Preprint
Article

This version is not peer-reviewed.

An Information-Geometric Framework for Probabilistic Grammars and Rank-Frequency Scaling Dynamics

Submitted:

14 November 2026

Posted:

16 November 2026

You are already at the latest version

Abstract
We develop an information-geometric framework modeling the continuous evolution of probabilistic grammars as variational trajectories on a statistical manifold endowed with the Fisher–Rao metric, and investigate when these flows exhibit power-law rank–frequency scaling. Idealized grammar evolution is postulated to follow Fisher geodesics, representing a variational modeling hypothesis rather than a consequence of the metric itself. Asymptotic scaling is characterized via rank-dilation covariance: if \( p(\lfloor\lambda r\rfloor)/p(r) \to \lambda^{-\alpha} \) for all \( \lambda > 0 \), Karamata’s regular variation theory implies \( p(r) = r^{-\alpha}L(r) \) with \( L \) slowly varying, independent of manifold geometry. To link parameter-space geometry and rank-space scaling, we formulate a falsifiable Geometric Scaling Conjecture. For a Gibbs family \( p_n(\beta) \propto e^{-\beta E_n} \) with an asymptotically logarithmic spectrum \( E_n \sim \log n \), rank-dilation covariance holds for \( \beta > 1 \) with \( \alpha = \beta \), and Fisher geodesics dictate the evolution speed of \( \alpha(s) = \beta(s) \). For the zeta distribution, \( g_{\beta\beta} \sim (\beta-1)^{-2} \) as \( \beta \to 1^+ \), placing the classical Zipf boundary \( \alpha=1 \) at infinite Fisher distance. Finally, regular and multi-agent grammar examples demonstrate how grammatical coupling induces non-flat Fisher curvature and prices state transitions.
Keywords: 
;  ;  ;  ;  ;  ;  ;  

1. Introduction

A parametric family of probability distributions can be naturally formalized as a differentiable manifold endowed with an intrinsic geometric structure, where local statistical distinguishability is uniquely quantified by the Fisher information metric [1,3]. Within the foundational framework of information geometry, a statistical model
P = { p ( x | θ ) ∣ θ ∈ Θ }
is conceptualized as a Riemannian manifold parameterized by the coordinate vector θ ∈ Θ . The metric tensor field g i j ( θ ) is defined via the expectation of the score functions as:
g i j ( θ ) = ∫ X p ( x | θ ) ∂ i log p ( x | θ ) ∂ j log p ( x | θ ) d x
Under Chentsov’s uniqueness theorem [4], the Fisher–Rao metric operates as the geometrically invariant Riemannian metric under sufficient statistics, providing an intrinsic measure of distance on the statistical manifold. Consequently, variational principles and geodesic flows dictated by this metric offer a coordinate-free framework to characterize the continuous deformation of probability measures, avoiding externally prescribed or heuristically driven cost functionals [2].
Probabilistic formal grammars—such as Probabilistic Context-Free Grammars (PCFGs) and stochastic automata—constitute a compelling domain for the application of this geometric paradigm [6]. A grammar structurally assigns probability weights to a discrete or continuous set of alternative productions X , where the parameter vector θ completely dictates the state of the grammar. The space of all admissible distributions induced by these production rules defines a statistical sub-manifold M ⊂ P .
Conventionally, the evolution or optimization of probabilistic grammars is modeled via discrete algorithmic updates, stochastic learning rules, or heuristic selection functions. In contrast, this work posits that the parametric trajectory θ ( t ) of a probabilistic grammar can be governed intrinsically by the underlying geometry of the probability model. By constructing an action functional based on the Fisher arc length, the continuous evolution of the grammar is derived as a variational problem on M :
S [ θ ] = ∫ t 0 t 1 g i j ( θ ) θ ˙ i θ ˙ j d t
The stationary curves of S [ θ ] yield the classical Euler–Lagrange equations, which manifest as the Levi-Civita geodesic equations:
θ ¨ k + Γ i j k θ ˙ i θ ˙ j = 0
where Γ i j k denote the Christoffel symbols of the second kind associated with the Fisher metric. This formulation yields an intrinsic statistical selection mechanism: the redistribution of probability mass among competing grammatical productions is driven strictly by the local minimization of statistical distinguishability cost, rather than by an arbitrary, externally imposed utility or fitness function.
While parameter-space dynamics are cleanly captured by Fisher geodesics, a longstanding mathematical disconnect persists when mapping these trajectories to empirical rank-frequency distributions. The Fisher metric lives natively on the parameter manifold Θ , whereas rank-frequency representations—such as Zipf’s law—are empirical regularities defined over an ordered discrete index space r ∈ N labeling the magnitude of the ordered probabilities [7]:
p ( 1 ) ≥ p ( 2 ) ≥ ⋯ ≥ p ( r ) ≥ ⋯
To construct a rigorous analytical bridge between parameter-space differential geometry and asymptotic rank space, we formulate the asymptotic response of the distribution under rank dilations. We say that the ranked sequence exhibits scale-preserving properties if it satisfies an asymptotic rank-dilation invariance condition for a scale factor λ > 0 :
lim r → ∞ p ( ⌊ λ r ⌋ ) ( s ) p ( r ) ( s ) = λ − α ( s )
By invoking Karamata’s regular variation theory [5], this limiting ratio implies that the ranked probability sequence belongs to the class of regularly varying functions at infinity with index − α ( s ) . Consequently, the tail of the distribution admits the canonical representation:
p ( r ) ( s ) = r − α ( s ) L ( r ; s )
where L ( r ; s ) is a slowly varying function satisfying lim r → ∞ L ( λ r ) / L ( r ) = 1 .
Through this formulation, generalized Zipf-type scaling is decoupled from ad hoc log-log regression assumptions and reframed as an asymptotic scale-covariance condition in rank space. Crucially, however, Fisher–Rao geometry on the parameter manifold does not by itself generate power-law rank distributions or select whether scaling occurs; rather, it governs the continuous trajectory and rate of evolution of the model parameters once a statistical family is specified. The primary objective of this paper is therefore to establish a coherent information-geometric framework and formulate a testable Geometric Scaling Conjecture linking geodesic parameter flows to regularly varying rank distributions. We demonstrate that for statistical families possessing an asymptotically logarithmic configuration spectrum, Fisher-geodesic evolution maps directly onto the continuous evolution of the rank-dilation exponent α ( s ) = β ( s ) , while the Fisher metric controls the speed profile and boundary accessibility of the flow. By separating intrinsic geometric properties from spectral and asymptotic assumptions, this work provides a falsifiable mathematical bridge between variational grammar dynamics and the evolution of power-law scaling in statistical systems.

1.1. Summary of Main Contributions to Be Made

To address the analytical disconnect between differential-geometric parameter flows and discrete rank–frequency regularities, this work will establish a rigorous, falsifiable framework connecting information geometry to asymptotic power-law scaling. Specifically, our primary contributions to be found below are as follows:
  • Variational Formulation of Continuous Grammar Dynamics: By equipping the parameter manifold M of a probabilistic formal grammar with the invariant Fisher–Rao metric g i j ( θ ) , we model idealized continuous grammar evolution as stationary curves of the Fisher arc-length functional. This yields Levi-Civita geodesic trajectories driven intrinsically by local statistical distinguishability rather than exogenous heuristic fitness functions.
  • Rank-Dilation Covariance and Regular Variation: Moving beyond ad hoc log–log linear regressions over finite samples, we formulate generalized Zipf scaling as an asymptotic rank-dilation covariance condition, lim r → ∞ p ( ⌊ λ r ⌋ ) ( s ) / p ( r ) ( s ) = λ − α ( s ) for all λ > 0 . Invoking Karamata’s representation theorem, we prove that this symmetry characterizes the class of regularly varying probability sequences p ( r ) ( s ) = r − α ( s ) L ( r ; s ) independently of the underlying manifold geometry.
  • The Geometric Scaling Conjecture and Explicit Boundary Delimitation: We formulate a testable Geometric Scaling Conjecture and establish exact analytical boundaries separating geometric kinematics from spectral assumptions. Using an infinite-state Gibbs family p n ( β ) ∝ e − β E n with an asymptotically logarithmic spectrum E n ∼ log n , we prove that Fisher geodesics govern the evolution speed of the scaling exponent α ( s ) = β ( s ) , and show that for the zeta distribution ( g β β ∼ ( β − 1 ) − 2 as β → 1 + ), the non-normalizable classical Zipf limit ( α = 1 ) lies at infinite Fisher distance and is unreachable in finite evolution parameter.
  • Curvature and Transition Pricing in Formal Grammars: Through closed-form worked examples—including a bifurcated regular grammar, a correlated multi-agent simplex model, and a Bernoulli grammar—we demonstrate how grammatical production rules and inter-agent couplings induce non-flat constant Fisher curvature ( K = ± 1 / 4 ) and determine the infinitesimal information-geometric cost of probability mass transfers between states.

1.2. Historical Context and Interdisciplinary Scope

Zipf’s law originated in empirical linguistics through the work of George Kingsley Zipf, who observed that in large text corpora, word frequency f decays inversely with rank r ( f ∝ r − 1 ). Mathematically, this regularity corresponds to a discrete power-law probability distribution whose tail behavior is foundational to statistical inference and information theory. Beyond quantitative linguistics, generalized rank–frequency power laws characterize city population sizes and income distributions in economics, critical phenomena and statistical mechanics in physics, and network topology, data compression, and web traffic modeling in computer science.

2. This Work’s Role in Information Geometry

The information-geometric character of the present approach lies in treating a probabilistic grammar as a statistical manifold whose intrinsic geometry is determined by the Fisher information metric. The parameters specifying the production probabilities are therefore not regarded merely as coordinates of a dynamical model, but as coordinates on a manifold endowed with a natural measure of local statistical distinguishability. Idealized continuous grammar evolution is represented by curves on this manifold, and the variational principle considered here selects Fisher geodesics as a distinguished class of constant-speed trajectories. The resulting geometric flow governs the evolution of the probability distributions generated by the grammar and, consequently, the parameter dependence of their ordered rank-frequency distributions. The central problem of this work is to determine the conditions under which this information-geometric evolution is accompanied by an asymptotic scaling structure in rank space. We formalize this link through the rank-dilation ratio and connect it to Karamata’s theory of regular variation, thereby establishing a precise analytical bridge between Fisher–Rao geometry in parameter space and Zipf-type scaling in probability space. Rather than postulating Zipf’s law as an external empirical input or claiming that geodesic flow alone produces power-law tails, this framework isolates the spectral conditions required for scale covariance and characterizes how Fisher geometry governs the trajectory and boundary accessibility of the resulting scaling exponent.

