Time-Delayed Network Models

• Time-delayed network models are dynamical systems where current states depend on past interactions through discrete, distributed, or virtual delays.
• They employ methodologies like delay differential equations, convolution kernels, and state augmentation to capture memory effects and induce oscillatory dynamics.
• These models have practical implications in neuroscience, control systems, and deep learning by enabling stability analysis, synchronization, and adaptive delay mechanisms.

Searching arXiv for the specified paper and closely related time-delayed network model literature to ground the article in cited work. A time-delayed network model is a networked dynamical system in which the present evolution depends on earlier states, earlier outputs, or earlier interactions rather than only on instantaneous variables. In the literature, this dependence appears in several mathematically distinct forms: edge-dependent discrete delays in delay differential equations, distributed delays represented by convolution kernels, age-structured population equations in which delayed activity modulates a renewal boundary condition, and state-augmented formulations in which delay is converted into auxiliary transport or shift-register dynamics. A particularly developed instance is the homogeneous fully connected time-elapsed neuron network, where the elapsed time since the last discharge is the structuring variable and the network input is a distributed delay of past discharges (Weng, 2015). Related formulations show that delays may encode cycle-level timing invariants in directed graphs, may be unfolded into “virtual nodes” of a single delayed system, or may themselves become adaptive variables through plasticity (Lücken et al., 2012, Stelzer et al., 2020, Ruschel et al., 22 May 2026).

1. Conceptual scope and principal mathematical forms

In delayed network theory, the delay need not be tied to a single formalism. One standard form is a directed network of subsystems with edge-specific discrete delays,

$\dot{\boldsymbol{x}_{j}(t)=f_{j}\left(\boldsymbol{x}_{j}(t),\left(\boldsymbol{x}_{k}\left(t-\tau_{jk}\right)\right)_{k\in P_{j}\right),$

where $\tau_{jk}\ge 0$ is the propagation delay from node $k$ to node $j$. In that setting, the delays are attached to edges of a graph, and the network is a system of delay differential equations with heterogeneous discrete delays (Lücken et al., 2012).

A second form is the distributed-delay formulation, in which past network output enters through convolution with a probability kernel rather than through a single retarded argument. In the time-elapsed neuron model, the delayed mean-field activity is

$m(t):=\int_0^\infty p(t-y)\,b(dy),$

so the current firing propensity depends on a filtered history of collective discharge rather than on one discrete lag (Weng, 2015). This distinction is substantive: distributed delay preserves memory over an interval and is naturally compatible with transport and semigroup methods.

A third form replaces spatially distinct nodes by temporally sampled “virtual nodes.” A single delayed dynamical system,

$$\dot{x}(t) = f(x(t), z(t), \mathcal{M}_1(t)x(t-\tau_1),\ldots , \mathcal{M}_D(t)x(t-\tau_D)),$$

can be unfolded over a clock cycle $T$ into $N$ sampled states $x_n^\ell$, with delays determining allowed offsets and piecewise-constant masks $\mathcal{M}_d(t)$ determining coupling weights. In that construction, temporal delay is converted into effective network topology (Stelzer et al., 2020).

A fourth form treats delay as a slowly evolving network parameter. In adaptive axonal delay models, delays are given by

$\tau_{jk}\ge 0$0

with conduction velocities $\tau_{jk}\ge 0$1 obeying slow adaptive dynamics. Here the timing of interactions is itself plastic, and delay participates in attractor selection rather than merely perturbing a fixed network (Ruschel et al., 22 May 2026).

These formulations make clear that a common misconception is to identify a time-delayed network model exclusively with fixed edgewise lags. The literature instead supports a broader definition in which delay may be discrete, distributed, virtualized by time multiplexing, or embedded in adaptive internal state.

