← All Cracked Math projects
PM-I05NOT STARTEDINTERMEDIATEMATHPYTHONIMPACT ★★★☆☆

Is it really a power law? Clauset-grade fitting

Half the 'power laws' in papers fail a proper test. Build the test.

EST.12 h
REWARD+300 XP
AFTERClauset + 2

ML-ARENA BENCHMARKS

PROOF: FINDING
METRICINDUSTRYMINE
alpha-hat (fitted exponent), x_min-hat (fitted cutoff) and accept/reject verdicts vs the paper's 24 datasets
GOAL · alpha-hat within the paper's reported error on >= 20/24 datasets; the same plausibility verdicts on >= 20/24; one new dataset written up
Clauset, Shalizi & Newman 2009 (arXiv 0706.1062) and the `powerlaw` package · measured locally—
no results yet
YOUR RESULTS

Your recorded results appear here.

Brief

A power law says the probability of a value x falls off like x^(-alpha) above some cutoff x_min. Many papers claim one from a straight line on a log-log plot. Implement continuous and discrete power-law MLE (maximum-likelihood estimation of alpha), using the Hurwitz zeta function as the normalising constant in the discrete case. Choose x_min by minimising the KS distance (the largest gap between the data's and the model's cumulative distributions). Compute goodness-of-fit p-values with a semiparametric bootstrap: simulate many synthetic datasets from the fitted model and see how often they fit worse than the real data. Add Vuong likelihood-ratio tests (which say whether one distribution fits significantly better than another) against lognormal, exponential and stretched exponential. Reproduce the paper's datasets, then test one fresh claim, for example GitHub stars, PyPI downloads or BPE token frequencies.

VICTORY CONDITIONS

  1. alpha-hat within the paper's reported error on >= 20/24 datasets; the same plausibility verdicts on >= 20/24; one new dataset written up (alpha-hat (fitted exponent), x_min-hat (fitted cutoff) and accept/reject verdicts vs the paper's 24 datasets).
  2. A public repository with a README that shows your result and how to reproduce it.

Terms in plain words

Power law
A distribution whose probability falls off as a fixed power of x, giving a heavy tail with occasional very large values.
MLE (maximum-likelihood estimation)
Choosing the parameter value that makes the observed data most probable under the model.
KS distance
The largest vertical gap between two cumulative distribution functions, used here to judge how well the fit matches the data.
Semiparametric bootstrap
Generating fake datasets that follow the fitted power law above x_min and resample the real data below it, to get a fair p-value.
Vuong test
A likelihood-ratio test that compares two non-nested models and says which fits better, or that the data cannot tell.

Why it rates ★★★☆☆

Rigorous scepticism of heavy-tail claims is rare and valued, even though the method is a replication.