Benford's Law
In many natural collections of numbers the leading digit is 1 about 30% of the time, not 11% — a logarithmic bias that holds across populations, areas, and physical constants.
Benford's Law
Pick a number at random from a long list of real-world quantities — the populations of countries, the lengths of rivers, the areas of nations, the figures buried in a tax return — and look only at its first digit. Naively you might expect each of the digits 1 through 9 to lead about \tfrac{1}{9} \approx 11\% of the time. They do not. The digit 1 leads roughly 30% of the time, 2 about 18%, and the frequencies fall away smoothly to 9, which leads under 5% of the time. This lopsided pattern is Benford's law, also called the first-digit law.
The rule is precise. The probability that the leading digit equals d is
Because \sum_{d=1}^{9} \log_{10}(1 + 1/d) = \log_{10}(10) = 1, this is a genuine Probability Distribution over the nine possible leading digits. Plugging in d=1 gives \log_{10} 2 = 0.301 — the famous "30%."
Why a logarithm?
The deep reason is scale invariance. If a collection of numbers obeys some universal first-digit law at all, that law cannot care whether the quantities are measured in miles or kilometres, dollars or euros — multiplying every value by a constant must leave the digit distribution unchanged. The only distribution with that property is the logarithmic one above.
The intuition: lay the numbers out on a logarithmic ruler, where the distance from 1 to 2 is much wider than from 8 to 9. A quantity that grows by steady percentages — the way populations, prices, and bank balances do — spreads itself uniformly across that logarithmic ruler. The stretch of the ruler labelled with a leading 1 occupies \log_{10} 2 \approx 30\% of each decade; the stretch labelled 9 occupies only \log_{10}(10/9) \approx 4.6\%. The digits inherit the widths of their logarithmic homes. This is why the law loves data that spans many orders of magnitude and arises from multiplicative growth — the same engine behind a Scale-Free Network's power-law degrees.
Testing it on real data
Talk is cheap; the law is checkable. The dataset data/world-figures.csv holds 292 genuine reference figures — the 2023 population and the land area in km² of 146 countries. These span from Luxembourg's 2,586 km² to Russia's 17 million km², and from Iceland's 375 thousand people to China's 1.4 billion: many orders of magnitude, exactly the regime where Benford should bite. Stripping each value down to its leading digit and tallying gives:
How we know this
| digit | count | observed | benford |
|---|---|---|---|
| 1 | 85 | 29.1% | 30.1% |
| 2 | 53 | 18.2% | 17.6% |
| 3 | 37 | 12.7% | 12.5% |
| 4 | 25 | 8.6% | 9.7% |
| 5 | 26 | 8.9% | 7.9% |
| 6 | 23 | 7.9% | 6.7% |
| 7 | 12 | 4.1% | 5.8% |
| 8 | 16 | 5.5% | 5.1% |
| 9 | 15 | 5.1% | 4.6% |
| N=292 | chi2=3.22 | maxdev=1.7% |
# evidence/benford-leading-digits/{inputs: data/world-figures.csv, run}
LC_ALL=C awk -F, '
NR>1 { v=$3; sub(/^[^1-9]*/, "", v); d = substr(v,1,1) + 0
if (d>=1 && d<=9) { c[d]++; n++ } }
END { split("0.301 0.176 0.125 0.097 0.079 0.067 0.058 0.051 0.046", b, " ")
print "digit count observed benford"; chi2=0; maxdev=0
for (d=1; d<=9; d++) { obs=c[d]/n; ex=n*b[d]
chi2 += (c[d]-ex)*(c[d]-ex)/ex
dev=obs*100-b[d]*100; if(dev<0)dev=-dev; if(dev>maxdev)maxdev=dev
printf "%d %5d %6.1f%% %5.1f%%\n", d, c[d], obs*100, b[d]*100 }
printf "N=%d chi2=%.2f maxdev=%.1f%%\n", n, chi2, maxdev }' world-figures.csvThe single most visible signature of the law is the dominance of the digit 1:
How we know this
# evidence/benford-digit-one/{inputs: data/world-figures.csv, run}
LC_ALL=C awk -F, '
NR>1 { v=$3; sub(/^[^1-9]*/, "", v); d=substr(v,1,1)+0
if (d>=1 && d<=9) { n++; if (d==1) ones++ } }
END { printf "%.1f%%\n", 100*ones/n }' world-figures.csvThe fit is not a coincidence of these particular numbers — pool populations and areas measured in different units and the same shape appears, exactly as scale invariance demands.
Where it is used
Because honest natural data obeys Benford so reliably, departures from it are suspicious. Forensic accountants and auditors run Benford tests on ledgers, expense reports, and tax filings: fabricated figures, invented by people who unconsciously spread their leading digits too evenly, tend to show a deficit of 1s and a surplus of middle digits. The same test has been turned on reported election tallies, scientific datasets, and macroeconomic statistics as a cheap first-pass screen for manipulation. It is not proof — some clean datasets fail the law and some doctored ones pass — but a strong Benford violation is a flag worth pulling.
The law also reaches into pure mathematics: the leading digits of the powers of 2, of the Fibonacci numbers, and of many sequences tied to the Prime Number counting function are Benford-distributed, again because their logarithms fill the unit interval uniformly.