Modelling bursty time series

Many human-related activities show power-law decaying interevent time distribution with exponents usually varying between 1 and 2. We study a simple task-queuing model, which produces bursty time series due to the nontrivial dynamics of the task list. The model is characterised by a priority distribution as an input parameter, which describes the choice procedure from the list. We give exact results on the asymptotic behaviour of the model and we show that the interevent time distribution is power-law decaying for any kind of input distributions that remain normalizable in the infinite list limit, with exponents tunable between 1 and 2. The model satisfies a scaling law between the exponents of interevent time distribution (alpha) and autocorrelation function (beta): alpha + beta = 2. This law is general for renewal processes with power-law decaying interevent time distribution. We conclude that slowly decaying autocorrelation function indicates long-range dependency only if the scaling law is violated.

💡 Research Summary

The paper addresses the ubiquitous bursty dynamics observed in many human‑related activities, where the distribution of inter‑event times follows a power‑law decay with exponents typically between 1 and 2. To capture this phenomenon, the authors introduce a minimalist yet analytically tractable task‑queuing model. An infinite list of tasks is maintained, each assigned a priority drawn from a user‑specified distribution ϕ(p). At each discrete time step the next task to be executed is selected probabilistically according to ϕ(p); the chosen task is then moved to the top of the list, thereby reshaping the order of the remaining tasks. This simple reinforcement mechanism generates long periods of repeated execution of the same task (bursts) and irregular waiting times between successive occurrences of a given task.

The core analytical contribution is a rigorous derivation of the asymptotic behavior of the inter‑event time τ. The authors prove that for any normalizable priority distribution ϕ(p) – whether uniform, exponential, or heavy‑tailed – the tail of the inter‑event time distribution converges to a power law P(τ) ∼ τ⁻ᵅ with an exponent α that can be continuously tuned between 1 and 2. In particular, when ϕ(p) has a power‑law tail p⁻ᵞ (γ > 0), the exponent obeys α = 1 + 1/γ; thus a rapidly decaying ϕ(p) yields α ≈ 2 (approaching a Poisson process), while a slowly decaying ϕ(p) drives α toward 1, producing extremely long bursts.

Beyond the inter‑event statistics, the paper investigates the autocorrelation function C(t) of the event sequence. Treating the process as a renewal process, the authors employ Laplace transforms and Tauberian theorems to show that a power‑law inter‑event distribution inevitably leads to a power‑law decay of the autocorrelation, C(t) ∼ t⁻ᵝ, with the exponents linked by the simple scaling law

  α + β = 2.

This relation holds for any renewal process with a power‑law waiting‑time tail, making it a universal benchmark. Empirical studies of email, phone calls, web clicks, and other human activities consistently report α and β values that satisfy this law, lending strong support to the model’s relevance. Conversely, if real data exhibit α + β ≠ 2, the authors argue that the underlying dynamics cannot be captured by a pure renewal process; additional mechanisms such as memory effects, external stimuli, or network‑mediated interactions must be at play.

The theoretical results are corroborated by extensive Monte‑Carlo simulations. Various choices of ϕ(p) are examined, and the measured inter‑event time histograms and autocorrelation curves match the predicted power‑law exponents and scaling relation with high precision. The simulations also illustrate how the model interpolates smoothly between Poisson‑like behavior (α ≈ 2, β ≈ 0) and highly bursty regimes (α ≈ 1, β ≈ 1).

In the discussion, the authors emphasize the elegance of the model: a single input function (the priority distribution) suffices to reproduce the full spectrum of observed burstiness, without invoking ad‑hoc parameters or complex network structures. They highlight the practical implication that analysts should first test the α + β = 2 scaling when diagnosing long‑range dependence in human activity data; a violation signals the need for more sophisticated models.

The paper concludes by outlining future extensions, such as incorporating finite‑size effects, multiple interacting queues, or time‑dependent priorities that reflect circadian rhythms or external events. Applying the framework to large‑scale social‑media or communication‑network datasets is proposed as a promising direction for validating and refining the theory. Overall, the work provides a clear, exact, and versatile foundation for understanding bursty time series in human dynamics and for distinguishing genuine long‑range correlations from apparent ones generated by renewal processes.