels).

Figure 5. Evolution of the normalized total variations as functions of the coupling strength ε\varepsilon for several representative values of the interaction range pp. The blue, red, and green curves correspond to V(⟨θ⟩)V(\langle\theta\rangle), V(⟨a⟩)V(\langle a\rangle), and V(⟨ω⟩)V(\langle\omega\rangle), respectively. The background color indicates the dynamical regime assigned by the clustering procedure: coherent state (blue), amplitude-mediated chimera (orange), and phase chimera (brown). This combined representation allows one to track the dynamical transitions as ε\varepsilon increases.

In this way we can observe the transitions from one dynamical behavior to another one, once we vary ε\varepsilon. For instance for p=18p = 18 we observe a transition from amplitude-mediated chimera to coherent state for ε∼0.25\varepsilon\sim0.25, and indeed numerical simulations for ε=0.15\varepsilon= 0.15 (see Figure 6(aFigure 1-dFigure 1) ε=0.65\varepsilon= 0.65 (see Figure 6(aFigure 2-dFigure 2), allow to visually confirm this claim.

Figure 6. Time series and Fourier features for the case p=18p = 18. The top row (ε=0.15\varepsilon= 0.15) corresponds to an amplitude-mediated chimera, shown through the spatiotemporal diagram (a1), the amplitude profile (b1), the frequency profile (c1), and the phase profile (d1). The bottom row (ε=0.65\varepsilon= 0.65) corresponds to a coherent state, as evidenced by the corresponding snapshots in (a2)–(d2). This figure confirms the transition identified in the clustered parameter plane, from amplitude-mediated chimera to coherence. The remaining parameters are ω=2\omega= 2, N=200N = 200, Δ=1\Delta= 1, and α=1\alpha= 1.

LIMITS OF SOME CURRENT METRICS FOR ASSESSING CHIMERA STATES

Several metrics have been proposed in the literature to characterize chimera states. Among them, the incoherence strength (SI) introduced by Gopal et al. [15] is one of the most widely used. Although SI provides a simple scalar classification: coherent (SI = 0), chimera (0 < SI < 1), and incoherent (SI = 1), it has several limitations. First, the measure depends on an arbitrary threshold δ\delta and on the choice of the grouping parameter MM, both of which can affect the classification result for the same trajectory: if the incoherent domain is small compared to the group size n=N/Mn = N/M, the incoherent nodes are absorbed into mostly coherent groups and the chimera behavior goes undetected, while a group size that is too small may lead to misclassifying phase-shifted coherent nodes as incoherent. In addition, δ\delta is usually chosen as a percentage of the signal range yi,max⁡−yi,min⁡y_{i,\max}-y_{i,\min}, which changes with the model parameters. More fundamentally, SI reduces all the spatiotemporal dynamics to a single number, losing all information about the spatial profiles of frequency, amplitude, and phase across the network; those are precisely the quantities that distinguish, for example, amplitude chimeras from phase chimeras. A similar criticism applies to the metrics introduced by Provata [53], Δr=rmax⁡−rmin⁡\Delta r=r_{\max}-r_{\min} and Δω=ωmax⁡−ωmin⁡\Delta\omega=\omega_{\max}-\omega_{\min}, where rmax⁡r_{\max} and rmin⁡r_{\min} denote the maximum and minimum asymptotic amplitudes over all nodes after the transient, and ωmax⁡\omega_{\max}, ωmin⁡\omega_{\min} the corresponding extreme mean phase velocities. Although convenient, these purely global scalars reduce all the spatial structure to two numbers; in particular, a single deviating oscillator in an otherwise perfectly synchronized network is enough to produce nonzero values of Δr\Delta r or Δω\Delta\omega, which can lead to a wrong