3. Regular Variation and Asymptotic Scale Covariance

To establish a rigorous mathematical link between the differential geometry of a parameter manifold and the emergent scaling properties in rank space, it is necessary to formalize the behavior of probability measures under dilatation operators. While a statistical model’s local geometry is governed by its metric tensor, its macroscopic, large-rank structure is fundamentally characterized by asymptotic scale covariance. In this section, we develop the framework of regular variation as the precise mathematical vehicle for this asymptotic structural analysis.

3.1. Asymptotic Scale Covariance

In standard statistical analysis, structural properties are often evaluated via the localized covariance of random vectors. Let ( X , Σ , μ ) be a probability space, and let X = [ X 1 , X 2 , ⋯ , X n ] T denote a random vector mapping into R n . The centered second-moment structure is captured by the symmetric, positive semi-definite covariance operator Σ ( X ) ∈ GL ( n , R ) . Under a linear change of scale parameterized by a non-singular transformation matrix A ∈ GL ( n , R ) , the covariance tensor exhibits strict scale covariance satisfying the algebraic transformation law:
Σ ( A X ) = A Σ ( X ) A T
While this relation dictates how localized variance scales under finite coordinate transformations, it does not capture the behavior of probability tails or ordered sequences in infinite-dimensional spaces. To bridge this gap, regular variation relaxes strict algebraic scale covariance into an asymptotic condition. This enables the characterization of functions and sequences that become structurally invariant under multiplicative dilations only as their arguments approach infinity, isolating the leading scaling behavior from sub-leading structural perturbations.

3.2. Mathematical Foundations of Regular Variation

Let f : R + → R + be a measurable function that is eventually positive. The function f is said to be regularly varying at infinity with index ρ ∈ R (denoted by f ∈ RV ρ ) if it satisfies the asymptotic scale-covariance relation:
lim x → ∞ f ( λ x ) f ( x ) = λ ρ ∀ λ > 0
The parameter ρ operates as the scaling exponent, determining the fundamental response of the function to a multiplicative scaling or dilation of its argument by the factor λ . If ρ = 0 , the function belongs to the class RV 0 of slowly varying functions, typically denoted by L ( x ) . By definition, a slowly varying function satisfies:
lim x → ∞ L ( λ x ) L ( x ) = 1 ∀ λ > 0
The structural decomposition of regularly varying functions is governed by Karamata’s Representation Theorem [5]. Under standard local integrability assumptions, any function f ∈ RV ρ admits a unique canonical factorization into a pure scaling factor and a slowly varying correction:
f ( x ) = x ρ L ( x ) = x ρ · c ( x ) exp ∫ x 0 x ε ( t ) t d t
where c ( x ) → c ∈ ( 0 , ∞ ) and ε ( t ) → 0 as x , t → ∞ . This representation demonstrates that a regularly varying function may deviate substantially from a pure power law over any compact interval [ x 0 , x 1 ] , yet its asymptotic trajectory under multiplicative dilation is strictly dominated by the power-law exponent ρ . Prototypical examples of such slowly varying corrections include iterated logarithmic configurations, L ( x ) = ( log x ) γ , which satisfy the limit condition:
lim x → ∞ log ( λ x ) γ log ( x ) γ = lim x → ∞ 1 + log λ log x γ = 1

3.3. Regularly Varying Rank-Frequency Sequences

The mathematical machinery of regular variation translates naturally to discrete spaces by mapping the continuous argument x to the discrete rank variable r ∈ N . Let p ( r ) r ≥ 1 denote a globally normalized probability sequence arranged in non-increasing order:
p ( 1 ) ≥ p ( 2 ) ≥ ⋯ ≥ p ( r ) ≥ ⋯ > 0
A rank-frequency distribution possesses a regularly varying tail with a scaling index − α if the discrete sequence satisfies the rank-dilation condition under the floor operator:
lim r → ∞ p ( ⌊ λ r ⌋ ) p ( r ) = λ − α ∀ λ > 0
Applying the discrete analogue of Karamata’s theorem, any sequence satisfying the above equation can be written in the generalized scaling form:
p ( r ) = r − α L ( r )
where L ( r ) is a discrete slowly varying sequence. This formulation provides a rigorous mathematical foundation for generalized Zipf scaling. It circumvents the empirical ambiguity of fitting finite sequences to linear regressions on log-log scales, which frequently convolutes sub-leading corrections or finite-size cutoffs with true asymptotic scaling behavior.

3.4. The Classical Zipf Paradigm and Normalization Constraints

The classical Zipf distribution constitutes a strict singular case within the broader family of regularly varying sequences, corresponding to a scaling exponent of:
α = 1
In this regime, the scale-covariance relation under rank dilation simplifies to:
lim r → ∞ p ( ⌊ λ r ⌋ ) p ( r ) = λ − 1
When evaluating this scaling relation on infinite supports ( X = N ), an inherent topological tension arises between the asymptotic scaling law and global probability normalization. A pure power law p ( r ) = C r − α is normalizable on an infinite support if and only if it satisfies the strict inequality:
α > 1
in which case the normalization constant is given by the inverse of the Riemann zeta function, C = ζ ( α ) − 1 . The classical Zipf limit ( α = 1 ) is non-integrable over an infinite discrete horizon, as the harmonic series diverges. Consequently, for an information-geometric dynamics to yield a well-defined, globally normalized probability model that asymptotically approaches the Zipf limit, the model must structurally incorporate either a finite effective support N, an exponential truncation cutoff, or a modulating slowly varying function L ( r ) such that ∑ r = 1 ∞ r − 1 L ( r ) < ∞ . By formulating Zipf scaling as a special case within the generalized framework of regular variation, we isolate the dynamics of the scaling exponent α from global normalization constraints. This structural independence allows us to construct an analytical bridge where continuous parameter deformations along Fisher geodesics map directly into the smooth evolution of the asymptotic rank-scaling exponent α ( s ) .

4. Probabilistic Grammars and Fisher Geometry

4.1. Probabilistic Grammars

A probabilistic formal grammar defines a generative probability measure over admissible symbolic configurations or sequences. Let G = ( Σ , N , P , S , π ) denote a probabilistic grammar, where Σ is the finite alphabet of terminal symbols, N is the set of non-terminal symbols, P is the set of admissible rewrite productions, S ∈ N is the start symbol, and π specifies the production probability assignments. The production probabilities are collected into a parameter vector
θ = ( θ 1 , ⋯ , θ d ) ∈ Θ ,
where Θ is the corresponding parameter space. For a given value of θ , the grammar induces a probability distribution over the set of admissible configurations x ∈ X ,
p ( x | θ ) , ∑ x ∈ X p ( x | θ ) = 1 .
The statistical description of the grammar is therefore determined not only by its symbolic production rules, but also by the probability distribution generated by those rules. We consider the family
M = p ( x | θ ) ; θ ∈ Θ
as a statistical manifold, provided the usual regularity conditions are satisfied. The parameter vector θ may represent independent production probabilities, a reduced set of independent parameters subject to normalization constraints, or any other coordinates describing the probabilistic grammar. The formulation below does not require a particular choice of grammar or of parameterization.

4.2. Statistical Manifold and Fisher Information

For a regular parametric family, neighboring probability distributions can be compared through their statistical distinguishability. The infinitesimal distance between two neighboring points of the statistical manifold is determined by the Fisher information metric,
d s 2 = g i j ( θ ) d θ i d θ j ,
where repeated indices are summed over. For a discrete configuration space, the Fisher information matrix is
g i j ( θ ) = ∑ x ∈ X p ( x | θ ) ∂ i log p ( x | θ ) ∂ j log p ( x | θ ) ,
with ∂ i ≡ ∂ ∂ θ i . For a continuous configuration space, the corresponding expression is
g i j ( θ ) = ∫ X p ( x | θ ) ∂ i log p ( x | θ ) ∂ j log p ( x | θ ) d x .
Under the standard regularity conditions, the score functions ∂ i log p ( x | θ ) satisfy
∑ x p ( x | θ ) ∂ i log p ( x | θ ) = 0 ,
or, in the continuous case,
∫ p ( x | θ ) ∂ i log p ( x | θ ) d x = 0 .
Consequently, the Fisher metric may equivalently be written as the covariance matrix of the score functions.
A statistical model is locally identifiable at a specific parameter point if there exists an open neighborhood around that point where no other parameter values yield the exact same probability distribution for the observed data. In simpler terms, while a globally identifiable model ensures that every single point in the entire parameter space maps to a completely unique data distribution, a locally identifiable model guarantees uniqueness only within a restricted, nearby region. When the statistical model is locally identifiable, the metric is positive definite. More generally, the Fisher information matrix is positive semidefinite and may possess null directions associated with redundant or statistically indistinguishable parameter directions. A null vector has a size of exactly zero. Because it has no length, mathematicians define its direction as arbitrary, undefined, or pointing everywhere and nowhere at once. It is essentially a single point in space.
For sufficiently regular models one also has the alternative representation
g i j ( θ ) = ∑ x p ( x | θ ) ∂ i ∂ j log p ( x | θ ) ,
with the analogous integral expression for continuous variables. The Fisher metric is invariant under smooth reparameterizations of the statistical model. Thus, although the coordinates θ used to describe the grammar are not unique, the corresponding information geometry is an intrinsic property of the statistical family.

4.3. Statistical Distinguishability and Infinitesimal Evolution

Consider two nearby parameter values, θ and θ + d θ . The Kullback–Leibler divergence between the corresponding probability distributions has the local expansion
D KL p ( x | θ ) | p ( x | θ + d θ ) = 1 2 g i j ( θ ) d θ i d θ j + O ( d θ 3 ) .
Thus the Fisher metric quantifies the leading-order statistical distinguishability generated by an infinitesimal change of the grammar parameters. A trajectory θ = θ ( s ) therefore has an information-geometric length
L = ∫ s 0 s 1 g i j ( θ ( s ) ) d θ i d s d θ j d s d s .
This geometric length will be used in Sec. Section 5 to formulate the dynamical evolution of probabilistic grammars. It is useful to emphasize that the Fisher metric acts on the parameter manifold Θ . It does not directly define a metric on the rank variable r associated with an ordered probability distribution. The distinction between these two spaces is central to the rank-frequency problem considered below. The parameter-space geometry determines how the probability model evolves, whereas rank-frequency scaling concerns the asymptotic structure of the probabilities after they have been ordered.

