 Research
 Open Access
 Published:
Synchronization and resilience in the Kuramoto white matter network model with adaptive statedependent delays
The Journal of Mathematical Neuroscience volume 10, Article number: 16 (2020)
Abstract
White matter pathways form a complex network of myelinated axons that regulate signal transmission in the nervous system and play a key role in behaviour and cognition. Recent evidence reveals that white matter networks are adaptive and that myelin remodels itself in an activitydependent way, during both developmental stages and later on through behaviour and learning. As a result, axonal conduction delays continuously adjust in order to regulate the timing of neural signals propagating between different brain areas. This delay plasticity mechanism has yet to be integrated in computational neural models, where conduction delays are oftentimes constant or simply ignored. As a first approach to adaptive white matter remodeling, we modified the canonical Kuramoto model by enabling all connections with adaptive, phasedependent delays. We analyzed the equilibria and stability of this system, and applied our results to twooscillator and largedimensional networks. Our joint mathematical and numerical analysis demonstrates that plastic delays act as a stabilizing mechanism promoting the network’s ability to maintain synchronous activity. Our work also shows that global synchronization is more resilient to perturbations and injury towards network architecture. Our results provide key insights about the analysis and potential significance of activitydependent myelination in largescale brain synchrony.
Introduction
Synchronization, the mechanism by which oscillatory processes collectively organize to align their phase, has been the focus of intense research across the field of nonlinear dynamics [1], especially in the biological sciences due to its involvement in myriads of physiological processes. In the brain, such synchronized oscillatory patterns, observed in the rhythmic discharge of neuronal electrical impulses, play a central role in neural communication, information processing, and the implementation of higher cognitive function [2]. However, the mechanisms by which these oscillations emerge and interact within and across brain microcircuits and systems remain poorly understood.
In the mammalian brain, local synchronized oscillatory activity are coordinated through a vast network of myelinated axonal fibers called white matter. The intricate white matter structure is formed under a population of glial cells called oligodendrocytes. The oligodendrocytes produce an insulating myelin sheath coiling around axonal membranes to greatly increase the conduction velocities of propagating signals between neurons. As white matter develops to adopt a genetically programmed structure, oligodendrocytes determine to which extent specific axons are myelinated. The resulting myelin structure of the network is responsible for the emergence and evolution of a rich repertoire of spatiotemporal activity patterns, notably oscillations with various spectral properties [3].
While white matter is highly relevant in shaping brain network dynamics, the mechanisms governing myelin development and its relationship with neural activity remain mostly unknown [4]. Nevertheless, the general hypothesis surrounding this topic has been shifting away from traditional viewpoints. Recent studies suggest that white matter structure undergoes continuous changes past the stages of developmental myelination and well into adulthood, rather than remaining static as initially presumed [5]. Indeed, substantial myelin formation continues to occur within the fully mature central nervous system in an activitydependent manner [6, 7]. Furthermore, newly discovered evidence implies that white matter structure is responsive to experiences such as learning and social interactions [8]. These findings unveil potential ways white matter can restructure itself to facilitate neurological function over time. In particular, white matter characterized as a plastic, adaptive structure, can be critical in maintaining the essential oscillatory and synchronous processes found in many neural systems [9].
Despite the complexity of the neurophysiological processes involved, it has been hypothesized that myelin remodeling collectively reinforces synchronous dynamics and oscillatory coordination in largescale brain networks [8]. From this perspective, the temporal structure of these plastic changes can provide a higher level of corrective adjustments, improving the network’s ability to converge towards phase synchrony. As a first approach to this intricate problem, we wish to use a simplified model to determine whether activitydependent delays impact the phase coordination of coupled oscillators when compared to static delays. We have used the Kuramoto model, a phenomenological model which has a long history of applications in neuroscience, notably to study the collective organization of oscillatory neural systems [10]. The Kuramoto model, with or without time delay, has been extensively studied [1, 11, 12], and is increasingly used to model local oscillatory neural activity in brainscale simulations informed by anatomical white matter connectivity [13–15]. Appealing for its relative simplicity, we have enabled the Kuramoto model with phasedependent delays to examine the collective impact of adaptive conduction on coupled phase oscillators. Specifically, we performed the stability analysis of this modified system and examined its response to structural perturbations. While preliminary, these results can provide new insights into the largescale impact of white matter remodeling and its potential role in neural synchrony.
This paper is structured as follows. In Sect. 2, we first introduce our network model composed as a system of coupled phaseoscillators with statedependent delays. In Sect. 3, we derive the equations for an Noscillator network’s synchronous state and its respective stability criteria. Section 4 applies the derived equations from Sect. 3 to a reduced twooscillator network. Section 5 applies the derived equations from Sect. 3 to a largedimensional oscillator network through taking an Nlimit approximation by handling the asymptotic phases statistically. In Sect. 6, we conduct numerical experiments and provide evidence supporting the analysis in Sects. 4 and 5 through simulated results. Section 7 explores the stabilization property in the context of spontaneous changes in network connectivity and compares the resilience of the synchronized state with and without plastic delays.
Model
We consider a prototypical model involving a network of N weaklycoupled nonlinear and delayed Kuramoto oscillators [9] whose phases \(\theta _{i}(t)\) evolve according to the following system of differential equations:
where \(\omega _{i} \in \mathbb{R}\) is the natural frequency of oscillator i, and \(g > 0\) is a constant global coupling gain. The coefficients \(a_{ij}\) represent synaptic weights: \(a_{ij} = 1\) if there is a white matter connection between oscillator j to i; otherwise, \(a_{ij} = 0\). The axonal conduction delays \(\tau _{ij}\) correspond to the conduction time taken by an action potential along a myelinated fiber. The propagation speed \(c_{ij} = c_{ij}(t, \ell )\) fluctuates with respect to a propagating signal’s position ℓ along the length of an axon as saltatory conduction occurs at successive nodes of Ranvier. The conduction velocity \(c_{ij}(t, \ell )\) also changes in time. As such, each conduction delay may be written as
That is, a line integral along the axonal pathway \(P_{ij}\) connecting oscillator j to i. It is uncertain how neural activity and/or oscillatory brain dynamics influences myelin formation. However, it is known that myelination correlates with information transmission within and across brain areas [5], which are known to involve long range synchronization [16]. To represent this in our model, we echo previous work on oscillatory neural communication [8, 9, 17] and make axonal conduction delays phasedependent. Specifically, we assume that each conduction delay \(\tau _{ij}(t)\) evolves under the following sublinear plasticity equation:
where \(1 \leq i\), \(j \leq N\). Here, \(\alpha _{\tau }\) is the homeostatic rate constant that sets the time scale of the evolving delays. The plasticity coefficient \(\kappa > 0\) sets the gain of the conduction delay changes. The initial condition for delays \(\tau _{ij}(t)\) at \(t = 0\) is provided by baseline lags \(\tau _{ij}^{0} \geq 0\). That is, \(\tau _{ij}(0) = \tau _{ij}^{0}\) for all i, j. The Heaviside function \(H(\tau )\) is defined as a smooth function such that \(H(\tau ) = 0\) if \(\tau < 0\) and \(H(\tau ) = 1\) if \(\tau > \varepsilon \) for some small \(\varepsilon > 0\). The Heaviside function \(H(\tau )\) is present to preserve nonnegativity of the delays \(\tau _{ij}(t)\) at all times \(t \geq 0\). For details on the construction of the smooth Heaviside function, we refer the reader to Appendix A. According to this rule, fluctuations in conduction delays are governed by an interplay between a local drift component that represents metabolic inertia towards an initial baseline lag \(\tau _{ij}^{0}\) representing minimal myelination, and a forcing term that depends on the phase difference \(\Delta _{ij} := \theta _{j}  \theta _{i}\) between oscillators i and j.
A schematic of the network structure as well as the activitydependent plasticity rule are plotted in Fig. 1. There, given \(\theta _{j}  \theta _{i} < \pi \), one can see that whenever \(\theta _{j} > \theta _{i}\), the time delay increases due to an effective reduction in conduction velocity caused by myelin retraction. The opposite occurs if \(\theta _{j} < \theta _{i}\) and the time delay decreases: the connection speeds up due to myelin stabilization [8]. If the phase is the same, no changes in conduction velocity are required, and the delay remains stable at its initial lag \(\tau _{ij}^{0}\).