chimera classification, since the main feature of these states is the coexistence of spatially extended coherent and incoherent domains, rather than the existence of a single isolated deviating oscillator or a configuration that is, at best, a weak chimera. To overcome these limitations, we use a method based on the Fourier transform that extracts, at each node ii, the local amplitude ⟨ai⟩\langle a_i\rangle, frequency ⟨Ωi⟩\langle\Omega_i\rangle, and phase ⟨θi⟩\langle\theta_i\rangle, as well as their normalized total variation as a measure of regularity. This method works for periodic, quasi-periodic, and weakly aperiodic signals, and provides spatially resolved profiles that allow a clear identification of the chimera type: for example, when a single outlier node contributes a non-negligible deviation, the total variation remains small but nonzero, and combined with the fact that the difference between the maximum and minimum of the observed variable is also nonzero, this allows us to correctly identify the state as a weak chimera, while spatially extended incoherent domains produce significantly nonzero values of V(⟨a⟩)V(\langle a\rangle), V(⟨ω⟩)V(\langle\omega\rangle), and V(⟨θ⟩)V(\langle\theta\rangle), thus providing a clear, parameter-free criterion for chimera detection. Another advantage of this method is illustrated in Figure 7, which shows the coexistence of coherence and incoherence in the phases, amplitudes, and frequencies at the same time. By focusing only on the centers of mass, as proposed in [38], one would classify this state as an amplitude chimera, since some oscillators have limit cycles centered at the origin while others do not; the same conclusion is obtained by using the mean number of inhomogeneous oscillators [41], even when varying the threshold Λ\Lambda in [0.2, 1], and also by combining the mean center of mass with the spatial correlation g0g_0 [47]. Each one of these classifications is however incorrect: by applying the Fourier method, we can clearly show the coexistence of coherent and incoherent domains in amplitude, phase, and frequency at the same time, a behavior that is characteristic of a amplitude-mediated chimera rather than a pure amplitude chimera, and a distinction that none of the scalar metrics above is able to capture.

Figure 7. Example of amplitude-mediated chimera, ε=0.834\varepsilon= 0.834, p=5p = 5 and the use of a rotational matrix. Panel (a) shows the space–time plots, panels (b), (c) and (d), display respectively amplitude, frequency and phase, computed with the Fourier method. The center-of-mass is reported in panel (e). From the data reported in those panels we can conclude that the system exhibits an amplitude-mediated chimera. The remaining parameters are ω=2\omega= 2, N=200N = 200, Δ=1\Delta= 1, ϕ=π/2−0.1\phi= \pi/2 - 0.1, and α=3\alpha= 3.

We also consider the case of a rotational matrix coupling previously considered in [47]. The results are reported in Figure 8. In the first row, the dynamical state can clearly be identified as an amplitude-mediated chimera, since one can observe the coexistence of coherent and incoherent domains in the phase, the amplitude, and frequency variation profiles. This state was also characterized by the authors of [47] as an amplitude-mediated chimera. In contrast, the second row requires a more refined interpretation. At first glance, the unequal amplitudes across the oscillators may suggest an amplitude chimera. However, a closer inspection of the phase profile reveals a clear coexistence of coherent and incoherent domains, while the frequency profile remains almost identical for all oscillators. For this reason, although the authors of [47] classified this regime as a amplitude chimera, we believe that this characterization is not accurate enough. Indeed, an amplitude chimera should be associated with the coexistence of coherence and incoherence in the amplitude profile itself. Since the present state exhibits coexistence of coherent and incoherent domains in the phase and the amplitude, we classify it more appropriately as a phase-amplitude chimera.

Figure 8. Example of amplitude-mediated chimera, ε=0.8\varepsilon= 0.8 (top row), phase-amplitude chimera ε=2.0\varepsilon= 2.0 (bottom row), p=5p = 5 and the use of a rotational matrix. The first column presents the space–time plots, followed by the amplitude (second column), frequency (third column), phase (fourth column), and center-of-mass (last column) profiles. From the analysis of the top row panels one can conclude that the system exhibits an amplitude-mediated chimera state, while bottom panels report the case of phase-amplitude chimera. The remaining parameters are ω=2\omega= 2, N=200N = 200, Δ=1\Delta= 1, ϕ=π/2−0.1\phi= \pi/2 - 0.1, and α=1\alpha= 1.

Conclusion

In this work, we proposed a method for classifying chimera states based on Fourier analysis and unsupervised learning. From the time series generated by the dynamics of a network of Rayleigh oscillators, we extracted local signal features, namely amplitude, phase, and frequency, by using an adapted Fourier-based method. From these quantities, we then computed their normalized total variations in order to quantify the spatial organization of the system and characterize the different dynamical regimes observed in a more coherent way.

By representing these features in a three-dimensional space and then by applying unsupervised clustering methods, we obtained an objective classification of the system states without relying on arbitrary thresholds. The proposed approach first allows to clearly distinguish coherent states from chimera states, and then, in a second step, to refine this classification by identifying two main subclasses of chimeras, namely amplitude-mediated chimeras and phase chimeras.

