Skip to content

Gibrat's Law

The claim that a firm's proportional growth rate is independent of its current size — which, iterated as multiplicative iid noise, makes log-size a random walk and drives the cross-sectional size distribution toward log-normal.

Core Idea

Gibrat's law (Robert Gibrat, 1931) is the claim that a firm's growth rate is statistically independent of its current size: if S(t) is firm size at time t, then S(t+1)/S(t) is an independent random variable whose distribution does not vary with S(t). The mechanical consequence of iterating this size-independent proportional random growth across many periods is that log(S) executes an additive random walk, and by the central limit theorem its cross-sectional distribution converges to approximately normal — meaning firm size itself converges to approximately log-normal. Gibrat's law thus simultaneously poses an empirical regularity about firm growth and identifies the generative mechanism — multiplicative iid noise — that produces the log-normal cross-sectional size distribution as its asymptotic outcome.

The construct operates at two levels that are easy to conflate. As an empirical claim, Gibrat's law holds approximately for large established firms but fails noticeably for small firms, which grow proportionally faster than Gibrat predicts, and fails at the upper tail, where size distributions tend toward Pareto rather than log-normal, indicating that preferential-attachment or merger dynamics, not size-independent noise, dominate there. As a generative mechanism, multiplicative iid shocks producing log-normal distributions is a canonical mathematical structure that extends beyond firms into city-size distributions, biological body-size distributions within taxa, word-frequency distributions, and asset prices under geometric Brownian motion — the continuous-time Gibrat analogue underlying Black–Scholes. In each application domain, Gibrat supplies a tractable null against which alternative generators (preferential attachment, size-dependent growth, minimum efficient scale) are evaluated as named deviations.

Structural Signature

Sig role-phrases:

  • the population of entities — a set of firms (or cities, organisms) each carrying a size S(t)
  • the multiplicative growth step — S(t+1) = S(t)·ε(t), so growth is proportional rather than additive
  • the size-independence assumption — the distribution of the growth factor ε does not depend on current size S(t), the load-bearing claim
  • the iid-over-time iteration — the same shock distribution compounds across many periods
  • the log-additive random walk — taking logs converts the multiplicative process into an additive walk in log-size
  • the central-limit convergence — the CLT applied to the log-walk drives the cross-section toward normal in logs, i.e. log-normal in levels
  • the linearly-growing dispersion — variance of log-size accumulates linearly with elapsed time, a clock on the process
  • the empirical test — regress log-growth on log-size; a zero coefficient supports the law, a negative one flags fast small-firm growth
  • the named departures — the regime limits where the null fails: minimum-efficient-scale at the small-firm end, Pareto preferential-attachment at the upper tail, age effects for young firms

What It Is Not

  • Not the claim that all firms grow at the same rate. The law says the distribution of the proportional growth factor is independent of size, not that every firm grows identically. That distribution has positive variance, so firms diverge widely — indeed the law's whole point is that size-independent random growth spreads the population into a log-normal. "Equal rates for large and small firms" is one operationalization that holds only above a small-firm threshold, not the law itself.
  • Not a universal or exact empirical law. It holds only as a regime-restricted baseline — approximately, for large, established, surviving firms over a long horizon. It fails noticeably for small firms (which grow proportionally faster) and at the upper tail (which tends Pareto, not log-normal). Applying the null outside its regime manufactures spurious "violations" that are really regime-boundary artifacts.
  • Not a generator of power-law or Pareto distributions. Size-independent multiplicative iid noise produces a log-normal cross-section, which is thinner-tailed than Pareto. A heavy Pareto/Zipf upper tail is precisely the signature that Gibrat has failed and a different generator — preferential attachment or mergers — dominates the top. Conflating Gibrat's log-normal output with a power law inverts what the deviation diagnoses.
  • Not a causal law or a deterministic scaling relation. It is a claim about a stochastic process — proportional random growth — not a force that makes firms grow, and not a deterministic power-scaling between physical quantities. A log-normal size distribution is evidence about an underlying proportional-growth process, not proof of it: survival selection, truncation, or too-short a horizon can mimic or mask the shape.
  • Not a single claim conflating dynamics and distribution. Three things are routinely fused and must be kept apart: that growth is size-independent (the law proper), that the size distribution is log-normal (its asymptotic consequence), and that large and small firms grow at equal rates (one testable operationalization). Evidence speaking to one level does not automatically settle the others.

