Strong Benford Behavior

• Strong Benford behavior is the rigorous form of Benford’s law, requiring that the complete significand distribution follows a logarithmic model, equivalent to uniformity of logarithms mod 1.
• It is analyzed through multiplicative models, Fourier analysis, and Mellin transforms, with mechanisms such as fragmentation and dependence driving asymptotic convergence.
• Recent research distinguishes formal significand-level convergence from mere first-digit conformity, emphasizing the practical implications of near-Benford phenomena in various applications.

Searching arXiv for recent and foundational papers on strong Benford behavior, fragmentation, dependence, and related asymptotic models. Strong Benford behavior is the stronger, significand-level form of Benford’s law. In base $B$, a positive quantity $X$ is strongly Benford when its significand $S_B(X)\in[1,B)$ satisfies

$\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B,$

so that the entire significand distribution, not merely the first digit, matches the logarithmic Benford model (Fang et al., 18 Aug 2025). Equivalently, $\log_B X \bmod 1$ is equidistributed on $[0,1)$, or continuously uniformly distributed mod $1$ in the terminology of linear-flow theory (Durst et al., 2017, Berger, 2015). Across the literature, however, the phrase is not completely uniform: in some empirical work, “strongly Benford” denotes exceptionally small first-digit deviation, often measured by a very low sum of squared deviations rather than by significand-level uniformity (Kossovsky, 2014). The topic therefore spans a precise distributional notion, several asymptotic mechanisms that generate it, and a set of statistical practices that sometimes use the same label more loosely.

1. Definition, equivalences, and terminological variation

The formal core of strong Benford behavior is the significand law. For a positive real number written as $x=S_B(x)B^k$ with $S_B(x)\in[1,B)$, strong Benford behavior requires

$\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B.$

This implies the familiar first-digit law

$X$0

but the converse does not hold: weak Benford behavior concerns only the first digit, whereas strong Benford behavior fixes the full significand distribution (Fang et al., 24 Aug 2025, Durmić et al., 2023).

A standard equivalent formulation is logarithmic. A sequence or family is strongly Benford in base $X$1 if and only if $X$2 converges to being equidistributed mod $X$3 (Fang et al., 24 Aug 2025). This equivalence is the basic bridge used in work on fragmentation, copulas, recurrence relations, Markov chains, and linear flows. In the dynamical setting, the same equivalence is stated as

$X$4

which recasts strong Benford behavior as a uniform-distribution property of observable signals (Berger, 2015).

The literature also contains a second usage. In “Arithmetical Tugs of War and Benford’s Law,” strong Benford behavior refers to empirical first-digit frequencies that are very close to the Benford proportions, with low

$X$5

That paper treats $X$6 below $X$7 as “ideally Benford,” while values above $X$8 indicate substantial deviation (Kossovsky, 2014). This is not the same definition as significand-level strong Benford behavior. A reasonable synthesis is that the formal theory reserves “strong” for the full significand law, while some empirical papers use it for unusually accurate first-digit conformity.

2. Principal mechanisms: logarithms, multiplicative structure, and skewness

The dominant theoretical mechanism is multiplicative. Repeated multiplication pushes logarithms into additive form, and Benford behavior emerges when those logarithms become equidistributed mod $X$9. This perspective appears explicitly in product models, fragmentation models, and recurrence relations, and analytically it is often implemented through Fourier analysis, Mellin transforms, or Poisson summation (Becker et al., 2013, Cuff et al., 2014).

A complementary explanatory language emphasizes distributional shape. “Arithmetical Tugs of War and Benford’s Law” argues that strong Benford behavior emerges when data are both highly variable in magnitude and strongly skewed to the right, with multiplication acting as the pro-Benford force and addition as the anti-Benford force (Kossovsky, 2014). The paper defines

$S_B(X)\in[1,B)$0

and associates Benford behavior with high $S_B(X)\in[1,B)$1 or $S_B(X)\in[1,B)$2, roughly $S_B(X)\in[1,B)$3 or $S_B(X)\in[1,B)$4, together with positive skewness. It also states that multiplication increases variability through relations such as

$S_B(X)\in[1,B)$5

whereas addition does not create comparable growth in order of magnitude (Kossovsky, 2014).