4.4. Fisher Geometry of a Probabilistic Grammar

The construction above associates with a probabilistic grammar the geometric object ( M , g i j ) . A point on M corresponds to a particular probabilistic grammar, while a curve on M represents a continuous deformation of the probability assignments defining that grammar. The Fisher line element d s 2 = g i j d θ i d θ j then provides a measure of the statistical cost of such an infinitesimal deformation. In this sense, Fisher geometry supplies a natural notion of distance between neighboring probabilistic grammars. The present paper considers the possibility that grammatical evolution can be modeled by distinguished trajectories on this manifold. In particular, the next section introduces a variational formulation in which admissible continuous trajectories are stationary curves of the Fisher length. The resulting construction gives the following hierarchy:
G ⟶ p ( x | θ ) ⟶ M ⟶ g i j ( θ ) .
The rank-frequency analysis developed later requires one additional step. Given a probability distribution generated at a point θ ∈ M , its probabilities are rearranged in decreasing order,
p ( 1 ) ≥ p ( 2 ) ≥ ⋯ .
The resulting sequence defines the rank-frequency distribution. Establishing a mathematical connection between the Fisher-geodesic dynamics of θ and the asymptotic behavior of p ( r ) is the central problem addressed in Sec. Section 6.

5. Variational Dynamics on Statistical Manifolds

5.1. Fisher Length and Admissible Trajectories

Let θ ( s ) = θ 1 ( s ) , ⋯ , θ d ( s ) be a smooth curve on the statistical manifold M introduced in Sec. Section 4. The parameter s is an evolution parameter and need not coincide with physical time. The corresponding probability distribution is p s ( x ) = p x | θ ( s ) . The Fisher line element along this curve is d s F 2 = g i j ( θ ) d θ i d θ j , and its associated length is
L [ θ ] = ∫ s 0 s 1 g i j ( θ ( s ) ) θ ˙ i ( s ) θ ˙ j ( s ) d s ,
where θ ˙ i ≡ d θ i d s . Equation (32) measures the accumulated statistical distinguishability associated with the trajectory. In the present framework, a variational description is introduced by requiring the admissible continuous evolution of the probabilistic grammar to be stationary with respect to this functional. It is important to distinguish this assumption from a statement about arbitrary grammatical dynamics. A probabilistic grammar may in general evolve according to learning rules, stochastic dynamics, external constraints, or other mechanisms. The geodesic dynamics considered here corresponds to the specific modeling hypothesis that the relevant continuous evolution is selected by the Fisher geometry of the statistical family.

5.2. The Information-Geometric Evolution Principle

We formulate the dynamical assumption used in this work as follows.
Information-Geometric Evolution Principle. Given a regular statistical manifold associated with a probabilistic grammar, the continuous idealized evolution of the grammar is represented by a stationary curve of the Fisher length between prescribed endpoints.
Thus,
δ L = 0 ,
where the variation is taken over smooth curves θ ( s ) satisfying fixed endpoint conditions, δ θ ( s 0 ) = δ θ ( s 1 ) = 0 . The principle is a modeling postulate. It does not follow from the definition of the Fisher metric alone, nor does it assert that empirical grammatical evolution is necessarily geodesic. Its role is to provide a well-defined dynamical model whose consequences for rank-frequency distributions can subsequently be analyzed.

5.3. Affine Parameters

An affine parameter is a specific way of measuring your progress along a path—like a straight line or a geodesic—such that you move at a constant “speed” with zero acceleration relative to the geometry of the space. In physics and differential geometry, it is the exact variable required to make the standard geodesic equation hold true without any extra mathematical friction. We can regard an affine parameter as a variable used to describe the position of a point along a straight line—or more broadly, a geodesic (the straightest possible path in a curved space)—in such a way that the path satisfies the standard geodesic equation. To understand why this parameter matters, we can break the definition down into three core concepts:
  • The Path (Geodesics): In flat space, the shortest path between two points is a straight line. In curved space (like the surface of the Earth or warped spacetime), this straightest possible path is called a geodesic.
  • The Parameter (The Clock): A parameter is just a variable (like time t or distance s) that tells you where you are along that path. As the parameter increases, you move forward along the line.
  • The Affine” Quality (No Added Acceleration): The geodesic equation describes a ”state of free fall” or natural motion where no external forces are pushing you off course. If you use an affine parameter (usually denoted as λ ), your mathematical acceleration along the path is exactly zero.
Put simply, it is a parameter that ensures a particle traveling along that path is moving at a “constant speed” without acceleration or deceleration relative to the space itself. When we track a path using an affine parameter, the path’s tangent vectors remain parallel and proportional in length along the entire curve. If you choose a random, non-affine parameter (let’s call it u), it acts like a clock that ticks at a changing rate. Even if you are walking in a straight line, your math will falsely claim you are speeding up or slowing down. Mathematically, this forces a messy correction term into your equations: d 2 x μ d u 2 + Γ α β μ d x α d u d x β d u = f ( u ) d x μ d u When you switch to an affine parameter, that extra force term on the right side f ( u ) becomes exactly zero, simplifying the expression into the standard geodesic equation. Once you find one affine parameter ( λ ), any linear transformation of it is also affine. If you scale it or shift it ( λ ′ = a λ + b ), the path still perfectly satisfies the standard geodesic equation. In general relativity, proper time ( τ ) is the most famous example of an affine parameter for objects traveling through spacetime. If a curve is parameterized by an arbitrary variable s, its tangent velocity vector u μ = d x μ d s might change in magnitude, yielding the equation: u μ ∇ μ u ν = f ( s ) u ν , where f ( s ) is a function representing an acceleration parallel to the velocity. However, if we reparameterize the curve using an affine parameter, the acceleration becomes zero, meaning the tangent vector is perfectly parallel-transported. Affine parameters are unique up to a linear (or “affine”) transformation. If s 1 is an affine parameter, then s 2 = a s 1 + b is also a valid affine parameter. Scaling or shifting the parameter does not destroy its affine nature. An affine parameter t is uniquely determined by specifying its value at two points, or its value and d x μ / d t at a single point.

5.4. Geodesic Equation

The square-root functional of the Fisher action length is invariant under reparameterizations of the curve. It is therefore convenient to introduce an affine parameter and consider the equivalent energy functional
E [ θ ] = 1 2 ∫ s 0 s 1 g i j ( θ ) θ ˙ i θ ˙ j d s .
For an affine parameter, stationary curves satisfy the Euler–Lagrange equations
d d s g i j θ ˙ , j − 1 2 ∂ i g j k θ ˙ , j θ ˙ , k = 0 .

5.5. Cotangent Space

In differential geometry, the cotangent space T p * M at a point p on a deformable manifold M is the dual vector space of the tangent space T p M . While the tangent space contains all geometric vectors (such as velocities or directional arrows) at that point, the cotangent space is composed of covectors (or linear 1-forms), which are functions that take a tangent vector and transform it into a real number. If the manifold has dimension n, the cotangent space will also have exactly that dimension. To understand the cotangent space without getting lost in linear algebra, it is helpful to compare it to the tangent space:
  • Tangent Space ( T p M ): Imagine an ant walking on a sphere. The tangent space at its current position contains all possible directions and velocities in which it can move. Physically, it represents maps of the type γ : R → M (time-dependent curves on the surface).
  • Cotangent Space ( T p * M ): Imagine the sphere has a temperature distribution. The cotangent space is where the differentials of scalar functions live (maps of the type f : M → R ). The gradient or differential of the temperature at that point is not an arrow, but an “operator” that devours a direction of motion and immediately spits out the infinitesimal change in temperature along that direction.
Formally, an element ω ∈ T p * M is defined as a linear functional: ω : T p M → R . If in a local coordinate system ( x 1 , ⋯ , x n ) the tangent space has partial derivatives as its mathematical basis: ∂ ∂ x 1 , ∂ ∂ x 2 , ⋯ , ∂ ∂ x n , the cotangent space possesses a dual basis denoted by the differentials of the coordinate functions: d x 1 , d x 2 , ⋯ , d x n . Both bases are perfectly related by the Kronecker delta ( δ i j ), meaning that the differential d x i measures the component in the direction of its associated vector and annihilates the others: d x i ∂ ∂ x j = δ i j = 1 if i = j 0 if i ≠ j . The disjoint union of all cotangent spaces at every point of the geometric manifold gives rise to the Cotangent Bundle ( T * M ), which is the fundamental mathematical structure upon which analytical mechanics (the phase space in Hamiltonian mechanics) and symplectic geometry are built.

5.6. Parameter-Space Trajectory

An inverse metric tensor (denoted with superscript indices as g μ ν or g i j ) is the algebraic matrix inverse of the standard metric tensor ( g μ ν or g i j ). While the standard metric tensor is used to calculate distances, angles, and lower vector indices, the inverse metric tensor is primarily used to raise indices of tensors and define inner products on cotangent spaces (dual vectors). Multiplication by the inverse metric g ℓ i gives the standard geodesic equation
θ ¨ ℓ + Γ ℓ i j θ ˙ i θ ˙ j = 0 ,
where
Γ ℓ i j = 1 2 g ℓ k ∂ i g k j + ∂ j g k i − ∂ k g i j
are the Christoffel symbols of the Fisher metric. Such equation determines the parameter-space trajectory once the Fisher metric, the initial point, and the initial tangent vector have been specified. In particular, the geometry determines how a constant initial statistical velocity is transported along the manifold. For a geodesic parameterized affinely, the Fisher speed is constant:
g i j θ ˙ i θ ˙ j = constant .
As a consequence, the affine parameter may be chosen to coincide with Fisher arc length up to an overall constant.

