Delayed Vicsek Model: Dynamics & Delay Effects

• The Delayed Vicsek Model is a retarded-interaction extension of the classic Vicsek model where agents align based on past neighbor information rather than instantaneous values.
• It has been formulated in various settings (2D, 3D, vectorial-noise, and asynchronous updates) to analyze signal propagation, finite-size scaling, and phase transitions.
• Studies reveal that incorporating delay introduces a trade-off between increased stability and reduced responsiveness, leading to oscillatory behavior and altered collective motion.

Searching arXiv for papers on the delayed Vicsek model and closely related variants. The delayed Vicsek model is a retarded-interaction extension of the Vicsek model in which orientational alignment depends on past neighbor information rather than only on contemporaneous headings. In its standard explicit-delay form, particles move at constant speed and update their orientations from a vector sum containing their own current velocity and neighbors’ velocities evaluated at time $t-\tau$, where $\tau$ is a non-negative integer measured in time steps. Across the literature, this modification has been studied in 2D and 3D, with metric and topological interactions, and with both explicit discrete delays and minimal within-timestep reaction-time asymmetries. The central reported consequences are slower but more persistent collective response, oscillatory memory effects, a crossover of information transport from diffusive to ballistic-like, and nontrivial changes in ordering, finite-size scaling, and phase behavior (Geiß et al., 2022, Holubec et al., 2021, Packard et al., 15 May 2025, Horton et al., 7 Aug 2025).

1. Microscopic definition and principal formulations

In the standard metric Vicsek model, each particle $i$ has position $\mathbf{r}_i(t)$ and velocity $\mathbf{v}_i(t)$ of fixed magnitude $v_0$, with discrete-time updates consisting of a local orientational averaging step followed by advection. The delayed Vicsek model introduces retardation into the alignment field. In the 2D formulation analyzed for signal propagation and linear response, the update rule is

$$\mathbf{v}_i(t+1) = v_0\,\mathcal{R}_\eta\!\left[ \mathbf{v}_i(t) + \sum_{j\neq i} n_{ij}(t-\tau)\,\mathbf{v}_j(t-\tau) \right],$$

with

$$\mathbf{r}_i(t+1)=\mathbf{r}_i(t)+\mathbf{v}_i(t+1),$$

interaction radius $R=1$, and $\Delta t=1$. Here $\tau$0 is the metric adjacency matrix and $\tau$1 rotates the normalized vector by a random angle in $\tau$2. Orientations are often written as $\tau$3 (Geiß et al., 2022).

A closely related 3D vectorial-noise formulation uses

$\tau$4

with $\tau$5, fixed speed $\tau$6, $\tau$7, $\tau$8, and noise strength $\tau$9. In that study, the control parameter was the mean nearest-neighbor distance $i$0 rather than density directly (Holubec et al., 2021).

A high-speed 2D vectorial formulation likewise retains the delayed neighborhood $i$1,

$i$2

with $i$3, $i$4, and reduced delay $i$5 (Horton et al., 7 Aug 2025).

A distinct Vicsek-like delayed variant replaces an explicit scalar delay by an index-ordered asynchronous update within each time step. In that model, the deterministic torque on particle $i$6 depends on $i$7 for neighbors with $i$8 and on $i$9 for neighbors with $\mathbf{r}_i(t)$0, thereby implementing a minimal reaction-time asymmetry without introducing a separate delay parameter (Packard et al., 15 May 2025).

Formulation	Interaction structure	Reported emphasis
Explicit discrete delay in 2D	Metric neighbors at $\mathbf{r}_i(t)$1	Signal propagation, leader response
Explicit discrete delay in 3D	Metric neighbors at $\mathbf{r}_i(t)$2, vectorial noise	Finite-size and dynamic scaling
Index-ordered delayed Vicsek-like rule	Within-timestep delayed/advanced topological torque	Patterned phase in ordered flock

The global polarization is common across these formulations. In the 2D explicit-delay work it is

$\mathbf{r}_i(t)$3

Values $\mathbf{r}_i(t)$4 and $\mathbf{r}_i(t)$5 correspond respectively to strongly ordered and disordered states. Delay does not alter this formal definition, but it changes the dynamics of $\mathbf{r}_i(t)$6, including response lags, oscillations, and fluctuation structure (Geiß et al., 2022).