2. Age-structured mean-field delayed neuron networks

A canonical delayed network model in mathematical neuroscience is the time-elapsed neuron network, where the state variable is the density $\tau_{jk}\ge 0$2 of neurons with elapsed time $\tau_{jk}\ge 0$3 since last discharge. The evolution equation is

$\tau_{jk}\ge 0$4

with discharge activity

$\tau_{jk}\ge 0$5

and delayed network activity

$\tau_{jk}\ge 0$6

The transport term $\tau_{jk}\ge 0$7 expresses deterministic aging between spikes; the boundary condition $\tau_{jk}\ge 0$8 encodes reset to elapsed time $\tau_{jk}\ge 0$9 after firing; and the coupling parameter $k$0 measures mean-field connectivity strength (Weng, 2015).

This model is homogeneous and fully connected in a mean-field sense: neurons are not distinguished by graph position or class, but by elapsed-time state. Delay enters only through the interaction term $k$1, and therefore only through the firing rate $k$2. When $k$3, there is no delay and $k$4. When $k$5, the model has distributed delay. The same framework was analyzed in weak connectivity under general assumptions on the firing rate and delay distribution, with uniqueness of the steady state and nonlinear exponential stability near equilibrium (Mischler et al., 2015).

Stationary states are pairs $k$6 satisfying

$k$7

with $k$8. Writing

$k$9

the stationary density has explicit form

$j$0

and the fixed-point condition is

$j$1

An important structural point is that the delay does not alter the algebraic stationary equation itself, because convolution against a probability kernel leaves constants invariant. In this class of models, delay changes dynamics and linearized stability, not the equilibrium fixed-point equation (Weng, 2015).

The firing-rate assumptions split into a smooth monotone case,

$j$2

with bounded asymptotic rates $j$3, and a discontinuous threshold case

$j$4

Both versions admit global weak solutions with positivity and mass conservation; in the smooth case,

$j$5

and

$j$6

Thus the delayed time-elapsed network model is simultaneously a renewal equation, an age-structured transport PDE, and a delayed mean-field system.

3. Delay as an autonomous state variable

A recurrent theme in delayed network modeling is the replacement of explicit memory terms by autonomous state augmentation. In the delayed time-elapsed neuron model, the convolution

$j$7

prevents immediate closure of the linearized equation in the fluctuation $j$8. The remedy is to introduce an auxiliary transport variable $j$9 solving

$m(t):=\int_0^\infty p(t-y)\,b(dy),$0

so that

$m(t):=\int_0^\infty p(t-y)\,b(dy),$1

The delayed system thereby becomes an autonomous PDE on the product space

$m(t):=\int_0^\infty p(t-y)\,b(dy),$2

which is the setting used for semigroup and spectral analysis (Weng, 2015).

An analogous conversion appears in discrete-time spiking models. In a state-space approach to delays in spiking neural networks, each neuron is augmented by a delay state $m(t):=\int_0^\infty p(t-y)\,b(dy),$3 with shift dynamics

$m(t):=\int_0^\infty p(t-y)\,b(dy),$4

where $m(t):=\int_0^\infty p(t-y)\,b(dy),$5 is a lower shift matrix and $m(t):=\int_0^\infty p(t-y)\,b(dy),$6. With $m(t):=\int_0^\infty p(t-y)\,b(dy),$7, the delay state becomes

$m(t):=\int_0^\infty p(t-y)\,b(dy),$8

Delay is therefore represented as explicit finite history inside the neuron state, rather than as a separate synaptic or axonal transmission operator (Karilanova et al., 1 Dec 2025).

Hybrid oscillator–spiking architectures use yet another autonomous embedding. In the spiking-by-synchronization neural network $m(t):=\int_0^\infty p(t-y)\,b(dy),$9-Net, the top layer obeys vector Sakaguchi–Kuramoto dynamics,

$$\dot{x}(t) = f(x(t), z(t), \mathcal{M}_1(t)x(t-\tau_1),\ldots , \mathcal{M}_D(t)x(t-\tau_D)),$$0

while delayed rhythmic gating is defined by

$$\dot{x}(t) = f(x(t), z(t), \mathcal{M}_1(t)x(t-\tau_1),\ldots , \mathcal{M}_D(t)x(t-\tau_D)),$$1

The delay is not inserted directly into the oscillator coupling term; it enters as a delayed readout from the oscillator state into the spiking layer (Dan et al., 3 May 2026).