Synchronized state and stability with plastic delays
We are interested in characterizing the influence of the delay plasticity rule Eq. (3) towards the stability of global phaselocked solutions and, more generally, in stabilizing synchrony between neural populations (i.e. oscillators). Enabled with adaptive delays, our model corresponds to an \(N + N^{2}\)dimensional functional differential equation with statedependent delays. The analysis of such systems is technically challenging, especially in terms of stability where a modified approach must be used [18–20]. Mathematically, we analyze the network’s ability to asymptotically achieve the following synchronized state:
for all \(i \leq N\) as \(t \rightarrow \infty \), where Ω is the global fixed frequency of the oscillators and \(\phi _{i}\) is the asymptotic phaseoffset of oscillator i as time \(t \rightarrow \infty \). As we will see, adaptive delays following the plasticity rule equation (3) require nonzero phase differences \(\Delta _{ij}(t) := \theta _{j}(t)  \theta _{i}(t)\) in order to maintain its equilibrium values, from which the network can no longer become inphase. Nevertheless, the oscillators are able to become entrained to a common frequency Ω. Hence, we say that our network becomes synchronized when the nodes \(\theta _{i}(t)\) globally oscillate at some stable frequency Ω and become phaselocked under individual offsets \(\phi _{i}\). As we are primarily concerned with the effects of delays changes under the plastic Eq. (3), we assume that all oscillators share the same natural frequency \(\omega _{i} = \omega _{0}\) for all i in Eq. (1). To simplify the convergent behaviour of delays \(\tau _{ij}(t)\) following Eq. (3), we set \(\alpha _{\tau }= 1\).
Applying the ansatz Eq. (4) onto both the phases in Eq. (1) and the delays in Eq. (3), we obtain the following expressions for the global frequency \(\varOmega \in \mathbb {R}\) and individual offsets \(\boldsymbol {\phi }= (\phi _{1}, \ldots , \phi _{N}) \in \mathbb {R}^{N}\):
for all i, j, where \(\Delta _{ij} = \phi _{j}  \phi _{i}\). It must be true that if Ω satisfies Eq. (5), then \(\varOmega \in [\omega _{0}  g, \omega _{0} + g]\). To assess the network’s stability at the synchronized state \((\varOmega , \boldsymbol {\phi })\) satisfying Eqs. (5) and (6), we introduce the perturbation terms \(\epsilon _{i}(t)\), \(\eta _{ij}(t)\) around the equilibrium states Eq. (4) and Eq. (6) for \(\theta _{i}(t)\) and \(\tau _{ij}(t)\), respectively. That is, we write
and examine the stability at the equilibrium point \(\epsilon _{i}(t) = \eta _{ij}(t) = 0\). The delay perturbations \(\eta _{ij}(t)\) abide by the linearized form of Eq. (3) around all positive delays \(\tau _{ij}^{E} = \tau _{ij}^{0} + \kappa \sin (\Delta _{ij}) > 0\). For the equilibrium delays \(\tau _{ij}^{E} = 0\), we proceed by assuming the corresponding perturbation terms will act in an asymptotically stable manner. Specifically, for all phaseoffset differences such that \(\tau _{ij}^{E} = 0\), by the nature of the Heaviside cutoff function \(H(\tau )\) in Eq. (3) there exists some \(t_{\mathrm{asy}} \in \mathbb {R}\) such that for all \(t > t_{\mathrm{asy}}\), \(\tau _{ij}(t) = 0\) and consequently \(\eta _{ij}(t) = 0\). That is, if \(\tau _{ij}^{E} = 0\), the perturbation term is asymptotically \(\eta _{ij}(t) = 0\). This condition will be satisfied given the synchronous state \((\varOmega , \boldsymbol {\phi })\) is indeed stable. Since the purpose of linearization is to determine stability, the above assumption is valid to make before proceeding. Taken together, the linearized equations for the terms \(\eta _{ij}(t)\) as \(\tau _{ij}(t) \rightarrow \tau _{ij}^{E}\) are given by
where \(C_{ij}^{0} := \cos (\Delta _{ij})\). We can use the linearization approach in the context of equations with statedependent delays [18–20] as follows. Inserting Eq. (7) and Eq. (8) into the Kuramoto equation (1),
where we have made the approximation \(\epsilon _{j}(t  \tau _{ij}(t)) \approx \epsilon _{j}(t  \tau _{ij}^{E})\) when \(\tau _{ij}(t)\) is near \(\tau _{ij}^{E}\). With \((\varOmega , \boldsymbol {\phi })\) satisfying Eq. (5), and by taking a firstorder expansion around the term \(\varOmega \tau _{ij}^{E}\), we obtain the linear system for the phase perturbation term given by
where we denote \(C_{ij} = \cos (\varOmega \tau _{ij}^{E} + \Delta _{ij})\) and \(H_{ij} = H(\tau _{ij}^{E})\). Hence, the network synchronizes at \((\varOmega , \boldsymbol {\phi })\) given that all perturbation terms \(\epsilon _{i}(t)\), \(\eta _{ij}(t)\) following the linear system of fixed delay differential Eqs. (9) and (11) converges to 0. From here, we can analyze the stability at the synchronous point \((\varOmega , \boldsymbol {\phi })\) by setting the ansatz \(\epsilon _{i}(t) = v_{i}e^{\lambda t}\), \(\eta _{ij}(t) = w_{ij} e^{ \lambda t}\) with respect to eigenvalue \(\lambda \in \mathbb{C}\). By Eq. (9), the coefficients \(w_{ij}\) satisfy
for all i, j, \(\lambda \neq 1\). Applying Eq. (12) to the coefficients \(w_{ij}\),
That is, λ is an eigenvalue if there exists an eigenvector \(\vec{v} = (v_{1}, \ldots , v_{N}) \in \mathbb {C}^{N}\) satisfying Eq. (13). In matrix form, Eq. (13) can be expressed as \(M_{\lambda }\vec{v} = \vec{0}\), where \(M_{\lambda }= (M_{ij})\) is the Ndimensional square matrix with entries
In summary, we acquire the following stability criterion: The system is stable around the synchronized state \((\varOmega , \boldsymbol {\phi })\) if for all eigenvalues \(\lambda \in \mathbb{C}\) satisfying \(\det M_{\lambda }= 0\), \(\operatorname {Re}(\lambda ) < 0\). Appendix A discusses the existence and uniqueness of solutions \(\theta _{i}(t)\), \(\tau _{ij}(t)\), as well as the justification of the above linearization and respective stability analysis.
Of note is that our approach is an extension of the nonplastic case. Indeed, without plasticity \(\kappa = 0\), the delays remain fixed at the baseline lag \(\tau _{ij}^{0}\), and as a result there is no need for the delays to establish an equilibrium with positive phaseoffset differences \(\Delta _{ij}\). Hence, the oscillators become perfectly inphase with \(\phi _{i} = 0\) during synchronization. Consequently, Eq. (5) for the global frequency Ω reduces to the onedimensional fixedpoint expression
By employing similar steps as above, the phase perturbation terms \(\epsilon _{i}(t)\) follow the linear system
resulting in the corresponding eigenvalue matrix \(M_{\lambda }\) defined with entries
With nonplastic delays, stability analysis at \(\theta _{i}(t) = \varOmega t\) has been accomplished through other means. Lyapunov functionals [21] have shown that a sufficient criterion for synchronization around \(\theta _{i}(t) = \varOmega t\) is \(\cos (\varOmega \tau _{ij}^{0}) > 0\) for all active connections i, j such that \(a_{ij} =1\), and Ω satisfies Eq. (15). This criterion becomes necessary given a unique baseline lag \(\tau _{ij}^{0} = \tau ^{0}\) [22]. That is, it can be shown that the oscillators synchronize if and only if \(\cos (\varOmega \tau ^{0}) > 0\).
Synchronization of two coupled oscillators with plastic delays
As a first illustrative approach to the general case when the number of coupled oscillators is large, let us first consider a reduced twooscillator network. The following setup is similar to the system analyzed in [23]. We first let \(N = 2\) with \(a_{12} = a_{21} = 1\) and remove all selfintegration terms by setting \(a_{11} = a_{22} = 0\). We also set the baseline delays to be equal with \(\tau _{12}^{0} = \tau _{21}^{0} = \tau ^{0}\). Then the phases \(\theta _{1}(t)\), \(\theta _{2}(t)\) follow the Kuramoto system,
with respective plasticity equations
By symmetry we may let \(\Delta _{12} = \phi _{2}  \phi _{1} > 0\). We proceed by setting a sufficiently large plasticity gain with \(\kappa \gg \tau ^{0}\). Then it follows that the respective equilibrium delays are \(\tau _{12}^{E} := \tau ^{0} + \kappa \sin (\Delta _{12})\) and \(\tau _{21}^{E} = 0\), resulting in a single positive equilibrium delay. Hence, by Eq. (5) the frequency Ω and offset \(\Delta _{12}\) satisfies
Solving for \(\Delta _{12}\) above, we obtain
where positivity holds only when \(\varOmega < \omega _{0}\). This leads to the oscillators synchronizing at a lower frequency \(\varOmega \in [\omega _{0}  g, \omega _{0})\). Substituting \(\Delta _{12}\) in Eq. (20) with Eq. (22), if we define the root function
then \(\theta _{1}(t)\), \(\theta _{2}(t)\) synchronizes at a common frequency Ω implicitly satisfying \(R_{\kappa }(\varOmega ) = 0\). Since it was assumed that \(\tau _{21}^{E} = 0\), for selfconsistency the synchronous frequency Ω must also satisfy \(\tau ^{0}  \kappa \sin (\Delta _{12}) < 0\), which can be written as \(\varOmega < \omega _{0}  g\kappa ^{1}\tau ^{0} \approx \omega _{0}\) since we set \(\kappa \gg \tau ^{0}\). Without plasticity \(\kappa = 0\), the oscillators become inphase with \(\Delta _{12} = 0\) and synchronizes at a frequency Ω that is a root of the function
as stated by Eq. (15). Figure 2(A) plots functions \(\mathcal{R}_{\kappa }(\varOmega )\) on the interval \([\omega _{0}g, \omega _{0})\) with respect to representative values \(\kappa = 0, 20, 30\) of the plasticity gain. We observe that higher plasticity gains \(\kappa > 0\) generally leads to a greater number of potential synchronization frequencies Ω for our system. In Fig. 2(A), one can see that the functions \(\mathcal{R}_{0}(\varOmega )\) and \(\mathcal{R}_{20}(\varOmega )\) have single roots within the interval \([\omega _{0}g, \omega _{0})\), while \(\mathcal{R}_{30}(\varOmega )\) has five.