2. Order parameters, curvature response, and approximate orientation conservation

For response studies, a central observable is the discrete curvature of the global orientation,

$\mathbf{r}_i(t)$7

which in 2D is a scalar proportional to the sine of the mean turning angle per time step. For small changes, $\mathbf{r}_i(t)$8, where $\mathbf{r}_i(t)$9 is the particle-averaged heading increment (Geiß et al., 2022).

The standard leader protocol selects a fraction of particles as leaders after the system reaches a strongly ordered steady state. Leaders obey the same delayed alignment as followers but receive an additional deterministic rotation $\mathbf{v}_i(t)$0 per step, and the control parameter is

$\mathbf{v}_i(t)$1

The stationary response is quantified through the mean curvature $\mathbf{v}_i(t)$2 and its variance $\mathbf{v}_i(t)$3. In the dense, highly ordered regime, the reported linear-response law is

$\mathbf{v}_i(t)$4

so the mean response is linear in the applied bias and the susceptibility decreases approximately as $\mathbf{v}_i(t)$5. In the same regime, $\mathbf{v}_i(t)$6, and the delay dependence is well fitted by

$\mathbf{v}_i(t)$7

with $\mathbf{v}_i(t)$8; accordingly the variance first increases with $\mathbf{v}_i(t)$9 and then saturates (Geiß et al., 2022).

The interpretation advanced for this behavior is approximate orientation conservation in the ordered phase. For $v_0$0, high density and low noise make the total orientation approximately conserved in the absence of external bias, which yields $v_0$1 and hence $v_0$2. For $v_0$3, the injected orientation is spread over $v_0$4 steps, leading in the linear-response regime to

$v_0$5

consistent with the observed $v_0$6 (Geiß et al., 2022).

This approximate conservation principle has a limited domain of validity. Linear response breaks down at large bias $v_0$7 and large delay $v_0$8, where curvature-response curves bend and saturate, density inhomogeneities become important, and coherent turning gives way to fragmentation and complex trajectories. The product $v_0$9 is suggested as a control parameter for the onset of this breakdown, although no single sharp threshold is reported (Geiß et al., 2022).

3. Signal propagation, oscillatory memory, and dynamic scaling

The delayed Vicsek model supports several distinct modes of information transport. In pulse-response experiments, information about a leader’s turn is represented by the local turning rate

$$\mathbf{v}_i(t+1) = v_0\,\mathcal{R}_\eta\!\left[ \mathbf{v}_i(t) + \sum_{j\neq i} n_{ij}(t-\tau)\,\mathbf{v}_j(t-\tau) \right],$$0

