Yule–Simon Distribution¶

A one-parameter distribution on positive integers with beta-function probability mass and a power-law tail, associated with cumulative-advantage frequency models.

Version: v1 · 2026-08-30 · History
Domain-specific #: 3134
Origin domain: statistics

Core Idea¶

The Yule–Simon distribution is a discrete probability law on the positive integers. With shape parameter \(\rho>0\), its probability mass function is

\[ \Pr(K=k)=\rho B(k,\rho+1) =\rho\frac{\Gamma(k)\Gamma(\rho+1)}{\Gamma(k+\rho+1)}, \qquad k=1,2,\ldots. \]

The beta-function law produces a power-law mass tail,

\[ \Pr(K=k)\sim \rho\Gamma(\rho+1)k^{-(\rho+1)}. \]

SciPy and Stan document the same standardized positive-integer law and parameter domain.^[1]^[2] Its stable identity combines support, formula, one shape parameter, and tail behavior. A classic generative interpretation comes from cumulative advantage: new types enter while existing types are selected in proportion to their current frequency, yielding the distribution as a limiting frequency law under the Simon model.^[3] The law is autonomous from that one story, but the mechanism explains why it repeatedly models “counts of occurrences per type.”

Structural Signature¶

Sig role-phrases:

the positive-integer variate — \(K\in\{1,2,\ldots\}\), typically a count conditional on a type having appeared;
the shape parameter — \(\rho>0\), controlling concentration and tail decay;
the beta-function mass law — \(p_k=\rho B(k,\rho+1)\);
the recurrence — \(p_{k+1}/p_k=k/(k+\rho+1)\), fixing successive masses;
the power-law tail — exponent \(\rho+1\) for the probability mass;
the moment thresholds — mean finite only for \(\rho>1\), variance finite only for \(\rho>2\);
the cumulative-advantage interpretation — optional process semantics linking repeated selection to the limiting law.

Recognition test: verify support begins at one, fit or specify a single positive \(\rho\), and check the beta-function mass—not just a visually straight log-log tail. A generic heavy-tailed count distribution or any preferential-attachment process is not enough.

What It Is Not¶

The distribution is not Zipf's law, which usually states a rank-frequency relation rather than this probability mass over positive counts. It is not the zeta distribution: both have power-law tails, but their exact normalizations and probability ratios differ. It is not a continuous Pareto distribution, whose support and density are continuous.

It is also not identical to the Simon stochastic process. A process specifies sequential arrival and reinforcement; the Yule–Simon distribution is a marginal or limiting law. The same law can be generated through other representations, including a geometric mixture with an exponential latent variable used in random-variate generation.^[1] Nor is every observed heavy tail evidence of cumulative advantage; alternative mechanisms and misspecification must be considered.

Scope of Application¶

The law belongs to probability theory, frequency modeling, and discrete heavy-tail statistics. Simon developed a family of skew distributions for phenomena such as word frequencies, city sizes, and biological genera under cumulative advantage.^[3] Yule's earlier evolutionary model studied the distribution of species among genera and supplied the historical branching context from which the name derives.^[4]

Modern uses treat \(K\) as the frequency or size of a type selected from a population of types: occurrences per word, members per category, or other positive counts. The same formula may be used phenomenologically without claiming the literal Simon process, but the analyst must preserve the sampling unit. A distribution over type frequencies is not automatically the distribution of ranks or of token identities.

Clarity¶

The name distinguishes three layers: the exact distributional family, its asymptotic tail, and a possible generating mechanism. A dataset may approximate a power law without following the beta-function probabilities at small and moderate \(k\). Conversely, fitting the Yule–Simon law does not prove preferential attachment caused the observations.

Parameterization must be explicit. This entry uses \(\rho\) in \(p_k=\rho B(k,\rho+1)\). Some software calls the shape parameter alpha.^[1] In the classic innovation formulation with probability \(a\in(0,1)\) of a new type and proportional selection otherwise, the corresponding limit uses \(\rho=1/(1-a)\); that mechanistic parameterization occupies only \(\rho>1\).^[5] Reusing the symbol \(\alpha\) for both roles without the map creates errors.