A further variant appears in the Equilibrated Recurrent Neural Network, where time-delayed self-feedback is represented implicitly through a per-time-step fixed-point equation

$$\dot{x}(t) = f(x(t), z(t), \mathcal{M}_1(t)x(t-\tau_1),\ldots , \mathcal{M}_D(t)x(t-\tau_D)),$$2

and approximated by an inexact Newton iteration

$$\dot{x}(t) = f(x(t), z(t), \mathcal{M}_1(t)x(t-\tau_1),\ldots , \mathcal{M}_D(t)x(t-\tau_D)),$$3

Here the delay is an inner equilibration process rather than a retarded argument in physical time (Zhang et al., 2019).

These constructions share a common principle: delay can often be reformulated as transport, memory register, delayed gate, or implicit recurrent state. A plausible implication is that the distinction between “delayed network” and “higher-dimensional autonomous network” is frequently representational rather than structural.

4. Stability, spectral structure, and delay-induced oscillation

Delayed network models exhibit sharply different long-time regimes depending on connectivity, delay size, and the form of the coupling. In the general time-elapsed neuron network, the principal stability theorem states that under either smooth monotone firing-rate assumptions or the step-rate assumptions, and with either no delay $$\dot{x}(t) = f(x(t), z(t), \mathcal{M}_1(t)x(t-\tau_1),\ldots , \mathcal{M}_D(t)x(t-\tau_D)),$$4 or smooth distributed delay, there exist $$\dot{x}(t) = f(x(t), z(t), \mathcal{M}_1(t)x(t-\tau_1),\ldots , \mathcal{M}_D(t)x(t-\tau_D)),$$5, $$\dot{x}(t) = f(x(t), z(t), \mathcal{M}_1(t)x(t-\tau_1),\ldots , \mathcal{M}_D(t)x(t-\tau_D)),$$6, $$\dot{x}(t) = f(x(t), z(t), \mathcal{M}_1(t)x(t-\tau_1),\ldots , \mathcal{M}_D(t)x(t-\tau_D)),$$7, $$\dot{x}(t) = f(x(t), z(t), \mathcal{M}_1(t)x(t-\tau_1),\ldots , \mathcal{M}_D(t)x(t-\tau_D)),$$8, and $$\dot{x}(t) = f(x(t), z(t), \mathcal{M}_1(t)x(t-\tau_1),\ldots , \mathcal{M}_D(t)x(t-\tau_D)),$$9 such that for $T$0 or $T$1,

$T$2

for sufficiently close unit-mass initial data. The delayed model therefore inherits local exponential relaxation in both weak and strong connectivity regimes after reformulation as an autonomous semigroup problem (Weng, 2015).

The weak-connectivity precursor established the same picture near $T$3: uniqueness of the steady state and nonlinear exponential stability in the delayed model under general smooth assumptions on the firing rate and delay kernel. In that perturbative regime, delay changes the linearized generator but not the uniqueness of the asynchronous stationary regime (Mischler et al., 2015).

The large-delay regime can be qualitatively different. For the strongly nonlinear age-structured time-elapsed model

$T$4

the no-delay inhibitory network converges to the unique stationary state and is interpreted as desynchronized, but large delays reduce the dynamics to iterates of the scalar fixed-point map

$T$5

If

$T$6

then $T$7 has a nontrivial period-two orbit $T$8, and the delayed PDE generates periodic oscillations with physical period approximately $T$9 through alternation between two pseudoequilibria (Perthame et al., 12 Mar 2025).