Scope of Application

Gibrat's law has two scopes that must be kept apart. As an empirical claim about firm growth, it is a mechanism bounded by industrial organization and firm dynamics. As a generative mechanism — multiplicative iid noise iterated over time producing a log-normal cross-section — it is portable mathematical machinery that applies literally wherever that proportional-random-growth structure genuinely obtains; the quantitative fields below are real uses of the identical generator, not metaphor (the loose "big things grow proportionally" analogy without the multiplicative-noise structure stays out).

  • Industrial organization and firm dynamics — the home turf of the empirical law: firm-size and market-share evolution, age-vs-size relationships, and tests of Gibrat against size-dependent growth (small firms grow faster) and against the Pareto upper tail (preferential attachment, mergers).
  • Urban economics — city-size distributions, log-normal in the body and Zipf at the top, with proportional random growth the baseline against which Zipf deviations are measured.
  • Quantitative linguistics — word-frequency distributions in the Estoup-Zipf lineage, generated by the same multiplicative-growth process.
  • Biology — within-taxon body-size distributions, where iterated proportional growth produces the log-normal bulk.
  • Finance — asset prices under geometric Brownian motion, the continuous-time Gibrat analogue underlying Black-Scholes.
  • Macroeconomic distributions — income, wealth, and establishment-size cross-sections, where the multiplicative-shock baseline shapes expectations of the distributional form.

Clarity

Naming Gibrat's law gives industrial organization a sharp answer to a foundational question that informal discussion leaves muddy: does being large confer a growth advantage? By positing "no — proportional growth is size-independent and random," it converts a vague intuition into a falsifiable null, and the empirical literature inherits a clean test (regress log-growth on log-size; a zero coefficient supports the law, a negative one flags that small firms grow faster). Deviations stop being anomalies and become named departures — a significant slope at the small-firm end points to minimum efficient scale and the rapid early growth of survivors; a heavy Pareto upper tail points to preferential attachment or merger dynamics rather than size-independent noise. The law thus organizes the field's puzzles by telling the analyst exactly which observations are consistent with the baseline and which demand a different generator.

Its second clarifying move is to make legible the link between micro-level dynamics and macro-level shape: a log-normal cross-sectional size distribution becomes evidence about an underlying proportional-growth process, and departure from log-normality becomes evidence of departure from Gibrat. And it disentangles three claims routinely conflated — (a) growth rates are independent of size, (b) the size distribution is log-normal, © large and small firms grow at the same rate. Gibrat is properly (a); (b) is its asymptotic consequence; © is one operationalization of (a) that holds only above a small-firm threshold. Keeping the generative claim distinct from its distributional consequence and from its testable operationalization is what lets a practitioner reason cleanly about which level a given piece of evidence speaks to.

Manages Complexity

The behavior an industrial-organization analyst would otherwise have to track is daunting: a whole population of firms, each entering, exiting, merging, and growing through its own idiosyncratic history of demand shocks, financing rounds, and managerial decisions, and the question is what determines the shape of the cross-sectional size distribution that emerges from all of it. Gibrat's law collapses that high-dimensional population dynamics to two scalars and an initial condition. Once growth is posited size-independent and multiplicative, the entire process is a log-additive random walk governed by the mean μ and variance σ² of the per-period growth distribution; the cross-sectional shape is then derivable rather than separately observed — approximately log-normal, with a spread that widens predictably (variance in log-size growing linearly with elapsed time). The analyst stops modeling firms one at a time and tracks the drift and dispersion of a single random walk, reading the asymptotic distribution off those parameters.