5.7. Evolution of the Probability Distribution

The geodesic equation is written in terms of the statistical parameters, but its physical object in the present formulation is the probability distribution generated by the grammar.
Along a trajectory, p s ( x ) = p x | θ ( s ) , and therefore
d d s p s ( x ) = θ ˙ i ∂ i p ( x | θ ) .
Equivalently, whenever p ( x | θ ) > 0 ,
d d s log p s ( x ) = θ ˙ i ∂ i log p ( x | θ ) .
The normalization condition is preserved automatically:
∑ x p s ( x ) = 1 → d d s ∑ x p s ( x ) = 0 .
The evolution of the probability model is thus induced by the trajectory θ ( s ) on the statistical manifold. Combining equations provides the dynamical chain
θ ( s ) → geodesic p s ( x ) → ordering p ( r ) ( s ) .
The final step is not a geometric operation on the statistical manifold. It is the ordering of the probabilities generated by the evolving model. This distinction is essential for establishing a mathematical connection between Fisher geometry and rank-frequency scaling.

5.8. Geodesic Dynamics and the Rank-Frequency Problem

Let p ( 1 ) ( s ) ≥ p ( 2 ) ( s ) ≥ ⋯ denote the probabilities generated by the grammar at evolution parameter s, arranged in decreasing order. The rank-frequency problem is to determine the asymptotic behavior of this sequence for large rank. The geodesic equation by itself does not imply a particular large-r form for p ( r ) . The Fisher metric is defined on the parameter coordinates θ , whereas the rank r labels ordered probability values. Consequently, an additional structural property of the probability model is required in order to connect the two descriptions. The present work introduces this connection through asymptotic rank self-similarity. For λ > 0 , define the rank-dilation ratio
R λ ( r ; s ) = p ( ⌊ λ r ⌋ ) ( s ) p ( r ) ( s ) .

5.9. Rank Scaling

When speaking of a change of rank scale, we are concerned with these items:
  • Scale refers to the size, resolution, or magnification level at which you are looking at a system.
  • Rank scale (or scaling of rank) typically implies that the system is organized hierarchically, or measured in ordered levels (like ranks in a network, tensor ranks, or a sequence of sizes). Therefore, a “change of rank scale” means you are magnifying, shrinking, or shifting the operational level of the system by a constant factor λ (e.g., x → λ x ).
In this context, the term covariance means that an object or property changes in a predictable, structurally preserved way when the underlying coordinate system or scale changes. It implies that the rule governing the system does not break down when you zoom in or out; instead, the output “moves along” with the change according to a strict mathematical transformation. Enters then a definite exponent: this is the exact mathematical power (let’s call it p) that dictates how fast the response changes. It signifies that the relationship is a power law, rather than a linear or exponential one. If you put these pieces together, the statement is describing a function where if you scale the input x by a factor λ , the response f ( λ x ) is directly proportional to the original response f ( x ) multiplied by λ raised to a fixed exponent p: f ( λ x ) = λ p f ( x ) . This specific type of covariance is the defining characteristic of scale invariance and self-similarity. One sees this in statistical physics (during phase transitions), network science / linguistics (Zipf’s Law) and fractal geometry.

5.10. Rank-Dilation Scaling and Regular Variation

A central difficulty in relating information geometry to Zipf-type behavior is that the Fisher metric is defined on the parameter manifold, whereas a rank-frequency distribution is defined on an ordered set of probabilities. A direct statement of the connection therefore requires a scaling criterion that refers explicitly to the rank variable. Let p ( 1 ) ( s ) ≥ p ( 2 ) ( s ) ≥ ⋯ denote the probabilities generated by the statistical model at evolution parameter s, arranged in non-increasing order. For a positive scale factor λ , consider the ratio
R λ ( r ; s ) = p ( ⌊ λ r ⌋ ) ( s ) p ( r ) ( s ) .
This quantity compares the probability at rank r with the probability at a rank that has been dilated by the factor λ . It therefore probes the response of the rank-frequency distribution to a multiplicative change of scale in the rank variable. Unlike a fit of p ( r ) to a power law over a finite interval, the above relation defines an asymptotic property of the distribution itself. We say that the ranked distribution exhibits an asymptotic rank-dilation scaling regime if there exists an exponent α ( s ) such that we have the dilation-relation
lim r → ∞ R λ ( r ; s ) = λ − α ( s ) for every λ > 0 .
The significance of this definition can be seen immediately by considering an exact power law, p ( r ) ( s ) = C ( s ) r − α ( s ) . For such a distribution,
R λ ( r ; s ) = C ( s ) ( ⌊ λ r ⌋ ) − α ( s ) C ( s ) r − α ( s ) ,
and hence R λ ( r ; s ) ⟶ λ − α ( s ) as r → ∞ . The normalization constant C ( s ) cancels from the ratio. Thus the exponent is determined by the response of the distribution to a dilation of rank rather than by its overall normalization. The ratio formulation is more general than an exact power-law ansatz. Indeed, the dilation-relation allows for asymptotic corrections that do not change under a finite rescaling of the rank variable. To make this explicit, suppose that p ( r ) ( s ) = r − α ( s ) L ( r ; s ) , where L ( r ; s ) is a positive slowly varying function. By definition,
lim r → ∞ L ( ⌊ λ r ⌋ ; s ) L ( r ; s ) = 1 , for every λ > 0 ,
it follows that
R λ ( r ; s ) = ( ⌊ λ r ⌋ ) − α ( s ) L ( ⌊ λ r ⌋ ; s ) r − α ( s ) L ( r ; s ) = ⌊ λ r ⌋ r − α ( s ) L ( ⌊ λ r ⌋ ; s ) L ( r ; s ) .
Since ⌊ λ r ⌋ r → λ and the slowly varying factor satisfies Eq. (47), one obtains R λ ( r ; s ) → λ − α ( s ) . Thus the rank-dilation criterion does not require the distribution to be an exact power law. It identifies a broader class of asymptotic distributions whose leading scaling is power-like. This class is precisely the class of regularly varying sequences. Under the usual positivity and regularity assumptions of the theory of regular variation, the converse also holds: if the limiting ratio
lim r → ∞ p ( ⌊ λ r ⌋ ) ( s ) p ( r ) ( s ) = λ − α ( s )
exists for every λ > 0 , then the ranked probabilities are regularly varying with index − α ( s ) . Consequently, they admit the representation
p ( r ) ( s ) = r − α ( s ) L ( r ; s ) ,
where L ( r ; s ) is slowly varying in the sense of the above equation. This is the standard characterization of regular variation: asymptotic covariance under multiplicative changes of the argument corresponds to a power-law index multiplied by a slowly varying correction. The distinction between regular variation and an exact power law is important for the present work. A slowly varying function need not approach a constant. Examples include L ( r ) = ( log r ) γ , or more generally products of powers of iterated logarithms. Such factors modify the finite- and intermediate-rank behavior while leaving the rank-dilation exponent unchanged. For example, p ( r ) ∼ r − α ( log r ) γ satisfies p ( ⌊ λ r ⌋ ) p ( r ) ⟶ λ − α , because log ( λ r ) log r ⟶ 1 . Consequently, the exponent α characterizes the asymptotic scaling under rank dilation, whereas the slowly varying function L contains subleading information about the detailed shape of the rank-frequency distribution. This distinction is useful when comparing analytical models with finite numerical data, for which slowly varying corrections can produce substantial deviations from a straight line on a log-log plot. The rank-dilation formulation also avoids an ambiguity associated with estimating a power-law exponent from finite data. A straight segment in a log-log representation, log p ( r ) ≃ − α log r + constant , does not by itself establish an asymptotic power law. Finite-size distributions, crossovers, exponential cutoffs, and slowly varying corrections can all produce approximately linear regions over restricted ranges of r. By contrast, the regular variation ratio asks whether the same scaling exponent governs the response to arbitrary fixed multiplicative changes of rank as r becomes large. There is also a useful composition property. If the rank-dilation relation holds, then for λ , μ > 0 , R λ μ ( r ; s ) = p ( ⌊ λ μ r ⌋ ) ( s ) p ( r ) ( s ) . In the asymptotic scaling regime, the corresponding limits satisfy λ − α ( s ) μ − α ( s ) = ( λ μ ) − α ( s ) . Thus the limiting scaling function G s ( λ ) = lim r → ∞ R λ ( r ; s ) has the multiplicative covariance property
G s ( λ μ ) = G s ( λ ) G s ( μ ) .
Under the usual regularity conditions, the positive solutions of this multiplicative functional equation are power functions, G s ( λ ) = λ − α ( s ) . The power-law exponent therefore emerges as the parameter characterizing the asymptotic representation of the multiplicative rank-dilation operation. This observation provides a useful conceptual interpretation of generalized Zipf behavior. The essential property is not that the probability is literally equal to a power of its rank at every rank, but that its asymptotic response to a change of rank scale is covariant with a definite exponent: when you rescale the system’s input or structural scale, the system’s output (response) changes proportionally by that scaling factor raised to a specific power. Regular variation provides the mathematical language for this asymptotic scale covariance.

5.11. Relation to Information-Geometric Evolution

The rank-dilation criterion is formulated entirely in rank space, whereas the dynamics introduced in Sec. Section 5 takes place on the statistical manifold. The central problem of the present paper is therefore to determine when these two structures become related. At each value of the evolution parameter s, the Fisher-geodesic trajectory determines a probability distribution θ ( s ) ⟶ p ( x | θ ( s ) ) . Ordering its probabilities produces p ( 1 ) ( s ) , p ( 2 ) ( s ) , ⋯ , and hence the rank-dilation ratio R λ ( r ; s ) . The information-geometric problem can therefore be expressed as the following implication:
Fisher - geod . evol . ⟶ asymp . rank - dilation cov .
This implication does not hold for an arbitrary statistical manifold and cannot be inferred from the Fisher metric alone. It depends on the functional form of the probability model and, in particular, on how the state or configuration labels enter the probability weights. Establishing conditions under which Eq. (51) is realized is therefore a substantive mathematical question, addressed in the following section. An exactly solvable many-state model provides such a setting. Assume that the probabilities are parameterized by an inverse-temperature-like variable β ,
p n ( β ) = exp [ − β E n ] Z ( β ) ,
and the Fisher metric determines the geometry of the β -manifold. When the state spectrum is asymptotically logarithmic, E ⌊ λ n ⌋ − E n ⟶ log λ , the rank-dilation ratio becomes p ⌊ λ n ⌋ ( β ) p n ( β ) → λ − β . In such a model, therefore, the asymptotic rank exponent is directly related to the statistical parameter, α = β , while β itself evolves according to the Fisher geometry. This provides the explicit bridge between the geometric dynamics and the rank-frequency scaling that is absent from the general formulation. The subsequent analysis separates two logically distinct statements. First, rank-dilation covariance implies regular variation of the rank-frequency sequence. Second, for a particular statistical model, the Fisher-geodesic dynamics can be shown to generate the rank-dilation covariance. Only the combination of these two results yields the proposed information-geometric mechanism for generalized Zipf scaling.