Delay may also be reduced rather than expanded. In a second-order reduction of time-delayed Kuramoto–Daido networks,

$N$0

the delayed first-order network is approximated, to second order in weak coupling and weak heterogeneity, by a delay-free second-order system in which delay appears as effective inertia and delay-induced triadic interactions. In the single-harmonic case the reduced equation is

$N$1

with $N$2. In that view, delay is not exhausted by a static phase lag; it also induces higher-dimensional intrinsic dynamics and multibody coupling (Smirnov et al., 11 Dec 2025).

At the level of synchronization transitions, delay can either preserve synchrony or produce desynchronization depending on topology and eigenstructure. In the regular-ring delayed Kuramoto model,

$N$3

the delay creates synchronized states, helical phase-locked states, random phase-locked states, and moving-turbulent chimera states. A synchrony-forbidden region appears roughly for $N$4, and narrow transition windows near its boundaries support chimera behavior (Ameli et al., 2 Feb 2025). In a related Wilson–Cowan network with homeostatic inhibitory plasticity and distributed delay kernels, each eigenvalue $N$5 of the connectivity matrix generates its own characteristic equation

$N$6

There the synchronous eigenvalue $N$7 controls synchronous Hopf onset, while the nontrivial eigenvalues control asynchronous Hopf onset. Bi-directional rings with real spectra support only asymptotically stable synchronous oscillations, whereas uni-directional rings with complex eigenvalues admit double Hopf points, stable asynchronous limit cycles, and torus-like solutions; increasing network size or mean time delay makes the intersections of synchronous and asynchronous Hopf curves more likely (Al-Darabsah et al., 2023).

A common misconception is that delay merely slows convergence. The literature instead shows that delay can preserve equilibrium stability, destroy it, replace it by periodic behavior, or be reducible to effective inertia and higher-order interactions, depending on the regime and the modeling scale.

5. Topology, equivalence classes, and virtual-node realization

Time-delayed network models are constrained not only by local node dynamics but also by graph topology and by the representation chosen for delay. In directed networks with edge-dependent discrete delays, a componentwise timeshift transformation

$N$8

changes the delays to

$N$9

The central invariants are the semicycle roundtrip sums

$x_n^\ell$0

and for a connected graph with $x_n^\ell$1 links and $x_n^\ell$2 nodes the number of generic essential delays is

$x_n^\ell$3

Thus many heterogeneous edge delays are dynamically redundant: the invariant content is carried by the cycle space rather than by each individual delay (Lücken et al., 2012).

A complementary viewpoint constructs networks directly from delayed dependencies in a time series. In the reduced autoregressive model

$x_n^\ell$4

each selected lag $x_n^\ell$5 becomes a directed edge $x_n^\ell$6, and the coefficient is converted into an edge “distance” through

$x_n^\ell$7

This produces a network with explicit time structure, in which shortest indirect paths can dominate direct delayed links (Nakamura et al., 2012).

The opposite construction eliminates many physical nodes altogether. A single delayed system with clock period $x_n^\ell$8, $x_n^\ell$9 subintervals of length $\mathcal{M}_d(t)$0, commensurate delays $\mathcal{M}_d(t)$1, and piecewise-constant modulation masks $\mathcal{M}_d(t)$2 can emulate multilayer feed-forward networks and recurrent networks of coupled discrete maps. Under

$\mathcal{M}_d(t)$3

all diagonals of an $\mathcal{M}_d(t)$4 weight matrix are present, so arbitrary recurrent adjacency matrices can be realized by choosing the modulation step heights appropriately (Stelzer et al., 2020). This shows that a time-delayed network model need not be spatially instantiated as many separate nodes.

Geometry provides a further topological layer. In a Hindmarsh–Rose network with fixed ring-lattice adjacency and diffusive membrane coupling,

$\mathcal{M}_d(t)$5

the delays are determined by Euclidean embedding: $\mathcal{M}_d(t)$6 A Metropolis–Hastings spatial evolution on node positions and currents then optimizes the power spectrum of selected output neurons, leading to the claim that a system with time delays corresponding to arrangement in physical space and with specific output properties has a specific spatial dimension that allows it to function properly (Łepek et al., 2018).

