The First Digit Law: Why 1 Leads About 30% of the Time

Pick any collection of real-world numbers — the populations of every country on Earth, the GDP of every nation, the lengths of rivers, the physical constants in your textbook. Count how often the first digit is a 1. You expect roughly one in nine, about 11%. Instead you get about 30%. Count the 9s: you expect 11%, you get 4.6%. The distribution is not flat. It never is, for naturally occurring numbers. And the shape of that skew is always the same.

This is Benford’s Law. It is one of the stranger gifts mathematics has handed to fraud investigators — and once you understand why it holds, you will never look at a column of numbers quite the same way.

30.1%30.1%30.1%P(leading digit = 1)the exact prediction

17.6%17.6%17.6%17.6%P(leading digit = 2)not 11.1% uniform

4.6%4.6%4.6%4.6%P(leading digit = 9)the rarest first digit

20,22920,22920,229numbers Benford tested20 datasets, 1938

The observation: worn pages

Simon Newcomb noticed it in 1881 — not in data, but in a physical object. He was a mathematician and astronomer at the time, working at the Nautical Almanac Office, and he used logarithm tables constantly. Everyone did, before calculators. The tables were shared, and he noticed something odd about the wear patterns.

That the ten digits do not occur with equal frequency must be evident to anyone making much use of logarithmic tables, and noticing how much faster the first pages wear out than the last ones.

The pages starting with 1 were worn and grimy. The pages for 8s and 9s were barely touched. Someone was looking up numbers beginning with 1 far more often than any other digit — not because of any particular bias in the user, but because the numbers they were computing genuinely started with 1 more often.

Newcomb wrote it up in two pages, stated the correct formula, and was entirely ignored.¹1S. Newcomb, “Note on the Frequency of Use of the Different Digits in Natural Numbers,” American Journal of Mathematics 4 (1881), 39–40. The paper is two pages long. Newcomb gives the formula — “the law of probability of the occurrence of numbers is such that all mantissae of their logarithms are equally probable” — which is the log-uniform statement that implies Benford’s distribution exactly. The paper was ignored for 57 years. Newcomb 1881 Fifty-seven years later, Frank Benford — a physicist at General Electric — made the same observation about logarithm tables, apparently unaware of Newcomb’s paper. Where Newcomb had stated the principle and moved on, Benford went further: he collected 20,229 numbers from 20 completely different datasets — river drainage areas, population figures, street addresses, atomic weights, death rates, physical constants, molecular weights — and confirmed the same distribution across all of them.²2F. Benford, “The Law of Anomalous Numbers,” Proceedings of the American Philosophical Society 78, no. 4 (1938), 551–572. Benford found P(1) = 30.6% across his combined sample of 20,229 numbers — within rounding of the theoretical 30.1%. His datasets ranged from American League baseball statistics to areas of drainage basins to numbers found in front pages of newspapers. Benford 1938 The law carries his name.

The formula both men arrived at is compact. For a leading digit $d \in \{1, 2, \ldots, 9\}$:

$P(d) = \log_{10}\!\left(1 + \frac{1}{d}\right)$The probability that a number from a Benford-obeying dataset has leading digit d. P(1) = log10(2) ≈ 0.301. P(9) = log10(10/9) ≈ 0.046. The nine probabilities sum to log10(10) = 1.

The predicted values are exact: P(1) = 30.1%, P(2) = 17.6%, declining to P(9) = 4.6%. Before you see whether real data matches this, pause and ask yourself: what does your gut expect for a collection of “random” numbers?

Test your intuition first

The sandbox below generates 500 random integers between 1 and 10,000 and shows their leading-digit distribution. This is the intuitive baseline — numbers drawn uniformly from a range. Run it a few times. The shape is stable: near-flat, hovering around 11% for every digit. Now hold that shape in your mind.

runnable · js▸ run

output

Run the code to see its output here.

The uniform baseline hovers near 11% for every digit. Now look at what real-world data actually does.

