Beyond Zipf's Law: Equifinality and Mechanistic Inference from Scaling Laws

Abstract

Scaling laws summarize complex systems through low-dimensional regularities, but the same marginal law can arise from different stochastic dynamics. We examine this ambiguity for Zipf rank--frequency scaling. An i.i.d. finite-Zipf process, a persistent Markov chain, and canonical sample-space reduction (SSR) are constructed to have exactly the same stationary marginal, $p_j=(jH_V)^{-1}$, while a latent-scale mixture produces a similar marginal through aggregation. The first three models therefore hold the population rank distribution fixed while changing temporal organization. Markov dependence shifts the finite-sample distribution of fitted exponents; at moderate persistence, a block-adjusted effective sample size reproduces most of this shift. Excess lag-1 mutual information detects serial dependence relative to a shuffle null, whereas transition direction separates reversible Markov persistence from the directional contraction of SSR. Conditioning on latent scale reveals the aggregation mechanism. Fit-window, alphabet, and sequence-boundary analyses quantify sensitivity to the observation design. These examples separate the population marginal, the finite-sample behavior of a fitted exponent, and temporal structure. Matching a scaling law is therefore a compatibility condition, not a mechanism identifier: discrimination requires observables on which candidate processes make different predictions.