Finally, this study highlights the relevance of a fully data-driven approach for identification and classification, which makes it possible to better capture the diversity of dynamical behaviors present in the network under consideration. Hence, the method allows to overcome the issue related to the choice of the threshold to identify chimera states. In its present form, the method is not capable to determine the number and the size of regular regions; however, it can be improved to solve this task by identifying a reliable feature of the signals, so to enlarge the dimension of the data-set.

Let us observe that the method is also versatile: indeed, we can use several methods to extract the dynamical features beyond Fourier analysis, as well as other classification algo-

rithms. It therefore provides a robust and promising framework for the analysis of chimera states in coupled oscillator systems, and could be extended to other network topologies, including higher-order ones, as well as to more complex dynamics.

Acknowledgments

The authors acknowledge Patrick Louodop for preliminary discussions about the model. R.M. acknowledges JSPS KAK-ENHI 24KF0211 for financial support.

Author contributions

R.T.D.: conceptualization, software, investigation, visualization, formal analysis, validation, writing – original draft, writing – review and editing. R.M.: conceptualization, methodology, supervision, writing – review and editing. T.N.: writing – review and editing. T.C.: conceptualization, methodology, visualization, formal analysis, supervision, writing – original draft, writing – review and editing. All authors read and approved the manuscript.

Gaussian Mixture Model

Gaussian Mixture Model (GMM). The Gaussian Mixture Model can be viewed as a probabilistic extension of kk-means. It assumes that the data are generated by a mixture of kk Gaussian distributions, so that the marginal probability density takes the form

p(x)=∑c=1kπcN(x∣μc,Σc),(9)p(\mathbf{x}) = \sum_{c=1}^{k} \pi_c \mathcal{N}(\mathbf{x} \mid\boldsymbol{\mu}_c, \boldsymbol{\Sigma}_c), \tag*{(9)}

where πc≥0\pi_c \ge0 are the mixing weights satisfying ∑c=1kπc=1\sum_{c=1}^{k}\pi_c=1, and N(x∣μc,Σc)\mathcal{N}(\mathbf{x}\mid\boldsymbol{\mu}_c,\boldsymbol{\Sigma}_c) denotes the Gaussian density with mean μc\boldsymbol{\mu}_c and covariance matrix Σc\boldsymbol{\Sigma}_c:

N(x∣μc,Σc)=1(2π)d/2∣Σc∣1/2exp⁡[−12(x−μc)TΣc−1(x−μc)].(10)\mathcal{N}(\mathbf{x}\mid\boldsymbol{\mu}_c,\boldsymbol{\Sigma}_c)=\frac{1}{(2\pi)^{d/2}|\boldsymbol{\Sigma}_c|^{1/2}}\exp\left[-\frac{1}{2}(\mathbf{x}-\boldsymbol{\mu}_c)^{\mathsf{T}}\boldsymbol{\Sigma}_c^{-1}(\mathbf{x}-\boldsymbol{\mu}_c)\right]. \tag*{(10)}

The model parameters Θ={πc,μc,Σc}c=1k\boldsymbol{\Theta}=\{\pi_c,\boldsymbol{\mu}_c,\boldsymbol{\Sigma}_c\}_{c=1}^{k} are estimated by maximizing the log-likelihood

ln⁡p(X∣Θ)=∑b=1Bln⁡(∑c=1kπcN(xb∣μc,Σc)),(11)\ln p(\mathbf{X}\mid\boldsymbol{\Theta})=\sum_{b=1}^{B}\ln\left(\sum_{c=1}^{k}\pi_c\mathcal{N}(\mathbf{x}_b\mid\boldsymbol{\mu}_c,\boldsymbol{\Sigma}_c)\right), \tag*{(11)}

using the Expectation–Maximization (EM) algorithm [54].

In the E-step, the posterior responsibilities are computed as

γbc=πcN(xb∣μc,Σc)∑r=1kπrN(xb∣μr,Σr),(12)\gamma_{bc}=\frac{\pi_c\mathcal{N}(\mathbf{x}_b\mid\boldsymbol{\mu}_c,\boldsymbol{\Sigma}_c)}{\sum_{r=1}^{k}\pi_r\mathcal{N}(\mathbf{x}_b\mid\boldsymbol{\mu}_r,\boldsymbol{\Sigma}_r)}, \tag*{(12)}

where γbc\gamma_{bc} is the posterior probability that the data point xb\mathbf{x}_b belongs to the Gaussian component cc.