The compression has a second, organizing face: it converts the field's open-ended catalog of size-and-growth puzzles into a small branch structure of named departures from one baseline. Every alternative generator becomes a labeled deviation indexed to where it bites and which sign it carries. A negative slope of log-growth on log-size at the small-firm end is the minimum-efficient-scale / fast-survivor branch; a heavy upper tail thicker than log-normal is the preferential-attachment-or-merger branch; an age effect overlaid on the size effect is the young-firm branch. The same regression — log-growth on log-size — serves as the universal probe, and its coefficient (zero, negative, or tail-localized) tells the analyst which branch a given dataset has entered. So rather than re-deriving the generative story for each industry, panel, or country, the practitioner carries one null (multiplicative iid noise → log-normal) and a short list of qualitatively distinct ways to depart from it, reading the operative mechanism off a single fitted slope and the tail shape — the move from population-scale bookkeeping to a two-parameter walk plus a finite menu of signed deviations.

Abstract Reasoning

Gibrat's law licenses a tight set of inferential moves, each running off the same generative core (size-independent multiplicative noise → log-normal cross-section) but pointing in a different direction.

Diagnostic — from cross-sectional shape back to growth process. The signature move reads the static size distribution as testimony about the dynamic process that produced it. Observe a roughly log-normal body and the analyst infers that growth in that range was proportional and size-independent — Gibrat held, and no size advantage operated. Observe a departure and the form of the departure names the culprit: a heavier-than-log-normal upper tail (Pareto/Zipf) says preferential-attachment or merger dynamics, not iid noise, govern the top; a body fatter at small sizes than log-normal predicts says small firms grew disproportionately fast. Crucially, the inference also runs the other way — from a fitted growth-size relationship to a prediction of the cross-sectional shape — so the two faces (a regression slope and a histogram) are mutually constraining and either can falsify a claim about the other. A zero slope of log-growth on log-size that coexists with a non-log-normal body is a contradiction the analyst must resolve (typically by finding the confound: survival selection, truncation, or too-short a time window).

Interventionist / counterfactual — what moves the distribution. Because the asymptotic shape is governed by just the drift μ and dispersion σ² of the per-period growth distribution, the analyst can reason about how a policy or shock that alters growth moments will reshape the population. Raise σ² (more volatile proportional shocks) and the predicted move is a wider log-normal — the size distribution spreads, inequality among firms grows, and it does so at a known rate (variance of log-size accumulating linearly in elapsed time). Introduce a genuine size advantage (make ε depend positively on S) and the prediction is a qualitative regime change: the tail thickens toward Pareto as the process leaves the Gibrat basin for preferential attachment. So an intervention is evaluated not firm-by-firm but by its effect on two growth moments and on the size-independence assumption itself.

Boundary-drawing — where the null is licensed. The law is explicitly a regime-restricted baseline, and a central reasoning move is deciding whether a given dataset sits inside that regime. Gibrat is licensed for large, established, surviving firms over a long-enough horizon for the random walk to wash out initial conditions; it is not licensed at the small-firm end (minimum efficient scale and fast early survivor growth break size-independence), not at the extreme upper tail (where heavy-tail generators dominate), and not for young firms (age effects overlay size effects). The analyst therefore brackets the population before testing — choosing the size band, excluding entrants, checking the horizon — because applying the null outside its regime manufactures spurious "violations" that are really regime-boundary artifacts.

Order-of-events / temporal. The law also supports reasoning about transients. Because dispersion in log-size grows monotonically with elapsed time under iid multiplicative noise, a cross-section that is too tight to be the asymptotic log-normal is read as a young or recently-reset population that has not yet diffused; convergence toward log-normality is itself a clock. This lets the analyst distinguish "Gibrat is false here" from "Gibrat is true but the process has not run long enough," reading the degree of spread as elapsed process-time rather than as evidence against the mechanism.

