Characterizing and modeling citation dynamics

Citation distributions are crucial for the analysis and modeling of the activity of scientists. We investigated bibliometric data of papers published in journals of the American Physical Society, searching for the type of function which best describes the observed citation distributions. We used the goodness of fit with Kolmogorov-Smirnov statistics for three classes of functions: log-normal, simple power law and shifted power law. The shifted power law turns out to be the most reliable hypothesis for all citation networks we derived, which correspond to different time spans. We find that citation dynamics is characterized by bursts, usually occurring within a few years since publication of a paper, and the burst size spans several orders of magnitude. We also investigated the microscopic mechanisms for the evolution of citation networks, by proposing a linear preferential attachment with time dependent initial attractiveness. The model successfully reproduces the empirical citation distributions and accounts for the presence of citation bursts as well.

💡 Research Summary

The paper conducts a comprehensive statistical analysis of citation dynamics using the extensive bibliographic records of the American Physical Society (APS) journals spanning 1950–2008. The authors first examine which functional form best describes the empirical citation distribution P(k) (where k is the number of citations a paper receives). They compare three candidate families—log‑normal, simple power‑law, and shifted (or “offset”) power‑law—by fitting each model with maximum‑likelihood estimation and evaluating goodness‑of‑fit with the Kolmogorov‑Smirnov (KS) statistic. Synthetic data generated from each fitted model are used to compute p‑values that quantify how plausible each hypothesis is. The results show that the simple power‑law only fits the high‑citation tail (k > 20) and the log‑normal works reasonably well only for early years (pre‑1970). In contrast, the shifted power‑law P(k) ∝ (k + k₀)^{‑γ} provides a consistently high p‑value across all time windows, with the exponent γ decreasing from about 5.6 in the 1950s to 3.1 in 2008, indicating a gradual flattening of the citation distribution over time.

Next, the authors investigate “citation bursts,” defined as the relative increase Δk/k in a one‑year interval. The burst‑size distribution spans several orders of magnitude and is heavily skewed toward the early years after publication: more than 90 % of large bursts (Δk/k > 3) occur within the first four years. This rapid early growth contradicts the smooth, cumulative advantage implied by classic preferential‑attachment models.

To reconcile the observed burstiness with the overall shifted‑power‑law shape, the authors propose a microscopic growth model that augments linear preferential attachment with a time‑dependent initial attractiveness term. The attachment probability for a paper i at time t is taken as Π_i(t) ∝ k_i(t) + A_i e^{‑δt}, where A_i represents the paper’s intrinsic “visibility” at publication and decays exponentially with rate δ. Empirically, they estimate the average attractiveness ⟨A⟩ ≈ 7 using the cumulative kernel Π⁾(k) = ∑_{k′≤k} Π(k′), which exhibits a quadratic dependence on k plus a linear term consistent with the model. When the analysis is restricted to papers older than five or ten years, the linear term disappears, confirming that attractiveness only matters in the first few years.

Simulations of the proposed model reproduce (i) the shifted‑power‑law citation distribution, (ii) the heavy‑tailed burst‑size distribution and its early‑time concentration, and (iii) the observed temporal evolution of the exponent γ and average degree. Parameter calibration yields reasonable values for the initial attractiveness A₀ and decay rate δ (≈ 0.5–0.8 yr⁻¹).

The study concludes that citation dynamics cannot be captured by pure preferential attachment alone; a transient attractiveness component that drives early bursts is essential. Moreover, the shifted power‑law emerges as the most robust description of citation distributions across decades, outperforming both log‑normal and simple power‑law alternatives. These findings have implications for bibliometric modeling, research evaluation, and the design of policies that aim to predict or influence scientific impact.