5.12. Scope of the Variational Construction

The variational formulation developed above should be interpreted at the level of an idealized continuous model. Actual grammar transformations are typically discrete and may involve stochastic or externally imposed changes of production probabilities. The Fisher-geodesic description represents a continuous limiting dynamics in which the statistical cost of an infinitesimal deformation is determined by the Fisher metric. Accordingly, no claim is made at this stage that all probabilistic grammars follow geodesics, or that every Fisher-geodesic trajectory possesses a power-law rank-frequency distribution. Such conclusions would require additional assumptions on the structure of the probability family. The purpose of the subsequent analysis is precisely to determine a concrete class of many-state models for which these additional assumptions can be made explicit and the relation between the Fisher geometry and the rank-frequency exponent can be established analytically.

6. Rank-Frequency Scaling and the Geometric Scaling Conjecture

6.1. Ranked Probability Distributions

At each point of the statistical trajectory θ ( s ) , the probabilistic grammar defines a normalized distribution p ( x | s ) = p x | θ ( s ) . Suppose that the set of admissible configurations can be ordered according to their probabilities. We denote the resulting sequence by
p ( 1 ) ( s ) ≥ p ( 2 ) ( s ) ≥ ⋯ ,
where p ( r ) ( s ) is the probability associated with rank r. The notation ( r ) emphasizes that rank is determined by ordering the probabilities and does not necessarily coincide with a particular microscopic label of the configuration. In models for which the probabilities are already monotone in a state label n, one simply has p ( r ) = p r . In the general case, however, the ordering operation must be performed explicitly. The central question of this section is whether the large-rank behavior of the ranked distribution can be characterized by an asymptotic scaling relation. We will formulate this question independently of any particular choice of grammar.

6.2. Rank-Dilation Covariance

For a fixed scale factor λ > 0 , define the rank-dilation ratio
R λ ( r ; s ) = p ( ⌊ λ r ⌋ ) ( s ) p ( r ) ( s ) ,
whenever the two probabilities are nonzero. A rank-frequency distribution is said to possess an asymptotic rank-dilation exponent α ( s ) if
lim r → ∞ R λ ( r ; s ) = λ − α ( s ) for every λ > 0 ,
a ratio formulation of generalized Zipf scaling. It is preferable to assuming a power-law form at the outset, because the exponent is defined directly through the asymptotic behavior of the ranked probabilities. If p ( r ) ( s ) ∼ C ( s ) r − α ( s ) , then the ratio formulation of generalized Zipf scaling follows immediately: p ( ⌊ λ r ⌋ ) ( s ) p ( r ) ( s ) → λ − α ( s ) . The converse requires an asymptotic regularity condition and leads, in general, not to a pure power law but to a regularly varying sequence.

6.3. Regular Variation and Generalized Zipf Scaling

We now state the mathematical implication of the rank-dilation property.
Proposition 1. 
Let p ( r ) ( s ) r ≥ 1 be eventually positive and eventually non-increasing. The statement means that as you move far enough out in the sequence, the terms will always be greater than zero (positive) and will either decrease or stay the same (non-increasing), never getting larger. In mathematics, the word “eventually” means that a property becomes permanently true after a certain point.
  • It does not matter what the sequence does at the beginning—the first few terms could be negative, zero, or alternating.
  • However, there must exist some specific starting index (let’s call it N) such that for every single step after it ( r ≥ N ), the value of p r ( s ) is strictly greater than zero ( p r ( s ) > 0 ).
Similarly, ”non-increasing” means the sequence moves downward or stays flat, but never climbs. Combined with eventually:
  • After passing a certain point in the sequence, each subsequent term must be less than or equal to the term before it.
  • Mathematically, there is an index M such that for all r ≥ M , the relationship is p r + 1 ( s ) ≤ p r ( s ) .
When a sequence is both eventually positive and eventually non-increasing, it means that if you look far enough down the line, the sequence behaves in a highly stable and predictable way: it is trapped above 0 and is steadily decaying or flattening out. This specific set of conditions is very common when applying convergence tests in calculus and analysis. Suppose then that, for some α ( s ) ∈ R ,
lim r → ∞ p ( ⌊ λ r ⌋ ) ( s ) p ( r ) ( s ) = λ − α ( s )
for every λ > 0 . Then the ranked probability sequence is regularly varying with index − α ( s ) ; that is,
p ( r ) ( s ) = r − α ( s ) L ( r ; s ) ,
where L ( r ; s ) is slowly varying:
lim r → ∞ L ( ⌊ λ r ⌋ ; s ) L ( r ; s ) = 1 for every λ > 0 .
The proposition is a standard characterization of regularly varying sequences. The significance for the present problem is that the rank dilation provides a precise mathematical definition of the asymptotic scaling regime without requiring the slowly varying factor to be constant. The special case L ( r ; s ) ⟶ C ( s ) gives the ordinary power law p ( r ) ( s ) ∼ C ( s ) r − α ( s ) . Thus the generalized scaling form p ( r ) ( s ) = r − α ( s ) L ( r ; s ) contains the usual Zipf-type law as a special case.

6.4. Classical Zipf Scaling

The classical Zipf form corresponds to p ( r ) ∝ 1 r , and therefore to α = 1 . In the present formulation, the defining asymptotic relation is
lim r → ∞ p ( ⌊ λ r ⌋ ) p ( r ) = 1 λ .
It is important, however, to distinguish the infinite-support scaling statement from normalization of the full distribution. A pure p ( r ) ∝ r − 1 law on an infinite set is not normalizable. Therefore, when α = 1 , a normalized model must contain either a finite effective support, a finite-size cutoff, or a slowly varying factor that modifies the global normalization. The asymptotic ratio relation above can nevertheless remain meaningful over an increasing range of ranks. For an infinite-support normalized pure power law, p ( r ) = r − α Z ( α ) , one requires
α > 1 , Z ( α ) = ∑ r = 1 ∞ r − α = ζ ( α ) .
This distinction will be relevant for the exactly solvable many-state model introduced below.

6.5. Time-Dependent Rank Exponents

Under the information-geometric evolution introduced in Sec. Section 5, the ranked distribution depends on the evolution parameter s: p ( r ) = p ( r ) ( s ) . If the rank-dilation limit exists at each value of s, we define the instantaneous asymptotic exponent by
α ( s ) = − lim r → ∞ log p ( ⌊ λ r ⌋ ) ( s ) / p ( r ) ( s ) log λ ,
where the limit is required to be independent of the chosen λ > 0 , consistently with the rank dilation. This definition allows the rank exponent to evolve as the probability model moves along a Fisher-geodesic trajectory. The exponent is therefore not assumed to be a universal constant of all grammars. Rather, it is a property of the asymptotic rank distribution associated with a particular point of the evolving statistical model. A central question is then whether the geometric evolution of θ ( s ) can induce a systematic evolution of α ( s ) . In general this cannot be inferred from the Fisher metric alone. An explicit relation requires a model in which the state dependence of the probabilities can be related to the geometric parameter controlling the trajectory.

6.6. Geometric Scaling Conjecture

The preceding results motivate a more precise formulation of the central hypothesis of this paper.
Geometric Scaling Conjecture. Consider a probabilistic grammar whose continuous evolution is represented by a Fisher geodesic θ ( s ) . Suppose that, in an asymptotic regime, the ordering of the probabilities is stable and the induced ranked distribution possesses a rank-dilation covariance of the form: lim r → ∞ p ( ⌊ λ r ⌋ ) ( s ) p ( r ) ( s ) = λ − α ( s ) . Then the rank-frequency distribution is regularly varying, p ( r ) ( s ) = r − α ( s ) L ( r ; s ) , with L slowly varying.
Unlike the conjecture in the original formulation, this statement does not claim that geodesic evolution alone implies a Zipf law. The regular-variation conclusion follows from the explicitly stated rank-dilation condition. The nontrivial part of the hypothesis is therefore the existence of a model in which Fisher-geodesic evolution produces that asymptotic covariance explicitly. This permits the conjectural part of the argument to be tested rather than assumed.

6.7. The Missing Geometric-to-Rank Bridge

The mathematical structure developed so far can be summarized as
θ ( s ) → geodesic p n ( s ) → ordering p ( r ) ( s ) → r → ∞ r − α ( s ) L ( r ; s ) .
The first arrow is determined by the variational principle of Sec. Section 5. The second is an ordering operation. The third requires a specific asymptotic structure of the probability model. The essential additional ingredient will be an asymptotically logarithmic dependence of the state label on the statistical weight. Suppose, for example, that the probabilities have the form
p n ( β ) = exp [ − β E n ] Z ( β ) ,
where Z ( β ) = ∑ n = 1 ∞ exp [ − β E n ] . If the state spectrum satisfies
E ⌊ λ n ⌋ − E n → log λ , n → ∞ ,
then
p ⌊ λ n ⌋ ( β ) p n ( β ) = exp − β E ⌊ λ n ⌋ − E n → λ − β .
Consequently, provided the state ordering agrees asymptotically with the rank ordering, α = β . This observation identifies the structural β bridge required in the present approach: Fisher-geodesic evolution leads to an asymptotically logarithmic state spectrum, which implies rank-dilation scaling. The normalization factor Z ( β ) cancels from the rank ratio and therefore does not affect the asymptotic exponent. This cancellation makes the model particularly useful for isolating the relation between statistical geometry and rank-frequency scaling.

6.8. Testable Consequence

The formulation above leads to a direct numerical and analytical test. For a given evolving probability model, one should calculate R λ ( r ; s ) = p ( ⌊ λ r ⌋ ) ( s ) p ( r ) ( s ) for several values of λ and increasingly large r. If the Fisher-geometric scaling hypothesis is realized, one should observe R λ ( r ; s ) ⟶ λ − α ( s ) as r increases. The exponent may then be estimated independently for different values of λ and compared with the exponent obtained from a direct fit of the ranked distribution. This provides a criterion that does not depend on visual inspection of a log-log plot. In particular, a finite-range approximate straight line is not by itself sufficient evidence for asymptotic Zipf scaling. The ratio test above directly probes the defining scale-covariance property. The exactly solvable model considered next provides a setting in which both the Fisher geometry and the rank-dilation ratio can be evaluated explicitly. It will therefore allow the proposed geometric mechanism to be examined without assuming the desired rank-frequency form in advance.

