An empirical regularity for rank–frequency relations stating that the frequency f(r) of the rth most common word is approximately proportional to 1/r^s (often with s ≈ 1), producing a heavy‑tailed distribution where few words are very frequent and many are rare.

Principle

Word usage in natural texts concentrates probability mass in a small set of high‑frequency items while generating a long tail of low‑frequency types; rank–frequency follows an approximately scale‑free power law, so relative frequency decreases predictably with rank within ranges where the fit holds.

Demonstration

Illustrative scenario: In a sample corpus, tokens are counted and ordered by decreasing frequency; plotting log(frequency) against log(rank) yields an approximately linear segment whose slope corresponds to −s, indicating an inverse power‑law relation. Recognition: common function words occupy top ranks with disproportionately high mass while most distinct types occur rarely.

Misapplication

Assuming an exact inverse proportion (s=1) for all corpora, interpreting deviations as errors, or applying raw Zipf predictions to small or highly specialised corpora without checking fit are incorrect. Treating Zipf’s empirical shape as a causal linguistic law rather than a statistical regularity is a category error.

Consequence

Zipf’s law informs expectations about vocabulary concentration, influences smoothing and backoff choices in language models, and explains why handling rare words is a persistent engineering challenge; causally, heavy tails imply that additional data continuously reveals new low‑frequency types, complicating exhaustive lexicon construction.

Reversal

Many corpora exhibit systematic deviations from a pure Zipf form: low‑rank flattening, high‑rank cutoffs, or better fit to variants (e.g., Zipf–Mandelbrot, double Pareto); in small samples or highly curated registers the power‑law segment may be absent. Different languages and tokenisation strategies also change the fitted exponent s.

Boundary

Clearly within: large token frequency lists aggregated from natural language samples where counts span many orders of magnitude. Boundary case: genre‑specific or very small corpora where rank–frequency shows irregularities. Clearly outside: measures of semantic importance, collocation strength, or context‑sensitive importance that are not raw token frequency by rank.

Semantic Tension

Zipf’s descriptive compactness competes with explanatory ambitions: statistical regularity is useful for modeling but does not itself adjudicate between competing linguistic accounts (communicative efficiency, least‑effort, sampling processes), generating tension between description and causal explanation.

Synthesis

Zipf’s law summarises a robust statistical pattern—concentration of mass in few tokens and a long tail of rare types—that shapes both theoretical accounts of language and practical decisions in corpus processing and language modelling, but it should be treated as an empirical regularity whose parameters and exact form depend on scope and preprocessing.