In the M-step, defining the effective membership of component cc as

Γc=∑b=1Bγbc,(13)\Gamma_c=\sum_{b=1}^{B}\gamma_{bc}, \tag*{(13)}

the parameters are updated according to

πc=ΓcB,μc=1Γc∑b=1Bγbcxb,Σc=1Γc∑b=1Bγbc(xb−μc)(xb−μc)T.(14)\pi_c=\frac{\Gamma_c}{B},\qquad\boldsymbol{\mu}_c=\frac{1}{\Gamma_c}\sum_{b=1}^{B}\gamma_{bc}\mathbf{x}_b,\qquad\boldsymbol{\Sigma}_c=\frac{1}{\Gamma_c}\sum_{b=1}^{B}\gamma_{bc}(\mathbf{x}_b-\boldsymbol{\mu}_c)(\mathbf{x}_b-\boldsymbol{\mu}_c)^{\mathsf{T}}. \tag*{(14)}

The E-step and M-step are repeated until convergence, typically until the increase in log-likelihood becomes smaller than a prescribed tolerance.

Unlike kk-means, the GMM provides a soft probabilistic assignment of points to clusters and, when full covariance matrices are used, can model ellipsoidal cluster geometries through the covariance matrices Σc\boldsymbol{\Sigma}_c.

Additional applications of the proposed metrics

Let us observe that the variations defined in (3) can be employed to estimate the transient life associated with the emergence or disappearance of chimera states. Assume that the numerical simulations are performed over the time interval [t0,tmax⁡][t_0,t_{\max}]. After the transient time of the network dynamics, we subdivide the remaining time interval into nn consecutive time windows and apply the Fourier reconstruction procedure described above within each window. From this reconstruction we obtain the local quantities ⟨ai⟩\langle a_i\rangle, ⟨Ωi⟩\langle\Omega_i\rangle, and ⟨θi⟩\langle\theta_i\rangle, from which we compute the corresponding normalized total variations V(⟨a⟩)V(\langle a\rangle), V(⟨Ω⟩)V(\langle\Omega\rangle), and V(⟨θ⟩)V(\langle\theta\rangle). By observing the evolution of these quantities across successive windows, one can determine whether the system remains in a chimera regime or evolves toward a coherent state. In particular, if at a certain instant all normalized total variations converge toward zero, this indicates that the spatial profiles of amplitude, frequency, and phase have become uniform across the network, implying that the system has transitioned from a chimera configuration to a coherent state. The corresponding instant therefore provides an estimate of the transient time ttrt_{\mathrm{tr}} of the chimera.

Different dynamics for the nonlinear coupling with the rotational coupling matrix

Motivated by the work of [47], in which the authors showed the coexistence of several types of chimera states in the case of linear coupling, we wondered the impact nonlinear coupling, as opposed to linear coupling, would have on the overall network dynamics, and more specifically on pattern formation. To answer to this question we performed two numerical experiments; first we fixed the coupling range while varying the coupling strength, and then fixed the coupling strength while varying the coupling range. Figure A2 presents a series of spatio-temporal diagrams obtained for a network of nonlinearly coupled oscillators with α=3\alpha= 3 by employing the rotational matrix. Depending on the values of the coupling parameters pp and for ϵ=3\epsilon= 3, the system exhibits a variety of dynamical regimes. For p=3p = 3, the network tends to settle into coherent oscillation death states, as shown in panel (a1). By slightly increasing pp, multicluster structures emerge, for instance the 10-cluster state in panel (b1), or more complex chimera states such as the nine-chimera state in panel (c1). For intermediate parameter values, the system develops different types of chimera death states, ranging from the weak 7-cluster chimera death (a2) to the weak 3-cluster chimera death (b2), and finally to the weak 2-cluster chimera death (c2). At larger pp, the system converges to highly organized chimera death states, as shown in panel (a3), in particular the 1-cluster chimera death, and also to coherent structures illustrated in panel (b3), corresponding to the one-cluster oscillation death. These observations highlight the intricate transition between multi-cluster oscillation death formation and chimera dynamics induced by nonlinear coupling when only the parameter pp is varied. We also remark that, in the case of coherent cluster or chimera death, an increase of pp tends to reduce the number of clusters. For instance, in the case of chimera death, the system evolves from the weak 7-cluster chimera death to the 1-cluster chimera death. This observation has already been reported in previous works, such as [38] where an increase in the coupling range reduced the number of clusters.