Two real datasets: populations and GDP

The World Bank publishes population and GDP figures for every country and territory. Both span roughly six orders of magnitude — from Tuvalu’s 9,646 people (and Nauru’s $163 million GDP) up to India’s 1.45 billion population and the United States’ $29.2 trillion output. That range is exactly the condition Benford’s Law requires.

Here is the leading-digit distribution of 217 country populations, World Bank 2024 vintage, overlaid on the Benford prediction.³3World Bank, World Development Indicators, SP.POP.TOTL, 2024 vintage. Retrieved via the World Bank public API (api.worldbank.org/v2/country/all/indicator/SP.POP.TOTL) on 2026-06-26. 217 country and territory entries. Leading-digit frequencies computed from raw figures: 67 of 217 begin with 1 (30.9%), 10 begin with 9 (4.6%). World Bank 2024

chart: line of pct by digit (series) — 18 points

The fit is not perfect — 217 is a modest sample, and population data has its own clustering structure — but the dominant shape is unmistakably the Benford curve. Digit 1 leads at 30.9%; digit 9 lands at 4.6%, exactly on prediction.

Now the same test on a completely different quantity: country GDP.⁴4World Bank, World Development Indicators, NY.GDP.MKTP.CD, 2024 vintage. Retrieved via the same API on 2026-06-26. 206 country-level entries in current USD (regional and income-group aggregates excluded by iso3-code filter). Leading-digit frequencies: 65 of 206 begin with 1 (31.6%), 9 begin with 9 (4.4%). World Bank 2024

chart: line of pct by digit (series) — 18 points

Two completely different quantities — people and dollars — one shared shape. The pattern holds for river lengths, stock prices, physical constants, earthquake magnitudes, and the transaction amounts in a legitimate company’s accounts payable ledger. Benford assembled evidence from 20 disparate sources in 1938 and found it each time. The question worth sitting with is why.

Why? The logarithmic number line

The key is how numbers distribute on a logarithmic scale rather than a linear one.

On a linear scale, the integers 1 through 9 each occupy the same width. On a log scale — where equal distances represent equal ratios rather than equal differences — the distance from 1 to 2 is the same as the distance from 2 to 4, or from 4 to 8. Each doubling is the same step.

Now look at what this means for leading digits within a single decade $[10^k, 10^{k+1}]$. The interval $[1 \times 10^k, 2 \times 10^k)$ covers numbers with leading digit 1. On the log scale, this interval has width $\log_{10}(2) - \log_{10}(1) = \log_{10}(2) \approx 0.301$. The interval for digit 9 — from $[9 \times 10^k, 10^{k+1})$ — has width $\log_{10}(10) - \log_{10}(9) \approx 0.046$. Digit 1 occupies more than six times as much log-scale real estate as digit 9.

chart: barY of logspan by digit — 9 points

Benford’s Law is exactly this: the probability of leading digit $d$ equals the fraction of the log-decade that $d$ occupies — which is $\log_{10}\!\left(\frac{d+1}{d}\right) = \log_{10}\!\left(1 + \frac{1}{d}\right)$These are the same expression. log10((d+1)/d) = log10(d+1) − log10(d) = the width of the interval [d, d+1) on the log10 number line. For d=1: log10(2) ≈ 0.301. For d=9: log10(10/9) ≈ 0.046..

This reframes the question: why do real-world numbers behave as if they were sampled uniformly on a log scale? Country populations and GDP figures were not drawn with that intent.

Scale invariance: the deep reason

The most illuminating answer comes from a symmetry argument. Ask: what property must a distribution have if its leading-digit histogram looks the same regardless of what units you express it in?

Measure river lengths in kilometers. Convert to miles. If the dataset truly has no preferred scale — no natural unit that makes it “special” — then multiplying all values by any constant $c$ should leave the leading-digit histogram unchanged. This is scale invariance.