Knowledge Transfer

The honest account of Gibrat's law's transfer must keep its two levels apart, because they travel in opposite ways. As an empirical claim about firms — that a firm's growth rate is independent of its current size — it transfers as mechanism within industrial organization and firm dynamics: the diagnostic (regress log-growth on log-size; a zero coefficient supports the law), the menu of named departures (a negative small-firm slope flags minimum efficient scale and fast-survivor growth; a heavy Pareto upper tail flags preferential attachment or mergers; an age effect flags young firms), and the regime restriction (licensed for large, established, surviving firms over a long horizon) all carry intact across firm panels, market-share studies, and industrial-demography models of entry, exit, and merger. There the firm-specific content — establishments, employees, minimum efficient scale, survival selection — is load-bearing and stays home.

As a generative mechanism, however — size-independent multiplicative iid noise iterated over time produces a log-normal cross-section, because log-size executes an additive random walk and the central limit theorem applies — Gibrat's law is a piece of portable mathematical machinery, and it transfers literally wherever its precondition holds. This is the case where "mechanism within the home domain, metaphor beyond" does not apply: the construct is not a causal story bound to firms but a theorem about stochastic processes, so when the same multiplicative-shock structure genuinely obtains in another quantitative substrate, the log-normal conclusion follows for the same mathematical reason, not by analogy. And it does obtain elsewhere — in city-size distributions (log-normal in the body, Zipf at the top), word-frequency distributions (the Estoup-Zipf lineage), within-taxon biological body sizes, and asset prices under geometric Brownian motion, the continuous-time Gibrat analogue underlying Black-Scholes. In each, what transfers is not "Gibrat's law of firms" but the proportional-random-growth generator, applied to a new content domain that shares the substrate. The boundary to mark here is therefore not mechanism-versus-metaphor but instrument-reach versus over-reading: the generator's literal claim is "multiplicative iid noise → log-normal," and the standing error is to read a log-normal body as proof of size-independent growth when survival selection, truncation, or too-short a horizon can mimic or mask it, or to expect the generator to govern the upper tail, where it characteristically fails and a heavy-tail generator takes over.

This split also fixes where the cross-domain lesson properly lives in the catalog. The transferable substance is the generative-class logic — the choice between additive and multiplicative noise determines the distributional shape — which sits at the same level as the parent constructs it is a sibling of: the central_limit_theorem (the additive counterpart generating Gaussians), the preferential-attachment / Yule mechanism (generating Pareto), and the broader heavy_tailed_distributions and power_law families against which Gibrat supplies the thin-tailed log-normal null. When a cross-domain analyst needs "the shape this process should produce," that lesson is carried by those general generators, with Gibrat as the named multiplicative-noise-to-log-normal member; the firm-growth empirical law, with its size regressions and minimum-efficient-scale departures, stays in industrial organization. Invoking "Gibrat's law" loosely for any system where big things grow proportionally to small ones, without the multiplicative-iid-noise structure actually in place, is the one genuinely metaphorical use — it borrows the size-independence slogan while dropping the stochastic-process machinery that makes the original a derivable result (see Structural Core vs. Domain Accent).

Examples

Canonical

Gibrat's own object was the firm. Start a cohort of firms and let each period multiply size by an independent random factor whose distribution does not depend on current size: S(t+1) = S(t)·ε(t). Take logs and the process becomes an additive random walk, log S(t+1) = log S(t) + log ε(t); iterated over many periods, the central limit theorem drives the sum of the log-shocks toward a normal distribution, so log-size is approximately normal and size itself approximately log-normal, with the variance of log-size widening linearly in elapsed time. The empirical test falls straight out: regress each firm's log-growth on its log-size across a panel; a coefficient of zero says growth is size-independent (Gibrat holds), a negative coefficient says small firms grew faster. Studies of large established manufacturing firms recover a slope near zero and a roughly log-normal body, the law's home success.

Mapped back: The firm cohort is the population of entities; S(t+1) = S(t)·ε is the multiplicative growth step, and ε's distribution not depending on S is the size-independence assumption. Logging yields the log-additive random walk, and the central-limit convergence delivers the log-normal cross-section, its spread the linearly-growing dispersion. The zero regression slope is the empirical test passing.

