Benford's Law
About 30.1% of numbers in real datasets start with 1. Fewer than 4.6% start with 9. If first digits were evenly spread, each would get 11.1%. The world often does not count that way — and when it does not, the gap is a forensic signal.
How it works
The probability that the first digit is d is log₁₀(1 + 1/d). That gives 30.1% for 1, 17.6% for 2, 12.5% for 3, falling to 4.6% for 9.
The intuition is multiplicative. A number must grow 100% to move from 1 to 2, but only 11.1% to move from 9 to 10. If values stretch across many orders of magnitude, they spend more logarithmic space with low leading digits. This is why the law survives unit changes: river lengths in miles or kilometers, market caps in dollars or yen, follow roughly the same curve. The pattern lives in ratios, not in raw quantities.
Frank Benford published the empirical version in 1938 with 20,229 observations — areas of rivers, atomic weights, street addresses, baseball statistics, numbers picked from newspaper front pages. Simon Newcomb had noticed the same shape in 1881 by observing that the early pages of logarithm books were more worn than the later ones.
Where it shows up
| First digit | Benford share |
|---|---|
| 1 | 30.1% |
| 2 | 17.6% |
| 3 | 12.5% |
| 4 | 9.7% |
| 9 | 4.6% |
Populations of countries, market capitalizations, file sizes on a hard drive, half-lives of radioactive isotopes, citation counts, earthquake magnitudes, Fibonacci numbers, and powers of 2 all track the curve closely.
It has been put to work. In 1993, Arizona accountant Wayne Nelson caught state treasurer manager Wayne James Nelson stealing $1.88 million in vendor payments — the forged check amounts had too many 7s, 8s, and 9s in leading positions. Mark Nigrini formalized the audit method in the 1990s; the U.S. IRS and large accounting firms have used Benford screens on tax returns and ledger entries ever since. In 2011, researchers flagged Greek macroeconomic data submitted to the EU as the worst Benford-fit in the eurozone, consistent with the country's later admitted misreporting.
A clean Benford match does not prove honesty. A mismatch does not prove fraud. It says: this dataset deserves a closer look.
What's contested
Election forensics is the loudest fight. Walter Mebane and others have applied Benford-style tests (often on the second digit, not the first) to precinct vote counts in Iran 2009, Russia 2011, and Venezuela. Critics — including Joseph Deckert and a 2011 Political Analysis paper — argue that vote totals are shaped by district sizes, turnout caps, and party geography in ways that violate Benford's preconditions. The math is clean; the inference is contested.
The deeper unsettled question is when Benford should apply. There is no clean theorem that tells you in advance whether a given dataset will obey it. Hill's 1995 result shows that mixing distributions sampled across many scales converges to Benford, but that is a sufficient condition, not a diagnostic. Narrow-range data (human heights, exam scores), assigned numbers (phone numbers, ZIP codes), and capped values (percentages, ratings) systematically break it. Whether a real ledger should follow Benford requires a judgment call before you run the test, not after.
Why this has to do with other realms
Benford's Law is a small doorway from probability into auditing: a logarithm becomes a fraud detector. The same multiplicative structure shows up in power laws — Zipf for word frequencies, Pareto for wealth, Gutenberg-Richter for earthquakes. All three are statements that the world is more often built by ratios than by sums.
The link runs further into measurement. A regularity that survives miles-to-kilometers conversion is telling you that the meaningful information lives in the ratio between values, not the values themselves. This is the same insight that underwrites the richter scale, stellar magnitudes, decibels, pH, and compound interest. When humans invent a number system to describe a multiplicative world, logarithms keep appearing — not as a trick, but as the right coordinate system.
An open question
If Benford's Law is best at catching numbers that humans invented to look random, what other mathematical fingerprints do fabricated datasets leave behind — and could a fraudster who knows the law simply Benford-shape their lies?
Key sources
- Frank Benford, "The Law of Anomalous Numbers" (1938) — the classic empirical paper with 20,229 observations across 20 datasets.
- Simon Newcomb, "Note on the Frequency of Use of the Different Digits in Natural Numbers", American Journal of Mathematics (1881) — the worn-logarithm-book origin.
- Theodore P. Hill, "A Statistical Derivation of the Significant-Digit Law", Statistical Science (1995) — the modern probability foundation via scale and base invariance.
- Mark J. Nigrini, Digital Analysis Using Benford's Law (2012) — practical audit and fraud-detection methodology.
- to verify: Joseph Deckert, Mikhail Myagkov, Peter Ordeshook critique of Benford-based election forensics, Political Analysis (~2011).
Further reading
- Benford's Law: Theory and Applications, edited by Steven Miller (2015) — the comprehensive academic survey if you want the proofs.
- Walter Mebane's working papers on election forensics — read alongside the Deckert critique to see both sides of the contested case.
- Terence Tao's blog post on Benford's Law (2009) — a working mathematician's intuition for why log-scale invariance forces the digit distribution.
- The Greek deficit data analysis by Bernhard Rauch et al., German Economic Review (2011) — Benford screening applied to sovereign accounting.
See Also
- power laws (same multiplicative skeleton, different domain)
- forensic accounting (where the law became a working tool)
- richter scale (another logarithmic coordinate system for a multiplicative world)
- compound interest (the ratio-thinking that makes Benford inevitable)
- probability
- measurement