This line of work is careful not to identify Benford behavior with lognormality. Products of several low-$S_B(X)\in[1,B)$6 uniforms can remain non-Benford even when the log-histogram becomes roughly Normal, and products of only two or three high-$S_B(X)\in[1,B)$7 distributions can become nearly Benford before the log-density is fully Normal. The paper therefore explicitly says that “Benford's Law and Normality of log are not two sides of the same coin” (Kossovsky, 2014). That distinction is important because it separates the Benford problem from a simple multiplicative-CLT heuristic.

In parametric continuous families, the same mechanism appears in more analytic form. For the Weibull distribution, the density of $S_B(X)\in[1,B)$8 is written as a uniform main term plus a rapidly decaying Fourier series,

$S_B(X)\in[1,B)$9

so deviation from strong Benford behavior is encoded in nonzero Fourier modes (Cuff et al., 2014). For the inverse gamma distribution, a parallel Fourier-series representation shows that closeness to Benford is governed primarily by the shape parameter $\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B,$0, while $\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B,$1 mainly induces a phase effect because the deviation is unchanged under $\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B,$2 (Durst et al., 2016).

3. Dependence, fragmentation, and the failure of independence-based intuition

A major theme in the modern literature is that independence is not essential, but dependence must be handled structurally. “Benford's Law and Continuous Dependent Random Variables” studies fragmentation models in which the resulting random variables are dependent, yet the empirical significand distribution still converges to the Benford cdf. In the unrestricted one-dimensional decomposition model, the paper proves

$\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B,$3

so the empirical proportion of fragment lengths with significand at most $\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B,$4 concentrates around the Benford value (Becker et al., 2013). The proof strategy organizes each fragment length as a product of many cut variables and controls dependence by counting how many factors two fragments share.

The same paper identifies a sharp arithmetic threshold in the fixed-proportion model. If

$\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B,$5

then the fragment lengths converge in distribution to Benford’s law if and only if $\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B,$6; when $\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B,$7 has finite irrationality exponent, the paper gives a power-saving rate of convergence (Becker et al., 2013). This irrational-versus-rational dichotomy reappears in later deterministic stick-fragmentation work. In the multi-proportion model, the lengths

$\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B,$8

converge to strong Benford behavior if and only if

$\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B,$9

for some $\log_B X \bmod 1$0 (Fang et al., 24 Aug 2025, Fang et al., 18 Aug 2025). Here the multinomial combinatorics reduce the high-dimensional model to a one-dimensional equidistribution problem.

Higher-dimensional fragmentation introduces a second route to strong Benford behavior. In unrestricted box fragmentation, fragment volumes are products of many independent cut proportions and complements, and a Mellin-transform condition forces

$\log_B X \bmod 1$1

yielding strong Benford behavior for $\log_B X \bmod 1$2-dimensional fragment volumes (Durmić et al., 2023). In a later box-fragmentation result, order statistics are used to prove that, under mild conditions, the maximum $\log_B X \bmod 1$3-dimensional face volume and then the total $\log_B X \bmod 1$4-volume converge to strong Benford behavior as $\log_B X \bmod 1$5 (Fang et al., 18 Aug 2025).

Dependence can also destroy Benford behavior. “Benford's Law Beyond Independence: Tracking Benford Behavior in Copula Models” studies products of dependent random variables modeled by an $\log_B X \bmod 1$6-copula $\log_B X \bmod 1$7. The product is Benford precisely when

$\log_B X \bmod 1$8

is uniform on $\log_B X \bmod 1$9, and the paper derives a central approximation formula for the corresponding density (Durst et al., 2017). Its central conclusion is that the convergence toward Benford behavior familiar in the independent case is not necessarily preserved under dependence, and that preservation depends more on the copula structure than on the marginals themselves (Durst et al., 2017). This is an important correction to the common intuition that “many products become Benford-like” regardless of dependence structure.