7. Worked Examples: Geodesic Flow on Grammatical Manifolds

To demonstrate the mathematical viability of the proposed framework, we present an analytical construction that explicitly connects information-geometric evolution with parametric trajectory dynamics. The two-state models frequently utilized in introductory literature collapse into flat, trivial statistical spaces that fail to capture the structural complexities of formal languages. In this section, we construct a multi-dimensional, non-flat statistical manifold generated by a probabilistic formal grammar, demonstrating how grammatical rules impose geometric curvature and dictate intrinsic selection along Levi-Civita geodesics.

7.1. A Bifurcated Regular Grammar

Consider the probabilistic regular grammar G with terminal alphabet Σ = { a , b , c } , one non-terminal S, and productions
R 1 : S → θ 1 a S , R 2 : S → θ 2 b S , R 3 : S → θ 3 c , θ 3 ≡ 1 − θ 1 − θ 2 .
The parameter space is the open 2-simplex
Θ = { ( θ 1 , θ 2 ) ∈ R 2 : θ 1 > 0 , θ 2 > 0 , θ 1 + θ 2 < 1 } .

Configuration space and normalization.

A derivation consists of k ≥ 0 applications of R 1 or R 2 in some order, followed by one application of R 3 . The generated configurations are therefore all words x = w c with w ∈ { a , b } * , so X = { a , b } * c , and
p ( x | θ ) = ( θ 1 ) n ( x ) ( θ 2 ) m ( x ) θ 3 ,
where n ( x ) and m ( x ) count the letters a and b in w. Normalization follows from summing over the 2 k words of length k:
∑ x ∈ X p ( x | θ ) = θ 3 ∑ k = 0 ∞ ( θ 1 + θ 2 ) k = 1 .
(If one instead restricts to the sorted words a n b m c and uses p = ( θ 1 ) n ( θ 2 ) m θ 3 , the total mass is θ 3 / [ ( 1 − θ 1 ) ( 1 − θ 2 ) ] ≠ 1 . Equivalently, one may work with the counts ( n , m ) and the weight n + m n ( θ 1 ) n ( θ 2 ) m θ 3 , which is normalized and has the same Fisher metric, since the binomial factor does not depend on θ .)

Fisher metric.

From log p = n log θ 1 + m log θ 2 + log θ 3 one finds
∂ 1 log p = n θ 1 − 1 θ 3 , ∂ 2 log p = m θ 2 − 1 θ 3 .
The number of productions R 1 , R 2 is geometrically distributed, and each is R 1 with probability θ 1 / ( θ 1 + θ 2 ) , so E θ [ n ] = θ 1 / θ 3 and E θ [ m ] = θ 2 / θ 3 . Using g i j = − E θ [ ∂ i ∂ j log p ] ,
[ g i j ] = 1 θ 3 1 θ 1 + 1 θ 3 1 θ 3 1 θ 3 1 θ 2 + 1 θ 3 , [ g i j ] = θ 3 θ 1 ( 1 − θ 1 ) − θ 1 θ 2 − θ 1 θ 2 θ 2 ( 1 − θ 2 ) .
Since E θ [ # productions ] = 1 / θ 3 , the metric is the Fisher metric of a single categorical draw g ^ i j = δ i j / θ i + 1 / θ 3 multiplied by the expected number of draws:
g i j = 1 θ 3 g ^ i j .

Curvature.

The metric g ^ is that of an octant of a 2-sphere of radius 2 (Gaussian curvature K ^ = 1 / 4 ), and Eq. (72) is a conformal rescaling g = e 2 φ g ^ with φ = − 1 2 ln θ 3 . In two dimensions K = e − 2 φ K ^ − Δ ^ φ . With θ 3 = cos 2 χ , where χ is the polar angle from the third axis, Δ ^ φ = 1 4 ( sec 2 χ + 1 ) and therefore
K = θ 3 − 1 4 sec 2 χ = − 1 4 , R = 2 K = − 1 2 .
The manifold ( Θ , g ) is thus non-flat, with constant negative curvature, and is locally isometric to a hyperbolic plane of curvature − 1 / 4 . The grammatical coupling through R 3 therefore curves the manifold, but the curvature is uniform.

Geodesic equations.

The Christoffel symbols of the first kind, Γ i j , k = 1 2 ( ∂ i g j k + ∂ j g i k − ∂ k g i j ) , are, with q ≡ θ 3 ,
Γ 11 , 1 = 1 q 3 + 1 2 θ 1 q 2 − 1 2 ( θ 1 ) 2 q , Γ 22 , 2 = 1 q 3 + 1 2 θ 2 q 2 − 1 2 ( θ 2 ) 2 q ,
Γ 11 , 2 = 1 q 3 − 1 2 θ 1 q 2 , Γ 22 , 1 = 1 q 3 − 1 2 θ 2 q 2 ,
Γ 12 , 1 = 1 q 3 + 1 2 θ 1 q 2 , Γ 12 , 2 = 1 q 3 + 1 2 θ 2 q 2 .
Raising the index with g i j of Eq. (71) ( Γ i j k = g k ℓ Γ i j , ℓ ) gives the geodesic system
θ ¨ k + Γ i j k ( θ ) θ ˙ i θ ˙ j = 0 , g i j ( θ ) θ ˙ i θ ˙ j = ∥ v ∥ F 2 = const .
Given an initial point θ ( 0 ) ∈ Θ and initial velocity v, the trajectory of production probabilities is determined, and it is the choice of ( θ ( 0 ) , v ) , or of two endpoints, that remains an external input to the model.

Remark on rank statistics.

Words of length k number 2 k and have probability ( θ 1 ) n ( θ 2 ) k − n θ 3 . The counting function N ( p ) = # { x : p ( x | θ ) ≥ p } then grows as p − α , where ( θ 1 ) α + ( θ 2 ) α = 1 (so α > 1 exactly when θ 1 + θ 2 < 1 ). For θ 1 = θ 2 = θ this gives α = − ln 2 / ln θ for the inverse rank-frequency exponent p ( r ) ∼ r − 1 / α . Because the probabilities take a lattice of values (random-typing model), p ( ⌊ λ r ⌋ ) / p ( r ) oscillates log-periodically and no strict limit exists; only the coarse-grained exponent does. The regular-variation analysis is therefore applied to the Gibbs family, not to this grammar.

7.2. A Gibbs Family with Logarithmic Spectrum

The model.

Let X = N and consider the one-parameter exponential family
p n ( β ) = e − β E n Z ( β ) , Z ( β ) = ∑ n = 1 ∞ e − β E n , β > 1 ,
with spectrum
E n = log n + η n , η n = η ∞ + O ( n − δ ) for some η ∞ ∈ R , δ > 0 .
For η n ≡ 0 this is the zeta distribution p n = n − β / ζ ( β ) . The logarithmic spectrum is an assumption of the model (it is the cost structure of Mandelbrot’s argument [15]); the Fisher geometry does not generate it. The series converges for β > 1 , so M β = ( 1 , ∞ ) is a regular one-dimensional statistical manifold.

Rank-dilation exponent.

Assume E n is non-decreasing for n large, so that the rank is r = n asymptotically. For every λ > 0 ,
p ⌊ λ n ⌋ ( β ) p n ( β ) = ⌊ λ n ⌋ n − β e − β ( η ⌊ λ n ⌋ − η n ) → n → ∞ λ − β .
The factor Z ( β ) cancels, and η ⌊ λ n ⌋ − η n → 0 by Eq. (79). Hence p ( r ) is regularly varying with index − β , i.e. p ( r ) = r − β L ( r ; β ) , and
α = β .
This identity holds at every  β > 1 and is a property of the spectrum Eq. (79); it does not depend on any dynamics. What the Fisher geometry determines is how β , and hence α , moves along an evolution.

Fisher metric.

For an exponential family with sufficient statistic E,
g β β ( β ) = Var β ( E ) = d 2 d β 2 log Z ( β ) > 0 ,
where positivity holds because E n is non-constant. Since M β is one-dimensional, the arc length
u ( β ) = ∫ β 0 β g β β ( b ) d b
is a global flat coordinate, d s F 2 = d u 2 . The geodesic equation reduces to u ¨ = 0 . Every constant-speed curve is a geodesic,
u ( β ( s ) ) = u 0 + v s ,
and every monotone curve β ( s ) is a reparametrization of one. The Fisher metric therefore fixes the speed profile  β ( s ) via g β β β ˙ = const , but it does not select a direction or endpoints; these are external inputs.
Proposition 2 
(Divergence of the Fisher distance to the Zipf point). For the zeta distribution ( η n ≡ 0 ),
g β β ( β ) = 1 ( β − 1 ) 2 + O ( 1 ) , β → 1 + ,
and consequently, for fixed β 0 > 1 ,
d F ( β , β 0 ) = ln ( β 0 − 1 β − 1 ) + O ( 1 ) ⟶ ∞ ( β → 1 + ) .
Proof. 
The Laurent expansion ζ ( β ) = 1 β − 1 + γ E + O ( β − 1 ) may be differentiated termwise, so ζ ′ = − 1 ( β − 1 ) 2 + O ( 1 ) and ζ ′ ′ = 2 ( β − 1 ) 3 + O ( 1 ) . Then
g β β = ζ ′ ′ ζ − ζ ′ ζ 2 = 2 ( β − 1 ) 2 − 1 ( β − 1 ) 2 + O ( 1 ) = 1 ( β − 1 ) 2 + O ( 1 ) .
Hence g β β = 1 β − 1 + O ( β − 1 ) , and integrating gives Eq. (86). □
Interpretation. Under p n ∝ n − β , ln n has an asymptotically exponential tail with rate β − 1 , so its variance behaves as ( β − 1 ) − 2 . As β → 1 + the normalization fails ( Z → ∞ ) and the distribution becomes infinitely hard to distinguish from its neighbours in the Fisher sense.
For spectra satisfying Eq. (79), the same conclusion holds if the asymptotics of Z and its derivatives are obtained as above from the Dirichlet series (this is where the condition η n − η ∞ = O ( n − δ ) is used). I give the proof only for the zeta case.

Consequence for the geodesic flow.

