In 1999, Albert-László Barabási and Réka Albert argued in Science that the World Wide Web, film-actor collaborations, and the Western US power grid shared one shape: a few hubs with vastly more links than the rest, following a power law. The paper has been cited more than 30,000 times. Twenty years later, a test of nearly 1,000 networks found that shape rare.
A power law is a relationship in which one quantity varies as a fixed power of another. In a power-law distribution, the chance of very large values falls off slowly, leaving a heavy tail. Each time the size doubles, the relevant quantity changes by the same ratio, as long as the exponent holds.
A straight line convinced a whole field
When Barabási’s team sampled the web in 1998, a few pages drew far more links than others, and the plotted counts looked like a power law. As Quanta Magazine recounts, the Science paper then extended the claim to other networks. In his 2002 book Linked, Barabási wrote that “amazingly simple and far-reaching natural laws govern the structure and evolution of all the complex networks that surround us.”
The method behind many follow-up papers was a chart. “There must be a thousand papers,” Aaron Clauset told Quanta, “in which people plot the degree distribution, put a line through it and say it’s scale-free without really doing the careful statistical work.”
In 2019, Anna Broido and Clauset published “Scale-free networks are rare” in Nature Communications. They tested nearly 1,000 social, biological, technological, transportation, and information networks. Nearly half (49 percent) showed no evidence of scale-free structure. In 88 percent, a log-normal distribution fit as well as or better than a power law.
Barabási disagreed. He argued that real networks mix several processes, which nudge them away from a pure power law and make them fail strict tests. That debate continues. Both sides agree that many networks have heavy tails. The open question is whether the specific power-law claim holds.
Density and tail are different claims
The general relationship can be written as y = C × x^k, where C is a constant and k is an exponent. For a declining relationship, the exponent is negative. If x doubles, y changes by the factor 2^k.
A distribution needs more precise language. A probability density describes how probability is spread locally. A tail probability, P(X ≥ x), is the chance of a value at least as large as x. Their exponents differ. For an ideal continuous density proportional to x^-alpha above a threshold, the tail is proportional to x^-(alpha-1), with alpha greater than 1.
The threshold matters too. Data may follow a power law only above a certain size.
Why a log-log line can fool you
A power law and an exponential decay both decline, but their scaling rules differ.
| Relationship | Example form | What stays constant |
|---|---|---|
| Power-law decline | y = C × x^-2 | Doubling x divides y by four |
| Exponential decline | y = C × exp(-lambda × x) | Adding a fixed amount to x multiplies y by a fixed factor |
A power law forms a straight line on ideal log-log axes because log(y) is a constant plus k × log(x). An exponential forms a straight line on ideal semilog axes because log(y) is a constant minus lambda × x.
This is the trap many network papers fell into. Small samples, a restricted range, noisy tails, and flexible plotting choices can make other distributions look straight over a short interval. A log-normal, in particular, can mimic a power law across a wide range. Clauset, Shalizi, and Newman’s 2009 method replaces eyeballing with fitting, goodness-of-fit checks, and comparison with alternatives. It is the method Broido and Clauset applied.
Doubling the threshold, dividing by four
The following distribution is constructed for teaching. It is not fitted to wealth, cities, networks, or any market.
Let X have the hypothetical tail probability P(X ≥ x) = x^-2 for x at least 1.
| Threshold x | Probability of a value at least x | Change when threshold doubles |
|---|---|---|
| 1 | 1 | Starting point |
| 2 | 0.25 | Divided by 4 |
| 4 | 0.0625 | Divided by 4 |
| 8 | 0.015625 | Divided by 4 |
The repeated ratio is the scaling feature. This tail corresponds to a continuous density 2 × x^-3 above 1. The tail exponent is 2; the density exponent is 3.
A hypothetical exponential tail exp(-(x-1)) also equals 1 at x = 1, but doubling x does not give the same ratio at every threshold.
Run the evidence test before the label
Before calling real observations a power law, use this original review checklist:
- Define the variable and unit. “Popularity” is vague; a specified count over a specified period is assessable.
- Check how observations were collected. Missing small values, censored extremes, or changing units can distort the shape.
- Specify the range. Identify where the proposed relationship is meant to apply.
- Distinguish density, frequency, rank, and tail probability. A plot’s axis labels must say which quantity is shown.
- Fit and test a stated model. Keep the fitting method and assumptions visible.
- Compare alternatives. A plausible power law need not be the best explanation.
- State uncertainty. Sparse tail observations make confident claims difficult.
Most early network papers stopped after a chart, which skipped steps 5 to 7. Clauset’s companion data page offers real data sets for practicing the full test. If a chart already matches what you expected, read confirmation bias before you trust it.
When the power-law idea backfires
- You treat 80/20 as proof. “20 percent of customers bring 80 percent of revenue” is a concentration summary. It gives no exponent, threshold, or full shape.
- You predict a person from a population. A distribution describes many cases under stated conditions. It does not say which individual lands in the tail. Critical thinking keeps that inference separate.
- You drop heavy tails when the power law fails. A log-normal still produces rare, very large values. Losing the label does not remove the risk.
Mental models and systems thinking help frame the question. Neither can identify a distribution without data.
Questions
Does a power law mean a few things dominate?
A heavy-tailed distribution can include rare large values, but concentration alone does not prove a power law. Specify and test the relationship.
Is a straight line on a log-log plot enough?
No. It fits ideal power-law scaling, but other distributions can look straight over limited ranges. Fitting, uncertainty, and alternative models matter.
Are all power laws probability distributions?
No. A power law can describe a relationship between quantities without defining a probability distribution. Say which meaning you are using before reading the exponent.
Your 5-minute power-law check
The claim I saw (e.g., "X follows a power law"):
Variable and unit actually measured:
Range where it is claimed to hold:
Density, rank, or tail probability?
Was a fit tested, or just a line drawn?
Alternatives compared (log-normal, exponential):
What changes for me if it's only heavy-tailed:Next time a chart claims a power law, paste it into a prompt from Save Effort and run these seven questions.