Benford's Law

The Dirty Pages

Open a book of logarithm tables — if you can find one, which in 1881 was like saying "open a web browser" — and flip through it. Go ahead. I'll wait.

Simon Newcomb did exactly this in his observatory at the United States Naval Almanac Office, and he noticed something that nobody had any business noticing: the first pages were filthy. Worn, smudged, soft at the corners. The pages listing logarithms of numbers starting with 1 had been thumbed through so often they were practically translucent. The pages for 8 and 9? Crisp as the day they were printed.1

This is the kind of observation most people would file under "huh, weird" and then go back to calculating the orbit of Neptune. But Newcomb was not most people. He was the kind of person who, upon noticing uneven page wear in a reference book, sat down and derived a mathematical law.

He published a two-page note in the American Journal of Mathematics proposing that the probability of a number's first digit being d was not 1/9 — not equal for all digits — but followed a logarithmic distribution. It was a genuine mathematical insight, published in a prestigious journal, by a serious astronomer.

And then everybody forgot about it for fifty-seven years.

The Physicist Who Counted Everything

Enter Frank Benford, a physicist at General Electric, who in 1938 independently made the same observation Newcomb had — and then did what physicists do when they suspect a pattern: he tested it against everything.2

Benford collected over 20,000 observations from 20 different datasets: areas of rivers, population figures, street addresses, numbers from newspaper front pages, molecular weights, death rates, baseball statistics. An ecumenical survey of numbers as they exist in the wild. He wasn't picky. He was thorough.

The result was almost spooky. About 30.1% of all first digits were 1. About 17.6% were 2. The digit 9 showed up first only 4.6% of the time. This wasn't a quirk of one dataset — it was a pattern woven into the fabric of how numbers behave in nature.

Benford also gave the law its stickiest feature: his name. Newcomb had priority by 57 years, but Benford had data — twenty datasets' worth — and in science, data talks louder than priority. Thus does a physicist at a lightbulb company become the namesake of one of mathematics' most elegant surprises.3

The Formula

The law itself is one line. The probability that the first digit of a number is d:

Plug in the numbers and you get a staircase descending from about 30.1% down to 4.6%:

But why logarithms? Here's the intuition that makes everything click. Think about what it takes for a number's first digit to change. To go from a leading 1 to a leading 2, a number has to double — from 1,000 to 2,000, say. That's a 100% increase. But to go from a leading 8 to a leading 9? That's only a 12.5% increase. And from 9 back to 1 (meaning the number crosses from 9,999 to 10,000)? Just 11%.

The digit 1 gets more "territory" on the number line because it takes longer to traverse. On a logarithmic scale — where equal distances represent equal ratios rather than equal differences — the interval from 1 to 2 is genuinely, physically wider than the interval from 8 to 9.

Scale Invariance, or Why Numbers Don't Know What Units They're In

Here's a thought experiment that should bother you. Take a table of country populations. The first-digit distribution follows Benford's Law. Now convert every population from people into thousands of people — divide everything by 1,000. Does the first-digit distribution change?

It shouldn't. And it doesn't. If a genuine law about first digits exists, it must be invariant under multiplication by a constant. River lengths don't know whether you're measuring in miles or kilometers. Your bank balance doesn't care whether it's denominated in dollars or yen.

In 1961, the mathematician Roger Pinkham proved that Benford's distribution is the only distribution on first digits that's invariant under scaling.4 Multiply every number in a Benford-distributed set by any constant, and the first digits still follow Benford's Law. This isn't a coincidence — it's a mathematical necessity for any dataset that spans multiple orders of magnitude and doesn't care about units.

Why It Works: Multiplication All the Way Down

The deepest explanation for Benford's Law comes from thinking about how numbers grow. Most quantities in the real world grow multiplicatively. A city doesn't add 10,000 residents per year — it grows by 2%. A stock doesn't rise $1 a day — it rises 0.5%. River lengths, molecular weights, financial figures — they're all products of many compounding factors.

When you multiply lots of random numbers together, you get a lognormal distribution — bell-curved on a logarithmic scale. And any distribution spread across several orders of magnitude on a log scale will, almost inevitably, produce first digits that follow Benford's Law.5

There's a beautiful theorem by Theodore Hill (1995) that makes this even more general: if you pick probability distributions "at random" and then pick numbers from those distributions, the combined dataset follows Benford's Law.6 It's a central limit theorem for first digits. Mix enough different sources, and Benford emerges like statistical gravity.

Catching the Cheaters

Here's where mathematics meets the crime beat. If naturally occurring numbers follow Benford's Law, then fabricated numbers usually don't. Humans are terrible at faking randomness. When someone invents financial figures off the top of their head, they spread first digits too evenly — too many 5s, 6s, 7s, 8s, and 9s, not enough 1s and 2s. The numbers feel "too random." And that's a red flag.

Today, Benford's Law is a standard tool in forensic accounting. "Digital analysis" of first digits — and second digits, and first-two-digit combinations — screens tax returns, corporate filings, and expense reports. Several countries' tax authorities use it as a first-pass filter for suspicious returns.8

Election forensics is a more controversial application. After the 2009 Iranian election, statisticians noted deviations from Benford's Law in certain vote tallies. But — and this is important — Benford's Law doesn't apply cleanly to all election data. Precinct-level vote counts are often constrained to a narrow range (a few hundred to a few thousand), which can naturally violate the distribution. The lesson: Benford's Law is a useful diagnostic, not a lie detector. Context matters enormously.9

When It Doesn't Work (and Why That Matters Too)

Benford's Law is not magic. Understanding when it fails is just as important as knowing when it works.

The law tends to hold when:

• Data spans several orders of magnitude
• Data arises from multiplicative or exponential processes
• Data isn't constrained to a narrow range
• There are no artificial cutoffs or caps

It tends to fail for:

• Assigned numbers — phone numbers, social security numbers, zip codes. These are identifiers, not measurements.
• Constrained ranges — human heights in centimeters (150–200) will have mostly 1s as leading digits, but for the wrong reason.
• Uniform distributions — lottery numbers, for instance.
• Small datasets — you need hundreds of observations at minimum for the pattern to emerge reliably.

Benford's Law sits at the intersection of pure mathematics and the grubby empirical world. It says something profound: numbers in nature are not uniformly random. They carry structure — a fingerprint left by the logarithmic geometry of multiplication. And when humans try to invent numbers from scratch, they miss this structure, because our intuition about "randomness" is linear, and the world's randomness is logarithmic.

Simon Newcomb noticed it in 1881. Frank Benford proved it in 1938. Roger Pinkham explained it in 1961. Ted Hill generalized it in 1995. And right now, somewhere in the world, a forensic accountant is running a first-digit analysis on a set of expense reports — and the numbers are about to rat out their author.

The next time you see a table of numbers — financial results, scientific data, census figures — remember: those numbers have a signature. And if the signature is wrong, someone has something to hide.