These results jointly undermine the idea that time-delayed network topology is simply “the graph plus one delay per edge.” Depending on the representation, the essential object may instead be a cycle-space normal form, a lag network inferred from data, a time-multiplexed virtual adjacency, or a geometry-induced delay matrix.

6. Contemporary extensions in neural computation, photonics, control, and learning

Recent work has extended time-delayed network models into several computational settings. In spiking neural networks, the state-space delay mechanism adds $\mathcal{M}_d(t)$7 memory variables per neuron and couples them into membrane dynamics through a learned or fixed matrix $\mathcal{M}_d(t)$8. The delayed LIF form is

$\mathcal{M}_d(t)$9

On the Spiking Heidelberg Digits dataset, the mechanism matches the performance of existing delay-based SNNs while remaining computationally efficient, and the strongest gains appear in small networks (Karilanova et al., 1 Dec 2025).

Delay can also operate as a coordination mechanism between scales. In $\tau_{jk}\ge 0$00-Net, the delayed oscillator gate

$\tau_{jk}\ge 0$01

modulates a discrete-time spiking network through

$\tau_{jk}\ge 0$02

with default $\tau_{jk}\ge 0$03, $\tau_{jk}\ge 0$04, and $\tau_{jk}\ge 0$05. Sensitivity analysis reported that $\tau_{jk}\ge 0$06 is worse, $\tau_{jk}\ge 0$07 is worse due to “over-frustration,” and $\tau_{jk}\ge 0$08 gives the best default trade-off (Dan et al., 3 May 2026).

In recurrent deep learning, time-delayed self-feedback has been recast as equilibrium seeking. The Equilibrated Recurrent Neural Network defines the hidden state implicitly by

$\tau_{jk}\ge 0$09

and solves the fixed-point condition approximately by an inexact Newton residual iteration with local linear convergence under

$\tau_{jk}\ge 0$10

Here delay appears as a short inner settling process rather than as a conventional retarded external interaction (Zhang et al., 2019).

Photonic hardware implements yet another minimal delayed architecture. A 4-channel time-delayed complex perceptron uses taps

$\tau_{jk}\ge 0$11

complex weights $\tau_{jk}\ge 0$12, and square-law readout

$\tau_{jk}\ge 0$13

Experimentally, the device was tested on 10 Gbps IMDD signals after propagation through up to 125 km optical fiber; the gain in transmitted signal equalization compensated the excess losses for links longer than 100 km (Staffoli et al., 2023).

Control-oriented models explicitly combine delayed history with latent linear dynamics. In the FRIB RFQ study, a long-short term memory-based Koopman network learns from 300 seconds of historical inputs sampled every 5 seconds and predicts the next 300 seconds of frequency detuning. The lifted state concatenates recent detuning history with a latent observable $\tau_{jk}\ge 0$14, and the predictor is embedded into model predictive control with Newton–Raphson optimization, reducing control time by half compared to a PID controller (Wan et al., 2024).

Delayed information is also a network constraint in distributed learning. In wireless online federated learning with over-the-air aggregation, both local loss information and channel state information are delayed by one round, and each device is subject to a time-varying power constraint

$\tau_{jk}\ge 0$15

COMUDO introduces a lower-and-upper-bounded virtual queue to handle the delayed information and hard constraint control, with closed-form local model updates and regret bounds (Wang et al., 10 Jan 2025).

Finally, adaptive axonal delay models place delay plasticity itself at the center of network organization. In delay-coupled phase oscillators with

$\tau_{jk}\ge 0$16

activity-dependent myelination changes conduction speeds and therefore changes the delays. On brain connectivity data and fully coupled ring networks, this mechanism yields collective frequency selection, chimera selection, and explosive network relaxation oscillations (Ruschel et al., 22 May 2026).

Across these extensions, the same structural lesson recurs: delay is not confined to passive latency. It may be a dynamical state, a gating signal, a hardware tap line, a control horizon variable, or a plastic network degree of freedom.