For a static flock with $$\mathbf{v}_i(t+1) = v_0\,\mathcal{R}_\eta\!\left[ \mathbf{v}_i(t) + \sum_{j\neq i} n_{ij}(t-\tau)\,\mathbf{v}_j(t-\tau) \right],$$1, the neighbor network is fixed and the perturbation spreads conductively. At $$\mathbf{v}_i(t+1) = v_0\,\mathcal{R}_\eta\!\left[ \mathbf{v}_i(t) + \sum_{j\neq i} n_{ij}(t-\tau)\,\mathbf{v}_j(t-\tau) \right],$$2, the signal propagates in discrete shells of thickness $$\mathbf{v}_i(t+1) = v_0\,\mathcal{R}_\eta\!\left[ \mathbf{v}_i(t) + \sum_{j\neq i} n_{ij}(t-\tau)\,\mathbf{v}_j(t-\tau) \right],$$3, and $$\mathbf{v}_i(t+1) = v_0\,\mathcal{R}_\eta\!\left[ \mathbf{v}_i(t) + \sum_{j\neq i} n_{ij}(t-\tau)\,\mathbf{v}_j(t-\tau) \right],$$4 decays with distance because the perturbation is repeatedly averaged. For $$\mathbf{v}_i(t+1) = v_0\,\mathcal{R}_\eta\!\left[ \mathbf{v}_i(t) + \sum_{j\neq i} n_{ij}(t-\tau)\,\mathbf{v}_j(t-\tau) \right],$$5, signal onset is delayed, approximately $$\mathbf{v}_i(t+1) = v_0\,\mathcal{R}_\eta\!\left[ \mathbf{v}_i(t) + \sum_{j\neq i} n_{ij}(t-\tau)\,\mathbf{v}_j(t-\tau) \right],$$6, and $$\mathbf{v}_i(t+1) = v_0\,\mathcal{R}_\eta\!\left[ \mathbf{v}_i(t) + \sum_{j\neq i} n_{ij}(t-\tau)\,\mathbf{v}_j(t-\tau) \right],$$7 exhibits oscillations with period $$\mathbf{v}_i(t+1) = v_0\,\mathcal{R}_\eta\!\left[ \mathbf{v}_i(t) + \sum_{j\neq i} n_{ij}(t-\tau)\,\mathbf{v}_j(t-\tau) \right],$$8, including sign changes; these oscillations were identified as a generic delayed-system echo effect (Geiß et al., 2022).

For moving flocks with $$\mathbf{v}_i(t+1) = v_0\,\mathcal{R}_\eta\!\left[ \mathbf{v}_i(t) + \sum_{j\neq i} n_{ij}(t-\tau)\,\mathbf{v}_j(t-\tau) \right],$$9, the leader’s interaction disk moves through the flock, producing anisotropic propagation. Spatial discontinuities are smeared into kinks, and spreading is stronger in the direction of motion because agents ahead enter the interaction zone while agents behind leave it. Defining the signal arrival time $$\mathbf{r}_i(t+1)=\mathbf{r}_i(t)+\mathbf{v}_i(t+1),$$0 as the time at which $$\mathbf{r}_i(t+1)=\mathbf{r}_i(t)+\mathbf{v}_i(t+1),$$1 reaches its maximum and the effective propagation speed as $$\mathbf{r}_i(t+1)=\mathbf{r}_i(t)+\mathbf{v}_i(t+1),$$2, the measured dispersion relation $$\mathbf{r}_i(t+1)=\mathbf{r}_i(t)+\mathbf{v}_i(t+1),$$3 versus $$\mathbf{r}_i(t+1)=\mathbf{r}_i(t)+\mathbf{v}_i(t+1),$$4 shows a crossover from

$$\mathbf{r}_i(t+1)=\mathbf{r}_i(t)+\mathbf{v}_i(t+1),$$5

at small $$\mathbf{r}_i(t+1)=\mathbf{r}_i(t)+\mathbf{v}_i(t+1),$$6 and $$\mathbf{r}_i(t+1)=\mathbf{r}_i(t)+\mathbf{v}_i(t+1),$$7, to an approximately linear regime

$$\mathbf{r}_i(t+1)=\mathbf{r}_i(t)+\mathbf{v}_i(t+1),$$8

for larger $$\mathbf{r}_i(t+1)=\mathbf{r}_i(t)+\mathbf{v}_i(t+1),$$9 and-or larger $R=1$0. The slope $R=1$1 matches well the leader’s speed relative to the flock, $R=1$2. For large $R=1$3, the full dispersion relation oscillates with period $R=1$4: the front half-plane is almost purely ballistic, whereas the rear half-plane remains diffusive (Geiß et al., 2022).

A simplified delayed spin-wave theory captures the same tendency. Linearizing small angular fluctuations $R=1$5 about a uniform direction yields

$R=1$6

and Fourier transformation gives the transcendental dispersion relation

$R=1$7

For $R=1$8, $R=1$9, which is purely diffusive. For $\Delta t=1$0, the Lambert-$\Delta t=1$1 solution contains branches with nonzero real part, implying oscillatory, wave-like propagation superposed on damping (Geiß et al., 2022).