Figure A2. Time series for nonlinear coupling with α=3\alpha= 3, obtained using the rotational coupling matrix. The coupling strength is fixed at ϵ=3\epsilon= 3. The panels correspond to the following regimes: (a1) p=3p = 3, coherent oscillation death; (b1) p=5p = 5, multicluster state, specifically a 10-cluster state; (c1) p=6p = 6, weak 9-chimera state; (a2) p=8p = 8, weak 7-cluster chimera death; (b2) p=15p = 15, weak 3-cluster chimera death; (c2) p=20p = 20, weak 2-cluster chimera death; (a3) p=54p = 54, one-cluster chimera death; and (b3) p=56p = 56, single coherent-cluster oscillation death. The remaining parameters are ω=2\omega= 2, N=200N = 200, Δ=1\Delta= 1, and ϕ=π/2−0.1\phi= \pi/2 - 0.1.

Those results demonstrate that by considering nonlinear couplings leads to the emergence of interesting dynamics, which were not observed in the case of linear coupling (see [47]).

Let us now consider the inverse scenario, where the coupling range is fixed while the coupling strength is varied. Figure A3 and A4 present the different dynamical behaviors observed when the parameter p=9p = 9 is held constant. In Figure A3, the first column displays the spatiotemporal diagrams, the second the amplitudes, the third the frequencies, the fourth the phases, and the fifth the center of mass, while Figure A4 shows the spatio-temporal diagrams. For ϵ=0.2\epsilon= 0.2, we observe an amplitude-mediated chimera (first row) with V(⟨a⟩)≈0.013V(\langle a\rangle) \approx0.013, V(⟨ω⟩)=0.0069V(\langle\omega\rangle) = 0.0069, and V(⟨θ⟩)=0.165V(\langle\theta\rangle) = 0.165; as ϵ\epsilon increases, the system transitions through various states, such as a traveling wave (second row) with V(⟨a⟩)≈7.352×10−4V(\langle a\rangle) \approx7.352 \times10^{-4}, V(⟨ω⟩)=2.92×10−4V(\langle\omega\rangle) = 2.92 \times10^{-4}, and V(⟨θ⟩)=0.0078V(\langle\theta\rangle) = 0.0078, before reaching the amplitude-mediated chimera state with V(⟨a⟩)≈0.0492V(\langle a\rangle) \approx0.0492, V(⟨ω⟩)=0.1798V(\langle\omega\rangle) = 0.1798, and V(⟨θ⟩)=0.1042V(\langle\theta\rangle) = 0.1042. A progressive increase in the coupling strength leads the system to evolve from oscillatory states to oscillation-death states, as illustrated in Figure A4. Specifically, the dynamics transition from incoherent oscillation death (A4 (a1)) to the weak 7-cluster chimera death (A4 (b1)), then to the weak 5-cluster chimera death (A4 (c1)), followed by coherent states such as the 4-cluster oscillation death (A4 (d1)), and finally the pattern that we name traveling oscillation death, characterized by oscillation suppression combined with a traveling pattern. This analysis demonstrates the strong impact of coupling strength variation and highlights the richness of dynamical behaviors emerging in the system. Note the presence of oscillation-death states that were not present in the linear case, this thus highlights the richness and diversity that nonlinear coupling can provide.

Figure A3. Time series and Fourier-derived features for p=9p = 9 in the nonlinear coupling case with α=3\alpha= 3, obtained using the rotational coupling matrix. The first column shows the spatiotemporal diagrams, followed by the amplitude profiles (second column), frequency profiles (third column), phase profiles (fourth column), and center-of-mass profiles (last column). In all cases, the coupling range is fixed at p=9p = 9, while the coupling strength takes the values ϵ=0.2\epsilon= 0.2 in panels (a1)–(e1), corresponding to an amplitude-mediated chimera; ϵ=0.48\epsilon= 0.48 in panels (a2)–(e2), corresponding to a traveling wave; and ϵ=0.65\epsilon= 0.65 in panels (a3)–(e3), corresponding to an amplitude-mediated chimera. The remaining parameters are ω=2\omega= 2, N=200N = 200, Δ=1\Delta= 1, and ϕ=π/2−0.1\phi= \pi/2 - 0.1.