Manages Complexity¶

One parameter compresses an infinite probability table while preserving exact mass ratios, tail exponent, and moment-existence thresholds. Instead of estimating a separate probability for every count, an analyst estimates \(\rho\) and obtains a coherent normalized law. The recurrence supports stable numerical evaluation, and the tail approximation supports qualitative risk reasoning.

The compression hides heterogeneity, cutoffs, finite-population effects, and time variation. A one-parameter law cannot separately adjust head shape and tail exponent. Mixtures, truncations, or two-parameter generalizations may fit data better but are no longer the ordinary Yule–Simon identity. Model parsimony must not be mistaken for universal adequacy.

Abstract Reasoning¶

Gamma-function asymptotics give \(\Gamma(k)/\Gamma(k+\rho+1)\sim k^{-(\rho+1)}\), establishing the mass-tail exponent. The adjacent-probability ratio follows directly:

\[ \frac{p_{k+1}}{p_k} =\frac{\Gamma(k+1)}{\Gamma(k)} \frac{\Gamma(k+\rho+1)}{\Gamma(k+\rho+2)} =\frac{k}{k+\rho+1}. \]

The mean and variance are

\[ \mathbb E[K]=\frac{\rho}{\rho-1}\quad(\rho>1), \qquad \operatorname{Var}(K)= \frac{\rho^2}{(\rho-1)^2(\rho-2)}\quad(\rho>2), \]

with divergence outside those stated domains.^[1] These thresholds permit immediate diagnostics: a fitted \(\rho\le2\) makes ordinary variance-based summaries theoretically unstable even though every observation is finite.

Knowledge Transfer¶

The distribution transfers literally among datasets whose observational unit is a positive count per type and whose mass shape is compatible with the beta law. Formula, estimation, random generation, and moment diagnostics then carry unchanged. A biological genus size and a word frequency can instantiate the same probability family without claiming the substrates are otherwise alike.

The process interpretation transfers more cautiously. Cumulative advantage requires a flow of innovations and reinforcement proportional to current frequency. If a domain lacks those roles, only the distributional fit transfers. Conversely, a preferential-attachment process with aging, fitness, memory, or nonlinear reinforcement may yield a different law. Mechanism and marginal law must be mapped separately.

Examples¶

Parameter \(\rho=2\). The first probabilities are

\[ p_1=2B(1,3)=\frac23, \quad p_2=2B(2,3)=\frac16, \quad p_3=2B(3,3)=\frac1{15}. \]

The recurrence confirms \(p_2/p_1=1/4\) and \(p_3/p_2=2/5\). The mass tail has exponent \(3\). The mean equals \(2\), but the variance is not finite because the variance condition is strict: \(\rho>2\), not \(\rho\ge2\).^[1]

Innovation mapping. If a Simon process introduces a new type with probability \(a=0.2\), the standard limiting parameter is \(\rho=1/(1-0.2)=1.25\).^[5] The associated mean exists, while variance does not because \(1.25\le2\). This is a process-to-law inference under stated stationarity and proportional-selection assumptions, not a claim about every innovation process.

Nonexample. A continuous Pareto measurement beginning at \(x_{\min}>0\) may have a power-law density but lacks positive-integer support and beta-function mass. Tail resemblance alone does not establish Yule–Simon identity.

Structural Tensions¶

Exact law versus tail resemblance. Many distributions look linear on a log-log plot. Diagnostic: compare the beta-function probabilities or likelihood across the full observed support, not only the far tail.
Fit versus causal story. A good fit does not prove cumulative advantage. Diagnostic: test for innovation and frequency-proportional selection using process data independently of marginal fit.
Parsimony versus head flexibility. One parameter couples early masses and tail exponent. Diagnostic: inspect systematic residuals at small \(k\) before accepting the family.
Autonomy versus reduction. Probability Distribution and Heavy-Tailed Distribution describe the genus and broad shape but do not entail this beta law, support, recurrence, or moment thresholds. Diagnostic: if those identities affect inference, an autonomous Yule–Simon residual remains.