The 3D finite-size-scaling study reached an analogous conclusion from correlation dynamics. At the edge of disorder, the normalized correlation $\Delta t=1$2 acquires oscillations of period $\Delta t=1$3, the envelope decays more slowly as $\Delta t=1$4 increases, and the dynamic exponent in

$\Delta t=1$5

drops from $\Delta t=1$6 at $\Delta t=1$7 to $\Delta t=1$8 around $\Delta t=1$9. This was interpreted as a crossover from diffusive to ballistic dynamic scaling, and as an alternative explanation of empirical swarm traits previously attributed to inertia (Holubec et al., 2021).

4. Ordering, finite-size scaling, and phase behavior

The effect of delay on flocking thresholds is non-monotonic and depends on both geometry and operating regime. In the 3D vectorial-noise model with $\tau$00, $\tau$01, and $\tau$02 to $\tau$03, the control parameter is the mean nearest-neighbor distance $\tau$04. The susceptibility peak defines a pseudo-critical point $\tau$05, and finite-size scaling is performed using

$\tau$06

The extrapolated critical distance $\tau$07 is non-monotonic: it increases at small delay, reaches a maximum at $\tau$08, and then decreases below the no-delay value at larger delay. At the same time, both $\tau$09 and $\tau$10 increase and eventually saturate roughly when $\tau$11. The authors explicitly caution that these large values should not be naively interpreted as standard critical exponents of an equilibrium universality class (Holubec et al., 2021).

In a large but fixed 2D high-speed system with $\tau$12, $\tau$13, and densities $\tau$14, the explicit-delay model retains the same three phases as the standard Vicsek model: an ordered homogeneous flock, a liquid-gas coexistence state with traveling bands, and a disordered state. Delay nonetheless shifts both phase boundaries. The upper transition $\tau$15 from coexistence to disorder increases monotonically and saturates with delay, while the lower transition $\tau$16 from homogeneous order to coexistence is non-monotonic, increasing at short delay and decreasing strongly at long delay. For $\tau$17, the representative values reported are

$\tau$18

at $\tau$19, and

$\tau$20

over the same range, so the coexistence interval broadens with delay (Horton et al., 7 Aug 2025).

The same high-speed study characterizes the transition region using the polarization

$\tau$21

its variance

$\tau$22

and the Binder cumulant

$\tau$23

At the upper transition, the susceptibility peak decreases sharply as $\tau$24 increases from $\tau$25 to about $\tau$26 and then saturates, while the Binder cumulant displays pronounced negative dips consistent with bistability in finite systems (Horton et al., 7 Aug 2025).

Delay also modifies band morphology and kinetics. The maximum number of parallel bands in the coexistence regime increases with delay at both $\tau$27 and $\tau$28. Band-formation time, extracted from the density-correlation length $\tau$29 through

$\tau$30

decreases substantially from $\tau$31 to $\tau$32, then increases somewhat at $\tau$33. This acceleration is linked to transient swirling arcs whose typical radius grows with increasing delay and which serve as precursors of straight traveling bands (Horton et al., 7 Aug 2025).

5. Collective maneuvers, robustness, and the cost of delay

Persistent turning by a leader exposes a central trade-off of the delayed Vicsek model. In deterministic maneuver experiments, a fully polarized flock is initialized in a circular region and a single leader at the front executes a prescribed constant turn,

$\tau$34

with no noise. For small $\tau$35 and small $\tau$36, the flock initially turns coherently, but information spreading is weak and predominantly diffusive, so particles nearer the leader track it more closely than distant particles. This causes defocusing and possible break-up. The dependence on delay is non-monotonic: increasing $\tau$37 initially worsens cohesion, but beyond a threshold $\tau$38 the leader’s dwell time inside the flock becomes so short that its net effect is reduced and break-up becomes less likely (Geiß et al., 2022).

The same study reports that delay diminishes the ability of strongly aligned swarms to follow a leader while increasing their stability against random orientation fluctuations. With persistent noise, randomly turning particles act as additional leaders and make splitting events more frequent. For fixed $\tau$39 and fixed initial density $\tau$40, increasing $\tau$41 enlarges the effective flock size and makes splitting more likely because information to distant regions is strongly damped. For fixed $\tau$42, increasing $\tau$43 increases density, stabilizes the flock against noise, but reduces responsiveness to a single leader because individual perturbations carry less weight (Geiß et al., 2022).