Applied / In Practice

Urban economics uses the identical generator to explain city sizes. Cities grow by roughly proportional random increments — a metro that is twice as large tends to add population at a similar percentage rate, size-independently — so the same multiplicative-noise machinery predicts a log-normal body to the city-size distribution. Xavier Gabaix (1999) showed that adding a lower reflecting barrier (cities do not shrink below some floor) to Gibrat-style proportional growth generates Zipf's law — a Pareto upper tail with exponent near one — at the top of the distribution, which matches U.S. city data strikingly well. Gibrat supplies the log-normal null; the observed heavy tail is the named departure that a modified generator explains.

Mapped back: Cities are the population of entities under a multiplicative growth step with the size-independence assumption (percentage growth independent of city size). The central-limit convergence gives the log-normal body — a literal, non-metaphorical use of the generator on a new substrate. The Zipf/Pareto upper tail is one of the named departures: the generator characteristically fails at the extreme tail, where a boundary-modified or preferential-attachment mechanism takes over.

Structural Tensions

T1: The empirical regularity versus the mathematical theorem (one name, two incompatible epistemic statuses). Gibrat's law is at once a contingent empirical claim about firms — size-independent proportional growth, which holds only approximately and fails at both the small-firm and upper-tail edges — and an exact theorem about stochastic processes: multiplicative iid noise iterated over time yields a log-normal cross-section, true by the central limit theorem wherever its precondition holds. These statuses are opposite in kind: one is falsifiable and regime-restricted, the other unfalsifiable and universal. The standing error the entry flags is treating evidence at one level as settling the other — the empirical failure of Gibrat for small firms does not touch the theorem, and the theorem's truth does not confirm the empirical law for any actual firm panel. The tension is that a single label carries a testable regularity and an untestable mathematical identity, and reasoning slides between them unless the level is fixed. Diagnostic: Is the claim being tested an empirical regularity about these firms, or invoked as the multiplicative-noise theorem that is true by construction regardless of the data?

T2: A log-normal body as evidence of size-independent growth versus equifinality (the shape underdetermines the process). The signature diagnostic reads a roughly log-normal cross-section backward to a proportional, size-independent growth process. But distributional shape is weak testimony about its generator: survival selection, truncation, and too-short a horizon can each mimic or mask log-normality, and several distinct processes converge to a near-log-normal body, so observing the shape is far from proof that Gibrat's specific mechanism produced it. The backward inference the concept most relies on — histogram to process — is exactly the inference equifinality undermines, so a log-normal fit is consistent with, but not confirmation of, size-independent growth. The tension is that the construct's central evidential move treats a shape as testimony about a mechanism when multiple mechanisms yield the same shape. Diagnostic: Is the log-normal body being read as confirmation of size-independent growth, or as one shape that survival selection, truncation, and rival generators could equally have produced?

T3: A reliable null in the body versus systematic blindness to the tail that matters (the concept fails where the action is). Gibrat supplies a clean, tractable baseline for the log-normal body of a size distribution, and its value is largest as the well-behaved null against which departures are measured. But it characteristically fails at the upper tail, where preferential-attachment or merger dynamics drive a heavy Pareto/Zipf shape — and that tail is usually the economically and socially decisive part: the largest firms, the biggest cities, the concentration of wealth. So the null is most dependable precisely in the middle, where the stakes are lowest, and breaks down exactly at the extreme where the consequential phenomena live. The tension is that Gibrat's reliability as a baseline and its structural inability to govern the tail are two sides of the same thin-tailed generator, so leaning on it risks a well-fitted body lulling the analyst past the tail it cannot capture. Diagnostic: Is the log-normal null being applied to the body, where it is licensed, or stretched over the heavy tail where a different generator takes over and the consequential concentration actually resides?

T4: The dispersion clock as discriminator versus as an unfalsifiability escape. Because variance in log-size accumulates linearly with elapsed time, a cross-section too tight to be the asymptotic log-normal can be read as a young or recently-reset population that has not yet diffused — a genuine tool for distinguishing "Gibrat is false here" from "Gibrat is true but the process has not run long enough." But the same move is double-edged: any failure of the data to match log-normality can be rescued by appealing to insufficient elapsed process-time, so the clock that discriminates true-but-young from false can also immunize the null against disconfirmation. The tension is that the temporal reasoning which correctly withholds a false-verdict from an under-diffused population is the same reasoning that, applied freely, lets the analyst attribute every misfit to a too-short horizon. Diagnostic: Is the appeal to "the walk hasn't run long enough" backed by an independent estimate of elapsed process-time, or is it a movable excuse that would absolve any degree of misfit?

