A family of probability distributions whose tail decays as a negative power of the variable: for large x, P(X ≥ x) ∝ x^{-α} with α>0. Power-law distributions are heavy-tailed and scale-free, so extreme values occur with non-negligible probability and no single characteristic scale dominates. In bibliometrics they describe phenomena where event frequency (e.g., citations, downloads) falls off approximately as a power of rank or size.

Principle

When a quantity follows a power law, aggregate behavior is dominated by the tail: a small fraction of items accounts for a large fraction of the total, and conventional sample averages may be unrepresentative or unstable unless tail parameters imply finite moments.

Demonstration

Illustrative scenario: A corpus of article citations shows many papers with few citations and a few with very many. Plotting rank versus frequency on log–log axes yields an approximately straight line for high ranks; estimating the slope gives α. Recognition → identify linear region on log–log plot and estimate α → action: treat top-ranked items as dominant in resource allocation → consequence: indexing, caching and recommendation policies prioritize those few high-impact items.

Misapplication

Fitting a straight line by ordinary least squares to log-transformed counts without testing alternative heavy-tail models or assessing the valid range can falsely claim a power law; confusing a log-normal or stretched exponential with a power law is a common semantic error.

Consequence

Modeling data as power-law implies different statistical tools (robust estimators, tail-risk measures), influences design decisions (focus on popular items), and cautions against using the arithmetic mean as a summary; it also predicts fragility to shocks affecting top items.

Reversal

The idealized mathematical implications (e.g., infinite variance when 0<α≤2) assume infinite support; empirical systems with finite population, measurement cutoffs, or exponential truncation will deviate from pure power-law behavior and may exhibit finite moments and different scaling at extremes.

Boundary

Clearly within: the upper tail of citation counts that follows x^{-α} over several orders of magnitude. Boundary case: a dataset whose mid-range mimics a power law but whose tail curves downward (truncated). Clearly outside: Gaussian or exponential distributions where tails decay much faster and no scale-free property holds.

Semantic Tension

Representative statistics (mean/variance) versus tail-dominance: the desire to summarise data with simple moments conflicts with the reality that a few large observations determine totals under a power law.

Synthesis

Identifying a power law shifts attention from central tendency to tail behavior and to mechanisms that generate skew (e.g., preferential attachment); correct practice pairs tail-aware statistics with scrutiny of fitting range and alternative heavy-tail explanations.