Dynamical Neural Field Theory

• Dynamical Neural Fields are spatially extended models that describe neuronal population dynamics through integro-differential equations.
• They capture complex phenomena such as wave propagation, oscillatory patterns, and bifurcations, unifying microscopic and mesoscopic scales.
• Field-theoretic frameworks offer robust methodologies to analyze connectivity, plasticity, and emergent cortical assemblies.

Dynamical neural fields are spatially extended models of neuronal populations in which the temporal evolution of neural activity and, often, synaptic connectivity are modeled by integro-differential or field-theoretic equations. These frameworks generalize classic mean-field rate models to continuous cortical domains, inherently capturing wave propagation, spatiotemporal pattern formation, stimulus integration, plasticity, and the emergence of mesoscopic assemblies. Dynamical neural field theory now constitutes a rigorous multiscale formalism, ranging from microscopic spiking models through continuum field equations to effective actions for emergent collective states and connectivity dynamics.

1. Foundations: Field-Theoretic Formulations and Neural Field Equations

The modern dynamical neural field framework derives macroscopic equations for population activity from underlying stochastic neuron models, often by a path integral or action-functional approach. A paradigmatic construction starts from stochastic quadratic integrate-and-fire neurons with noise; their membrane potential and firing rate dynamics are recast into a path integral, yielding an Onsager–Machlup action in terms of firing rate and membrane state variables. After suitable mean-field, continuum, or saddle-point approximations—such as projecting the microscopic activity field $\psi(\theta,Z,w)$ onto a reduced set $\varphi(\theta,Z)$—the theory produces a mesoscopic neural field equation of the form

$$\tau \partial_t u(x,t) = -u(x,t) + \int w(x,x')f[u(x',t)]\,dx' + I(x,t),$$

where $w(x,x')$ encodes spatial connectivity, $f$ is the single-neuron transfer function (such as an $f$–$I$ curve), $I(x,t)$ is external input, and $\tau$ is the effective time constant derived from noise or adaptation (Gosselin et al., 2020). This hierarchy of reductions allows consistent derivation of neural field equations from first principles, ensures faithful retention of fluctuation effects, and bridges the micro- and mesoscopic levels.

2. Structure and Dynamics of Neural Fields

The core dynamical elements of neural fields are: (1) local or nonlocal population activity fields $\phi(x,t)$; (2) possibly plastic, dynamically evolving connectivity fields $\varphi(\theta,Z)$0; and (3) response kernels or transfer functions governing spatial and temporal propagation (Gosselin et al., 2023, Gosselin et al., 28 Oct 2025). The field-theoretic models feature actions or Lagrangians that include quadratic kinetic terms, spatial coupling (as Laplacians or convolution kernels), nonlinear activation potentials, and explicit noise/dissipation contributions.

Stationary states and stability: Setting $\varphi(\theta,Z)$1 in the typical field equation produces a spatial integral equation

$\varphi(\theta,Z)$2

where $\varphi(\theta,Z)$3 is the coupling kernel and $\varphi(\theta,Z)$4 is the transfer function; this governs the fixed points and background patterns. Linearization about steady states yields dispersion relations generalizing the Klein–Gordon or wave equation, allowing explicit analysis of the stability and eigenmode spectra of neural patterns (Gosselin et al., 2020, Cooray et al., 2023).

Wave propagation and pattern formation: Traveling and standing wave solutions, as well as oscillatory or stationary patterns (such as bumps or stripes), arise naturally within neural field equations with either local or nonlocal coupling. For example, a damped wave equation for firing rate

$\varphi(\theta,Z)$5

supports both oscillatory and wave-like activity, with dispersion relation dictated by kernel coefficients and biophysical time constants (Gosselin et al., 2020).