T5: Autonomy versus reduction (a firm-growth law or a member of the generative-class family). Uniquely, Gibrat's law transfers in two ways that resolve the autonomy question at different levels. As the empirical firm-growth claim — with its size regressions, minimum-efficient-scale departures, and survival selection — it is bounded to industrial organization, where that content is load-bearing. As a generative mechanism it transfers literally, not metaphorically, to city sizes, word frequencies, within-taxon body sizes, and asset prices under geometric Brownian motion, because it is a theorem, not a firm-specific story. What actually recurs cross-domain is the generative-class logic — the choice between additive and multiplicative noise fixes the distributional shape — which sits alongside its siblings central_limit_theorem (additive → Gaussian), the preferential-attachment/Yule mechanism (→ Pareto), and the heavy_tailed_distributions/power_law families, with Gibrat as the named multiplicative-noise-to-log-normal member. The tension is that the firm law is domain-bound while the generator is a general family member, and the loose "big things grow proportionally" invocation without the multiplicative-iid structure is the one genuinely metaphorical use. Diagnostic: Resolve toward the generative-class parents (CLT, Yule/preferential attachment, the power-law family) when carrying the shape-of-the-distribution lesson to a new substrate; toward Gibrat's firm law when testing size-independent growth in an actual firm panel — and refuse the name entirely where the multiplicative-noise structure is absent.

Structural–Framed Character

Gibrat's law sits toward the structural end of the spectrum but stops short of the pole — best read as mixed-structural, and closely analogous to how isostasy is characterized, though for a different reason. On four of the five criteria its structural credentials are strong. Its evaluative_weight is nil: size-independent proportional growth is neither good nor bad, and "Gibrat's law" convicts and praises nothing — it names a stochastic regularity, not a verdict. On human_practice_bound it points structural: the process runs observer-free — the tell is that the identical generator governs within-taxon biological body-size distributions, a substrate with no economist and no institution in it, so the mechanism plainly runs on populations and clocks rather than on a judging agent; no research practice constitutes it (a size regression merely tests it). On institutional_origin it also points structural: Gibrat (1931) named a regularity that firm populations already exhibit, the way Airy named isostatic balance, not an artifact of a survey or agency — with the mild caveat that firms themselves are human-economic entities. And on import_vs_recognize it is, if anything, more structural than isostasy: where isostasy travels beyond geophysics only by analogy, Gibrat's generative mechanism transfers literally — the entry is emphatic that "mechanism within the home domain, metaphor beyond" does not apply, because the generator is a theorem, so the log-normal conclusion follows for the same mathematical reason in city sizes, word frequencies, and asset prices as in firms. That is recognition of the same mechanism, not import-by-analogy.

What keeps it off the structural pole is the remaining criterion, vocab_travels, which it only half-passes — and that half-pass is what fixes the placement. Gibrat's vocabulary is split. The firm-level content — minimum efficient scale, survival selection, establishments, the age-vs-size departures — is irreducibly pinned to industrial organization and is exactly what makes "Gibrat's law" the named firm-growth empirical law; it does not travel. The generative-mechanism vocabulary — multiplicative iid noise, the log-additive random walk, central-limit convergence to log-normal — does float free and travels literally. But that free-floating part is precisely the portable skeleton the named law instantiates from a more general parent, not what makes "Gibrat's law" itself distinctive: as the entry insists, what recurs cross-domain is the generative-class logic, carried by the sibling parents central_limit_theorem (additive → Gaussian), the Yule/preferential-attachment mechanism (→ Pareto), and the heavy_tailed_distributions/power_law family, with Gibrat merely the named multiplicative-noise-to-log-normal member. So the criterion resolves the same way it does for isostasy: the substrate-spanning content already belongs to the umbrella, while the entry's own distinctive cargo — the regime-restricted, contingent, falsifiable firm law with its size regressions and named departures — stays home.