The stability in the twodimensional case can be easily determined. Indeed, as derived in Sect. 3, the stability of the oscillators at the synchronization state \((\varOmega , \Delta _{12})\) is determined by the distribution of eigenvalues \(\lambda \in \mathbb {C}\) that satisfy
where \(C_{12} = g \cos (\varOmega \tau _{12}^{E} + \Delta _{12})\), \(C_{21} = g \cos (\Delta _{12})\), \(C_{0} = \cos (\Delta _{12})\). This results in the root equation \(P_{\varOmega }(\lambda ) + Q_{\varOmega }(\lambda )e^{\lambda \tau } = 0\), where \(\tau = \tau _{12}^{E}\) is the single positive delay and \(P_{\varOmega }(\lambda )\), \(Q_{\varOmega }(\lambda )\) are polynomials given by
where \(\tilde{\kappa } = 1  \varOmega \kappa C_{0}\). Note that \(P_{\varOmega }(0) + Q_{\varOmega }(0) = 0\), which means we have neutral stability. If we make the approximation \(\tau \approx 0\), the eigenvalues λ correspond to exact cubic roots of \(P_{\varOmega }(\lambda ) + Q_{\varOmega }(\lambda )\). It is possible for the stability of the system under \(\tau > 0\) to align with the stability under \(\tau = 0\), particularly for small τ. If we ignore the neutral stability of our zero transcendental equation, Theorem 1 of [24] states that under certain conditions, the stability of the system at \(\tau = 0\) and at any \(\tau > 0\) does not change. For rigorous purposes however, we are still interested in finding the distribution of eigenvalues \(\lambda \in \mathbb {C}\) solving Eq. (25), particularly if it is accessible through numerical approximation.
Setting the plasticity gain \(\kappa = 30\), Fig. 2(B) shows the real parts of the two nonzero branches \(\lambda _{1}\), \(\lambda _{2}\) of the roots \(P_{\varOmega }(\lambda ) + Q_{\varOmega }(\lambda ) = 0\), plotted with respect to the global frequency Ω. If any of the branches are positive, it implies by the above discussion that the oscillators will not synchronize at the state \((\varOmega , \Delta _{12})\), where \(\Delta _{12} = \Delta _{12}(\varOmega )\) is given by Eq. (22). We denote \(\varOmega _{1} < \varOmega _{2} < \varOmega _{3}\) to be the three largest synchronization frequencies corresponding to equilibria solutions \(\varOmega _{i} < \omega _{0}\) of the root equation \(\mathcal{R}_{\kappa }(\varOmega _{i}) = 0\) as presented in Fig. 2(A). The eigenvalue branches imply that frequencies \(\varOmega _{1}\), \(\varOmega _{3}\) are stable with \(\operatorname {Re}(\lambda _{1}), \operatorname {Re}(\lambda _{2}) < 0\) at \(\varOmega = \varOmega _{1}, \varOmega _{3}\), while \(\varOmega _{2}\) is an unstable frequency with \(\operatorname {Re}(\lambda _{1}) > 0\) at \(\varOmega = \varOmega _{2}\). Figure 2(C), (D) shows heatmap approximations of complex roots λ of Eq. (24) at \(\varOmega = \varOmega _{1}\), \(\varOmega = \varOmega _{2}\), respectively. We can see that at \(\varOmega = \varOmega _{1}\), both nonzero cubic roots are located on the leftside of the imaginary axis, and the error heatmap shows that the eigenvalues are located on the lefthand side of the imaginary axis. At \(\varOmega = \varOmega _{2}\), all cubic roots are real with a single positive root \(\lambda _{1} > 0\). Consistent with the stability of our system at \(\tau = 0\), the error heatmap implies there exists an eigenvalue \(\lambda \in \mathbb {C}\) satisfying Eq. (25) near the positive root \(\lambda _{1}\). This highlights that under sufficiently large plasticity gain, adaptive delays introduce multiple stable states \((\varOmega , \Delta _{12})\) in a twooscillator system, whose stability can be assessed under the zero delay approximation \(\tau _{12}^{E} \approx 0\). Our stability analysis reveals that the oscillators can synchronize at multiple possible frequencies, which suggests a greater degree of adaptability in our system.
Synchronization in largescale oscillator networks with plastic delays
Using inspiration by the twodimensional case in Sect. 4, we consider here a largedimensional system using Nlimit approximations. Again for simplicity, we set the baseline lags to be constant with \(\tau _{ij}^{0} \equiv \tau ^{0}\) and the connection topology to be alltoall with \(a_{ij} \equiv 1\). We approach the synchronization state \((\varOmega , \boldsymbol {\phi })\) of the network with adaptive delays, given by Eqs. (5) and (6), in the following statistical sense. Suppose that the phaseoffsets \(\phi _{i}\) are i.i.d. under some distribution. We can set the offsets to be centered at 0 by defining the recentered offsets \(\Delta _{i} := \phi _{i}  \overline{\phi }\), where ϕ̅ is the mean of ϕ. Then \(\Delta _{ij} = \phi _{j}  \phi _{i} = \Delta _{j}  \Delta _{i}\). We take \(\Delta _{i}\) to be i.i.d. under some density function \(\rho (\Delta )\). Setting a random \(\Delta _{i}\) in the global frequency equation (5) for each \(i \leq N\), and then taking the limit \(N \rightarrow \infty \), we obtain the following Nlimit approximation for frequency Ω and density \(\rho (\Delta )\):
for all fixed \(\Delta ' \in \mathbb {R}\), where \(\tau ^{E}(\Delta ) = H(\tau ^{0} + \kappa \sin \Delta ) \cdot (\tau ^{0} + \kappa \sin \Delta )\). In order to apply Eq. (28) to obtain the global frequencies Ω, we parametrize the unknown density \(\rho (\Delta )\) by assuming it is Gaussian under some small phaseoffset variance \(\delta ^{2}\). That is, we have
and we set \(\rho (\Delta ) = \rho _{\delta }(\Delta )\). If the variance \(\delta ^{2}\) is small, we can approximate the fixed offset in Eq. (28) as \(\Delta ' \approx 0\), as well as make the approximation \(\tau ^{E}(\Delta ) \approx H(\tau ^{0} + \kappa \Delta ) \cdot (\tau ^{0} + \kappa \Delta )\) since Δ is small. Hence, our largescale network synchronizes near a global frequency Ω with Gaussian distributed phaseoffsets with variance \(\delta ^{2}\) given that \((\varOmega , \delta ^{2})\) is a root of the function
Without plasticity, we recall that we have inphase synchronization at global frequency Ω which is a solution to \(\mathcal{R}_{0}(\varOmega , 0) = 0\). That is,
Figure 3(A) plots the curve \(\mathcal{R}(\varOmega , \delta ^{2})\) for various fixedvalues of \(\delta > 0\) on \(\varOmega \in [\omega _{0}  g, \omega _{0} + g]\). We find that there is a unique but different root Ω to \(\mathcal{R}(\varOmega , \delta ^{2}) = 0\) at each fixed variance \(\delta ^{2} > 0\). Hence, we can obtain an implicit curve \(\varOmega = \varOmega (\delta )\) by parametrizing the level curve \(\mathcal{R}(\varOmega , \delta ^{2}) = 0\) with respect to \(\delta > 0\). The curve is plotted in Fig. 3(B), and shows a continuous range of potential synchronization states \((\varOmega (\delta ), \delta ^{2})\) along \(\delta > 0\) to a large Ndimensional system of oscillators. The graph of the level curve \(\mathcal{R}(\varOmega , \delta ^{2}) = 0\) also shows that \(\varOmega (\delta ) < \omega _{0}\) for all \(\delta > 0\), implying that the oscillators are drawn to synchronize at a lower frequency from their natural frequency.