Structural–Framed Character¶

The probability law is strongly structural: support, mass formula, recurrence, asymptotic exponent, and moments are mathematical invariants. The Yule–Simon name is historically framed by Yule's branching-frequency work and Simon's cumulative-advantage distributions.^[4]^[3] Application labels such as “word frequency” are examples, not components of the definition.

The distribution itself is descriptively neutral. Evaluative choices enter through sampling, model selection, and causal interpretation. The same random-variable law survives a substrate change from taxa to lexical types; the causal model survives only when its innovation and reinforcement roles are also present.

Structural Core vs. Domain Accent¶

The portable core is a normalized heavy-tailed law on positive integers. The domain accent is probabilistic and statistical: random variable, beta and gamma functions, shape parameter, mass likelihood, moments, sampling unit, and limiting stochastic process.

It does not clear the prime bar. Its applications span subject matter, but they are literal uses of one probability distribution inside a single mathematical-statistical substrate. domain_specific:probability_distribution supplies the genus, while Heavy-Tailed Distribution supplies a broad neighbor. The named formula is a domain-specific family member.

The proposed parent is domain_specific:probability_distribution, with strict specialization: the Yule–Simon law completely specifies probability mass over a discrete outcome space. It relates to Heavy-Tailed Distributions through polynomial decay and to Cumulative Advantage through the Simon mechanism.

Heavy-Tailed Distribution is not used as the minimal parent because it is a qualitative class shared by many exact families; Probability Distribution is the literal taxonomic genus. Cumulative Advantage describes one generator, not every use of the law. The proposal makes no live DAG mutation.

Relationships to Other Abstractions¶

Current abstraction Yule–Simon Distribution Domain-specific

Parents (1) — more general patterns this builds on

Yule–Simon Distribution is a kind of Probability Distribution Domain-specific

The proposed parent is domain_specific:probability_distribution, with strict specialization: the Yule–Simon law completely specifies probability mass over a discrete outcome space.

Neighborhood in Abstraction Space¶

Yule–Simon Distribution sits in a sparse region of the domain-specific corpus (73^rd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Random Variable — 0.85
Monotone Likelihood Ratio Property — 0.84
K-Distribution — 0.84
Empirical Measure — 0.83
Probability Distribution — 0.83

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Simon model: a sequential innovation-and-reinforcement process; distinguish generator from resulting law.
Yule process: a branching process with related historical distributions, not automatically the standardized Yule–Simon mass.
Zipf law: a rank-frequency scaling relation rather than this count probability mass.
Zeta distribution: a different exact discrete power-law family with normalization by the zeta function.
Pareto distribution: a continuous heavy-tailed law.
Generalized Yule–Simon distribution: a multi-parameter or cutoff extension, not the one-parameter family defined here.

References¶

[1] SciPy Community, “Yule-Simon Distribution,” SciPy v1.17.0 Manual, accessed 2026-08-29, https://docs.scipy.org/doc/scipy/tutorial/stats/discrete_yulesimon.html. registry ↩a ↩b ↩c ↩d ↩e

[2] Stan Development Team, “Yule Simon Distribution,” Stan Functions Reference 2.39, accessed 2026-08-29, https://mc-stan.org/docs/2_39/functions-reference/unbounded_discrete_distributions.html. registry ↩

[3] Herbert A. Simon, “On a Class of Skew Distribution Functions,” Biometrika 42, nos. 3–4 (1955): 425–440, https://doi.org/10.1093/biomet/42.3-4.425. registry ↩a ↩b ↩c

[4] G. Udny Yule, “A Mathematical Theory of Evolution, Based on the Conclusions of Dr. J. C. Willis,” Philosophical Transactions of the Royal Society B 213 (1925): 21–87, https://doi.org/10.1098/rstb.1925.0002. registry ↩a ↩b

[5] Fabrizio Leisen, Luca Rossini, and Cristiano Villa, “Objective Bayesian Analysis of the Yule–Simon Distribution with Applications,” Computational Statistics 33 (2018): 99–126, https://doi.org/10.1007/s00180-017-0735-1. registry ↩a ↩b