Figure A4. Time series for nonlinear coupling with α=3\alpha= 3, obtained using the rotational coupling matrix. The coupling range is fixed at p=9p = 9, while the coupling strength ϵ\epsilon is varied. The panels show the following regimes: (a1) ϵ=1.2\epsilon= 1.2, incoherent oscillation death; (b1) ϵ=3\epsilon= 3, weak 7-cluster chimera death; (c1) ϵ=4\epsilon= 4, weak 5-cluster chimera death; (d1) ϵ=5\epsilon= 5, weak 4-cluster oscillation death; and (e1) ϵ=7\epsilon= 7, traveling oscillation death. The remaining parameters are ω=2\omega= 2, N=200N = 200, Δ=1\Delta= 1, and ϕ=π/2−0.1\phi= \pi/2 - 0.1.

References

  1. [1]K. Kaneko. Period-doubling of kink-antikink patterns, quasiperiodicity in antifero-like structures and spatial intermittency in coupled logistic lattice: towards a prelude of a “field theory of chaos”. Progress of Theoretical Physics, 72(3):480–486, 1984.DOI
  2. [2]Kunihiko Kaneko. Clustering, coding, switching, hierarchical ordering, and control in a network of chaotic elements. Physica D: Nonlinear Phenomena, 41(2):137–172, 1990.DOI
  3. [3]Vincent Hakim and Wouter-Jan Rappel. Dynamics of the globally coupled complex ginzburg-landau equation. Physical Review A, 46(12):R7347, 1992.pubmed.ncbi.nlm.nih.gov/9908163
  4. [4]Naoko Nakagawa and Yoshiki Kuramoto. Collective chaos in a population of globally coupled oscillators. Progress of Theoretical Physics, 89(2):313–323, 1993.DOI
  5. [5]Marie-Line Chabanol, Vincent Hakim, and Wouter-Jan Rappel. Collective chaos and noise in the globally coupled complex ginzburg-landau equation. Physica D: Nonlinear Phenomena, 103(1-4):273–293, 1997.DOI
  6. [6]Yoshiki Kuramoto. Scaling behavior of turbulent oscillators with non-local interaction. Progress of Theoretical Physics, 94(3):321–330, 1995.DOI
  7. [7]Yoshiki Kuramoto and Hiroya Nakao. Origin of power-law spatial correlations in distributed oscillators and maps with nonlocal coupling. Physical review letters, 76(23):4352, 1996.pubmed.ncbi.nlm.nih.gov/10061268
  8. [8]Yoshiki Kuramoto and Hiroya Nakao. Power-law spatial correlations and the onset of individual motions in self-oscillatory media with non-local coupling. Physica D: Nonlinear Phenomena, 103(1-4):294–313, 1997.
  9. [9]Yoshiki Kuramoto, Dorjsuren Battogtokh, and Hiroya Nakao. Multiaffine chemical turbulence. Physical review letters, 81(16):3543, 1998.DOI
  10. [10]Yoshiki Kuramoto, Hiroya Nakao, and Dorjsuren Battogtokh. Multi-scaled turbulence in large populations of oscillators in a diffusive medium. Physica A: Statistical Mechanics and its Applications, 288(1-4):244–264, 2000.
  11. [11]Yoshiki Kuramoto and Dorjsuren Battogtokh. Coexistence of coherence and incoherence in nonlocally coupled phase oscillators. arXiv preprint cond-mat/0210694, 2002.DOI
  12. [12]Daniel M Abrams and Steven H Strogatz. Chimera states for coupled oscillators. Physical review letters, 93(17):174102, 2004.DOI
  13. [13]Stefan Ulonska, Iryna Omelchenko, Anna Zakharova, and Eckehard Schöll. Chimera states in networks of van der pol oscillators with hierarchical connectivities. Chaos: An Interdisciplinary Journal of Nonlinear Science, 26(9):094825, 2016.pubmed.ncbi.nlm.nih.gov/27781460
  14. [14]Sergey A Bogomolov, Andrei V Slepnev, Galina I Strelkova, Eckehard Schöll, and Vadim S Anishchenko. Mechanisms of appearance of amplitude and phase chimera states in ensembles of nonlocally coupled chaotic systems. Communications in Nonlinear Science and Numerical Simulation, 43:25–36, 2017.DOI
  15. [15]R Gopal, VK Chandrasekar, A Venkatesan, and M Lakshmanan. Observation and characterization of chimera states in coupled dynamical systems with nonlocal coupling. Physical review E, 89(5):052914, 2014.DOI
  16. [16]Iryna Omelchenko, Yuri Maistrenko, Philipp Hövel, and Eckehard Schöll. Loss of coherence in dynamical networks: spatial chaos and chimera states. Physical review letters, 106(23):234102, 2011.DOI