3. Temporal Kernels, Delay, and Bifurcations

Temporal memory and signal delay in neural fields are introduced through convolutions with temporal kernels or explicit transmission delays. For instance, an exponential temporal kernel $\varphi(\theta,Z)$6 with distance-dependent transmission delay $\varphi(\theta,Z)$7 leads to equations of the form

$\varphi(\theta,Z)$8

where bifurcation analysis reveals the absence of static Turing bifurcations and the existence of Hopf and Turing–Hopf bifurcations for oscillatory and traveling wave states (Shamsara et al., 2019).

Delayed coupling in ring and higher-dimensional architectures: In spatially periodic domains (rings, sheets), delay-coupled neural fields produce quantized families of traveling “bump” solutions, with propagation velocities and stability determined by delay, kernel form, and nonlinear threshold. Interface dynamics reductions yield low-dimensional DDEs for bump positions and lead to discrete attractor dynamics (Parks et al., 27 Jan 2026, Spek et al., 2020).

4. Connectivity Fields and Plasticity

Neural field theory has been extended to treat the connectivity as a dynamic field (“connectivity field” $\varphi(\theta,Z)$9), itself subject to slow dynamics and plasticity rules. The resulting two-field or multilevel models possess actions

$$\tau \partial_t u(x,t) = -u(x,t) + \int w(x,x')f[u(x',t)]\,dx' + I(x,t),$$0

and define coupled evolution equations for both activity and connectivity (Gosselin et al., 2023, Cooray et al., 2024, Gosselin et al., 2023). The dynamics of $$\tau \partial_t u(x,t) = -u(x,t) + \int w(x,x')f[u(x',t)]\,dx' + I(x,t),$$1 can encode Hebbian or non-Hebbian plasticity, implement emergent assemblies via interference and bound states, and track evolving connection topologies.

Plasticity mechanisms: In the weak-coupling limit, field-theoretic constructions recover both linear and nonlinear Hebbian learning rules, such as

$$\tau \partial_t u(x,t) = -u(x,t) + \int w(x,x')f[u(x',t)]\,dx' + I(x,t),$$2

where $$\tau \partial_t u(x,t) = -u(x,t) + \int w(x,x')f[u(x',t)]\,dx' + I(x,t),$$3 denote presynaptic and postsynaptic activities (Cooray et al., 2024). Multi-layer systems with SU(2) symmetry admit cross-layer gains and oscillatory rotations in activity space.

5. Multiscale and Hierarchical Extensions

Recent field-theoretic work develops frameworks to describe the emergence, interaction, and stability of composite collective states—“assemblies”—from underlying micro- and mesoscopic connectivity and activity fields. This synthesis encompasses:

• Hierarchies of activity and connectivity variables, e.g., $$\tau \partial_t u(x,t) = -u(x,t) + \int w(x,x')f[u(x',t)]\,dx' + I(x,t),$$4 (activity), $$\tau \partial_t u(x,t) = -u(x,t) + \int w(x,x')f[u(x',t)]\,dx' + I(x,t),$$5 (connectivity), and emergent fields for collective states $$\tau \partial_t u(x,t) = -u(x,t) + \int w(x,x')f[u(x',t)]\,dx' + I(x,t),$$6, indexed by activation class and spatial support (Gosselin et al., 28 Oct 2025).
• Stability analyses of assemblies via effective actions and eigenvalue problems, generalizing to the classification of transitions (binding/unbinding, fragmentation, modulation of oscillation modes).
• Use of Čech cohomology to encode discrete class labels for assembly interactions and substructure assignment, yielding a sheaf-like state-space structure over the cortical domain.

Phase-space portrait and cascading transitions: The formalism supports cascade diagrams describing formation or dissolution of hierarchical structures and predicts attractors, saddles, and transition paths among functional states (Gosselin et al., 28 Oct 2025, Gosselin et al., 2023).

