Timeliness Criticality in Complex Systems

The paper addresses how timing (timeliness) governs the functionality and value of schedule-based socio-technical and economic systems (e.g., transport, food, healthcare, emergency response, computing). Competitive incentives push operators to reduce temporal buffers to increase efficiency (e.g., tighter schedules, fewer standby resources), which risks propagating delays. The research question is whether there exists a critical point in such systems, controlled by the size of delay-mitigating buffers, beyond which delays propagate system-wide and accumulate without bound. The study proposes a minimal stylized model of delay propagation on temporal networks to demonstrate and characterize a novel form of critical behaviour (timeliness criticality), quantify fluctuations (avalanches), and connect to known frameworks in statistical physics.

The work situates itself within statistical physics of phase transitions and critical phenomena, relating delay accumulation to depinning/unbinding transitions and to the bounded Kardar–Parisi–Zhang (KPZ) equation describing driven interfaces against a hard wall. Analogies are drawn to mode-coupling theory of the glass transition and K-core percolation. Prior studies on cascading delays in transport, supply-chain disruptions, and inventory dynamics motivate the application to production networks (e.g., just-in-time practices), where buffers (inventories, redundancy) mitigate propagation but can be depleted. The paper notes depinning phenomena across domains (magnet domain walls, fracture, yielding) as conceptual parallels.

• Model formulation on temporal networks: For component i at schedule step t, delay τ_i(t) follows τ_i(t) = max_{t'<t} [∑_j A_ij(t,t') τ_j(t') − B_i(t)] + ε_i(t), where A is the temporal adjacency (whether j at t' affects i at t), B_i(t) is buffer, and ε_i(t) is random noise; [x] = max(0,x). The general framework captures transport and production networks (buffers correspond to dwell times or inventories).
• Simplifications: Assume Markovian dynamics (only t−1 matters), constant uniform buffer B across components, and consider the thermodynamic limit N→∞. This yields τ(t) = max_i [A(t−1) τ(t−1) − B] + ε(t).
• Network variants: (1) Mean-field (MF): each component can be delayed by K randomly chosen independent components, redrawn each time step (A(t−1) effectively yields the maximum over K). (2) Synthetic temporal networks (STN): additional constraints enforce in-degree and out-degree K at each step.
• Noise: iid across nodes and time; analytical solution derived for MF with exponential noise P(ε)=exp(−ε) (setting v=1 for units), but simulations also consider half-Gaussian noise for robustness.
• Analytical treatment (MF, exponential noise): Define Ψ_t(τ) as the complementary CDF of delays and Q_t(τ)=K Ψ_t(τ)[1−Ψ_t(τ)] as the density of the maximum of K delays. This leads to an integral equation for Ψ_t (Eq. 3) and a differential recursion Ψ't(τ)+Ψ_t(τ)=K [1−Ψ{t−1}(τ+B)]^{−1} Ψ_{t−1}(τ+B) (Eq. 4). Analyze large-t behaviour assuming exponential tails. Two regimes emerge: a stationary distribution with exponential tail for sufficiently large B, and a propagating front with velocity v when B is below a critical threshold.
• Critical buffer and order parameter: A non-trivial critical buffer B_c is determined (expressible via the Lambert W function; asymptotically B_c ~ log K as K→∞). The order parameter is the average delay growth per step v = r(t) − r(t−1), predicted to satisfy v = B − B_c for B > B_c and v = 0 for B < B_c (second-order transition in time).
• Tail parameter behaviour: The inverse-tail parameter α of the delay distribution exhibits different forms above and below B_c with a square-root singularity at the transition.
• Simulations: Direct simulations of the stylized model (Eq. 2) for various N and K validate MF predictions and assess finite-size effects. For N=10,000 and K=5, v vs. B is fitted to identify finite-N B_c(N); α(B) is measured from exponential tails. Noise variations (e.g., half-Gaussian) are tested; heavy-tailed noise is noted as a potential modifier but not explored here.
• Temporal correlations and avalanches (MF): Analyze autocorrelation of mean delay per node (Στ/N) near criticality; collapse curves using B-dependent scale factors to extract divergence exponent γ. Define delay avalanches as contiguous intervals where mean delay per node exceeds B; measure persistence times and sizes; obtain their distributions and scaling of means versus (B−B_c(N)).
• Real-world temporal networks: Use publicly available high-resolution contact datasets (school and workplace) mapped into event-based temporal networks. Agents move between events; event delay is max of entering agents’ delays plus exponential noise; single-agent events have no noise. Simulate delay dynamics with uniform buffer B; extract v(B) and identify critical points and singular behavior.