The one portable structural skeleton is the multiplicative-noise-to-log-normal generator — size-independent multiplicative iid shocks iterated over time drive log-size into an additive random walk whose central-limit convergence yields a log-normal cross-section. It is what Gibrat's firm law instantiates from its generative-class umbrella (sibling to central_limit_theorem, Yule/preferential attachment, and the power-law family), not what makes "Gibrat's law" itself travel: the cross-substrate reach belongs to that generator-family, while the firm-specific empirical claim is the domain-accented specialization. Its character: an evaluatively neutral, observer-free, discovered statistical regularity whose generative core transfers literally rather than by analogy — structural in that multiplicative-noise-to-log-normal skeleton borrowed from its generative-class umbrella — but held to mixed-structural because the named Gibrat's law is the regime-bound firm-growth empirical claim, stated in industrial-organization vocabulary that pins it to its home domain.

Structural Core vs. Domain Accent

This section decides why Gibrat's law is a domain-specific abstraction and not a prime — an unusual case, because its generative mechanism transfers literally across a wide field, yet what travels is a generative-class theorem owned by its parent primes, while the named law is the regime-bound firm-growth empirical claim.

What is skeletal (could lift toward a cross-domain prime). Strip the firms and a thin mathematical structure survives: size-independent multiplicative iid shocks iterated over time drive log-size into an additive random walk whose central-limit convergence yields a log-normal cross-section. The portable pieces are abstract — a population of entities carrying a size, a proportional (not additive) growth step whose distribution is independent of current size, and the distributional shape that follows by theorem. That skeleton is genuinely substrate-portable, which is why the cross-domain content is already carried at prime level by the generative-class family Gibrat is a sibling of: central_limit_theorem (the additive counterpart generating Gaussians), the Yule / preferential-attachment mechanism (generating Pareto), and the heavy_tailed_distributions / power_law families against which Gibrat supplies the thin-tailed log-normal null. The general lesson — the choice between additive and multiplicative noise fixes the distributional shape — is the core Gibrat shares; it is not what makes "Gibrat's law" distinctive.

What is domain-bound. What makes it Gibrat's law in particular is the empirical firm-growth claim and its industrial-organization furniture: the firm cohort with establishments and employees; the regime restriction to large, established, surviving firms over a long horizon; the log-growth-on-log-size regression as the field's test; and the menu of named departures — minimum-efficient-scale and fast-survivor growth at the small-firm end, Pareto preferential-attachment or mergers at the upper tail, age effects for young firms. This content is contingent, falsifiable, and regime-restricted (it holds only approximately and fails at both edges), and it is exactly what does not travel. The decisive test: invoking "Gibrat's law" loosely for any system where big things grow proportionally to small ones, without the multiplicative-iid-noise structure actually in place, is the one genuinely metaphorical use — it borrows the size-independence slogan while dropping the stochastic-process machinery that makes the original a derivable result. The named law is bound to firm dynamics; the theorem beneath it is not the named law.

Why this does not clear the prime bar. A prime's vocabulary travels unconditionally and its transfer is recognition of the same mechanism. Gibrat's transfer is genuinely bilevel. As the empirical firm-growth claim it travels as mechanism only within industrial organization — the regression diagnostic, the named departures, and the regime restriction carry across firm panels, market-share studies, and industrial demography, where the firm-specific content is load-bearing and stays home. As a generative mechanism it transfers literally — the identical multiplicative-noise-to-log-normal theorem governs city sizes, word frequencies, within-taxon body sizes, and asset prices under geometric Brownian motion, for the same mathematical reason, not by analogy. But that literal reach is exactly what makes the case: what recurs across those substrates is not "Gibrat's law of firms" but the proportional-random-growth generator applied to new content — and that generator is a sibling member of a generative-class family, so the substrate-spanning content already belongs to central_limit_theorem, Yule/preferential attachment, and the power_law/heavy_tailed_distributions parents, with Gibrat merely the named multiplicative-noise-to-log-normal member. When a cross-domain analyst needs "the shape this process should produce," the lesson is carried by those general generators; the firm-growth empirical law, with its size regressions and minimum-efficient-scale departures, stays in industrial organization. The cross-domain reach belongs to the generative-class umbrella; the named law is its regime-bound firm-growth specialization — which is exactly what keeps it a domain-specific abstraction below the prime bar.