Pulse perturbations reveal the same stability-responsiveness trade-off in a cleaner form. When a leader makes a single abrupt turn and then resumes ordinary delayed dynamics, the polarization response in the noiseless system exhibits a lag

$\tau$44

and oscillatory spikes with period $\tau$45, with increasing amplitude as $\tau$46 grows. Under persistent noise, the relaxation time $\tau$47 first increases with $\tau$48, then passes through a peak, then decreases and saturates. The reported interpretation is competition between weakened effective alignment and increased memory or resistance to fast perturbations. Thus delay can stabilize the ordered phase up to intermediate $\tau$49, but only at the cost of lower agility and poorer leader-following (Geiß et al., 2022).

A frequent misconception is that delay simply reproduces inertial effects. The literature gives a more differentiated picture. In scaling behavior and correlation dynamics, delayed alignment can indeed reproduce ballistic-like exponents and non-exponential correlations previously associated with inertial models (Holubec et al., 2021). By contrast, in maneuvering and leadership, delay has the opposite qualitative effect from the inertial spin model: it reduces coherent leader-following rather than improving it (Geiß et al., 2022).

6. Reaction-time symmetry breaking, patterned phases, and broader interpretation

Recent work broadened the notion of a delayed Vicsek model by introducing a minimal reaction-time-symmetry-broken rule. In this topological, index-ordered variant, the interaction is not an explicit $\tau$50 memory term but a deterministic within-timestep ordering: low-index particles react mainly to neighbors still at time $\tau$51, whereas high-index particles react mainly to neighbors already updated to $\tau$52. This produces an effectively distributed reaction time across the population and breaks invariance under exchange of particle labels in the torque law (Packard et al., 15 May 2025).

This minimal temporal asymmetry preserves the usual disorder-to-order transition but also generates a second transition deep in the polar flocking phase to a state with spatially patterned transverse velocities. In that state, the global flocking direction remains well defined, yet the system forms spatially separated high-density bands with opposite transverse velocities. The transition is tracked through the Binder cumulant of relative angles,

$\tau$53

which is approximately zero for Gaussian fluctuations and positive for the bimodal angle distributions associated with the patterned phase. The effect is strongly size dependent: for small systems the cumulant can become negative before turning positive, while for large systems it becomes clearly positive at low noise (Packard et al., 15 May 2025).

The same work links this phase to a slow-relaxing “index-order field.” Direct measurement of such a field was reported to be prohibitively difficult because Voronoi neighborhoods are sparse and index gradients are noisy, but an indirect susceptibility to random index permutations was measured. In the ordinary polar flock, shuffling indices has essentially no effect on the global polarization. In the patterned phase, the same permutation destroys the subtle correlation between index ordering and spatial structure, producing a pronounced response. A coarse-grained non-reciprocal torque field,

$\tau$54

is found to have a sinusoidal longitudinal profile out of phase with the transverse-velocity profile, suggesting that the pattern is sustained by spatially organized non-reciprocal torques rather than by standard equal-time alignment alone (Packard et al., 15 May 2025).

Taken together, these studies support a broad physical interpretation of delay as sensing-processing-actuation lag, communication latency, or heterogeneous reaction time. Short delays favor responsiveness but can amplify perturbations; moderate delays can stabilize collective motion by smoothing fast fluctuations and increasing memory; long delays align agents to outdated information and can lead to oscillations, fragmentation, or patterned states, depending on how retardation is implemented (Geiß et al., 2022, Packard et al., 15 May 2025). A plausible implication is that “the delayed Vicsek model” is best understood not as a single microscopic rule but as a family of retarded-alignment Vicsek-type dynamics whose macroscopic behavior depends sensitively on whether the delay enters as an explicit discrete memory, an effective transport scale $\tau$55, or a symmetry-breaking update protocol (Holubec et al., 2021).