4. Arithmetic, combinatorial, and dynamical manifestations

Strong Benford behavior also appears in structured discrete systems. In Zeckendorf theory, a uniformly random integer $[0,1)$0 has a decomposition into non-consecutive Fibonacci summands, and the proportion of those summands lying in any subset $[0,1)$1 of Fibonacci numbers with asymptotic density $[0,1)$2 converges in probability to $[0,1)$3 (Best et al., 2014). Taking $[0,1)$4 to be the set of Fibonacci numbers with a given leading digit or digit block yields Benford behavior for the summands. A generalized version extends this density-transfer principle to positive linear recurrence sequences $[0,1)$5, where the proportion of summands from a positive-density subset converges to that density with probability $[0,1)$6 (Best et al., 2014). In these papers, Benford behavior is a corollary of a stronger density theorem.

For finite-state Markov chains, the relevant objects are not raw state counts but the matrix sequences $[0,1)$7 and $[0,1)$8. A chain is called Benford when every component of both sequences is Benford or eventually zero, and a simple sufficient spectral criterion, nonresonance, guarantees this componentwise asymptotic law (Kaynar et al., 2010). The same paper proves that if the rows of the transition matrix are chosen independently and continuously, then the resulting Markov chain is Benford with probability one (Kaynar et al., 2010). This is a strong asymptotic manifestation: every nontrivial component sequence obeys Benford’s law.

Linear flows on $[0,1)$9 admit an even sharper characterization. For a linear flow $1$0 and a linear observable $1$1, the signal $1$2 is $1$3-Benford for every nonzero observable if and only if the spectrum is exponentially $1$4-nonresonant (Berger, 2015). The same work shows that exponential resonance is a meagre nullset in $1$5, so most linear flows are Benford both topologically and measure-theoretically (Berger, 2015). In this setting, strong Benford behavior is entirely spectral.

Recurrence relations produce a related “dominant main term” mechanism. For constant-coefficient recurrences, Benford behavior follows when the root of largest modulus has irrational logarithm in the relevant base. For non-constant recurrences of the form

$1$6

the key condition is that $1$7 is nondecreasing and $1$8, so the sequence is asymptotically equivalent to a multiplicative main term $1$9 whose logarithms can be handled by equidistribution methods (Farris et al., 2019).

A different, finite-scale notion of strength appears in “The Surprising Accuracy of Benford's Law in Mathematics.” There the issue is not significand-level convergence but extraordinarily accurate first-digit counts for sequences such as $x=S_B(x)B^k$0. The paper formalizes lower perfect hits, upper perfect hits, and bounded Benford error, and proves, for example, that digit $x=S_B(x)B^k$1 in $x=S_B(x)B^k$2 satisfies

$x=S_B(x)B^k$3

(Cai et al., 2019). This is a different use of “strong”: exact finite-$x=S_B(x)B^k$4 agreement rather than the significand law.

5. Near-Benford continuous families and quantitative deviation

Not all important examples are exactly strongly Benford; many are almost Benford in a quantitatively controlled sense. The Weibull distribution is a central case. Because the mod-$x=S_B(x)B^k$5 density of $x=S_B(x)B^k$6 has a constant term $x=S_B(x)B^k$7 and rapidly decaying Fourier corrections, the distribution is often very close to uniform on $x=S_B(x)B^k$8, hence close to strong Benford behavior (Cuff et al., 2014). The paper reports that the fit is very close when $x=S_B(x)B^k$9 is near $S_B(x)\in[1,B)$0, worsens as $S_B(x)\in[1,B)$1 increases, and depends only weakly on the scale parameter $S_B(x)\in[1,B)$2, since replacing $S_B(x)\in[1,B)$3 by $S_B(x)\in[1,B)$4 does not change the mod-$S_B(x)\in[1,B)$5 distribution (Cuff et al., 2014). It evaluates deviation by the Kolmogorov–Smirnov distance, the $S_B(x)\in[1,B)$6-norm, and the $S_B(x)\in[1,B)$7-norm.