6. Applications: Physical and Cognitive Modeling

Dynamical neural fields model a range of phenomena beyond classical population activity patterns. Recent advances include:

• World model construction in physical prediction tasks: Neural fields with learnable lateral kernels and motor-gated channels serve as world models in visuomotor settings, outperforming latent-variable alternatives in sim-to-real transfer tasks and demonstrating unbiased, topology-preserving propagation of activity (“no teleportation”) (Nunley, 21 Feb 2026).
• Latent field discovery in interacting systems: Neural fields can serve as function approximators for latent force fields in systems of interacting agents or particles (e.g., N-body gravitation, traffic flow), disentangling global field effects from local equivariant interactions within graph-net architectures (Kofinas et al., 2023).
• Wavefronts and front-propagation thresholds: Interface dynamics approaches provide explicit threshold criteria for wave initiation in Heaviside neural fields, revealing all-or-nothing transitions and the dependence of traveling front initiation on initial activation region size and input duration (Faye et al., 2018).

7. Spectral Characterization and Canonical Cortical Field Theories

The macroscopic limit of coupled neural mass models—upon continuum reduction—generates coupled Klein–Gordon field equations across the cortical sheet. These universal canonical cortical field theories

$$\tau \partial_t u(x,t) = -u(x,t) + \int w(x,x')f[u(x',t)]\,dx' + I(x,t),$$7

are invariant under topological transformations of microscopic dynamics and admit analytic control of dispersion, spectrum, and the emergence of $$\tau \partial_t u(x,t) = -u(x,t) + \int w(x,x')f[u(x',t)]\,dx' + I(x,t),$$8 power laws observed in neural data (Cooray et al., 2023). Variation in parameters tunes resonance and propagation; anisotropic or non-Hermitian couplings modulate information coding and selective pattern transformation (Cooray et al., 2024).

---

Key References

First author(s)	Focus/Contribution	arXiv id
Gosselin, Lotz, Wambst	Stochastic field theory derivation of neural-field equations, multiscale	(Gosselin et al., 2020)
Cooray, Friston	Canonical Klein–Gordon cortical field theory, 1/f spectrum	(Cooray et al., 2023)
Gosselin, Lotz (series)	Two-field statistical theory (activity, plastic connectivity), assemblies	(Gosselin et al., 2023, Gosselin et al., 28 Oct 2025, Gosselin et al., 2023, Gosselin et al., 2023)
Amari, Wilson–Cowan et al.	Local (short-range) expansion, classical rate-based PDEs	(Gosselin et al., 2020)
Coombes; Ermentrout	Interface and front-propagation methods (traveling/thresh. dynamics)	(Faye et al., 2018)
Spek, Trofimchuk et al.	Delay, Hopf bifurcation, spectrum in 2D neural fields	(Spek et al., 2020)
Fung, Amari, Wu	Fluctuation–response unification, Goldstone modes, anticipation	(Fung et al., 2014)
Lotz, Gosselin	Lagrangian gauge theory for neural fields and connectivity	(Cooray et al., 2024)
Gök, Nonnemacher et al.	Memory kernels, bifurcation theory, spatiotemporal pattern analysis	(Shamsara et al., 2019)
Yahia, Kiani, Rutishauser	Delay-coupled ring attractors, stroboscopic motion illusions, DDE formulations	(Parks et al., 27 Jan 2026)
Medina, Zador et al.	Visual, sensorimotor, and world modeling with learnable neural fields	(Nunley, 21 Feb 2026, Kofinas et al., 2023)

Summary

Dynamical neural fields are a foundational and unifying concept spanning microscopic, mesoscopic, and macroscopic scales in neuroscience. The theory systematically derives spatially extended population models with nonlinear, noisy, and plastic dynamics; supports detailed analysis of stationary, oscillatory, and wave states; and incorporates modern field-theoretic, Lagrangian, and delay-based methodologies to explain emergent cognition, perception, and learning at the population level (Gosselin et al., 2020, Gosselin et al., 28 Oct 2025, Cooray et al., 2024, Gosselin et al., 2023, Cooray et al., 2023).