For small offset differences, the equilibrium delays are approximately \(\tau _{ij}^{E} \approx \tau ^{0} + \kappa \Delta _{ij}\) if \(\Delta _{ij} > 0\), and \(\tau _{ij}^{E} = 0\) if \(\Delta _{ij} < 0\). Before proceeding, we set κ to be sufficiently larger than the baseline lag \(\tau ^{0}\) such that \(\frac{\tau ^{0}}{\kappa } \approx 0\). Then most negative offset differences \(\Delta _{ij} < 0\) fall below the Heaviside cutoff \(\tau ^{0} + \kappa \Delta _{ij} < 0\). From this, the Heaviside term becomes approximately dependent to the sign of \(\Delta _{ij}\) with \(H_{ij} \approx H(\Delta _{ij})\) for all i, j. For large gain κ, two categories of equilibrium delays emerge that contribute to the global frequency Ω in Eq. (30): For half the connections with offset difference \(\Delta _{ij} < 0\), the corresponding delays \(\tau _{ij}(t)\) decay to \(\tau _{ij}^{E} = 0\) as we have \(\tau ^{0} < \kappa \Delta _{ij} < 0\) with \(\tau ^{0} \ll \kappa \). Otherwise, with \(\Delta _{ij} > 0\) the delay \(\tau _{ij}(t)\) establishes a positive equilibrium at \(\tau _{ij}^{E} = \tau ^{0} + \kappa \Delta _{ij}\). The positive delays \(\tau _{ij}^{E}\) become widely distributed under large standard deviation κδ.
The coupled network can synchronize towards any stable point \((\varOmega , \delta ^{2})\) along the curve \(R(\varOmega , \delta ^{2}) = 0\). To assess the stability at each state \((\varOmega , \delta ^{2})\), consider our Ndimensional eigenvalue stability criterion \(M_{\lambda }\vec{v} = 0\) for eigenvector \(\vec{v} \in \mathbb {C}^{N}\) and matrix \(M_{\lambda }\) whose entries are given by Eq. (13). That is,
for all \(i \leq N\), where \(\tilde{\kappa } = \varOmega \kappa \), \(C_{0}(\Delta ) = \cos (\Delta )\), \(C_{\varOmega }(\Delta ) = \cos (\varOmega \tau ^{E}(\Delta ) + \Delta )\), and \(\tau ^{E}(\Delta ) = H(\Delta )(\tau ^{0} + \kappa \Delta )\). Once again, if our offset variance \(\delta ^{2}\) is small for Eq. (31) we can approximate the terms \(\Delta _{ij} = \Delta _{j}  \Delta _{i} \approx \Delta _{j}\) by assuming \(\Delta _{i} \approx 0\) at each i. We derive an Nlimit version of the eigenvalue Eq. (32) as follows. Likewise to the frequency equation Ω, we can obtain a similar Nlimit approximation to Eq. (32) as follows. We define a continuous eigenfunction \(v:[0,1] \rightarrow \mathbb {C}\) such that \(v_{i} = v(i/N)\). Then, taking the limit \(N \rightarrow \infty \) to Eq. (32) with \(\Delta _{ij} \approx \Delta _{j} \sim N(0, \delta ^{2})\), we obtain the continuous eigenvalue criterion for λ with respect to eigenfunction \(v(x)\) given by
for every \(x \in [0,1]\). Justification regarding the Nlimit step to derive Eq. (33) is provided in Appendix B. Here, the only continuous eigenfunction solution to the above equations is the constant function \(v(x) = 1\).^{Footnote 1} Hence, Eq. (33) simplifies to the following Nlimit eigenvalue equation for λ:
We claim that the synchronization state \((\varOmega , \delta ^{2})\) is (neutrally) stable given that for all \(\lambda \in \mathbb {C}\) satisfying Eq. (34), \(\operatorname {Re}(\lambda ) \leq 0\). Note that for all \(\lambda = u + iv\) satisfying Eq. (34),
where \(I_{1} = \int _{0}^{\infty }C_{\varOmega }(\Delta )\rho (\Delta ) \, d \Delta \). By Eq. (35), if \(u \rightarrow \infty \), \(\lambda  \leq gI_{1}\), which is a contradiction. It follows that, as \(\lambda  \rightarrow \infty \), \(\operatorname {Re}(\lambda ) \rightarrow \infty \) for the distribution of eigenvalues λ satisfying Eq. (34). If \(\kappa = 0\), then the frequency Ω solving Eq. (31) is (neutrally) stable if all eigenvalues \(\lambda \in \mathbb {C}\) given by
have nonpositive real parts \(\operatorname {Re}(\lambda ) \leq 0\). As shown in [22], the nonplastic delay network synchronizes at Ω if and only if \(\cos (\varOmega \tau ^{0}) > 0\).
At this point, the Nlimit approximate eigenvalue Eq. (34) has done little to improve the Ndimensional criterion \(\det M_{\lambda }= 0\), due to exponential blowup that persists within the integrand term. We proceed to reduce the eigenvalue Eq. (34) into an exponential polynomial root equation as follows. Rescale \(\lambda \rightarrow R \lambda \) by some radius \(R > \kappa \),
with rescaled small delay term \(\tau ^{R} = \tau ^{0} / R\). Expressing the power series of the exponential λ term up to degree M in Eq. (37), we obtain the approximate exponential polynomial root equation
with polynomials
and power term coefficients
Choosing a large radius \(R > \kappa \), the rescaled delay term \(\tau ^{R}\) and coefficients \(I_{m}\) for larger degrees m become arbitrarily small. From this, we claim that the stability at \((\varOmega , \delta ^{2})\) is predominantly determined by the first few terms of the exponential power expansion. That is, the stability at synchronization state \((\varOmega , \delta ^{2})\) is determined by the finitely many polynomial roots \(\lambda \in \mathbb {C}\) satisfying Eq. (38) with \(\tau ^{R} \approx 0\) and some degree M. Denoting \(\varLambda (\varOmega , \delta ^{2}) = \{\lambda _{1}, \ldots , \lambda _{M} \}\) as the roots of Mdegree polynomial \(P(\lambda \mid \varOmega , \delta ^{2}) + Q(\lambda \mid \varOmega , \delta ^{2})\), the synchronization state \((\varOmega , \delta ^{2})\) is (neutrally) stable given that
We note that these analytic results are reminiscent of what we obtained for twodimensional systems in Sect. 4 with the cubic exponential polynomial equation \(P_{\varOmega }(\lambda ) + Q_{\varOmega }(\lambda )e^{\lambda \tau } = 0\) as defined by Eqs. (26) and (27).
As we experienced large scale fluctuations of stability term \(E(\varOmega , \delta ^{2})\), we processed its value with either the sign function \(\operatorname {sgn}(x)\) or the sign logorithm \(\text{s.} \log (x)\) defined as
In Fig. 3(B), we plot \(\operatorname {sgn}E(\varOmega , \delta ^{2})\) for all \(\varOmega \in [\omega _{0}  g, \omega _{0} + g]\) and small \(\delta > 0\). Stable regions are indicated where \(\operatorname {sgn}E(\varOmega , \delta ^{2}) = 1\). In computing \(E(\varOmega , \delta ^{2})\), eigenvalues satisfying \(\operatorname {Re}(\lambda ) > 10^{8}\) were considered to be nonzero. We notice that a section of the level curve \(\mathcal{R}(\varOmega , \delta ^{2}) = 0\) overlaps part of the stable region for \((\varOmega , \delta ^{2})\). Indeed, Fig. 3(C) plots the stability term \(\text{s.} \log E(\varOmega , \delta ^{2})\) along all points \((\varOmega , \delta ^{2})\) on the implicit solution curve \(\varOmega = \varOmega (\delta )\). As the plot shows, there is a small interval \((\delta _{l}, \delta _{r})\) such that \((\varOmega (\delta ), \delta )\) is stable for all \(\delta \in (\delta _{l}, \delta _{r})\) such that \(E(\varOmega , \delta ^{2}) < 0\). Figure 3(D) shows the transition of the eigenvalues \(\lambda \in \mathbb {C}\) of \(P(\lambda \mid \varOmega , \delta ^{2}) + Q(\lambda \mid \varOmega , \delta ^{2}) = 0\), corresponding to state \((\varOmega (\delta ), \delta ^{2})\) as δ leaves the interval \((\delta _{l}, \delta _{r})\). We see that the eigenvalues shift towards the rightside of the imaginary axis, which shows the transition from stability to instability along the solution curve \((\varOmega (\delta ), \delta ^{2})\).
