Benford’s law is usually introduced as a party trick:

In many real datasets, about 30% of numbers start with 1.

That is true, and also the least interesting way to say it.

The better version is geometric:

look at log10(x), throw away the integer part, and ask whether the remainder is
uniform on the unit interval

The first digit lives on that circle.

If the fractional part of (\log_{10} x) is uniform on ([0,1)), then the probability of first digit (d) is the length of the interval

\[[\log_{10} d,\log_{10}(d+1)).\]

So

\[\Pr(D=d)=\log_{10}(d+1)-\log_{10}(d) =\log_{10}\left(1+\frac{1}{d}\right).\]

That formula says digit 1 gets the interval from (1) to (2), digit 2 gets (2) to (3), and digit 9 gets only (9) to (10), all measured on a log scale. The low digits are not lucky. Their logarithmic intervals are wider.

The Dirty Log Tables Were a Clue

Simon Newcomb noticed the phenomenon in 1881 by looking at logarithm tables. The early pages were more worn than the later pages, suggesting that numbers beginning with small leading digits were looked up more often.1

Frank Benford returned to the observation in 1938 and collected about twenty thousand first digits from heterogeneous sources: physical constants, areas, populations, drainage rates, addresses, and other tables.2 The name stuck to Benford, but the clue was already in Newcomb’s dirty pages.

The clue matters because logarithm tables are organized by mantissa. When you look up a number in a common-log table, the decimal point has already been removed. 123, 12.3, and 0.00123 live in the same neighborhood because they share the same significant digits.

That is why Benford’s law is not primarily about base-10 notation. It is about not having a privileged scale.

Multiply every value by 3.7. A genuinely Benford dataset should still look Benford. The decimal points move. The units change. The fractional log positions only rotate around the circle.

Hill’s 1995 theorem makes this intuition precise from an invariance direction: assuming base-invariance of the underlying law leads to the Benford distribution; the core objects are mantissas and invariant measures on a circle.3

A Scale Lab

The lab below generates synthetic datasets under four contracts:

  • scale mixture: sizes drawn across several orders of magnitude;
  • growth path: multiplicative growth observed over time;
  • linear range: values uniformly sampled from a bounded linear interval;
  • ID window: serial-like values with a fixed prefix.

Then it measures three things:

  • first-digit mean absolute deviation from Benford;
  • Kolmogorov-Smirnov distance of log mantissas from uniform;
  • scale drift after multiplying every value by constants like 2, 5, and 7.

The default scale mixture has first-digit MAD about 0.30%, mantissa KS about 0.014, and scale drift about 3.0%. Switch to linear range and the digit histogram becomes almost flat; the MAD jumps to about 6.24% and scale drift to about 53.3%. Same sample size, different contract.

Synthetic data only. Threshold contamination snaps some values near endings like 1.99, 4.99, and 9.99 within their current decade.

A First-Digit Test Is a Unit Test

The multiplier panel is the one I trust most.

If the data are genuinely scale-free in the first-digit sense, multiplying by a constant should not change the diagnostic much. It is the same collection of sizes expressed in a different unit. Meters to feet. Dollars to yen. Seconds to milliseconds. The mantissas rotate, but a uniform circle stays uniform.

If the data are a linear range, a serial-number window, or a collection cut by policy thresholds, the rotation changes the shape. That is what the lab calls scale drift.

This is why “many orders of magnitude” is not the whole story. A linear uniform range from 1 to 100000 spans five decades, but its first digits are almost flat because the final decade contains most of the linear mass. Benford behavior comes from log-scale spread, not from a large maximum printed in the column.

Fraud Detection Is the Dangerous Use Case

Benford tests are popular in audit and forensic analytics because humans invent numbers poorly, and operational rules leave digit-shaped scars. Nigrini-style training material emphasizes exactly the right caveats: numbers should represent sizes of facts or events, should not be identifiers, and should not be dominated by built-in minimums or maximums.4

The lab’s human thresholds slider is deliberately crude. It does not model a real fraudster. It only shows how easy it is to produce non-Benford digits by placing mass near psychologically or administratively meaningful endings:

49, 99, 199, 499, 999, ...

At 25% threshold contamination, the default scale mixture flips from Benford-like to not Benford: the first-digit MAD is about 2.94%, the mantissa KS is about 0.105, and scale drift is about 23.9%.

That is a lead, not a conviction.

Miller and Nigrini’s fraud-detection slides put the caution plainly: not all fraud is detectable by Benford’s law, and a mathematical anomaly is not proof of fraud.5 The right interpretation is investigative:

this subset does not look like the contract I claimed generated it

Then the work starts. Was there a reporting threshold? A tax bracket? A maximum reimbursement? A minimum invoice size? A serial-number prefix? A unit conversion? A product catalog? A seasonal campaign? A missing negative sign? A filter that quietly removed small values?

Benford is useful when it makes you ask those questions earlier.

The Contract I Want Written Down

Before running a Benford test on real data, I want the data owner to answer:

  • Are these measurements of size, not labels, rankings, IDs, prices chosen from a catalog, or capped scores?
  • Do the values cover enough log-scale range?
  • Were zeros, negatives, refunds, censored values, or rounded values removed?
  • Are there policy thresholds that create mass near 49, 99, or 999?
  • Is the tested subset natural, or was it selected after looking at the digits?
  • What non-fraud process would create the same anomaly?

If those answers are fuzzy, the first-digit chart is decoration.

The strongest form of the test is not “the blue bars differ from the purple dots.” The strongest form is:

the data-generating story predicts log mantissas should be close to uniform,
but this audited subset is not stable under changes of scale

That is a falsifiable sentence. It names a mechanism. It can be checked against domain facts.

The Useful Lesson

Benford’s law has survived because it is both simple and slippery.

Simple:

\[\Pr(D=d)=\log_{10}\left(1+\frac{1}{d}\right).\]

Slippery because the formula is not a universal law of “real data.” It is the shadow cast by a stronger symmetry: no preferred unit, no privileged decade, no hard range that pins the mantissa in place.

The first digit is a scale test.

When it passes, the dataset is behaving as if its logarithms are spread around a circle. When it fails, the interesting question is not whether the number 1 appeared too often. The interesting question is:

what broke the circle?

  1. Simon Newcomb, “Note on the Frequency of Use of the Different Digits in Natural Numbers”, American Journal of Mathematics 4(1), 39-40, 1881. PDF. ↩

  2. Frank Benford, “The Law of Anomalous Numbers”, Proceedings of the American Philosophical Society 78(4), 551-572, 1938. PDF. ↩

  3. Theodore P. Hill, “Base-Invariance Implies Benford’s Law”, Proceedings of the American Mathematical Society 123(3), 887-895, 1995. Record and abstract. ↩

  4. Mark J. Nigrini, “Benford’s Law research” training slides, 2012. The slides summarize practical conditions for digit tests, including that numbers should represent sizes and not identification numbers. PDF. ↩

  5. Steven J. Miller and Mark Nigrini, “Theory and applications of Benford’s law to fraud detection”, IRS Boston Offices talk, 2008. PDF. ↩