From Eqs. (84) and (86), a geodesic moving toward the Zipf point has u ( β ( s ) ) = u 0 − v s with v > 0 , hence
β ( s ) − 1 ∼ K e − v s , s → ∞ ,
with K > 0 depending on the initial condition. Thus
α ( s ) = β ( s ) ⟶ 1 + only as s → ∞ :
the exact Zipf exponent α = 1 is at infinite Fisher distance and is not reached in any finite evolution parameter. This is consistent with the normalization constraint α > 1 : the Fisher geometry independently excludes the non-normalizable point. In the opposite direction ( β → ∞ ) the distribution concentrates on the lowest-energy state, g β β ∼ ( E 2 − E 1 ) 2 e − β ( E 2 − E 1 ) , the arc length is finite, and the end of the manifold is reached at finite s.

Dependence on the spectrum.

The divergence is not universal. Take E n = log n + γ log log n with γ > 0 ; then p ( r ) = r − β ( log r ) − β γ / Z is still regularly varying with α = β , but the slowly varying factor changes the boundary behavior. The sum Z ( 1 ) is finite for γ > 1 , and Var ( E ) is finite at β = 1 for γ > 3 . In that case the Fisher distance to β = 1 is finite and the geodesic can reach it. The slowly varying correction therefore decides whether the Zipf point is at finite or infinite distance, although it does not change the rank-dilation exponent.

Mixing.

Let p i ( ε ) = ( 1 − ε ) p i ( β ) + ε ∑ n K i n p n ( β ) with K a stochastic kernel such that ∑ n K i n p n ( β ) = o ( i − β ) as i → ∞ and p i ( ε ) is eventually non-increasing. Then Eq. (80) is unchanged for each fixed ε < 1 , so α = β . The two-parameter Fisher metric g ( β , ε ) is a separate object, and its geodesics need not keep ε fixed; the exponent along such geodesics is governed by Eq. (81) only while the stated condition on K holds.

What is and is not established.

(i) Rank-dilation covariance with α = β follows from the assumed logarithmic spectrum. (ii) The Fisher geometry governs the evolution of β ( s ) and shows that α = 1 is unreachable in finite Fisher length for η n → η ∞ . (iii) The statement “Fisher-geodesic evolution produces Zipf scaling” is not established: the scaling is a property of the spectrum, not of the geodesic flow.

7.3. Baselines

In Data Science and Machine Learning, a baseline distribution is a dataset or “statistical map” used as the expected standard or starting point to compare future data against. Typically, it represents the shape, mean, spread, and overall behavior of the data during the training or validation stage of a model. What is it used for?
  • Data Drift Detection: In production, real-world data changes over time. By comparing the distribution of incoming live data against the baseline distribution, monitoring tools can detect if the environment has changed (for example, if user behavior has shifted drastically).
  • Data Quality and Validation: It helps identify anomalies or outliers. If the new data looks nothing like the baseline distribution, the system triggers an alert.
  • Model Explainability: In interpretability techniques (such as SHAP values), a baseline distribution is often used to understand how changes in variables alter predictions compared to a “neutral” or average reference case.
A Simple Example: imagine a credit card approval model. The baseline distribution (obtained during training) shows that the average applicant has a monthly income of 3,000 USD. If a year later the system starts receiving applications from users with an average income of 1,200 USD, the current distribution diverges from the baseline. This alerts engineers that the model might fail and needs to be retrained. The term is also used in theoretical statistics (foundational probability distribution) and economics / transfer pricing (baseline distribution function).

7.3.1. A Concrete Baseline Distribution

For numerical implementation, a particularly useful baseline family is obtained by assigning the states an intrinsic score E i and defining
q i ( β ) = exp [ − β E i ] ∑ j = 1 N exp [ − β E j ] .
The energies E i need not be interpreted thermodynamically; they simply provide a controlled parametrization of a large probability simplex. The inverse-temperature parameter β can therefore be regarded as one coordinate of the statistical manifold. A second parameter can be introduced through the communication strength ε , giving θ = ( β , ε ) . The Fisher metric is then the 2×2 matrix
g ( β , ε ) = g β β g β ε g ε β g ε ε ,
with
g μ ν = ∑ i = 1 N p i ( β , ε ) ∂ μ log p i ( β , ε ) ∂ ν log p i ( β , ε ) , μ , ν ∈ { β , ε } .
This model has a useful conceptual advantage. The Fisher geometry is calculated in the parameter space ε , β , whereas the quantity R λ , N ( r ) is calculated in rank space. The numerical experiment therefore directly tests the missing bridge between these two descriptions rather than assuming it from the outset.

7.3.2. Numerical Criterion for the Geometric Scaling Conjecture

The many-state construction provides a direct falsifiable test of the conjecture. For each N, one should:
  • calculate the Fisher metric g μ ν ;
  • integrate the corresponding geodesic equations;
  • construct the probability distribution p i ( t ; ε ) ;
  • rank the probabilities to obtain p ( r ) ( t ; ε ) ;
  • calculate R λ , N ( r ; t , ε ) for several values of λ ;
  • examine its behavior as r increases and N is increased;
  • estimate α from the α estimator;
  • test whether the estimated exponent is stable with respect to λ , N, and the time interval used to define the asymptotic regime.
Evidence for rank-dilation covariance would consist of the simultaneous observations R λ , N ( r ; t , ε ) ⟶ λ − α over an expanding intermediate-rank regime as N increases, together with approximately consistent values of α for different λ . Conversely, if R λ , N ( r ; t , ε ) does not approach a λ − α form, or if the inferred exponent depends systematically on λ , N, or r, then the Fisher-geodesic mechanism considered here does not generate the proposed asymptotic rank-dilation symmetry for that model. Such a result would be informative: it would identify an explicit class of Fisher-geodesic models for which the proposed Zipf mechanism fails. The role of this model is therefore deliberately narrower than that of the general theory. It does not assume that geodesic evolution produces a Zipf law. Instead, it provides a quantitative test of the additional hypothesis required to establish such a connection: Fisher-geodesic evolution leads to asymptotic rank-dilation covariance. Only if this implication is supported by the many-state dynamics can the regular variation result of Section 3.3 be invoked to characterize the resulting rank-frequency distribution.

7.4. An Adaptive-Agent Grammar

The grammar.

Two agents each choose an action, C (cooperate) or D (defect). The joint outcome is generated by the one-step grammar
S → p i x i , x i ∈ { C C , C D , D C , D D } , ∑ i = 1 4 p i = 1 , p i > 0 .
Its statistical manifold is the open probability simplex Δ 3 , with coordinates θ = ( p 1 , p 2 , p 3 ) and p 4 = 1 − p 1 − p 2 − p 3 . Note that expressions such as C C → C D are not rewrite rules of the grammar; they denote deformations of the weights p, that is, tangent vectors on Δ 3 .

Fisher metric.

Since ∂ i log p i = 1 / p i and ∂ i log p 4 = − 1 / p 4 for i = 1 , 2 , 3 ,
g i j = ∑ k = 1 4 p k ∂ i log p k ∂ j log p k = δ i j p i + 1 p 4 , g i j = p i δ i j − p i p j .
The sign of the 1 / p 4 term is positive. Equivalently, for a tangent vector v = ( v 1 , ⋯ , v 4 ) with ∑ k v k = 0 , ∥ v ∥ F 2 = ∑ k = 1 4 v k 2 / p k .

Geometry and geodesics.

The map x k = 2 p k sends Δ 3 isometrically onto the open positive orthant of the sphere { x ∈ R 4 : | x | = 2 } , because ∑ k d p k 2 / p k = ∑ k d x k 2 . The Fisher geometry of Δ 3 is therefore that of a spherical region of radius 2, with constant sectional curvature K = 1 / 4 . Geodesics are arcs of great circles. In terms of p k , a unit-speed geodesic through p 0 is
p k ( s ) = q k cos s 2 + e k sin s 2 , q k = p k 0 ,
where e ∈ R 4 satisfies ∑ k e k 2 = 1 and ∑ k q k e k = 0 (the initial direction), and g i j θ ˙ i θ ˙ j = 1 . The Fisher distance between two distributions is
d F ( p , p ′ ) = 2 arccos ∑ k p k p k ′ .
A geodesic leaves Δ 3 when some component p k reaches zero, which happens at the finite value tan ( s / 2 ) = − q k / e k for any k with e k < 0 . Unlike the Zipf point of Sec. Section 7.2, the boundary of the simplex is at finite Fisher distance.

Cost of competing adaptive transitions.

A transfer of probability mass ϵ from state i to state j is the tangent vector v = ϵ ( e j − e i ) , with Fisher cost
∥ v ∥ F 2 = ϵ 2 1 p i + 1 p j .
Consider two competing responses from the same state C C : moving mass to C D or to D C . Both are admissible, and for equal ϵ their costs differ by ϵ 2 ( 1 / p C D − 1 / p D C ) . The transfer toward the more probable target is cheaper. This is a first-order comparison at the current point and for equal mass transfers. It is not a selection by geodesics: any two points of Δ 3 are joined by a geodesic, so a geodesic alone does not single out one response. Which transfers occur, and in which amount, is an external input, and so is the endpoint of any geodesic evolution.

Independent agents: a flat submanifold.

If the agents act independently, with cooperation probabilities a and b, then p = ( a b , a ( 1 − b ) , ( 1 − a ) b , ( 1 − a ) ( 1 − b ) ) . The Fisher metric of this two-parameter family is the product metric
d s 2 = d a 2 a ( 1 − a ) + d b 2 b ( 1 − b ) = 4 d ψ a 2 + d ψ b 2 , a = sin 2 ψ a , b = sin 2 ψ b ,
with ψ a , ψ b ∈ ( 0 , π / 2 ) . This metric is flat, and its geodesics are the straight lines ψ ( s ) = ψ ( 0 ) + w s with 4 | w | 2 = 1 . Thus the correlated four-outcome family has curvature K = 1 / 4 , while its independent-agent subfamily is intrinsically flat. The two geodesic problems are different: a geodesic of the submanifold is in general not a geodesic of Δ 3 , because the independent family is not totally geodesic in the sphere. The choice of statistical model therefore matters for the evolution, and correlation between agents is what makes the geometry curved.

Scope.

The model has finite support (4 states), so the rank-dilation criterion is not applicable and no rank–frequency scaling is claimed. Its purpose is to show how the Fisher metric prices transitions between grammatical states, and how the choice of statistical family changes the geodesic dynamics.