On the log scale, multiplying all values by $c$ means adding $\log_{10}(c)$ to every value — a rigid translation. The only distribution that is invariant under all such translations is the uniform one. And uniform on the log scale gives exactly Benford’s distribution.⁵5T. P. Hill, “A Statistical Derivation of the Significant-Digit Law,” Statistical Science 10, no. 4 (1995), 354–363. Hill proves scale invariance as a necessary and sufficient characterization of Benford’s Law: a distribution obeys Benford if and only if it is scale-invariant in the sense of being invariant under multiplication by any positive constant. Hill 1995

This is why river lengths obey Benford whether you measure in meters, feet, or furlongs. It is why stock prices obey it across bull and bear markets separated by factors of ten. Data that genuinely has no preferred scale must be log-uniform — and log-uniform means Benford.

Hill’s theorem: why aggregates converge

There is a second, complementary reason that explains why even data that doesn’t perfectly obey Benford in isolation tends to conform when mixed.

Theodore Hill’s 1995 theorem — a central-limit-theorem analogue for significant digits — states: if you randomly select distributions (in any unbiased way), then draw random samples from each, the combined sample’s leading digits converge to the Benford distribution, regardless of what the individual source distributions were.⁶6Hill’s main theorem: if $F_1, F_2, \ldots$ are chosen randomly from the space of all probability distributions (under any non-atomic measure), and $X_k \sim F_k$ independently, then the significant digits of $X_1, X_2, \ldots$ converge in distribution to Benford’s Law. The key insight is that the mixing of heterogeneous sources washes out the idiosyncrasies of each individual distribution. Hill 1995

This explains something practically important: a company’s accounts payable ledger mixes invoices from hundreds of vendors, across dozens of expense categories, spanning orders of magnitude in dollar value. The individual vendor datasets might not each obey Benford. But their mixture — the aggregate ledger — converges toward it. That convergence is what makes Benford analysis useful for auditing: legitimate financial data, by virtue of its heterogeneity, earns its Benford conformity. Fabricated data typically does not.

The law fails where these conditions fail. Numbers bounded into a single decade — human heights (150–200 cm), IQ scores (70–130), calendar years — do not obey Benford. Numbers with artificial cutoffs do not. Nigrini’s test is only valid when the underlying business process genuinely mixes heterogeneous sources across multiple orders of magnitude.

The interactive sandbox

Before going to the forensic mechanics, here is a sandbox to build intuition on your own data. The default dataset is the 45 largest country populations — paste in anything that spans a wide range: invoice amounts, stock prices, city populations, whatever you have. The output shows both your observed distribution and the Benford prediction side by side.

runnable · js▸ run

output

Run the code to see its output here.

Forensic mechanics: the MAD test

Mark Nigrini, an accounting professor, recognized in the 1990s that an auditor armed with Benford’s Law has an unexpected advantage: the expected distribution of first digits in legitimate financial data is known in advance, and departures from it are quantifiable.⁷7M. J. Nigrini, Benford’s Law: Applications for Forensic Accounting, Auditing, and Fraud Detection (Wiley, 2012). The book documents ten fraud case studies including Enron and AIG. The MAD thresholds were developed empirically by Nigrini across large samples of legitimate financial data. Nigrini 2012

The workhorse statistic is the Mean Absolute Deviation (MAD):

$\text{MAD} = \frac{1}{9} \sum_{d=1}^{9} |p_d^{\text{obs}} - p_d^{\text{Benford}}|$

where $p_d^{\text{obs}}$ is the observed fraction of leading digit $d$ and $p_d^{\text{Benford}} = \log_{10}(1 + 1/d)$. Nigrini’s empirically-calibrated thresholds for first-digit analysis:⁸8M. J. Nigrini, “Get M.A.D. with the Numbers,” Fraud Magazine, Sept/Oct 2000. Nigrini explicitly notes: “the thresholds of acceptability or conformity may vary with sample size and with the nature of the sample population.” The thresholds are calibrated for large transaction datasets (thousands of records), not the small country-level samples here. Nigrini 2000

filter4 rows

Nigrini's MAD conformity thresholds for first-digit Benford analysis. These are calibrated for large financial datasets — thousands of transactions. The thresholds tighten significantly for the two-digit test. Source: Nigrini (2000, 2012).

