<- Back to laws

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.

130.1%
217.6%
312.5%
49.7%
57.9%
66.7%
75.8%
85.1%
94.6%
EXPECTED FIRST DIGIT FREQUENCY
ObjectFirst significant digit
DistributionLogarithmic
FormulaP(d) = log10(1 + 1/d)
Strong propertyScale invariance
Professional useAnomaly screening
Not equivalent toProof of fraud
INTERACTIVE 01 / DATA FORENSICS LAB

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.

FIRST QUESTIONWere these values generated by a process for which a Benford comparison is defensible? A test statistic cannot answer that question for you.
FIRST-DIGIT PROFILEMULTIPLICATIVE GROWTH
Usable values480nonzero numeric observations
Mean abs. deviation0.0046average absolute share gap
Chi-square1.9descriptive, not automatic proof
30%20%10%0
130.1%
217.6%
312.5%
49.7%
57.9%
66.7%
75.8%
85.1%
94.6%
ExpectedObserved
FIRST-DIGIT FIT IS CLOSE

The generated logarithmic mantissas produce a close Benford profile. That demonstrates the mathematics, not the truthfulness of a real dataset.

LINEAR VIEW123910Every interval [d, d+1) has equal width.
LOG VIEW123910The interval from 1 to 2 occupies 30.1% of a decade.
01 / THE LAW

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.

FIRST DIGITP(D=d) = log10(1 + 1/d)
d1 through 9
Pexpected proportion
base10 in this page
130.10%
217.61%
312.49%
49.69%
57.92%
66.69%
75.80%
85.12%
94.58%

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.
02 / WHY LOGARITHMS

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]

VALUExtake a positive measurement
->
LOGARITHMlog10(x)separate scale from position
->
FRACTIONAL PART{log10(x)}locate the value within a decade
->
LEADING DIGITfloor(10{log})map position back to 1-9
SCALE INVARIANCE

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.

123456789x 3.28084
Important nuanceScale invariance characterizes the Benford distribution, but changing units will not magically make an unsuitable finite dataset conform. A particular sample can move closer or farther by chance.
03 / APPLICABILITY

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]

PLAUSIBLE CANDIDATES

Values with open-ended, mixed scales

  • Transaction amounts from varied activities
  • Populations across settlements
  • Multiplicative growth sequences
  • Mixtures of several distributions
USUALLY UNSUITABLE

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
REQUIRES DOMAIN REVIEW

Data changed by human systems

  • Approval thresholds and spending limits
  • Rounded or estimated reports
  • Filtered samples and aggregated totals
  • Dependent observations from one process
Before testing, document:What creates each number?Why is each record included?Are values bounded?Are observations independent?Could policy create digit preference?Which digit test was chosen, and why?
04 / HISTORY

From worn logarithm tables to modern data analytics.

1881Simon Newcomb

Newcomb noted that the early pages of logarithm tables appeared more worn and proposed equal probability for logarithmic mantissas.[1]

1938Frank Benford

Benford assembled 20,229 observations in 20 groups and published the "law of anomalous numbers."[2]

1995Theodore Hill

Hill showed how samples drawn from random mixtures of distributions can converge to the significant-digit law.[3]

1990s-nowForensic analytics

Digit tests entered auditing, data-quality work, scientific diagnostics, and public controversy, creating both useful applications and serious overclaims.

05 / PROFESSIONAL WORKFLOW

A screen is the beginning of an investigation.

01Define the population

Document the source system, date range, value field, exclusions, currency, signs, duplicates, and transformations.

->
02Justify applicability

Explain why this process could produce Benford digits and identify thresholds or controls that could create legitimate deviations.

->
03Compare profiles

Use plots and more than one measure. Preserve the full distribution rather than reducing everything to a pass/fail label.

->
04Drill into records

Trace unusual digits to vendors, users, dates, accounts, approval levels, and source documents.

->
05Corroborate

Combine digit evidence with controls testing, duplicate detection, outliers, documents, interviews, and independent records.

WEAK ALONE
Digit deviationRecord clusterControl exceptionDocument conflictCorroborated intent
STRONGER CASE

Professional guidance presents Benford analysis as one data-analysis technique among many. It helps prioritize attention; it does not establish motive or authorship.[8]

06 / STATISTICAL TESTS

Different diagnostics answer different questions.

PROFILE

Observed bars

Best first view. Shows which digits create the mismatch and whether the pattern is broad or localized.

Interpret visually with context.
DISTANCE

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|
HYPOTHESIS TEST

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.
LOCAL DIAGNOSTIC

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.
SMALL SAMPLENoisy proportions

Random variation can produce dramatic-looking bars.

LARGE SAMPLETiny gaps become significant

Practical importance may be small even when a formal null is rejected.

ANY SAMPLEBad scope stays bad

More observations do not repair an inappropriate reference model.

07 / CASE FILES

What a digit pattern can reveal - and what it cannot.

CASE A / APPROVAL THRESHOLD

Amounts gather just below a control limit.

$4,850$4,920$4,995$4,760

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 analysis
CASE B / PIPELINE CHANGE

A reporting system begins rounding values.

120340560780910

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 comparison
CASE C / PUBLIC CLAIM

Election totals are declared fraudulent from a chart.

INSUFFICIENT

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 design
08 / MISUSE

Four shortcuts that break the analysis.

01

"Deviation proves fraud."

No. It identifies structure that may deserve explanation. Legitimate controls, prices, bounds, seasonality, and software can all create deviations.

02

"All natural data conform."

No. Many natural measurements are narrow, additive, bounded, or otherwise non-Benford. "Natural" is not a statistical mechanism.

03

"Several orders of magnitude are required."

Wide range is often suggestive, but it is neither necessary nor sufficient. Mathematical counterexamples exist.[6]

04

"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.

10 / REFERENCES

Sources and further reading.

Original papers, peer-reviewed mathematical reviews, and professional audit sources are prioritized.

  1. 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
  2. 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
  3. 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
  4. 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
  5. 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
  6. 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
  7. 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
  8. Institute of Internal Auditors - Data Analysis TechnologiesProfessional guidance placing Benford analysis among a broader set of audit data-analysis methods.Global Technology Audit Guide
  9. 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
  10. 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
CONTINUE EXPLORING

Related laws, with the relationship made explicit.

These are editorial connections, not claims that the laws are mathematically equivalent.

CONTINUE READING

Place this law inside the collection.

LAW 004 / 100 PUBLISHED

SAME SUBCATEGORY

More in Distributions

SAME DOMAIN

More in Mathematics & Probability