7.5. The Bernoulli Grammar: An Explicit Geodesic

The grammar.

The simplest nontrivial grammar has one non-terminal and two productions, S → θ 1 and S → 1 − θ 0 , with θ ∈ ( 0 , 1 ) . It generates p ( x | θ ) = θ x ( 1 − θ ) 1 − x , x ∈ { 0 , 1 } . This is the one-parameter case of the independent-agent family of Sec. Section 7.4.

Metric and geodesic.

The Fisher metric is
g ( θ ) = 1 θ ( 1 − θ ) .
With θ = sin 2 ψ , ψ ∈ ( 0 , π / 2 ) , one has g d θ 2 = 4 d ψ 2 , so ψ is a flat coordinate and the total Fisher length of the manifold is π . The geodesic equation θ ¨ + Γ θ ˙ 2 = 0 , with Γ = 1 2 g − 1 g ′ = − 1 − 2 θ 2 θ ( 1 − θ ) , reads
θ ¨ = 1 − 2 θ 2 θ ( 1 − θ ) θ ˙ 2 , g ( θ ) θ ˙ 2 = v 2 = const .
Its solution with θ ( 0 ) = θ 0 , ψ 0 = arcsin θ 0 , and ψ ˙ > 0 is
θ ( s ) = sin 2 ψ 0 + v s 2 , L ( s ) = ∫ 0 s g ( θ ) | θ ˙ | d s ′ = v s .
This can be checked by substitution in Eq. (100): both sides equal 1 2 v 2 cos 2 ψ .

Entropy along the geodesic.

The Shannon entropy is S ( s ) = − θ ln θ − ( 1 − θ ) ln ( 1 − θ ) evaluated on Eq. (101). It is not monotone. It reaches its maximum ln 2 when θ = 1 2 , i.e. at s * = 2 ( π / 4 − ψ 0 ) / v , and decreases to zero as the geodesic approaches the boundary. For θ 0 = 0.2 , ψ 0 = arctan 1 2 and v = 1 :
s * ≃ 0.644 , s θ → 1 = π − 2 ψ 0 v ≃ 2.214 , s θ → 0 = 2 ψ 0 v ≃ 0.927 ,
where the last value applies to the geodesic with ψ ˙ < 0 . In both directions the boundary is at finite Fisher distance, in contrast with the Zipf point of Sec. Section 7.2.

Numerical check.

Table 1 gives values of the closed-form solution for θ 0 = 0.2 , v = 1 .

Remarks.

(i) Levi-Civita geodesics are not the exponential-connection geodesics. In the natural parameter η = ln [ θ / ( 1 − θ ) ] the latter are straight lines, η ( s ) = η 0 + c s , and run to η → ± ∞ without reaching the boundary at finite length. The two flows follow different curves between the same endpoints. (ii) As in Sec. Section 7.2, the dimension is one, so every monotone curve θ ( s ) is a reparametrization of a geodesic. The direction and the endpoints are external inputs. (iii) The support has two points, so no rank–frequency scaling is claimed.

8. Conclusions

We have developed an information-geometric framework that models the idealized continuous evolution of probabilistic formal grammars as variational trajectories on statistical manifolds. By adopting the intrinsic Fisher–Rao metric as a coordinate-free measure of local statistical distinguishability, we formulated grammatical parameter deformations as Levi-Civita geodesic flows, providing a geometric baseline that avoids exogenous heuristic cost functions. To address the structural disconnect between parameter-space differential geometry and ordered rank-frequency distributions, we characterized power-law behavior through asymptotic rank-dilation covariance and Karamata’s theory of regular variation. Our analysis clarifies both the scope and the limits of this geometric approach. Fisher–Rao geometry does not by itself generate power-law scaling, nor does it select initial states or evolutionary endpoints; rather, rank-frequency scaling arises from the asymptotic structure of the configuration spectrum, while the metric determines the speed profile and geometric accessibility of parameter trajectories. Within the explicit benchmark of an infinite-state Gibbs family with an asymptotically logarithmic spectrum, we showed that geodesic evolution maps directly onto the trajectory of the rank-dilation exponent α ( s ) = β ( s ) , and that the non-normalizable Zipf limit α = 1 lies at infinite Fisher distance for the zeta distribution. Furthermore, our worked examples on regular and finite-state grammars demonstrate how production rules and inter-agent couplings induce constant negative or positive Fisher curvature and determine the infinitesimal statistical cost of state transitions. Together, these results establish a self-consistent theoretical framework and a falsifiable numerical protocol for testing when variational statistical dynamics governs the evolution of scaling exponents in complex grammatical systems.

8.1. Limitations and Theoretical Boundary Conditions

While the developed framework establishes a mathematically clean mapping for the canonical and grammatical configurations analyzed, certain boundary conditions delimit its immediate generalization to arbitrary complex adaptive systems:
  • Asymptotic Monotonicity Trapping : The analytical mapping between the microscopic state index n and the macroscopic rank variable r relies on the preservation of asymptotic monotonicity under structural perturbations ( η n ) and Markovian mixing operators ( K i n ). In highly non-monotonic or disordered systems where the score functions exhibit severe oscillations, the ordering operator sort ↓ may become non-local, inducing topological discontinuities that disrupt the regular variation limit.
  • Finite Horizon Cutoffs : True regular variation is inherently an infinite-dimensional asymptotic property r → ∞ . Real-world linguistic corpora and discrete computational grammars operate strictly within finite supports N < ∞ . While our ratio formulation under rank dilations ( R λ ) handles sub-leading scale corrections far better than rigid log-log linear regressions, finite-size effects introduce structural truncations that require the introduction of boundary-layer metrics near the edge of the simplex.

References

  1. Amari, S.; Nagaoka, H. Methods of Information Geometry. In Translations of Mathematical Monographs; American Mathematical Society: Providence, RI, 2000; vol. 191. [Google Scholar]
  2. Amari, S. Information Geometry and Its Applications. In Applied Mathematical Sciences; Springer: Tokyo, 2016; vol. 194. [Google Scholar]
  3. Rao, C. R. Information and the accuracy attainable in the estimation of statistical parameters. Bull. Calcutta Math. Soc. 1945, vol. 37, 81–91. [Google Scholar]
  4. Chentsov, N. N. Statistical Decision Rules and Optimal Inference. In Translations of Mathematical Monographs; Original Russian edition: Nauka, Moscow, 1972; American Mathematical Society: Providence, RI, 1982; vol. 53. [Google Scholar]
  5. Bingham, N. H.; Goldie, C. M.; Teugels, J. L. Regular Variation. In Encyclopedia of Mathematics and its Applications; Cambridge University Press: Cambridge, 1989; vol. 27. [Google Scholar]
  6. Booth, T. L.; Thompson, R. A. Applying probability measures to abstract languages. IEEE Trans. Comput. 1973, vol. C-22(no. 5), 442–450. [Google Scholar] [CrossRef]
  7. G. K. Zipf, Human Behavior and the Principle of Least Effort: An Introduction to Human Ecology; Addison-Wesley: Cambridge, MA, 1949.
  8. Hill, B. M. The Rank-Frequency Form of Zipf’s Law. J. Am. Stat. Assoc. 1974, vol. 69(no. 348), 1017–1026. [Google Scholar] [CrossRef]
  9. Newman, M. E. J. Power laws, Pareto distributions and Zipf’s law. Contemp. Phys. 2005, vol. 46(no. 5), 323–351. [Google Scholar] [CrossRef]
  10. Mitzenmacher, M. A brief history of generative models for power-law and lognormal distributions. Internet Math. 2004, vol. 1(no. 2), 226–251. [Google Scholar] [CrossRef]
  11. Clauset, A.; Shalizi, C. R.; Newman, M. E. J. Power-law distributions in empirical data. SIAM Rev. 2009, vol. 51(no. 4), 661–703. [Google Scholar] [CrossRef]
  12. Barabási, A.-L.; Albert, R. Emergence of scaling in random networks. Science 1999, vol. 286(no. 5439), 509–512. [Google Scholar] [CrossRef] [PubMed]
  13. R. Gibrat, Les Inégalités Économiques; Sirey, Paris, 1931.
  14. Bak, P.; Tang, C.; Wiesenfeld, K. Self-organized criticality. Phys. Rev. Lett. 1987, vol. 59(no. 4), 381–384. [Google Scholar] [CrossRef] [PubMed]
  15. Mandelbrot, B. An informational theory of the statistical structure of language. In Communication Theory; Jackson, W., Ed.; Butterworths, London, 1953; pp. 486–502. [Google Scholar]
  16. Miller, G. A. Some effects of intermittent silence. Am. J. Psychol. 1957, vol. 70(no. 2), 311–314. [Google Scholar] [CrossRef] [PubMed]
  17. Ferrer-i-Cancho, R.; Solé, R. V. Least effort and the origins of scaling in human language. Proc. Natl. Acad. Sci. USA 2003, vol. 100(no. 3), 788–791. [Google Scholar] [CrossRef] [PubMed]
  18. Ay, N.; Jost, J.; Lê, H. V.; Schwachhöfer, L. Information Geometry, Ergebnisse der Mathematik und ihrer Grenzgebiete; Springer: Cham, 2017; vol. 64. [Google Scholar]
  19. Petz, D. Monotone metrics on matrix spaces. Linear Algebra Its Appl. 1996, vol. 244, 81–96. [Google Scholar] [CrossRef]
  20. Cafaro, C. Information geometry and chaos on curved statistical manifolds. Int. J. Geom. Methods Mod. Phys. 2008, vol. 5(no. 2), 209–232. [Google Scholar]
Table 1. Fisher geodesic of the Bernoulli grammar, θ 0 = 0.2 , v = 1 .
Table 1. Fisher geodesic of the Bernoulli grammar, θ 0 = 0.2 , v = 1 .
s θ ( s ) S ( s ) L ( s )
0.00 0.2000 0.5004 0.00
0.25 0.3083 0.6177 0.25
0.50 0.4285 0.6829 0.50
0.75 0.5531 0.6875 0.75
1.00 0.6745 0.6309 1.00
1.25 0.7850 0.5205 1.25
1.50 0.8778 0.3713 1.50
1.75 0.9471 0.2071 1.75
2.00 0.9886 0.0625 2.00
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.