  17. [17]Tatiana E Vadivasova, Galina I Strelkova, Sergey A Bogomolov, and Vadim S Anishchenko. Correlation analysis of the coherence-incoherence transition in a ring of nonlocally coupled logistic maps. Chaos: an interdisciplinary journal of nonlinear science, 26(9), 2016.pubmed.ncbi.nlm.nih.gov/27781464
  18. [18]N Semenova, A Zakharova, E Schöll, and V Anishchenko. Does hyperbolicity impede emergence of chimera states in networks of nonlocally coupled chaotic oscillators? Europhysics Letters, 112(4):40002, 2015.
  19. [19]M. Asllani, B.A. Siebert, A. Arenas, and J.P. Gleeson. Symmetry-breaking mechanism for the formation of cluster chimera patterns. Chaos, 32:013107, 2022.DOI
  20. [20]R. Muolo, J.D.O’Brien, T. Carletti, and M. Asllani. Persistence of chimera states and the challenge for synchronization in real-world networks. The European Physical Journal B, 97(1):6, 2024.DOI
  21. [21]Gaël R Simo, Thierry Njougouo, RP Aristides, Patrick Luodop, Robert Tchitnga, and Hilda A Cerdeira. Chimera states in a neuronal network under the action of an electric field. Physical Review E, 103(6):062304, DOI: 10.1103/PhysRevE.103.062304, 2021.DOI
  22. [22]C. Bick and E.A. Martens. Controlling chimeras. New Journal of Physics, 17(3):033030, 2015.DOI
  23. [23]Thomas Isele, Johanne Hizanidis, Astero Provata, and Philipp Hövel. Controlling chimera states: The influence of excitable units. Physical Review E, 93(2):022217, 2016.DOI
  24. [24]L.V. Gambuzza and M. Frasca. Pinning control of chimera states. Physical Review E, 94(2):022306, 2016.DOI
  25. [25]Giulia Ruzzene, Iryna Omelchenko, Eckehard Schöll, Anna Zakharova, and Ralph G Andrzejak. Controlling chimera states via minimal coupling modification. Chaos: an interdisciplinary journal of nonlinear science, 29(5), 2019.DOI
  26. [26]Riccardo Muolo, Lucia Valentina Gambuzza, Hiroya Nakao, and Mattia Frasca. Pinning control of chimera states in systems with higher-order interactions. Nonlinear Dynamics, pages 1–23, 2025.DOI
  27. [27]Mark R Tinsley, Simbarashe Nkomo, and Kenneth Showalter. Chimera and phase-cluster states in populations of coupled chemical oscillators. Nature Physics, 8(9):662–665, 2012.DOI
  28. [28]Aaron M Hagstrom, Thomas E Murphy, Rajarshi Roy, Philipp Hövel, Iryna Omelchenko, and Eckehard Schöll. Experimental observation of chimeras in coupled-map lattices. Nature Physics, 8(9):658–661, 2012.
  29. [29]Erik Andreas Martens, Shashi Thutupalli, Antoine Fourriere, and Oskar Hallatschek. Chimera states in mechanical oscillator networks. Proceedings of the National Academy of Sciences, 110(26):10563–10567, 2013.DOI
  30. [30]Mahesh Wickramasinghe and István Z Kiss. Spatially organized dynamical states in chemical oscillator networks: Synchronization, dynamical differentiation, and chimera patterns. PloS one, 8(11):e80586, 2013.
  31. [31]Mahesh Wickramasinghe and István Z Kiss. Spatially organized partial synchronization through the chimera mechanism in a network of electrochemical reactions. Physical Chemistry Chemical Physics, 16(34):18360–18369, 2014.
  32. [32]Lucia Valentina Gambuzza, Arturo Buscarino, Sergio Chessari, Luigi Fortuna, Riccardo Meucci, and Mattia Frasca. Experimental investigation of chimera states with quiescent and synchronous domains in coupled electronic oscillators. Physical Review E, 90(3):032905, 2014.pubmed.ncbi.nlm.nih.gov/25314502
  33. [33]David P Rosin, Damien Rontani, Nicholas D Haynes, Eckehard Schöll, and Daniel J Gauthier. Transient scaling and resurgence of chimera states in networks of boolean phase oscillators. Physical Review E, 90(3):030902, 2014.