Conformity level↕	MAD range↕	Practical implication↕
Close conformity	0.000 – 0.006	Data follows Benford; no concern
Acceptable	0.006 – 0.012	Minor deviations; normal variation
Marginal	0.012 – 0.015	Worth a closer look; not definitive
Nonconforming	> 0.015	Flag for investigation; not proof of fraud

In practice, an auditor runs first-digit analysis to scan the whole ledger, then two-digit analysis (which has tighter thresholds: close conformity requires MAD below 0.0012) to narrow to suspicious digit combinations, then drills into the individual transactions clustered around the anomalous pattern. The sequence is: filter → investigate → conclude. Benford is only the first filter.

A concrete example: when an employee is inventing fake invoices, they tend to choose amounts that feel realistic — numbers in the hundreds or thousands, psychologically “middle” digits like 3, 4, 5, 6. The real ledger has thousands of legitimate micro-transactions starting with 1 (a $12 office supply order, a $150 subscription, a $1,400 flight). The fake invoices distort the 1-digit downward and inflate the 3–7 range upward. That is the signal Nigrini’s test catches.

filter9 rows

Benford's Law predictions vs. typical human-generated 'random' numbers. People tend to avoid 1s (they feel 'not random enough') and overuse the middle digits. That systematic mismatch is the forensic signal. Approximate figures from Nigrini (2012).

Leading digit↕	Benford P(d) %↕	Human 'random' (approx%)↕	Human / Benford↕
1	30.1	11	0.37
2	17.6	12	0.68
3	12.5	12	0.96
4	9.7	12	1.24
5	7.9	13	1.65
6	6.7	13	1.94
7	5.8	13	2.24
8	5.1	12	2.35
9	4.6	12	2.61

The ratio column tells the story. Humans under-represent 1s by a factor of nearly 3 and over-represent 8s and 9s by more than 2.5. The fabricator’s fingerprint is systematic, and it runs in exactly the direction the MAD test is calibrated to detect.

The election-forensics controversy

The most visible — and most contested — application of Benford’s Law is to election data. After Iran’s disputed 2009 presidential election, several analysts applied Benford analysis to precinct-level vote counts and found anomalies. The story spread fast: a mathematical law had detected fraud.

The political science community pushed back hard, and the critique is worth understanding in full. Deckert, Myagkov, and Ordeshook published a careful analysis in Political Analysis in 2011 arguing that Benford’s Law is essentially useless for detecting election fraud.⁹9J. Deckert, M. Myagkov, and P. C. Ordeshook, “Benford’s Law and the Detection of Election Fraud,” Political Analysis 19, no. 3 (2011), 245–268. The authors examined data from Ohio, Massachusetts, and Ukraine, plus simulated elections, and concluded that deviations from Benford’s Law “can arise regardless of whether an election is free and fair” — and, crucially, that fraud can be designed to satisfy the law. Deckert et al. 2011 Their argument has three prongs:

First, election data rarely spans multiple orders of magnitude. A typical national election has precinct-level vote counts that might range from a few dozen to a few thousand votes — that is one or two decades at most, not the five-or-six that financial data spans. The convergence result Hill proved requires heterogeneous mixing across many scales. Precinct vote counts often have a structural scale imposed by precinct size.

Second, the Benford deviation can go either direction. Real elections in Ohio and Massachusetts — which no one alleged were fraudulent — showed deviations from Benford that would have been “flagged” by the test. And a sufficiently sophisticated electoral manipulation can be engineered to satisfy Benford while still being fraudulent. The law has no model of what fraud looks like; it only tests whether the distribution matches a mathematical ideal.

Third, the hypothesis is underdetermined. Benford conformity is consistent with both honest elections and fraud; Benford nonconformity is consistent with both fraud and legitimate structural features of the voting process. Without a specific model of how fraud changes digit frequencies, the test cannot distinguish signal from noise.

The accounting case is stronger precisely because the failure mode is better understood. Invoice fabricators have a known psychological bias. Precinct manipulators do not have a known digit-level signature.

What the law can’t do

Benford’s Law is a first-pass screening filter, not a verdict. Before applying it, three conditions must hold:

The data must span multiple orders of magnitude — at least three decades, preferably more. Heights (150–200 cm), ages (0–100), calendar years, and zip codes all fail this condition and produce leading-digit distributions shaped by their own range structure, not by Benford.

