Logarithmic significant-digit distribution
Benford's
Law
In many datasets produced across changing scales, lower leading digits occur more often than higher ones. A leading 1 appears about 30.1% of the time; a leading 9 about 4.6%. The pattern is mathematically exact for some processes, approximate for others, and irrelevant to many datasets.
Inspect the digits. Then inspect the process.
Choose a controlled example or paste your own positive and negative numbers. Zero is ignored because it has no nonzero leading digit. Decimals are read by their first significant digit: 0.0047 begins with 4.
The generated logarithmic mantissas produce a close Benford profile. That demonstrates the mathematics, not the truthfulness of a real dataset.
A probability distribution for significant digits.
For decimal first digit d in {1,...,9}, Benford's Law assigns probability P(D=d)=log10(1+1/d). The cumulative probability that a significand lies below x, for 1 less than or equal to x less than 10, is log10(x). The law extends to the first two digits, later digits, and numeral bases other than ten.
The word "law" can mislead. This is not a claim that every collection of numbers must follow one table. It is a precise distribution that arises under particular mathematical conditions and approximates many, but not all, empirical datasets.
The surprising part is not that small digits can dominate. It is that the same logarithmic profile can survive a change of units.
Numbers spend unequal time behind each leading digit.
Consider a quantity growing by a constant percentage. To move from 1 to 2 it must double. To move from 8 to 9 it rises only 12.5%. The logarithmic distance of [1,2) is therefore much larger than the distance of [8,9). If the fractional parts of base-10 logarithms are uniformly distributed, leading digits follow Benford exactly.[4]
Change the unit; keep the law.
Multiplying every value by a positive constant rotates the logarithmic mantissas. A truly uniform mantissa distribution remains uniform, so its leading-digit distribution remains Benford.
Scope comes before calculation.
There is no universal checklist that guarantees Benford behavior. The generating process, selection rules, boundaries, dependence, and measurement conventions matter more than a slogan such as "spans several orders of magnitude."[6]
Values with open-ended, mixed scales
- Transaction amounts from varied activities
- Populations across settlements
- Multiplicative growth sequences
- Mixtures of several distributions
Assigned or tightly constrained values
- Telephone numbers and account IDs
- Human heights within one population
- Prices fixed near a standard point
- Values truncated by minimums or maximums
Data changed by human systems
- Approval thresholds and spending limits
- Rounded or estimated reports
- Filtered samples and aggregated totals
- Dependent observations from one process
From worn logarithm tables to modern data analytics.
Newcomb noted that the early pages of logarithm tables appeared more worn and proposed equal probability for logarithmic mantissas.[1]
Benford assembled 20,229 observations in 20 groups and published the "law of anomalous numbers."[2]
Hill showed how samples drawn from random mixtures of distributions can converge to the significant-digit law.[3]
Digit tests entered auditing, data-quality work, scientific diagnostics, and public controversy, creating both useful applications and serious overclaims.
A screen is the beginning of an investigation.
Document the source system, date range, value field, exclusions, currency, signs, duplicates, and transformations.
Explain why this process could produce Benford digits and identify thresholds or controls that could create legitimate deviations.
Use plots and more than one measure. Preserve the full distribution rather than reducing everything to a pass/fail label.
Trace unusual digits to vendors, users, dates, accounts, approval levels, and source documents.
Combine digit evidence with controls testing, duplicate detection, outliers, documents, interviews, and independent records.
Professional guidance presents Benford analysis as one data-analysis technique among many. It helps prioritize attention; it does not establish motive or authorship.[8]
Different diagnostics answer different questions.
Observed bars
Best first view. Shows which digits create the mismatch and whether the pattern is broad or localized.
Interpret visually with context.Mean absolute deviation
Averages the absolute gaps between observed and expected proportions. Easy to compare, but thresholds are conventions, not laws of nature.
MAD = mean |pobs - pexp|Chi-square
Can be hypersensitive in large samples and rests on sampling assumptions that real administrative data may violate.