Relationships to Other Abstractions

Local relationship map for Gibrat's LawParents appear above the current abstraction, mutual partners to the right, and children below. Node labels state whether each abstraction is prime or domain-specific; colors identify relation types.Gibrat's LawDOMAINPrime abstraction: Multiplicative Random Growth — is a decomposition ofMultiplicativeRandom GrowthPRIME

Current abstraction Gibrat's Law Domain-specific

Parents (1) — more general patterns this builds on

  • Gibrat's Law is a decomposition of Multiplicative Random Growth Prime

    Stripping Gibrat's firm-growth framing leaves Multiplicative Random Growth's size-independent proportional shocks, log-additive random walk, and conditional log-normal limit.

Hierarchy path (1) — routes to 1 parentless root

Not to Be Confused With

  • Zipf's law / Pareto / power-law distributions. Heavy-tailed distributions with a power-law upper tail. Gibrat's generator produces a log-normal cross-section, which is thinner-tailed; a heavy Pareto/Zipf tail is precisely the signature that Gibrat has failed and a different generator dominates the top. Conflating the two inverts what the deviation diagnoses. Tell: is the distribution power-law/heavy-tailed at the top (Zipf/Pareto — Gibrat's failure signature), or log-normal in the body from size-independent multiplicative noise (Gibrat)?

  • Preferential attachment / the Yule process. A growth generator in which the already-large accrue faster (rich-get-richer), producing power-law distributions. It is the rival generator at the upper tail and a sibling generative-class member; the size advantage it encodes is exactly what Gibrat's size-independence denies. Tell: does growth favor the already-large, yielding a heavy tail (preferential attachment), or is proportional growth size-independent, yielding a log-normal (Gibrat)?

  • The central limit theorem (the additive parent). The theorem that sums of iid shocks converge to a Gaussian. Gibrat is its multiplicative sibling: products of size-independent iid shocks make log-size a random walk that, by the CLT applied in logs, converges to normal — i.e., log-normal in levels. Tell: is the noise added so the sum is Gaussian (CLT), or multiplied so log-size random-walks to log-normal (Gibrat)?

  • Geometric Brownian motion. The continuous-time analogue of Gibrat's proportional random growth — the process underlying Black-Scholes for asset prices. It is the same generator in continuous time; Gibrat's law is the discrete, firm-focused empirical claim and its log-normal consequence. Tell: is it the continuous-time proportional-random-growth process for prices (GBM), or the discrete firm-growth empirical claim (Gibrat's law)?

  • The log-normal distribution (the output). The cross-sectional shape Gibrat's process produces. Gibrat's law is the generative process/claim (size-independent multiplicative growth); the log-normal is its asymptotic consequence, and observing a log-normal is evidence about, not proof of, that process (survival selection or truncation can mimic it). Tell: is the object the static distributional shape (log-normal), or the dynamic size-independent-growth process that generates it (Gibrat's law)?

  • The generative-class parents (central_limit_theorem, Yule/preferential attachment, power_law/heavy_tailed_distributions). The family whose members' choice between additive and multiplicative noise fixes the distributional shape; Gibrat is the named multiplicative-noise-to-log-normal member. When a cross-domain analyst needs "the shape this process should produce," these carry it. Tell: strip the firm-specific content and what recurs across cities, words, and prices — the noise-type-fixes-shape logic — is carried by these parents, not by "Gibrat's law of firms." (Treated fully in a later section.)

Neighborhood in Abstraction Space

Gibrat's Law sits in a sparse region of the domain-specific corpus (85th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Statistical Paradoxes & Distributional Structure (11 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-07-12