The results above were obtained by setting the polynomial degree to \(M = 3\) for \(Q(\lambda \mid \varOmega , \delta ^{2})\), while the approximation failed for powers \(M > 3\). In addition, it remains unresolved whether the distribution of eigenvalues \(\lambda _{N}\) satisfying the Ndimensional Eq. (32) with some eigenvector \(\vec{v} \in \mathbb {C}^{N}\) generally converge to the Nlimit eigenvalues λ satisfying Eq. (33) with corresponding continuous eigenfunction \(v:[0,1] \rightarrow \mathbb {C}\). That is, whether for each \(\epsilon > 0\) there is some large N such that for all \(\lambda _{N} \in \mathbb {C}\) satisfying Eq. (32) each at N dimensions, there exists some eigenvalue \(\lambda \in \mathbb {C}\) satisfying Eq. (33) such that \(\lambda _{N}  \lambda  < \epsilon \). Integral equation theory has proven some relevant theorems. For instance, it can be proven [25] that there exists a nonzero eigenvalue \(\lambda \in \mathbb {C}\) to Eq. (33) such that \(\lambda _{N(k)} \rightarrow \lambda \) for some subsequence of eigenvalues \(\lambda _{N(k)} \in \mathbb {C}\) satisfying Eq. (32) at dimension \(N = N(k)\). Nevertheless, our Nlimit analysis highlights an approximate region of points \((\varOmega , \delta ^{2})\) where a largedimensional system of oscillators will synchronize. We demonstrate in the following section that this region aligns with synchronous behavior in numerical simulations.
Comparison to numerical simulations
Here, we validate the theoretical analysis committed in the previous sections through comparisons with numerical simulations. We obtain the global frequency Ω and asymptotic phaseoffsets \(\phi _{i}\) numerically, and systematically compare them with their analytical counterparts. Given numerical solution \(\theta _{i}(t)\), \(i \leq N\) to the Kuramoto equation (1) with corresponding derivative solution \(\theta _{i}'(t), i \leq N\), we can obtain the asymptotic frequencies for each oscillator by summing over the time interval \([t, t+h]\) and taking the limit \(t \rightarrow \infty \). That is,
and estimate the global frequency Ω as the sample mean of individual frequencies
Likewise, we can numerically estimate the asymptotic phaseoffsets \(\phi _{i}\) for each oscillator by taking the difference \(\widehat {\phi}_{i}(t) := \theta _{i}(t)  \widehat {\varOmega}t \) and defining the limit
After modding \(\widehat {\phi}_{i}\) so that \(\widehat {\phi}_{i} \in [\pi , \pi )\), we can estimate the offset variance \(\delta ^{2}\) by taking the sample variance
and the average (asymptotic) phase ϕ̅ by taking the sample mean,
If our solutions \(\theta _{i}(t)\) synchronize towards some synchronous frequency Ω, offsets \(\phi _{i}\), then Ω̂, \(\widehat {\phi}_{i}\) are estimators for Ω, \(\phi _{i}\), respectively. Numerically, we evaluate the estimators by taking the average over interval \([t, t+h]\) starting at some large time \(t > 0\).
We set all baseline delays to be a unique value \(\tau ^{0} > 0\), so that \(\tau _{ij}(0) = \tau _{ij}^{0} = \tau ^{0}\). Since the plasticity rule Eq. (3) is an ordinary differential equation, it suffices for all delays to have an initial value at \(t = 0\). Before \(t = 0\), we set the phases \(\theta _{i}(t)\) to be positioned in accordance to some initial frequency \(\varOmega _{0}\) and initial phaseoffsets \(\phi _{i}^{0}\). That is, we define the linear initial function \(\varphi _{i}(t) = \varOmega _{0} t + \phi _{i}^{0}\) and set \(\theta _{i}(t) = \varphi _{i}(t)\) on \(t \leq 0\).^{Footnote 2} In order to have reasonably behaving solutions \(\theta (t)\), we modify the function \(\varphi _{i}(t)\) so that it satisfies the necessary condition
for all i. This adjustment is needed in order to avoid numerical discontinuities for \(\theta _{i}'(t)\) at \(t = 0\). For details, refer to Appendix C.
Figure 4 plots the results of a series of numerical simulations in the reduced twooscillator network as set up in Sect. 4. All trials were run from 0–200 s, with estimated values Ω̂, \(\widehat {\phi}_{i}\) obtained by averaging the arrays over the last 20 seconds. The same parameter values used in Fig. 2 were applied here when running the numerical trials. We demonstrate the existence of two stable synchronization states with different frequencies. Figure 4(A), (B), (C) shows the asymptotic behaviour of two trials (purple, orange) that started with different initial functions, plotting over the first 50 seconds. Figure 4(A) plots the derivative arrays \(\theta _{1}'(t)\), \(\theta _{2}'(t)\) of each trial. We observe in Fig. 4(A) that each pair of oscillator frequencies entrain to a common frequency over time. The two trials converge to different common frequencies, estimated to be \(\widehat {\varOmega} = 0.916\) (purple lines) and \(\widehat {\varOmega} = 0.625\) (orange lines), respectively. Figure 4(B) plots the offset arrays \(\sin \widehat {\phi}_{1}(t)\), \(\sin \widehat {\phi}_{2}(t)\) of each trial. For each trial, the twooscillators become phaselocked as \(\widehat {\phi}_{2}(t)  \widehat {\phi}_{1}(t)\) converges asymptotically to estimated constant differences \(\widehat {\Delta}_{12} = 0.111\) (purple lines) and \(\widehat {\Delta}_{12} = 0.523\) (orange lines), where \(\widehat {\Delta}_{12} = \widehat {\phi}_{2}  \widehat {\phi}_{1}\). Figure 4(C) plots the connection delays \(\tau _{12}(t)\), \(\tau _{21}(t)\) over time. By choosing our plasticity gain \(\kappa = 30 \gg 0.1 = \tau ^{0}\), it follows that for both trials, \(\tau _{12}(t)\) converges to some positive equilibrium delay \(\tau ^{E}\) and \(\tau _{21}(t)\) decays to 0.
The above results imply that there exist at least two stable synchronization states, and that the frequency the system converges towards is dependent on the initial functions \(\varphi _{1}(t)\), \(\varphi _{2}(t)\). To provide further confirmation towards this proposition, the trials shown in Fig. 4 were repeated 80 times randomized across sampled initial frequency \(\varOmega _{0} \in [\omega _{0}  g, \omega _{0} + g]\) and initial phase difference \(\Delta _{0} \in (0,1)\). The point \((\varOmega _{0}, \Delta _{0})\) (circle markers) defines the initial functions \(\varphi _{1}(t) = \varOmega _{0} t\), \(\varphi _{2}(t) = \varOmega _{0} t + \Delta _{0}\). The theoretically stable states corresponding to frequencies \(\varOmega = \varOmega _{1}\) and \(\varOmega = \varOmega _{3}\) are plotted as orange and purple stars, respectively, which are both close to the asymptotic frequency of the matching coloured trial discussed above. The rest of the theoretical synchronous states given by the roots of Eq. (23) are plotted as blue stars. Each trial’s solution arrays synchronized near one of the two stable states, as shown by the connecting coloured lines. The matching colours indicate which of the two stable frequencies \(\varOmega _{i}\) the trial’s solution arrays synchronized towards, such that the \( \widehat {\varOmega}  \varOmega _{i}  < 5 \times 10^{3}\). Each of the two trials graphed in Fig. 4(A), (B), (C) are also plotted in Fig. 4(D) with a diamond marker of matching colour. Convergence of the points is suggestive of a separatrix curve between the basins of attraction of both stable fixed points. It was also observed that the system synchronized towards the frequency \(\varOmega = \varOmega _{3}\) faster than \(\varOmega = \varOmega _{1}\), which suggests that the state \(\varOmega = \varOmega _{3}\) has a greater force of attraction. Hence, the experimental results align with the analysis outlined in Sect. 4. We draw the conclusion that in our reduced twooscillator system, plastic delays are able to generate multiple synchronization states in comparison to nonplastic delays.
Figure 5 provides the numerical results of a similar experiment performed as in Fig. 4 with a largedimensional network \(N = 50\) and alltoall network. All trials were run from 0100 s, with estimated values Ω̂, \(\widehat {\phi}_{i}\) obtained by averaging the arrays over the last 10 seconds. The same parameter values are used as in Fig. 3. The initial function \(\varphi (t)\) for each trial was set up by choosing some frequency \(\varOmega _{0} \in [\omega _{0}  \frac{g}{4}, \omega _{0} + \frac{g}{4}]\) and deviation \(\delta _{0} \in (0, 0.5)\). The initial phases \(\phi _{i}^{0}\) were i.i.d. sampled uniformly from the interval \([\sqrt{3}\delta _{0}, \sqrt{3}\delta _{0}]\). The ith oscillator was equipped with the initial linear function \(\varphi _{i}(t) = \varOmega _{0}t + \phi _{i}^{0}\). Figure 5(A), (B), (C), (D) graphs a single numerical trial with \(\varOmega _{0} = 0.913\), \(\delta _{0} = 0.295\). Figure 5(A) plots the derivative arrays \(\theta _{i}'(t)\), which we note converges to some constant frequency estimated as \(\widehat {\varOmega} = 0.839\). Figure 5(B) plots the offset arrays \(\sin \widehat {\phi}_{i}(t)\), which shows that each \(\widehat {\phi}_{i}(t)\) converges to some constant phaseoffset estimated by \(\widehat {\phi}_{i}\). Hence, the oscillators become asymptotically phaselocked under distributed offsets with estimated variance \(\widehat {\delta}^{2} = 0.050^{2}\). Figure 5(C) plots a sample of 50 adaptive delays \(\tau _{ij}(t)\), which become part of the positive equilibrium distribution \(\tau _{ij}^{E} > 0\) or decay to 0. Figure 5(D) plots the density of centered phases \(\widehat {\Delta}_{i} = \widehat {\phi}_{i}  \overline{\phi }\). As we assumed that the centralized phaseoffsets follow a Gaussian distribution, we perform a normality test on the numerical asymptotic offsets \(\widehat {\Delta}_{i}\). The Shapiro–Wilk test for nonnormality returned a pvalue of 0.005 for \(\widehat {\Delta}_{i}\), which suggests another distribution would be more accurate. Visually, a Gaussian curve \(N(0, \widehat {\delta}^{2})\) (black line) is fit over the density in Fig. 5(D). Nevertheless, the relevance of the Gaussian approximation, which greatly simplifies the analysis, becomes apparent as the numerical and analytical results nearly coincide. Other approximations could be used to facilitate the analysis further, and are left for future work.