The inverse gamma distribution exhibits a parallel pattern. Its mod-$S_B(x)\in[1,B)$8 density has the Fourier expansion

$S_B(x)\in[1,B)$9

so again the constant term encodes exact Benford behavior and the nonzero modes encode deviation (Durst et al., 2016). The principal conclusion is that the inverse gamma distribution only approximates Benford behavior for small $\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B.$0; as $\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B.$1 increases, the deviation $\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B.$2 becomes substantially larger (Durst et al., 2016). By contrast, $\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B.$3 mainly affects phase, because the Benford deviation is unchanged under multiplication by powers of $\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B.$4 (Durst et al., 2016).

These distributional analyses matter for two reasons. First, they show that strong Benford behavior is often best understood at the mod-$\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B.$5 logarithmic level rather than through raw digits. Second, they provide explicit truncation bounds and rapidly convergent Fourier series, so “close to strong Benford” can be quantified rather than asserted qualitatively (Cuff et al., 2014, Durst et al., 2016).

6. Statistical assessment, empirical usage, and recurring misconceptions

The most persistent misconception is that first-digit conformity is equivalent to strong Benford behavior. It is not. Several empirical studies report strong first-digit agreement without establishing significand-level uniformity. In NMR, the analysis is carried out through a first-digit Benford goodness parameter,

$\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B.$6

and many time-domain and frequency-domain signals, chemical-shift databases, and optimal-control RF pulses have high $\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B.$7 values (Bhole et al., 2014). These are strong empirical signs of Benford conformity, but they are not formal proofs of strong Benford behavior in the significand sense.

The same caution applies to large first-digit datasets in the social sciences. The Jehovah’s Witnesses activity study tests twelve variables with Pearson’s chi-square statistic against the $\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B.$8 critical value $\Pr(S_B(X)\le s)=\log_B(s),\qquad 1\le s<B.$9, and finds that an overwhelming majority of yearly samples conform to the first-digit Benford law (Mir, 2014). Yet the paper explicitly does not test second-digit behavior, significand uniformity, or base invariance (Mir, 2014). It therefore documents robust first-digit conformity rather than formal strong Benford behavior.

A second misconception is that any statistically significant deviation from Benford’s law is substantively important. “Severe testing of Benford's law” argues that conventional goodness-of-fit tests suffer from the “large $X$00” or “excess power” problem: in large samples, tiny and practically unimportant departures can be detected easily (Cerqueti et al., 2022). The paper derives the asymptotic distribution of the mean absolute deviation statistic, defines the excess $X$01, and uses severity testing to assess the largest discrepancy from the null warranted by the data (Cerqueti et al., 2022). In that framework, “strong Benford behavior” becomes a practical notion of closeness: data may reject exact Benford conformity and still remain close enough to Benford that the warranted discrepancy is substantively small.

A third misconception is that first-digit tests are robust against informed manipulation. “Robust inference under Benford's law” addresses the Benford-savvy adversary who preserves the first-digit law while altering the rest of the significand (Barabesi et al., 11 Jul 2025). The manipulated-Benford model keeps the first digit from a genuine Benford variable but splices in a different fractional significand, thereby preserving weak Benford appearance while breaking the full digit structure (Barabesi et al., 11 Jul 2025). The paper derives the null distribution of the fractional part $X$02, introduces the tests $X$03, $X$04, and $X$05, and shows in simulation and customs applications that these procedures detect manipulation precisely when classical first-digit tests do not (Barabesi et al., 11 Jul 2025). This directly reinforces the distinction between weak and strong Benford behavior.

Taken together, these studies suggest a stable hierarchy. At the strongest end lies the exact significand law, or an asymptotic proof via equidistribution mod $X$06. Below that lie near-Benford families with explicit quantitative deviation. Separate again are empirical first-digit diagnostics, which may show excellent conformity but do not by themselves establish strong Benford behavior. The contemporary literature is therefore not merely about why digit $X$07 occurs often; it is about when entire logarithmic mantissas become uniform, when dependence preserves or obstructs that phenomenon, and how to distinguish genuine Benford structure from superficial first-digit agreement.