There must be no structural cutoffs or targets. Prices set at round numbers, salaries clustered below a round-number tax threshold, or budgets that must match a preset line item will all show anomalous patterns for structural reasons that have nothing to do with fraud.

The distribution must be heterogeneous across sources. A single vendor’s invoices, all for roughly the same service, will not obey Benford — the homogeneity of the source breaks the mixing condition Hill’s theorem requires. It is the aggregate of many heterogeneous sources that converges.

When these conditions hold, a nonconforming MAD is a valid signal to investigate further. What it cannot do, in any dataset, is prove fraud. It raises a question; it does not answer one.

None of that diminishes the elegance of the underlying mathematics. The numbers at the top of this piece — 30.1%, 17.6%, 4.6% — are not empirical accidents. They are a theorem about the log-scale geometry of the number line, confirmed independently by a 19th-century astronomer noticing worn pages and a 20th-century physicist collecting 20,229 numbers from 20 different worlds. That the same mathematical truth turns out to make fabricated numbers detectable — in some datasets, under some conditions — is one of the rarer gifts in applied statistics.

---

Methodology. All Benford probabilities are computed from $P(d) = \log_{10}(1 + 1/d)$; specific digit values (P(1) = 30.1%, P(9) = 4.6%) are exact to one decimal place, matching both Newcomb (1881) and Benford (1938). The StatBand figure “20,229 numbers” is Benford’s stated sample size from his 1938 paper. The Newcomb quote (“That the ten digits do not occur with equal frequency…”) is drawn verbatim from his 1881 paper via archive at math.stonybrook.edu.

Population dataset (first Chart): World Bank SP.POP.TOTL, 2024 vintage, 217 country and territory entries, retrieved via the World Bank public API (api.worldbank.org/v2/country/all/indicator/SP.POP.TOTL) on 2026-06-26. Leading-digit frequencies computed from raw figures: 67 of 217 begin with 1 (30.9%), 10 begin with 9 (4.6%). MAD vs. Benford = 0.0142 — marginal conformity under Nigrini’s thresholds. This is expected: MAD thresholds are calibrated for large (n ≫ 1,000) financial transaction datasets; n = 217 introduces sampling variance that inflates MAD independent of any conformity issue.

GDP dataset (second Chart): World Bank NY.GDP.MKTP.CD, 2024 vintage, 206 country-level entries in current USD (aggregates excluded by iso3-code filter), retrieved via the same API on 2026-06-26. Leading-digit frequencies: 65 of 206 begin with 1 (31.6%), 9 begin with 9 (4.4%). MAD = 0.0213 — formally nonconforming, again due to small-n sampling variance. Both real datasets show the correct gross shape (monotone descent, digit-1 near 30%) and are cited as illustrative demonstrations, not as Benford-test verdicts.

Log-span chart: shows exact $\log_{10}(1 + 1/d)$ values — the Benford predictions themselves displayed geometrically, not separately collected data.

Human “random” digit table: approximate values from Nigrini (2012)‘s synthesis of human number-generation experiments. These represent the directional tendency; they are not from a single controlled study and should not be taken as exact figures.

MAD conformity table: Nigrini’s 2000 ACFE article thresholds (close: 0–0.006; acceptable: 0.006–0.012; marginal: 0.012–0.015; nonconforming: >0.015). Nigrini explicitly notes these may vary with sample size. The 2012 book reports slightly different threshold values in some editions; the ACFE article values are used here.

Election-fraud section: Deckert, Myagkov, and Ordeshook (2011) is the primary citation; their critique is presented as established scholarly consensus, not as one view among many. The 2009 Iran anomaly is mentioned as context for why the application arose, not as evidence of fraud. The asymmetry between the accounting and election applications is editorially motivated by the difference in underlying behavioral models, not by a claim about Iranian electoral integrity.

Hill’s 1995 theorem: cited for the convergence result; the full proof requires measure theory and is not reproduced here.