The numerical simulation, as presented above, was repeated 10 times with randomized initial conditions \((\varOmega _{0}, \delta _{0}^{2})\). For each trial, \(\varOmega _{0}\), \(\delta _{0}\) was sampled uniformly from intervals \([\omega _{0}  \frac{g}{4}, \omega _{0} + \frac{g}{4}]\) and \((0, 1)\), respectively. Figure 5(E) plots the following convergence results. Each trial with initial condition \((\varOmega _{0}, \delta _{0}^{2})\) (yellow markers) synchronized near the respective estimated point \((\widehat {\varOmega}, \widehat {\delta}^{2})\) (magenta markers). We can see that every trial synchronized at approximately the same state \((\varOmega , \delta ^{2})\). To determine whether the numerical results align with the analysis discussed in Sect. 5, the numerical values were plotted on top of Fig. 3(B). We observe that for each trial, the network synchronizes near the portion of the level curve \(\mathcal{R}(\varOmega , \delta ^{2}) = 0\) within the stable region where \(E(\varOmega , \delta ^{2}) < 0\). Hence, the numerical experiment for \(N = 50\) generated results that validates the theoretical Nlimit stability analysis in Sect. 5.
Neuroscience application: resilience to injury with sparse and uniform connectivity
One of the most salient examples of white matter plasticity comes from neuroimaging in the presence of a lesion. In these cases, white matter remodeling takes place in order to restore and maintain function, a process that notably impacts neural synchronization [26]. Let us now investigate whether the plasticity mechanism can be used to stabilize phaselocked states in the presence of network damage. That is, we would like to know whether changes in time delays can be used to compensate for a reduction in effective connectivity and make the global synchronous state more resilient. To investigate this problem, we model injury as a loss in connections \(a_{ij}\). Defining \(\gamma \in [0,1]\) as the insult index, we introduce here the sparse synaptic connectivity weights given by
where \(p_{ij}\) is a uniformly distributed i.i.d. sampling on \([0,1]\). γ represents the connectivity damage through an increase in the sparseness of network connections. Note that if \(\gamma = 0\), then we obtain the alltoall connection topology \(a_{ij} \equiv 1\) corresponding to no injury in the system. If \(\gamma = 1\), we have the trivial system \(a_{ij} = 0\) which means no signals between the oscillators occur. For any γ we have the probability to retain the connection \(\mathbb {P}(a_{ij} = 1) = 1  \gamma \). Without plasticity, the mean phase dynamics for \(N \rightarrow \infty \) is given by
see Eq. (31). By observation, one can see that we are in the presence of the same network dynamics, but with an effective coupling coefficient given by \(g_{\text{eff}} := (1  \gamma )g\). Thus, damage simply reduces the net coupling. To demonstrate this point, we start with a strong coupling parameter and decrease it until stability is either lost or preserved by plasticity. According to Eq. (51), \(\varOmega \rightarrow \omega _{0}\) as \(\gamma \rightarrow 1\). Hence, if the baseline lag \(\tau ^{0}\) is chosen such that \(\cos (\varOmega \tau ^{0}) \geq 0\) and \(\cos (\omega _{0}\tau ^{0}) < 0\), the network without plastic delays is susceptible to injury destabilizing the synchronous state.
We ran numerical simulations by applying similar parameter values as Sect. 6 while introducing injury \(\gamma = 0.8\) to the connections at time \(t = t_{\mathrm{inj}}\), set at \(t_{\mathrm{inj}} = 80~\mbox{s}\). The initial condition of the network was fixed at \((\varOmega _{0}, \delta _{0}) = (\omega _{0}, 0.25)\) for all trials. Figure 6(A) shows the destruction of existing connections following injury, comparing connection grids for \(a_{ij}\) before and after \(t_{\mathrm{inj}}\). Figure 6(B) shows the distribution of the delays \(\tau _{ij}(t)\) with existing connections \(a_{ij} = 1\) at timestamps \(t = 0~\mbox{s}\), \(t = 79~\mbox{s}\) (preinjury), \(t = 160~\mbox{s}\) (postinjury). With fewer connections available, the ability for the surviving delays to adjust themselves are crucial in restabilizing the system’s synchrony. Figure 6(C) (no gain) and Fig. 6(D) (with gain) show that both networks entrain to a global frequency successfully before and after inflicted injury towards the network. The entrainment frequencies preinjury \(\varOmega _{\text{pre}} \approx \widehat {\varOmega}_{\text{pre}}\) and postinjury \(\varOmega _{\text{post}} \approx \widehat {\varOmega}_{\text{post}}\) were estimated by taking the average of \(\theta _{i}'(t)\) at times 148–160 s and 304–320 s, respectively, for each \(i \leq N\). Figure 6(E) (no gain) and Fig. 6(F) (with gain) plots the sine phaseoffsets \(\sin \widehat {\phi}_{i}(t)\) over time, given by
Following injury, from Fig. 6(E) the network without adaptive delays collectively falls out of phase. In contrast, Fig. 6(F) shows the network with adaptive delays demonstrating resilience against the injury as most oscillators are able to collectively phaselock within close proximity to each other.
Figure 7 examines the effect of gradually increasing the severity of injury towards the system’s global frequency \(\varOmega \approx \widehat {\varOmega}\) and its phaseoffset variance \(\delta ^{2} \approx \widehat {\delta}^{2}\). For each trial at injury γ, the same initial condition and parameters were used as in Fig. 6. Figure 7(A) shows that connection loss generally leads to the system’s synchronization frequency Ω becoming closer to the natural frequency \(\omega _{0}\). Figure 7(B) plots the estimated phase standard deviation δ̂ with respect to increasing injury. The network without plastic delays exhibits a significant loss in coherent synchrony with increasing \(\widehat {\delta} > 0\) as more connections are lost. In contrast, the network equipped with adaptive delays persistently displays phase coherence with \(\widehat {\delta} \ll 1\) until higher injury levels γ.
Discussion
Our goal was to provide a mathematical framework that captures the synchronizing properties of networks with adaptive delays. We sought to implement the activitydependent property of myelinated connection delays by modifying the Kuramoto model as proposed in [9]. The focus was to determine whether such adaptive delays significantly improve the oscillatory system’s ability to become inphase and to entrain to a global frequency. Given that this is the case, the results of the model’s study reinforces the proposition that myelin plasticity is essential in maintaining the synchrony in the developing or injured brain.
White matter plays a critical role in maintaining brain function through the coordination of neural dynamics across multiple temporal and spatial scales. Recent evidence has shown that through the action of glia, white matter properties evolve continuously in time. Specifically, conduction velocity within and across brain areas is adjusted to promote efficient neural signaling. While the mechanism remains poorly understood, the consequences of such plastic processes on brain dynamics and synchronization can be readily examined and characterized using simplified mathematical models.
To accomplish this, we here examined the influence of adaptive conduction delays on the synchronization of neural oscillators. We developed a repertoire of mathematical tools to better examine the stability of phaselocked solutions. In theory, we derived implicit equations for the global frequency Ω and eigenvalues \(\lambda \in \mathbb {C}\) that provide a rigorous criterion for the stability around the synchronous state in twodimensional and largedimensional settings. Based on our model, flexibility in the white matter structure introduces an additional corrective dynamic next to the phase interactions that can further drive the network’s phase alignment. Higher stability with adaptive delays was demonstrated as the Kuramoto model had higher resilience against injury perturbations. However, adaptive delays improve the system’s synchronous features only when the delays adjust with a sufficiently high degree of plasticity, as represented by the plasticity gain κ.
There are many limitations in the prototypical model we used and its corresponding results. Myelination is bound by many physiological constraints, some of which remain uncertain [4]. It is established that white matter restructures itself in response to ongoing neural activity [8]. We primarily incorporated this fact in our plasticity rule in a manner that promotes local synchrony. Indeed, each connection delay changes at a rate proportional to the sine of the oscillator’s phase difference. This rule remains a tentative construction, as more research is needed to develop more biological relevant models in activitydependent myelination. In addition, the use of phase oscillators to model local neural dynamics remain limited and is relevant mostly in the context of largescale neural systems. In our analysis, we relied heavily on i.i.d. parametrical frameworks in order to establish our Nlimit approach, which may not be feasible as network elements are correlative in nature.
Moving forward, we hope to build upon our analysis alongside newly found experimental results pertaining to myelin. Despite its shortcomings, the mathematical approaches used and its results can potentially be applied to more complex, biological relevant models. The conduction delays \(\tau _{ij}(t) \propto c_{ij}^{1}(t)\) can be alternatively modelled with respect to a system of adaptive conduction velocities \(c_{ij}(t)\). In the realm of temporal equations, other parametric avenues have yet to be explored. For instance, the delays \(\tau _{ij}(t)\) can exhibit slow convergence by setting the rate constant \(\alpha _{\tau }\ll 1\). The aforementioned concepts are some proposed examples that may further lead to uncharted dynamics in the scope of neurocomputational models.
Notes
 1.
We can represent any connection topology of the network as an integral operator \(a:[0,1]^{2} \rightarrow \{0,1\}\) such that \(a_{ij} = a(i/N, j/N)\), for which we find the eigenfunction \(v:[0,1] \rightarrow \mathbb {C}\) satisfying
$$ \sigma v(x) = \int _{0}^{1} a(x,y)v(y) \,dy $$with respect to eigenvalue σ.
 2.
Trials were done using other types of initial functions \(\varphi _{i}(t)\), and yielded similar convergence results. However, the synchronization occurred at a slower rate.
Abbreviations
 i.i.d.:

independent and identically distributed
References
 1.
Strogatz SH. From Kuramoto to Crawford: exploring the onset of synchronization in populations of coupled oscillators. Phys D, Nonlinear Phenom. 2000;143(1):1–20.
 2.
Buzsáki G. Rhythms of the brain. 2009.
 3.
Cabral J, Luckhoo H, Woolrich M, Joensson M, Mohseni H, Baker A, et al.. Exploring mechanisms of spontaneous MEG functional connectivity: how delayed network interactions lead to structured amplitude envelopes of bandpass filtered oscillations. NeuroImage. 2013;12:90.
 4.
de Hoz L, Simons M. The emerging functions of oligodendrocytes in regulating neuronal network behaviour. BioEssays. 2015;37:60–9.
 5.
Monje M. Myelin plasticity and nervous system function. Annu Rev Neurosci. 2018;41:61–76.
 6.
Bechler M, Swire M, FfrenchConstant C. Intrinsic and adaptive myelination—a sequential mechanism for smart wiring in the brain. Dev Neurobiol. 2017;78(2):68–79.
 7.
Almeida RG, Lyons DA. On myelinated axon plasticity and neuronal circuit formation and function. J Neurosci. 2017;37(42):10023–34.
 8.
Douglas Fields R. A new mechanism of nervous system plasticity: activitydependent myelination. Nat Rev Neurosci. 2015;16:756–67.
 9.
Pajevic S, Basser PJ, Fields RD. Role of myelin plasticity in oscillations and synchrony of neuronal activity. Neuroscience. 2014;276:135–47. Secrets of the CNS White Matter.
 10.
Breakspear M, Heitmann S, Daffertshofer A. Generative models of cortical oscillations: neurobiological implications of the Kuramoto model. Front Human Neurosci. 2010;4:190.
 11.
Izhikevich EM. Phase models with explicit time delays. Phys Rev E. 1998;58:905–8.
 12.
Yeung MKS, Strogatz SH. Time delay in the Kuramoto model of coupled oscillators. Phys Rev Lett. 1999;82:648–51.
 13.
Petkoski S, Jirsa V. Transmission time delays organize the brain network synchronization. Philos Trans R Soc A, Math Phys Eng Sci. 2019;377:20180132.
 14.
Petkoski S, Palva JM, Jirsa V. Phaselags in large scale brain synchronization: methodological considerations and insilico analysis. PLoS Comput Biol. 2018;14:e1006160.
 15.
PonceAlvarez A, Deco G, Hagmann P, Romani GL, Mantini D, Corbetta M. Restingstate temporal synchronization networks emerge from connectivity topology and heterogeneity. PLoS Comput Biol. 2015;11:e1004100.
 16.
Dhamala M, Jirsa VK, Ding M. Enhancement of neural synchrony by time delay. Phys Rev Lett. 2004;92:074104.
 17.
Fries P. A mechanism for cognitive dynamics: neuronal communication through neuronal coherence. Trends Cogn Sci. 2005;9(10):474–80.
 18.
Hale JK, Lunel SMV. Introduction to functional differential equations. Berlin: Springer; 1993.
 19.
Hartung F, Krisztin T, Walther HO, Wu J. Chapter 5: Functional differential equations with statedependent delays: theory and applications. In: Handbook of differential equations: ordinary differential equations. vol. 3. 2006. p. 435–545. http://www.sciencedirect.com/science/article/pii/S187457250680009X.
 20.
Cooke KL, Huang W. On the problem of linearization for statedependent delay differential equations. Proc Am Math Soc. 1996;124(5):1417–26.
 21.
Papachristodoulou A, Jadbabaie A. Synchronization in oscillator networks: switching topologies and nonhomogeneous delays. In: Proceedings of the 44th IEEE conference on decision and control. 2005. p. 5692–7.
 22.
Earl MG, Strogatz SH. Synchronization in oscillator networks with delayed coupling: a stability criterion. Phys Rev E. 2003;67:036204.
 23.
Atay F. Distributed delays facilitate amplitude death of coupled oscillators. Phys Rev Lett. 2003;91:094101.
 24.
Cooke D. On zeroes of some transcendental equations. Funkc Ekvacioj. 1986;29:77–90.
 25.
Russell RD, Shampine LF. Existence of eigenvalues of integral equations. SIAM Rev. 1971;13(2):209–19. https://doi.org/10.1137/1013038.
 26.
Bells S, Lefebvre J, Prescott SA, Dockstader C, Bouffet E, Skocic J, et al.. Changes in white matter microstructure impact cognition by disrupting the ability of neural assemblies to synchronize. J Neurosci. 2017;37(34):8227–38. https://www.jneurosci.org/content/early/2017/07/25/JNEUROSCI.056017.2017.
 27.
Park SH. Kuramoto plastic delay. https://github.com/parkseo7/KuramotoPlasticDelay.git. Accessed March 27 2020.
Acknowledgements
We would like to thank Adam Stinchcombe for his valuable feedback regarding the numerical methods.
Availability of data and materials
The datasets generated and/or analysed during the current study are available in the github repository [27].
Funding
This work has been funded by the National Science and Engineering Research Council (NSERC), Canadian Institute for Health Research (CIHR) and University of Toronto.
Author information
Affiliations
Contributions
SHP and JL performed the research, did the analysis and wrote the manuscript. All authors read and approved the final manuscript.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
Not applicable.
Competing interests
The authors declare they have no competing interests.
Consent for publication
Not applicable.
Appendices
Appendix A: Existence, uniqueness, and linearization
We sketch out the arguments needed to justify the existence, uniqueness, and stability analysis of our proposed Kuramoto model with adaptive delays. Our model can be expressed as an \(N + N^{2}\) system of oscillators \(\theta (t) = (\theta _{i}(t))_{1 \leq i \leq N} \in \mathbb {R}^{N}\) and delays \(\tau (t) = (\tau _{ij}(t))_{1 \leq i,j \leq N} \in \mathbb {R}_{+}^{N \times N}\), where \(\mathbb {R}_{+} := [0, \infty )\), defined by the statedependent functional differential equations
where \(\theta _{t} \in C := ([r,0], \mathbb {R}^{N})\) defined by \(\theta _{t}(s) = \theta (t+s)\), \(s \in [r,0]\), and C is the space of all continuous functions from interval \([r,0]\) for some \(r > 0\). Here, \(F: C \times \mathbb {R}_{+}^{N \times N} \rightarrow \mathbb {R}^{N}\) is the function whose ith coordinate \(F_{i}(\theta _{t}, \tau (t))\) is given by Eq. (1) and \(G: \mathbb {R}^{N} \times \mathbb {R}_{+}^{N \times N} \rightarrow \mathbb {R}_{+}^{N \times N}\) is the function whose ijth coordinate \(G_{ij}(\theta _{i}(t), \theta _{j}(t), \tau _{ij}(t))\) is given by Eq. (3). We note that, from Eq. (3), each delay solution \(\tau (t)\) is bounded for all \(t \geq 0\). This is clear since given \(\tau _{ij}(t) > \tau ^{0} + \kappa \), we must have \(\tau _{ij}'(t) < 0\). Hence, it suffices to define the space C with any \(r > \tau ^{0} + \kappa \).
Denote \(C^{1} \subset C\) to be the subspace of all continuously differentiable functions. Let \(\theta _{0}(s) = \varphi (s)\) for some \(C^{1}\) initial function \(\varphi : [r,0] \rightarrow \mathbb {R}^{N}\) and \(\tau (0) = \tau ^{0} \in \mathbb {R}_{+}^{N \times N}\) be the initial delays. Since Eq. (54) is a system of ordinary differential equations, \(\tau (t)\) is independent of initial values on \(t < 0\). Hence, we represent an initial function for \(\tau (t)\) with an initial point \(\tau ^{0} \in \mathbb {R}_{+}^{N \times N}\). Existence and uniqueness of solution \((\theta (t), \tau (t))\) follows from theorems in [18] given that both F, G are continuously differentiable functions over \((\varphi , \tau ^{0}) \in C^{1} \times \mathbb {R}_{+}^{N \times N}\). The ith coordinate \(F_{i}(\varphi , \tau ^{0})\) is a linear combination of terms \(\sin (\varphi _{j}(\tau _{ij}^{0})  \varphi _{i})\), which is continuously differentiable over \(\varphi \in C^{1}\). G is continuously differentiable given the Heaviside function \(H(\tau )\) is smooth. The Heaviside function used throughout the paper is constructed as follows. Setting some arbitrarily small \(\varepsilon > 0\), we define the smooth mollifier function \(h: \mathbb {R}\rightarrow \mathbb {R}\) by
Then our smooth Heaviside function \(H(\tau )\) is defined as
Then each coordinate \(G_{ij}\) is smooth with respect to terms \(\varphi _{i}\), \(\varphi _{j}\), \(\tau ^{0}\).
Linearization and stability of our system around the synchronized state \((\varOmega , \boldsymbol {\phi })\) can be justified as follows. We can express Eqs. (53) and (54) under perturbation terms \((\epsilon (t), \eta (t))\) of \((\theta (t), \tau (t))\) defined by Eqs. (7) and (8). That is, consider the autonomous equations
where the ith coordinate of \(\tilde{F}(\epsilon _{t}, \eta (t))\) is given by Eq. (10) and \(\tilde{G}(\epsilon (t), \eta (t)) = G(\varOmega t + \boldsymbol {\phi }+ \epsilon (t), \tau ^{E} + \eta (t))\). Then, if \((\varOmega , \boldsymbol {\phi })\) satisfies Eq. (5), \(\tilde{F}(0,0) = \tilde{G}(0,0) = 0\). Hence, the stability of \((\theta , \tau )\) around \((\varOmega , \boldsymbol {\phi })\) depends on the linearization of F̃, G̃ around \((0,0)\). One can compute that the corresponding linearized system \((\dot{\epsilon }(t), \dot{\eta }(t)) = L(\epsilon (t), \eta (t))\) is given by Eqs. (9) and (11). To prove that this linearization is valid and provides the eigenvalue stability criteria \(\det M_{\lambda }= 0\), we can replicate the proof for Theorem 2.1 in [20]. Indeed, F̃, G̃ are composed of Lipschitz continuous terms that allow the steps to be valid for our perturbation Eqs. (57) and (58).
Appendix B: Derivation of Nlimit equations
Here, we justify the limit steps taken in Sect. 5 to derive the limitN equations for global frequency Ω and stability eigenvalues λ around the synchronous state. First, denote \(\mathbb {E}g(X)\) as the expectation of random variable \(g(X)\). That is, given X has distribution \(\rho (x)\),
The main proposition we applied is as follows.
Theorem 1
Let \((v_{k})_{k \geq 1}\)be a bounded sequence of real numbers, and X be a random variable such that \(\mathbb {E}X = 0\), \(\mathbb {E}X < \infty \). Suppose \((X_{k})_{k \geq 1}\)is an i.i.d. sequence of random variables such that each \(X_{i}\)has the same distribution as X. Then
almost surely as \(N \rightarrow \infty \).
Proof
For some \(M > 0\), \(v_{k} < M\) for all k. By the classic Strong Law of Large Numbers,
almost surely as \(N \rightarrow \infty \). □
What directly follows from Theorem 1 is the following limit result, which we utilize to derive the limitN equations for the global frequency Ω and the stability eigenvalues λ.
Corollary 1
Let \((v_{k})_{k \geq 1}\)be a bounded sequence in \(\mathbb {C}\), and X be a random variable such that \(\mathbb {E}X < \infty \). Suppose \((X_{k})_{k \geq 1}\)is an i.i.d. sequence of random variables such that each \(X_{i}\)has the same distribution as X. Then
where v̅ is the welldefined asymptotic sample mean of the coefficients:
In particular, if \(v:[0,1] \rightarrow \mathbb {C}\)is a continuous function and \(v_{k} = v(k/N)\), then
Proof
This immediately follows by applying Theorem 1 to the i.i.d. sequence \(Y_{k} = X_{k}  \mathbb {E}X\). It remains to show that v̅ exists. Denoting \(\overline{v}_{N} = \frac{1}{N} \sum_{k=1}^{N} v_{k}\) as the sample mean among the first N coefficients with \(v_{k} \leq M\) for all k, we show that \((\overline{v}_{N})_{N \geq 1}\) is a Cauchy sequence. We have
for sufficiently large N. Hence, the limit \(\overline{v} = \lim_{N \rightarrow \infty } \overline{v}_{N}\) exists. In the case where \(v_{i} = v(i/N)\), it is clear that \(\overline{v} = \int _{0}^{1} v(x) \,dx\) by taking Riemann sums. □
We applied Corollary 1 to random sequences of the form \(X_{k} = f(\Delta _{k}, \tau ^{0})\) where \(\Delta _{k}\) is Gaussian distributed, and \(\tau ^{0}\) is the constant baseline lag. The eigenfunction coefficients \(v_{k}\) are from ansatz \(\epsilon _{k}(t) = v_{k} e^{\lambda t}\) as \(N \rightarrow \infty \).
Appendix C: Numerical setup
All numerical simulations in Sects. 6, 7 were done using MATLAB’s ddesd function. We have a system of delay differential equations for the phases \(\theta _{i}(t)\), \(i \leq N\) given by Eq. (1) and an \(N^{2}\) system of ordinary differential equations for the statedependent delays \(\tau _{ij}(t)\), \(1 \leq i\), \(j \leq N\) given by Eq. (3). The initial data before starting time \(t = 0\) are as follows. We have a history function \(\varphi : [r, 0] \rightarrow \mathbb {R}^{N}\) for the phases, such that \(\theta _{i}(t) = \varphi _{i}(t)\) for all \(t \leq 0\). We considered only the simple case where delays \(\tau _{ij}(t)\) have a unique baseline lag \(\tau _{ij}(0) = \tau ^{0} \in \mathbb {R}\). As shown in Appendix A, it is sufficient to consider \(C^{1}\) initial phase functions \(\varphi (t)\) to ensure unique solutions. However, we are not guaranteed to have reasonably behaving solutions unless we have continuity of \(\theta '(t)\) at \(t = 0\) [19]. This is also a necessary condition for providing accurate numerical simulations, as interpolation steps in ddesd rely on \(\theta '(t)\) being continuous everywhere.
For our numerical experiments in Sects. 6 and 7, we consider the linear initial functions \(\varphi _{i}(t) = \varOmega _{0} t + \phi _{i}^{0}\), \(i \leq N\) with respect to initial frequency \(\varOmega _{0} \in \mathbb {R}\) and phaseoffsets \((\phi _{1}^{0}, \ldots , \phi _{N}^{0}) \in \mathbb {R}^{N}\). As discussed above, we require that our linear initial function must satisfy the necessary condition Eq. (49). Given any initial \(C^{1}\) function \(\varphi (t)\) for \(\theta (t)\), we define the modified \(C^{1}\) function \(\tilde{\varphi }(t)\) of \(\varphi (t)\) that satisfies the necessary condition as follows. We can obtain the modified slope \(\vec{v} \in \mathbb {R}^{N}\) by imposing the condition on \(\varphi (t)\). That is,
Then for each i we define the cubic polynomial \(p_{i}: \mathbb {R}\rightarrow \mathbb {R}\) that interpolates \(\varphi (t)\) between \(t = 0, \tau ^{0}\) using the modified slope. That is, \(p_{i}(t) = \varphi _{i}(t)\) at \(t = 0, \tau ^{0}\) and \(p_{i}'(0) = v_{i}\), \(p_{i}'(\tau ^{0}) = \varphi _{i}(\tau ^{0})\). Then the modified initial function \(\tilde{\varphi }(t)\) is given by
for \(i \leq N\). All numerical trials were conducted using the corresponding modified functions \(\tilde{\varphi }_{i}(t)\) of \(\varphi _{i}(t) = \varOmega t + \phi _{i}^{0}\), \(i \leq N\).
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Park, S.H., Lefebvre, J. Synchronization and resilience in the Kuramoto white matter network model with adaptive statedependent delays. J. Math. Neurosc. 10, 16 (2020). https://doi.org/10.1186/s1340802000091y
Received:
Accepted:
Published:
Keywords
 White matter plasticity
 Kuramoto model
 Synchronization
 Delay differential equations