• Existence of timeliness criticality: The model exhibits a second-order phase transition controlled by the buffer size B. The order parameter v (mean delay growth per step) obeys v = B − B_c for B > B_c and v = 0 for B < B_c.
• Critical buffer scaling: B_c is given analytically in MF (involving Lambert W), with asymptotic scaling B_c ~ log K as K→∞.
• Tail behaviour: The delay distribution has exponential tails; the inverse-tail parameter α shows distinct forms above and below B_c with a square-root singularity as B → B_c^+ (α ∝ (B−B_c)^{−1/2}).
• Finite-size validation: Simulations for N=10,000, K=5 confirm qualitative predictions; finite-N effects shift apparent B_c(N) and α(N) away from infinite-N analytical values (analytical B_c ≈ 3.99431 and α_c ≈ 0.749644 for these parameters).
• Temporal correlations: The autocorrelation of mean delay per node collapses with a stretched-exponential form exp[−(t/τ)^β] with τ ≈ 5835.6 and β ≈ 0.823 (for N=10,000, K=5 MF). Scale factors fit yield B_c(N) ≈ 3.6755 and a divergence exponent γ ≈ 1.6936, indicating correlation time diverges as a power law as B → B_c(N)^+.
• Avalanche statistics (MF): Persistence-time distributions collapse under scaling with (B−B_c(N)); mean persistence time diverges as (B−B_c(N))^{−2}. Mean avalanche size diverges as (B−B_c(N))^{−1}. Avalanche size distributions show apparent power-law tails with exponent ~ −0.32 at the high end, extending further as B → B_c(N)^+.
• Robustness to noise form: Results are qualitatively similar with half-Gaussian noise; heavy-tailed noise is expected to modify the transition (not pursued here).
• Real-world temporal networks: Simulations on empirical contact networks (school, workplace) display critical transitions in v(B); the singularity in v near B_c is well fit by a power law (B−B_c)^{−γ} with γ ≈ 5/2. Disorder tends to increase the order of the transition. Critical points occur at lower B than in comparable STNs, likely due to sparsity.

The study demonstrates that efficiency-driven reductions in temporal buffers induce a critical phenomenon in schedule-based systems: below a critical buffer B_c, delays do not accumulate; above it, even small perturbations generate avalanches that span a broad range of sizes and durations, with diverging correlation times. This directly addresses the research question by identifying B as a control parameter and v as an order parameter exhibiting a second-order transition, quantifying critical exponents, and revealing universal features reminiscent of depinning, bounded KPZ interfaces, K-core percolation, and glassy dynamics. The findings underscore a fundamental trade-off between efficiency (small buffers) and resilience/timeliness (sufficient buffers) in socio-technical and production systems. The framework applies beyond transport to production networks, where inventories and supplier redundancy act as buffers; it also captures criticality on empirical, disordered temporal networks, highlighting the impact of sparsity and disorder on the transition’s nature. These insights inform operational decisions on buffer sizing to prevent systemic delay cascades.

The paper introduces a minimal temporal-network model that captures a novel phase transition—timeliness criticality—governed by delay-mitigating buffers. It identifies an analytical critical buffer B_c (scaling ~ log K), establishes the order parameter v with linear behaviour above B_c, characterizes temporal correlations and avalanche statistics with diverging scales near criticality, and validates these results via simulations, including on real-world temporal networks where strong singularities are observed. The framework bridges delay propagation with depinning and glassy dynamics, offering a unifying perspective for schedule-based systems. Future research should: (i) map composite real-world buffering strategies (overtime, inventories, supplier redundancy, procurement) onto model parameters; (ii) incorporate heterogeneity in delays and inventories and dynamic buffers; (iii) study heavy-tailed noise and richer temporal dependencies beyond Markovian assumptions; (iv) extend empirical validation to larger, more detailed datasets (e.g., production networks); and (v) apply the framework to distributed computing and biological input-output systems.

• Model simplifications: Markovian dynamics (only nearest temporal dependencies), uniform and time-invariant buffer B, and iid noise; real systems have memory, dynamic buffers, and heterogeneity in delays/inventories.
• Analytical tractability limited to MF with exponential noise; heavy-tailed noise and structured networks may alter critical behaviour.
• Finite-size effects: Simulated finite N and K deviate from infinite-size analytical predictions, requiring careful extrapolation.
• Empirical mapping: Real-world buffer measures aggregate diverse operational strategies, complicating one-to-one parameterization; empirical datasets used (contact networks) are proxies and may not reflect production network intricacies.
• Inventory dynamics and topology heterogeneity are not explicitly modelled, though important in real production networks; data sparsity limits comprehensive validation.