Do not read p as probability of fraud.Digit-specific z scores
Highlights individual digits but creates a multiple-testing problem when many fields, subsets, and digit positions are searched.
Predefine tests where possible.Random variation can produce dramatic-looking bars.
Practical importance may be small even when a formal null is rejected.
More observations do not repair an inappropriate reference model.
What a digit pattern can reveal - and what it cannot.
Amounts gather just below a control limit.
A digit test may flag the population, but a threshold plot is more direct. Inspect approvers, vendors, dates, and documentation. The anomaly could reflect policy-compliant purchasing or deliberate splitting.
BEST NEXT TEST: threshold and duplicate analysisA reporting system begins rounding values.
Digit structure can reveal a discontinuity without implying fraud. Compare the change date with software releases, measurement precision, and reporting guidance.
BEST NEXT TEST: time-segment comparisonElection totals are declared fraudulent from a chart.
Election units, turnout bounds, district sizes, and the selected digit position determine the expected distribution. A mismatch by itself does not establish manipulation; public claims require a validated election-specific model and corroborating evidence.
BEST NEXT TEST: election-specific forensic designFour shortcuts that break the analysis.
"Deviation proves fraud."
No. It identifies structure that may deserve explanation. Legitimate controls, prices, bounds, seasonality, and software can all create deviations.
"All natural data conform."
No. Many natural measurements are narrow, additive, bounded, or otherwise non-Benford. "Natural" is not a statistical mechanism.
"Several orders of magnitude are required."
Wide range is often suggestive, but it is neither necessary nor sufficient. Mathematical counterexamples exist.[6]
"A significant p-value settles it."
No. It can reflect a huge sample, dependence, an unsuitable null, or repeated searching. Report effect size, design, and alternatives.
Sources and further reading.
Original papers, peer-reviewed mathematical reviews, and professional audit sources are prioritized.
- Simon Newcomb (1881) - Note on the Frequency of Use of the Different Digits in Natural NumbersThe first published statement of the logarithmic significant-digit pattern.doi.org/10.2307/2369148
- Frank Benford (1938) - The Law of Anomalous NumbersThe classic empirical paper, reporting 20,229 observations across 20 groups.Proceedings of the American Philosophical Society 78(4), 551-572
- Theodore P. Hill (1995) - A Statistical Derivation of the Significant-Digit LawA foundational mixture-of-distributions result for explaining Benford convergence.doi.org/10.1214/ss/1177009869
- Berger and Hill (2021) - The Mathematics of Benford's Law: A PrimerAn accessible, rigorous review of foundations, examples, applicability, and recurring errors.Statistical Methods & Applications 30, 779-795
- Berger and Hill (2011) - Benford's Law Strikes Back: No Simple Explanation in Sight for Mathematical GemA modern review emphasizing the phenomenon's mathematical depth and diverse explanations.Statistical Science 26(3), 358-374
- Theodore P. Hill (2020) - A Widespread Error in the Use of Benford's Law to Detect Election and Other FraudA constructive correction to the claim that every Benford distribution must span several orders of magnitude.arXiv:2011.13015
- Mark J. Nigrini (2012) - Benford's Law: Applications for Forensic Accounting, Auditing, and Fraud DetectionProfessional methods, tests, examples, and limitations for forensic practice.doi.org/10.1002/9781119203094
- Institute of Internal Auditors - Data Analysis TechnologiesProfessional guidance placing Benford analysis among a broader set of audit data-analysis methods.Global Technology Audit Guide
- Mark J. Nigrini (2022) - Using Benford's Law to Reveal Journal Entry IrregularitiesA practitioner-oriented example of anomaly screening, including both suspicious and legitimate causes.Journal of Accountancy, September 2022
- Kossovsky (2021) - On the Mistaken Use of the Chi-Square Test in Benford's LawA focused critique of automatic chi-square use without a defensible sampling model.Stats 4(2), 419-453