  34. [34]Evgeny A Viktorov, Tatiana Habruseva, Stephen P Hegarty, Guillaume Huyet, and Bryan Kelleher. Coherence and incoherence in an optical comb. Physical review letters, 112(22):224101, 2014.pubmed.ncbi.nlm.nih.gov/24949771
  35. [35]Simbarashe Nkomo, Mark R Tinsley, and Kenneth Showalter. Chimera states in populations of nonlocally coupled chemical oscillators. Physical review letters, 110(24):244102, 2013.pubmed.ncbi.nlm.nih.gov/25165927
  36. [36]Gautam C Sethia, Abhijit Sen, and George L Johnston. Amplitude-mediated chimera states. Physical Review E, 88(4):042917, 2013.DOI
  37. [37]Umesh Kumar Verma and G Ambika. Amplitude chimera and chimera death induced by external agents in two-layer networks. Chaos: An Interdisciplinary Journal of Nonlinear Science, 30(4):043104, 2020.DOI
  38. [38]Anna Zakharova, Marie Kapeller, and Eckehard Schöll. Chimera death: Symmetry breaking in dynamical networks. Physical review letters, 112(15):154101, 2014.DOI
  39. [39]Liudmila Tumash, Anna Zakharova, Judith Lehnert, Wolfram Just, and Eckehard Schöll. Stability of amplitude chimeras in oscillator networks. Europhysics Letters, 117(2):20001, 2017.DOI
  40. [40]K Premalatha, VK Chandrasekar, M Senthilvelan, and M Lakshmanan. Stable amplitude chimera states in a network of locally coupled stuart-landau oscillators. Chaos: An Interdisciplinary Journal of Nonlinear Science, 28(3), 2018.DOI
  41. [41]K Sathiyadevi, VK Chandrasekar, and DV Senthilkumar. Stable amplitude chimera in a network of coupled stuart-landau oscillators. Physical Review E, 98(3):032301, 2018.DOI
  42. [42]E.R. Zajdela and D.M. Abrams. Phase chimera states: frozen patterns of disorder. Chaos: An Interdisciplinary Journal of Nonlinear Science, 35(8), 2025.
  43. [43]Riccardo Muolo, Thierry Nougouo, Lucia Valentina Gambuzza, Timoteo Carletti, and Mattia Frasca. Phase chimera states on nonlocal hyperrings. Physical Review E, 109(2):L022201, 2024.DOI
  44. [44]Rommel Tchinda Djeudjo, Timoteo Carletti, Hiroya Nakao, and Riccardo Muolo. Chimera states on m-directed hypergraphs. to appear in Physical Review E, 2026.
  45. [45]S Nirmala Jenifer, Riccardo Muolo, Paulsamy Muruganandam, and Timoteo Carletti. A robust method for classification of chimera states. arXiv preprint arXiv:2603.22026, 2026.
  46. [46]James B McQueen. Some methods of classification and analysis of multivariate observations. In Proc. of 5th Berkeley Symposium on Math. Stat. and Prob., pages 281–297, 1967.
  47. [47]Tanmoy Banerjee, Debabrata Biswas, Debarati Ghosh, Eckehard Schöll, and Anna Zakharova. Networks of coupled oscillators: From phase to amplitude chimeras. Chaos: An Interdisciplinary Journal of Nonlinear Science, 28(11):113124, 2018.DOI
  48. [48]Zhongkui Sun, Qifan Xue, and Nannan Zhao. Chimera states in fractional-order coupled rayleigh oscillators. Communications in Nonlinear Science and Numerical Simulation, 135:108083, 2024.DOI
  49. [49]Peshawa J Muhammad Ali. Investigating the impact of min-max data normalization on the regression performance of k-nearest neighbor with different similarity measurements. ARO-The Scientific Journal of Koya University, 10(1):85–91, 2022.DOI
  50. [50]Peter J Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20:53–65, 1987.DOI
  51. [51]David L Davies and Donald W Bouldin. A cluster separation measure. IEEE transactions on pattern analysis and machine intelligence, (2):224–227, 1979.pubmed.ncbi.nlm.nih.gov/21868852
  52. [52]Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning, volume 4. Springer, 2006.
  53. [53]Astero Provata. Amplitude chimeras and bump states with and without frequency entanglement: a toy model. Journal of Physics: Complexity, 5(2):025011, 2024.arxiv.org/abs/2310.19471
  54. [54]Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological), 39(1):1–22, 1977.DOI

Paper details

vlnc:W_5mjfe4mg
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents