PaperlayerView exact revision on arXiv

Source-linked AI summary

Mathematical frameworks for oscillatory network dynamics in neuroscience

Peter Ashwin, Stephen Coombes, Rachel Nicks

arXiv:1506.05828v1nlin.AOq-bio.NC

TL;DR

The paper addresses the limits of weakly coupled phase-oscillator theory for neural networks, including strong coupling, stochastic forcing, and complex attractors. It reviews dynamical-systems tools spanning phase reduction, symmetry, synchrony, bifurcation, and heteroclinic dynamics, and proposes extending phase-amplitude methods to strongly coupled heterogeneous networks.

  • Problem

    Weakly coupled phase reduction is powerful for neural-network dynamics but has limitations, while strongly coupled, stochastic, and heteroclinic systems require broader analysis.

  • Method

    The review synthesizes mathematical approaches including phase response and interaction functions, symmetry-based bifurcation, synchrony analysis, and network models.

  • Results

    The review presents a practical framework for analysing oscillatory neural networks beyond a standard phase-oscillator perspective.

  • Takeaways & Limitations

    Phase-amplitude coordinates and network structural symmetries are proposed as foundations for studying strongly coupled heterogeneous networks and their nontrivial amplitude dynamics.

  • Takeaways & Limitations

    Isochron foliations are rarely available in closed form and generally require numerical computation based on the limit cycle.

Abstract

from arXiv · show

The tools of weakly coupled phase oscillator theory have had a profound impact on the neuroscience community, providing insight into a variety of network behaviours ranging from central pattern generation to synchronisation, as well as predicting novel network states such as chimeras. However, there are many instances when this theory is expected to break down, say in the presence of strong coupling, or must be carefully interpreted, as in the presence of stochastic forcing. There are also surprises in the dynamical complexity of the attractors that can robustly appear - for example, heteroclinic network attractors. In this review we present a set of mathematical tools that are suitable for addressing the dynamics of oscillatory neural networks, broadening from a standard phase oscillator perspective to provide a practical framework for further successful applications of mathematics to understanding network dynamics in neuroscience.

1 Introduction

Oscillator theory has become central to neuroscience because neural oscillations support diverse functions and arise across cellular, ensemble, and brain scales. The review introduces mathematical tools that extend weakly coupled phase reduction to broader collective dynamics.

  • Motivation: Neural oscillations contribute to feature binding, cognition, memory, odour perception, information transfer, coordination, rhythmic motor output, and neurological disorders.They can reflect both intrinsic cellular mechanisms and interactions among neurons.
  • Weak coupling: Weakly coupled oscillator theory reduces networks of limit-cycle oscillators to phase equations using invariant manifold and averaging theory.The relative phase becomes the relevant dynamical variable, enabling analyses from small rhythmic networks to whole-brain models.
  • Scope: The review broadens standard phase-oscillator analysis to address collective behaviours including partial synchrony, turbulent states, and heteroclinic attractors.These topics are developed alongside phase interaction functions and phase-locked states.
  • Prerequisites: The material assumes familiarity with nonlinear differential equations, dynamical-systems concepts, and generic codimension-one bifurcation theory.The stated prerequisites include linear stability, phase-plane analysis, saddle-node and Hopf bifurcations, and periodic-orbit bifurcations.

2 Neurons and neural populations as oscillators

Neural oscillators can be modelled across biophysical, reduced single-neuron, coupled-neuron, and neural-mass levels. Dynamical-systems analysis links their variables and coupling mechanisms to excitability, repetitive firing, network rhythms, and multistability.

  • Single-neuron models: Conductance-based Hodgkin-Huxley models describe membrane currents through voltage-dependent sodium and potassium channels but are high-dimensional and difficult to analyse.Reduced planar models simplify the dynamics by treating fast or similar gating variables through approximations.
  • Single-neuron models: Positive external current can destabilise an excitable fixed point and allow a limit cycle that produces a train of action potentials.The reduced planar model captures these essential Hodgkin-Huxley features geometrically, with onset through a subcritical Hopf bifurcation.
  • Single-neuron models: Quadratic-recovery cortical models can exhibit a saddle-node on an invariant circle bifurcation, with firing frequency scaling like √(I − Ic) near threshold.These models also support oscillatory solutions through mechanisms shared with the FitzHugh-Nagumo model.
  • Neuronal coupling: Purely inhibitory reciprocal coupling between two reduced Hodgkin-Huxley neurons can generate anti-phase half-center oscillations.Anode break excitation can support the rhythm when hyperpolarising inhibition terminates.
  • Neural mass models: Jansen-Rit neural-mass models can produce Hopf-generated oscillations and coexistence of two stable periodic orbits.The coexisting rhythms include an alpha-band oscillation and a lower-frequency, higher-amplitude oscillation.

3 Dynamical systems approaches to collective behaviour

The review uses dynamical-systems concepts and network structure to analyse collective behaviour in coupled nonlinear systems. It emphasizes synchrony, symmetry, bifurcation, stability, coupling topology, delays, and the limits of weak-coupling reductions.

  • Network formulation: Coupled neural systems can be represented as nonlinear ODEs whose individual dynamics are driven by the states of other systems.The coupling may be general or decomposed into weighted pairwise interactions, with network structure encoded by an adjacency matrix.
  • Synchrony: Collective dynamics may produce attracting synchrony or more complex spatio-temporal patterns such as clustering.The review distinguishes exact, generalised, phase, and frequency synchrony, while cautioning that synchrony usually characterises particular solutions rather than whole systems.
  • Network structure: Network topology influences possible dynamics, motivating analysis of adjacency, coupling weights, degree distributions, connectivity, path lengths, centrality, and motifs.The review contrasts scale-free degree distributions with homogeneous structured networks and distinguishes directed from undirected connectivity.
  • Limits and extensions: Weak coupling enables perturbation theory and phase-space reduction for limit-cycle oscillators but can miss effects such as oscillator death.Time delay is another important factor in collective dynamics and is natural to include in neural models.
  • Synchrony: For chaotic oscillators, an attracting exactly synchronised state requires coupling sufficiently strong relative to the maximal Lyapunov exponent.This result follows from analysing linearised differences between the oscillators.
  • Symmetric dynamics: Symmetry-based bifurcation theory classifies network states and predicts symmetry-breaking scenarios together with stability information.A transitive permutation symmetry implies identical oscillator vector fields.

3.7 Permutation symmetries and oscillator networks

Permutation symmetries organize coupled-cell dynamics through group actions, orbits, isotropy subgroups, and invariant fixed-point spaces. These structures constrain oscillator identities and periodic-orbit symmetries while enabling dimension reduction and dynamical classification.

  • Permutation actions: Permutation groups act on coupled-cell state spaces, with equivariance imposing constraints on the network vector field.The symmetric group consists of all permutations of N objects, represented as permutation matrices in simple state spaces.
  • Permutation actions: If the symmetry group acts transitively on the cells, all oscillators must share the same intrinsic dynamics.The vector fields then take the form fi(xi) = F(xi) for every cell.
  • Orbits and isotropy: Group orbits collect symmetry-related states that have dynamically equivalent behaviour, while isotropy subgroups record the symmetries fixing a state.The isotropy subgroups form an inclusion lattice, organizing states by their symmetry.
  • Invariant subspaces: Fixed-point spaces are dynamically invariant, and greater isotropy corresponds to smaller fixed-point subspaces.If H ⊂ K, then Fix(H) ⊃ Fix(K), allowing synchrony solutions to be studied within reduced-dimensional invariant spaces.
  • Invariant subspaces: Conjugate isotropy subgroups have equivalent fixed-point spaces and essentially equivalent dynamics, reducing the number of cases requiring analysis.Conjugate subgroups have fixed-point subspaces of the same dimension.
  • Periodic and chaotic attractors: Periodic-orbit symmetries are restricted to cyclic extensions of isotropy subgroups, whereas chaotic attractors can exhibit a wider range of spatio-temporal symmetries.For periodic orbits, K must be an isotropy subgroup and H must fix a connected component of Fix(K)/L_K.

3.9 Bifurcations with symmetry and genericity

Symmetry changes bifurcation theory by creating simultaneous marginal modes, constrained normal forms, and richer low-codimension dynamics. The same framework extends to coupled-cell networks with groupoid structures and invariant synchrony subspaces, although many cases remain unresolved.

  • Bifurcation analysis: Symmetric bifurcation analysis identifies marginally unstable modes, reduces to a symmetry-preserving centre manifold, and studies its constrained normal form.Symmetry can make the centre manifold higher-dimensional and impose zero, equal, or algebraically related normal-form coefficients.
  • Bifurcation analysis: Symmetry permits a wider range of low-codimension bifurcations because one instability can occur simultaneously in multiple symmetry-related directions.Without symmetry, generic codimension-one equilibrium bifurcations are saddle-node and Hopf bifurcations.
  • Bifurcation outcomes: Symmetry-constrained normal forms can support heteroclinic cycles or chaos bifurcating from highly symmetric solutions even in codimension one.These possibilities make symmetric bifurcations especially rich compared with generic unconstrained cases.
  • Bifurcation outcomes: Numerical path-following may miss symmetric bifurcation branches because degeneracies can produce many branches emerging simultaneously.The theory therefore needs to be developed for the particular group action.
  • Robust invariant dynamics: Symmetries can robustly preserve heteroclinic networks of equilibria, periodic orbits, or other invariant sets under sufficiently small symmetry-preserving perturbations.In higher-dimensional neural oscillator models, related attractors include depth-two connections, cycling chaos, winnerless competition, chaotic itinerancy, and slow switching.
  • Coupled-cell and groupoid structures: Groupoids generalize symmetry constraints for coupled-cell networks whose coupling structure may lack global permutation symmetry but still forces invariant synchrony subspaces.Balanced colourings yield polydiagonals, and synchronised cells can be collapsed into quotient networks; many dynamical consequences remain open.
  • Coupled-cell and groupoid structures: A three-cell network can have no exact permutation symmetry yet possess an invariant subspace with cells 1 and 2 synchronised.Its quotient network identifies the synchronised cells, reducing the network to two cells.
  • Coupled-cell and groupoid structures: Invariant subspaces and related network structures do not by themselves guarantee attracting solutions.This limitation applies to structures that resemble symmetric networks without necessarily possessing their symmetries.

4 Coupled limit cycle oscillators

This section examines coupled limit-cycle oscillators, showing how coupling can produce synchrony, oscillator death, cluster states, and heteroclinic dynamics. It combines pulse-coupled and synaptically coupled integrate-and-fire models with tools including firing maps, spectral stability analysis, and the master stability function.

  • Weakly coupled oscillator theory reduces limit-cycle network dynamics to phase equations, but stronger coupling can alter amplitudes and approach the complexity of general nonlinear systems.Special progress remains possible for specific neuron models, including integrate-and-fire, piecewise-linear, caricature, and relaxation-oscillator models.
  • The master stability function determines synchronous-state stability in strongly coupled identical networks by evaluating the greatest Lyapunov exponent at scaled network eigenvalues.The synchronous state is stable when the master stability function is negative at α = σλ_l for every nontrivial eigenvalue λ_l.
  • For pulsatile coupling with a strictly increasing, concave voltage-phase relation, almost all initial conditions synchronize because absorbed oscillators remain synchronized and the return-map fixed point is unstable.The firing-map construction gives h(φ) = g(ϵ + f(1 − φ)) and return map R(φ) = h(h(φ)); the synchronization result applies to the stated class of f.
  • With delays, multiple cluster-state attractors can coexist; weak noise preserves inhibitory periodic cluster states but drives persistent switching among unstable Milnor attractors under excitatory coupling.These dynamics are related to heteroclinic switching, and a reset-function framework connects unstable-attractor networks continuously to heteroclinic cycles.
  • Weak-coupling stability for integrate-and-fire networks depends on ϵK′(0)[Re ν_p − γ] < 0, whereas strong coupling can produce Hopf bifurcations, oscillator death, or loss of frequency locking.For slow synapses, weak-coupling analysis may predict neutral stability when K′(0) = 0, while strong-coupling analysis can yield a different stability condition; asymmetric networks may undergo discrete Hopf bifurcations, whereas symmetric networks may lose frequency locking.

5 Reduction of limit cycle oscillators to phase-amplitude and phase models

The review develops phase and phase-amplitude reductions for attracting hyperbolic limit-cycle oscillators, then explains when these descriptions support analysis of forcing and complex dynamics. Phase-only models are convenient but can fail under strong perturbations, whereas phase-amplitude coordinates remain applicable finite distances from the cycle.

  • Phase reduction: An attracting hyperbolic periodic orbit supports a phase description on a topological circle, with phase conventions based on either period T or angular frequency 2π/T.The orbit has one zero Floquet exponent and all others with negative real parts.
  • Isochronal coordinates: Isochrons provide phase coordinates off the cycle, but their basin-wide foliation is usually available only numerically.Closed-form isochrons are rare because computing them requires knowledge of the limit cycle and its basin of attraction.
  • Isochronal coordinates: Fourier averages can recover isochrons as level sets for almost all observables when the first Fourier coefficient along the cycle is nonzero.For stable limit-cycle dynamics, nonzero Fourier averages occur only at frequencies ω_n = 2πn/T.
  • Phase-amplitude models: Phase-amplitude coordinates represent attraction or repulsion and shear near the cycle, extending modelling to finite-distance perturbations without assuming weak forcing.They can be used for driven systems in which perturbations push trajectories away from the limit cycle.
  • Shear-induced chaos: Shear and nonconstant kicks can stretch and fold trajectories during relaxation, producing strange attractors under suitable conditions.More shear, larger kick amplitude, and weaker attraction increase folding; kick geometry must vary transversely to isochrons.
  • Phase oscillator models: Phase-only models enable phase-response and locking analyses, but strong interactions can produce neural responses that phase alone cannot explain.Weak forcing yields dϑ/dt = 1 + ϵ⟨Q,g⟩, while pulsatile forcing of the Morris-Lecar model showed responses beyond a phase description.

6 Weakly coupled phase oscillator networks

Weakly coupled oscillator theory reduces interacting limit-cycle systems to phase dynamics, enabling analysis of synchronization, phase locking, and collective states. The review also emphasizes how interaction functions encode biological realism and where averaging theory is limited.

  • Network reduction: Weak coupling reduces limit-cycle oscillator networks to lower-dimensional phase dynamics through perturbation and averaging methods.The resulting variables are oscillator phases and, especially, their relative differences.
  • Interaction functions: The phase interaction function H is obtained from the oscillator’s infinitesimal phase response curve and the coupling waveform.For Fourier coefficients, H_n = ⟨Q_-n,G_n⟩.
  • Averaging limitations: Averaging guarantees closeness between unaveraged and averaged trajectories only for times O(ϵ^-1), except near hyperbolic fixed points associated with phase-locked states.This limits direct interpretation of averaged dynamics over arbitrarily long times.
  • Interaction functions: Biological realism in phase-network models typically resides in the phase interaction function, which can represent experimentally or numerically derived neural responses.The Kuramoto choice H(ψ) = sin(ψ) is a simpler special case.
  • Two-oscillator locking: For reciprocal inhibitory coupling, slow synapses stabilize synchrony, whereas sufficiently fast synapses stabilize the anti-synchronous state.Synchrony occurs when 1/α > 1, and anti-synchrony is stable as α →∞.

6.1 Phase, frequency and mode locking

The review distinguishes phase locking from frequency locking and describes how locking ratios characterize oscillator relationships. In globally coupled averaged networks, symmetry generically supports synchronous, splay, and cluster states.

  • Phase locking: Phase locking with ratio (n:m) requires the weighted phase difference nθ_j(t) − mθ_k(t) to remain bounded.The integers n and m are nonzero together and have no common factors.
  • Frequency locking: Frequency locking requires only that the corresponding weighted phase difference grow sublinearly with time.The long-time frequency ratio can exist without a bounded phase difference.
  • Phase and frequency locking: Phase locking implies frequency locking, but frequency locking does not necessarily imply phase locking.The distinction arises because a sublinearly growing phase difference need not remain bounded.
  • Symmetric network states: For globally coupled averaged networks, permutation and phase-shift symmetry generically supports synchronous, splay, and diverse cluster states.These networks can have arbitrarily high dimension while being effectively determined by one computable interaction function H.

6.2 Dynamics of general networks of identical phase oscillators

For globally coupled identical phase oscillators, network structure and the phase interaction function determine the existence and stability of synchrony, asynchrony, clusters, and more complex attractors. Degeneracies and symmetry can produce invariant tori, heteroclinic attractors, and bifurcation behavior beyond generic local theory.

  • Degeneracy: Zero Fourier components can create degenerate dynamics, including m-dimensional invariant tori foliated by neutrally stable periodic orbits when N is a multiple of m.Small nonzero Fourier components or non-pairwise coupling remove this degeneracy, which is typically relevant even near a Hopf bifurcation.
  • Phase-locked states: Phase-locked states have constant relative phases, while their collective frequency is determined alongside N−1 relative phases that are independent of coupling strength.The solution is written θ_i(t) = ϕ_i + Ωt, with one reference oscillator fixing the phase origin.
  • Phase-locked states: Stability of a phase-locked state is determined by the nontrivial eigenvalues of a Jacobian that combines network anatomy with derivatives of the phase interaction function.The Jacobian has graph-Laplacian form with components −w_ijH′(ϕ_j−ϕ_i), and stability requires all nontrivial eigenvalues to have negative real parts.
  • Asynchrony and clustering: Flipping one Fourier coefficient of the interaction function destabilizes the asynchronous state toward the corresponding m-cluster, showing that small interaction-shape changes can alter emergent dynamics.The unmodified example has a stable splay state, whereas changing H_m to −H_m makes the mth mode unstable.
  • Bifurcations: Local bifurcations can have global consequences, including degenerate heteroclinic attractors for networks with at least four oscillators.Symmetry and torus topology facilitate connections that can produce heteroclinic structures at bifurcations.

6.3 Phase waves

Phase-wave analysis links oscillator chains and spatially continuous networks to travelling waves, stability regions, and turbulent states. Frequency gradients, anatomical kernels, coupling signs, and higher harmonics determine whether waves lock, destabilize, or develop phase turbulence.

  • Motivation: Weakly coupled oscillator chains model travelling neural activity relevant to central pattern generators, locomotion, and peristalsis.The lamprey spinal cord is presented as a biological example containing roughly 100 rhythm-generating segments.
  • Phase waves: a lattice model: A frequency gradient can generate a stable travelling wave, but sufficiently steep gradients prevent phase-locking and promote clusters oscillating at different frequencies.For the sinusoidal interaction function, the stable phase-locked solution corresponds to a travelling wave when the frequency-gradient condition is satisfied.
  • Phase waves: a lattice model: Constant-speed waves in identical chains require equal oscillator frequencies except potentially at the chain ends.These waves use equal phase differences ϕ_i = ϕ and a collective frequency determined by the boundary equations.
  • Continuum phase waves: In continuum models, travelling phase waves θ(x,t) = Ωt + βx are stable when every nonzero perturbation wavenumber has eigenvalue with negative real part.The β = 0 case is synchrony, and the zero mode represents constant phase shifts.
  • Continuum phase waves: For exponential anatomical coupling and sinusoidal phase interaction, synchrony is unstable at small propagation speeds, while the wave-stability region depends on β and v.The stability region is shown in the (β,v) plane for W(y) = exp(−|y|)/2 and ϵ > 0.
  • Phase turbulence: Mixing positive and negative coupling strengths can create an unstable band of wave numbers, while higher harmonics and excitation–inhibition mixtures can support turbulent states.The illustrated spectrum has an unstable band from zero to a maximum wave number, and the phase-turbulent state is represented using a complex field.

7 Heteroclinic attractors

Heteroclinic attractors extend coupled-oscillator dynamics beyond periodic states, using invariant subspaces to support robust connections among saddle equilibria or periodic orbits. In neural oscillator networks, they produce slow switching, clustering changes, winnerless competition, and sensitivity to noise and heterogeneity.

  • Structure: For sufficiently complex phase interactions, globally coupled oscillators can exhibit chaotic dynamics and attractors more complex than periodic states.The dynamical complexity depends on the phase interaction function H and oscillator number N.
  • Structure: Heteroclinic attractors arise when invariant subspaces support robust connections between dynamically simple saddle nodes.The nodes are typically equilibria or periodic orbits, while the invariant subspaces may be imposed by symmetry, clustering, coupling structure, or model assumptions.
  • Phase oscillator networks: For large N, an open parameter region supports heteroclinic cycles with noise-dependent slow switching between two macroscopic oscillator clusters.Under weak additive noise, the approximate transition period scales logarithmically with noise amplitude.
  • Phase oscillator networks: N ≥4 is required for attracting robust heteroclinic attractors of the specified all-to-all phase-oscillator form, with trajectories lingering near two-by-two cluster states for N=4.Connections involve temporary cluster breaking before transitions to another cluster state.
  • Phase oscillator networks: For N=5, 30 symmetry-related cluster states form a single large heteroclinic network that typical initial conditions attract.Noise-driven trajectories can switch among cluster states, including changes in which clusters are stable or unstable to splitting.
  • Winnerless competition: Heteroclinic attractors can model winnerless competition and stable heteroclinic channels, but their singular dynamics make long-term averages sensitive to noise and heterogeneity.Noise and heterogeneity produce finite average transition times and can support input-output responses.

8 Stochastic oscillator models

Stochastic phase reduction must account for both state-dependent noise coupling and amplitude effects in the underlying limit-cycle oscillator. The resulting framework derives stochastic phase equations and shows that noise correlation time and attraction timescale can change the reduced dynamics.

  • Phase-amplitude reduction: Noise perturbs both phase and amplitude, so reducing a noisy planar oscillator requires a phase-amplitude coordinate transformation.The transformation yields coupled stochastic equations for phase ϑ and amplitude coordinate ρ.
  • Phase-amplitude reduction: The full coordinate transformation is not prescribed, limiting direct interpretation of the formal stochastic equations for a specific model.A concrete model requires specifying the transformation from x to (ϑ, ρ).
  • Stochastic reduction: The reduced Itō equations include noise-induced drift terms involving derivatives of the phase and amplitude coupling functions.The Fokker–Planck equation is then constructed for the joint phase-amplitude probability distribution with periodic phase boundary conditions.
  • Stochastic reduction: Naively adding noise to a deterministic phase equation misses multiplication by the infinitesimal phase response curve and an additional amplitude-response term Y(ϑ).Equations (49) and (50) provide the stochastic phase descriptions for a weakly white-noise-driven limit cycle.
  • Consequences: The framework computes steady-state phase distributions, mean frequency, and phase diffusion from the stochastic phase description.Fourier expansions provide a route to approximate the stationary distribution and its moments for small noise.
  • Colored noise: Finite noise correlation time can interact with limit-cycle attraction timescale, producing results different from Gaussian white-noise reduction.The white-noise and rapidly attracting limits recover distinct simplified forms depending on α = λ/γ.

9 Low dimensional macroscropic dynamics and chimera states

The review examines reductions from microscopic oscillator networks to low-dimensional macroscopic dynamics and the emergence of chimera states. It emphasizes that continuum reductions can differ substantially from finite-network behavior, where chimeras may be transient or attractor-like depending on architecture.

  • Low-dimensional macroscopic dynamics: Microscopic-to-macroscopic reduction is difficult because neural mass models track mean activity rather than higher-order correlations.Such rate descriptions are expected to reduce microscopic dynamics only in special large-network limits, and the spike-to-rate link is often phenomenological.
  • Low-dimensional macroscopic dynamics: The Winfree model can support incoherence, frequency locking, and oscillator death, while the Ott–Antonsen ansatz reduces its population dynamics to lower-dimensional equations.The resulting planar system can be analysed using numerical bifurcation methods.
  • Chimera states: The Ott–Antonsen ansatz has also been useful for analysing chimera states, in which synchronised and incoherent oscillator subpopulations coexist.This provides a low-dimensional route for studying nontrivial collective states in oscillator populations.
  • Chimera states: Continuum non-local oscillator models can exhibit chimera solutions for parameter ranges near α = π/2 and selected initial conditions.Related discretisations and coupling kernels reproduce similar behaviour, with exact reductions available in some settings.
  • Chimera states: Finite oscillator approximations can make chimeras transient, with typical lifetimes that grow exponentially with network size before collapse to coherent or incoherent oscillation.In these systems, chimeras appear to behave like chaotic saddles, although other finite architectures can support chimera attractors.

10 Applications

The review applies mathematical frameworks to neural-network examples including connectivity, central pattern generators, and perceptual rivalry. These applications show how local dynamics, coupling, symmetry, and network structure shape collective patterns.

  • Functional and structural connectivity in neuroimaging: Wilson-Cowan node dynamics and weak-coupling analysis connect local bifurcations to global synchrony loss in neural-mass networks.The network uses N globally coupled identical nodes; for ε ≪ 1, weak-coupling theory describes its dynamics.
  • Central Pattern Generators: Central pattern generators model rhythmic activities such as locomotion, heartbeat, respiration, and digestion through spatiotemporal neural activity patterns.They also provide conceptual minimal circuits for more complex systems and robotic actuator control.
  • Central Pattern Generators: Symmetry arguments applied to CPG structure yield model-independent conclusions for gait patterns in animals with different numbers of legs.A cited example uses 2n oscillators for animals with n legs.
  • Central Pattern Generators: Three-cell burster motifs are used to classify emergent spatiotemporal patterns as coupling parameters vary.
  • Central Pattern Generators: Groupoid-based approaches extend CPG analysis to constrained connection structures not determined purely by network symmetries.They provide a context for studying patterns in lattices and relating synchrony properties to adjacency-matrix spectra.
  • Perceptual rivalry: Perceptual rivalry demonstrates autonomous temporal switching between competing percepts despite temporally static input.Models include competition, bifurcation, neural-circuit, network-structure, and heteroclinic-attractor approaches.

11 Discussion

The discussion places weakly coupled oscillator theory within a broader framework while identifying important omissions and directions for stronger, more heterogeneous network theories. It emphasizes amplitude dynamics, symmetry, and computational group-theoretic tools as future ingredients.

  • Scope limitations: The review omits subcellular behavior and does not cover single-unit or population forcing in depth.
  • Scope limitations: The review gives limited treatment to systems with three or more interacting frequencies, envelope locking, synchronised bursting, and chaotic networks.
  • Scope limitations: Neural sparsity, dendritic structure, glial coupling, and neuromodulation complicate treating neurons as oscillators and influence emergent rhythms.
  • Discussion: The review broadens weakly coupled oscillator theory into a wider framework for addressing neuroscience-driven challenges in coupled-oscillator dynamics.
  • Future directions: Heterogeneity is largely side-stepped because weak coupling requires nearby oscillator frequencies, while heterogeneous phase-response curves may have limited relevance to real networks.
  • Future directions: A proposed future theory targets strongly coupled heterogeneous networks using phase-amplitude interaction functions and symmetry-based equivariant bifurcation analysis.The proposal emphasizes clustered phase-amplitude chaos and multiple attractors as generic amplitude-related phenomena.

Appendix A

Appendix A supplies the Hodgkin-Huxley model’s standard rate functions and biophysical parameter values. These definitions specify voltage-dependent gating rates and the units and constants used by the model.

  • Hodgkin-Huxley parameters: The Hodgkin-Huxley description is completed with voltage-dependent transition-rate functions α_m, α_h, α_n, β_m, β_h, and β_n.
  • Hodgkin-Huxley parameters: The model specifies membrane capacitance, leak, potassium, and sodium conductances, reversal potentials, and measurement units.The listed units include mV for potentials, ms for times, and µA per cm2 for currents.

Glossary

The glossary defines abbreviations used throughout the review for differential equations, oscillator models, bifurcations, heteroclinic dynamics, phase response, and stability analysis.

  • Phase and stability analysis: iPRC denotes the infinitesimal phase response curve, while MSF denotes the master stability function.
  • Differential equations: ODE and PDE denote ordinary and partial differential equations, respectively.
  • Neural oscillator models: QIF, IF, FHN, HH, LIF, and ML name neural oscillator models: quadratic integrate-and-fire, integrate-and-fire, FitzHugh-Nagumo, Hodgkin-Huxley, leaky integrate-and-fire, and Morris-Lecar.
  • Equation types: SDE and DDE denote stochastic and delay differential equations.
  • Heteroclinic dynamics: SHC and WLC denote stable heteroclinic channel and winnerless competition.
  • Bifurcations: SNIC denotes saddle-node on an invariant circle, a bifurcation type.
Loading 1506.05828v1…