Skip to content
C.W.K.
Stream
Lesson 05 of 06 · published

Power Laws: A Candidate Model for Extreme Inequality

~13 min · power-law, pareto, scale-free, inequality

Level 0Stats Novice
0 XP0/55 lessons0/14 achievements
0/100 XP to next level100 XP to go0% complete
"Some heavy-tailed data look approximately power-law over a range. Recognizing the candidate is useful; declaring it from a log-log line is not."

Where Power Laws Are Considered

City sizes, earthquake measures, wealth tails, book sales, online degree counts, and network flow sizes can show strong skew and heavy tails. Power-law behavior is one candidate for some ranges of some datasets. Lognormal, stretched-exponential, truncated power-law, and mixture models can look similar in finite samples.

Define the Exponent Before Using It

For a Pareto tail written as P(X > x) ∝ x⁻ᵅ, α is the tail index. Under this convention the mean is finite only when α>1 and the variance only when α>2. If a source defines the density exponent instead, the thresholds shift by one. Always check the parameterization.

Power-law tails decay more slowly than exponential or normal tails. Extreme values remain rare, but they are less rare than a thin-tailed model predicts.

The 80/20 Rule Is Not a Law of Nature

The Pareto principle is a mnemonic for concentration, not a universal empirical ratio. An exact Pareto distribution can be scale-free over its modeled range, but the share held by the top fraction depends on α and on cutoffs. Real systems often have lower and upper scales, regime changes, and finite-size limits.

Means, Variances, and Honest Summaries

When the theoretical mean or variance is infinite, familiar convergence and standard-error formulas can fail. In finite real datasets, the sample mean is still a computable number, but it may be highly unstable and dominated by a few observations. Even when moments are finite, convergence can be slow.

That does not make reporting a mean malpractice. Means answer questions about totals and expectations. Pair them with medians, quantiles, tail shares, uncertainty, and a fitted-tail analysis suited to the decision.

The Frame to Internalize

A power law is a statistical model, not a synonym for inequality and not a moral explanation for it. Test its range and alternatives. A distribution can be unequal without being power-law, and social choices can shape inequality regardless of the fitted curve.

Code

Top-X% share in a Pareto (power-law) distribution·python
import numpy as np
rng = np.random.default_rng(15)

# Pareto-distributed sample with shape parameter alpha=1.5.
N = 100_000
alpha = 1.5
x = (rng.pareto(alpha, size=N) + 1)   # Pareto draws + 1 (xmin = 1)

# Compute the share of total held by the top 1%, top 10%, top 20%, top 50%.
x_sorted = np.sort(x)[::-1]
total = x.sum()
for pct in (0.01, 0.10, 0.20, 0.50):
    k = int(N * pct)
    share = x_sorted[:k].sum() / total
    print(f"Top {pct*100:>4.1f}%  holds  {share*100:>5.1f}% of the total")

# A 1.5-Pareto shows roughly: top 1% holds ~25%, top 20% holds ~75%.
# The exact numbers depend on alpha. The shape is universal: a few hold most.
# A bell curve cannot produce this shape no matter what mean and std you choose.

External links

Exercise

Choose a strongly skewed dataset from your work. Plot the empirical complementary CDF, identify a defensible tail threshold, and compare a Pareto tail with at least one alternative such as lognormal or truncated power law. Report uncertainty rather than diagnosing from a straight-looking plot alone.
Hint
A large top-1% share suggests concentration, not proof of a power law. Model comparison and sensitivity to the threshold are essential.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue
💛 by Ttoriwarm

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.