Here is a reliable way to fool yourself with a graph that arrives wearing a lab coat:

  1. collect a skewed data set;
  2. sort the observations;
  3. draw a log-log plot;
  4. see something almost straight;
  5. announce a power law.

The little ceremony is seductive because power laws really are beautiful. If \(p(x) \propto x^{-\alpha}\), scale changes do not change the basic shape. Ten times larger is not a new regime; it is the same regime with fewer examples. That is why power laws show up in stories about city sizes, wealth, earthquakes, network degrees, file sizes, word frequencies, market moves, and scientific citations.1

But “heavy-tailed” and “power-law” are not synonyms. A lognormal distribution, a stretched exponential, a finite mixture, or a power law with a cutoff can all look persuasive over a narrow range. The visual slope is a clue. It is not a theory, and it is not a test.

Clauset, Shalizi, and Newman put the problem bluntly: empirical power-law detection is hard because the relevant evidence lives in the rare tail, and because one must identify the range over which the scaling claim is supposed to hold.2 Goldstein, Morris, and Yen showed with a simple experiment that linear fitting on log-log plots is biased and inaccurate compared with maximum likelihood estimation.3

So the useful question is not “is the plot straight?” It is:

Where does the tail begin?
How many observations are actually in that tail?
Is the fitted exponent stable when the cutoff moves?
Does a plausible non-power-law tail explain the same data?
What mechanism would make the exponent meaningful?

The lab below is a small machine for making that suspicion tactile.

Empirical CCDF Fitted Pareto tail Chosen cutoff Exponent instability

Deterministic synthetic experiment. The Pareto fit uses the continuous MLE above the selected cutoff. The lognormal comparison is a simple tail likelihood comparison, not a full Clauset-Shalizi-Newman p-value procedure.

The default generator is an actual Pareto distribution. Even there, the fitted exponent moves when the cutoff changes because the tail has few observations. Switch to Lognormal or Stretched exponential. You will often still see a respectable line on the log-log plot, especially with a finite sample. The diagnostics become less comforting: the fitted exponent drifts across cutoffs, the KS distance grows, or the lognormal tail likelihood becomes competitive.

This is the first lesson of heavy-tail work: finite samples are good actors. They can impersonate asymptotic laws without blushing.

My Audit Ritual

For continuous data above a chosen lower cutoff \(x_{\min}\), the maximum likelihood estimate for a Pareto density exponent is

\[\hat{\alpha} = 1 + n\left[\sum_{i=1}^{n}\log(x_i/x_{\min})\right]^{-1}.\]

This one line is already better than fitting a straight line through binned points. It uses the tail observations directly. It also forces the analyst to state \(x_{\min}\). That cutoff is not clerical. It defines the claim.

A good diagnostic report should include at least four objects:

  • the fitted exponent and number of observations above \(x_{\min}\);
  • a goodness-of-fit statistic, commonly a Kolmogorov-Smirnov distance between empirical and fitted tail CDFs;
  • a sensitivity curve showing how \(\hat{\alpha}\) changes as \(x_{\min}\) moves;
  • likelihood comparisons against alternatives such as lognormal, exponential, stretched exponential, or cutoff power-law tails.

Clauset, Shalizi, and Newman combine maximum-likelihood fitting with goodness-of-fit tests and likelihood-ratio comparisons, then show that some canonical empirical data sets are consistent with power-law behavior while others are not.2 Their companion page hosts implementations of the methods, which is a useful reminder that this is an algorithmic workflow, not a plotting style.4

When a Line Gets Expensive

Finance is where the distinction between “heavy” and “power” becomes expensive. Benoit Mandelbrot’s 1963 paper on speculative prices argued against the Gaussian model of price changes and proposed stable Paretian laws as a replacement for that setting.5 The enduring lesson is not that every return series is exactly stable Paretian. It is that tail assumptions are not harmless. They decide how often your risk system expects large moves.

If daily losses are Gaussian, a six-sigma event is practically mythological. If losses are heavy-tailed, it may be part of the business model. If the tail is a finite mixture of calm and crisis regimes, a single exponent can be a false sense of simplicity. If the tail has a cutoff, extrapolating a pure power law can overstate impossible extremes.

This is why the mechanism matters. Preferential attachment, multiplicative growth, self-organized criticality, and mixture processes can all create heavy-tailed observations, but they imply different interventions. A straight line does not tell you which machine made it.

My Power-Law Audit

I do not read a power-law claim as false. I read it as incomplete until it has survived a short audit:

show the raw CCDF, not only binned densities
declare xmin before interpreting alpha
fit by likelihood, not visual regression
report how many tail points support the claim
move xmin and show exponent stability
compare against lognormal and cutoff alternatives
explain the generative mechanism
say what decision changes if the law is true

The final line matters. If no decision changes, the power law may be ornamental. If a decision changes, the evidence must be stronger than a line.

Power laws are among the most interesting statistical objects we have because they connect measurement, mechanism, and scale. That is exactly why they deserve more care. The plot is an invitation. The law has to be earned.

Paper Trail

  1. M. E. J. Newman, “Power laws, Pareto distributions and Zipf’s law,” Contemporary Physics, 2005. arXiv. ↩

  2. Aaron Clauset, Cosma Rohilla Shalizi, and M. E. J. Newman, “Power-law distributions in empirical data,” SIAM Review, 2009. arXiv. ↩ ↩2

  3. Michel L. Goldstein, Steven A. Morris, and Gary G. Yen, “Problems with Fitting to the Power-Law Distribution,” European Physical Journal B, 2004. arXiv. ↩

  4. Aaron Clauset, “Power-law Distributions in Empirical Data” companion page, including implementations of fitting and goodness-of-fit methods. Project page. ↩

  5. Benoit Mandelbrot, “The Variation of Certain Speculative Prices,” The Journal of Business, 1963. RePEc. ↩