DEV Community

Cover image for FORM AND FUNCTION IN ANCIENT WRITING SYSTEMS
Bahaa Budargham
Bahaa Budargham

Posted on Edited on

FORM AND FUNCTION IN ANCIENT WRITING SYSTEMS

Author: Bahaa BuDargham
Affiliation: Engineering of Alphabets (EOA) Program, Neurolabs R&D, Montréal, Canada
Date: November 2026
Primary Reference: Zenodo DOI: 10.5281/zenodo.23858517
License: _Creative Commons Attribution 4.0 International (CC BY 4.0)
_

1. Executive Summary: Reconciling Form and Function
Recent advances in computational paleography (Zemene et al., 2026) applied neural networks to over 28,000 glyph images, identifying significant visual and structural alignments between ancient Ethiopic (Ge'ez) and historical scripts of the Caucasus (most notably Classical Armenian). That work established an objective metric for visual form: stroke curves, angles, and 2D geometric topology.
This report delivers the complementary analysis of linguistic function: evaluating writing systems as deterministic discrete dynamical systems. Under the Engineering of Alphabets (EOA) framework, letter names are treated as deterministic recursive sequence engines. Using exact arbitrary-precision integer arithmetic (BigInt Berlekamp–Massey), Frobenius normal form decomposition, and spectral graph theory (Massey, 1969; Lind & Marcus, 1995), we extract the minimal recurrences, characteristic polynomials, and dominant Perron eigenvalues (λ) that govern each script.
Across the five target writing systems, totaling 178 individual letters, we establish five distinct structural profiles:
• The Golden Ratio Core (Ge’ez, φ² ≈ 2.618): Features a 12-letter recurrent core whose sequences satisfy an exact second-order Fibonacci-bisection recurrence, coupled to a 14-letter transient periphery.
• *The Low-Collapse Network *(Caucasian Albanian, λ ≈ 2.956): Almost every letter generates its own distinct sequence (46 of 49 groups are single letters), yet all 52 letters share one asymptotic growth constant, consistent with a single strongly connected system.

• The High-Collapse Core (Georgian, λ ≈ 2.521): 33 letters collapse into 11 sequence groups, six of them multi-letter (including an 11-letter group and a 5-letter vowel group), all converging to a shared growth rate.
• The Decomposed Regime (Armenian): Fragments into 25 distinct internal sequence groups displaying bimodal growth (slow polynomial and base-2 exponential), demonstrating a divergence between visual form and recurrence dynamics.
• The Unrecursive Baseline (Old South Arabian): Forms an unrecursive baseline where all 29 letters collapse to the constant sequence 1 (λ = 1.000).

2. Methodological Framework & Verification Pipeline
Our analysis follows a structured, three-stage computational pipeline. The symbolic-dynamics apparatus follows Lind & Marcus (1995), and recurrence extraction uses the Berlekamp–Massey procedure (Berlekamp, 1968; Massey, 1969).

Figure 1: The EOA Verification Pipeline, from Letter Name to Algebraic Invariant

The Pipeline on Released Data: Letter ለ (la)
Initial Presentation: The operator is applied to the single letter ለ(la), producing a deterministic integer sequence Lₙ.
Integer Sequence Generation: The sequence, audited to 500 terms in exact arithmetic:
Lₙ = 1, 3, 8, 21, 55, 144, 377, 987, 2584, 6765, 17711, ….
Exact Recurrence Extraction: Applying the Berlekamp–Massey algorithm in arbitrary-precision arithmetic proves that Lₙ satisfies an exact second-order recurrence (Massey, 1969): Lₙ₊₂ = 3Lₙ₊₁ − Lₙ (for n ≥ 1)
Characteristic Polynomial & Dominant Root: P(x) = x² − 3x + 1 = 0 ⟹ λ = (3 + √5) / 2 = φ² ≈ 2.6180339887…
The sequence generation operates via the deterministic EOA operator (withheld under staged disclosure), while the recurrence extraction, polynomial proofs, and spectral classifications are standard, fully reproducible mathematics. A single trajectory does not determine the operator, nor its application to the remaining 177 letters in the census.

3. Master Census Overview: The 5 Target Scripts (178 Letters)
We evaluated all 178 letters across the five target scripts. The results partition the scripts into distinct algebraic archetypes (BuDargham, 2026).
Input Sourcing: Because the EOA operator evaluates literal letter names or labels as input, the resulting sequences depend directly on the chosen phonetic mapping. Across all five scripts, inputs are derived from standard academic transliterations and official Unicode character names: specifically, 1st-order base syllables for Ge'ez, Traditional Mkhedruli names for Georgian, Classical names for Armenian, and official Unicode Character Names for Caucasian Albanian and Old South Arabian. A formal sensitivity check using alternative historical naming conventions is planned for future releases.
Alphabet Sequence Collapse Profile:
• Letters (N): Total alphabet size. Every letter is counted exactly once with zero internal overlap.
• Group Definition: Letters that share identical sequences.
• Distinct Groups: Total distinct sequence groups formed within a script.
• Group Sizes: The exact partition of the alphabet based on the groups formed within a script, denoted in descending multiset notation (e.g., Georgian's {11, 5, 4(×2), 2(×2), 1(×5)} indicates one group of 11 letters, one group of 5, two distinct groups of 4 letters each, two groups of 2 letters each, and 5 single-letter groups).
• Cross-Script Matches: Letters whose sequence is term-for-term identical to a sequence already established by a preceding script in the table.
Table 1: Spectral Overview and Collapse Profiles Across the 5 Target Scripts

*Note on Armenian Accounting: Armenian's 25 internal groups account for 36 letters. The remaining two letters (ն and օ) generate the trivial sequence 1, 1, 1..., matching the Old South Arabian row, totaling 38 letters (36 + 2 = 38). This match is trivial and is recorded solely to maintain strict sequence-row accounting.

Detailed Letter-by-Letter Breakdown (See Attachment):
For a complete letter-by-letter mapping of these groups, please refer to the attached supplementary report (EOA_OEIS_report_All_5_Scripts_Sequences.pdf).
Figure 2: Dominant Asymptotic Growth Spectrum (Perron Eigenvalue λ) Across the 5 Scripts

4. Deep Dive: Ge’ez as a Golden Ratio (φ²) Recurrence System
In recent paleographic machine learning benchmarks (Zemene et al., 2026), Ethiopic was selected as the comparative baseline. Under the EOA audit, Ge'ez resolves into a two-tier structure where all 26 letters asymptotically converge to the Golden Ratio squared (φ² ≈ 2.618034). The 26-letter matrix carries a single spectral radius, shared by the 12-letter recurrent core and the 14-letter transient periphery.
The recurrent core accounts for twelve letters across three sequence groups, all satisfying the minimal polynomial
x² − 3x + 1 = 0 from the first term:
Hub 1 (The Primary 8-Letter Hub):
Letters: ለ, ሠ, ረ, ሰ, ተ, ወ, ጠ, ጰ (la, sza, ra, sa, ta, wa, tha, pha)
Recurrence: Lₙ₊₂ = 3Lₙ₊₁ − Lₙ, characteristic polynomial x² − 3x + 1 = 0.
Identity: Exact term-for-term match to the Fibonacci Bisection F(2n)
(OEIS A001906: 1, 3, 8, 21, 55, 144, 377...).
Hub 2 (The Secondary 3-Letter Hub):
Letters: በ, አ, ፐ (ba, glottal a, pa)
Recurrence: Characteristic polynomial x² − 3x + 1 = 0.
Identity: Exact term-for-term match to the Odd Fibonacci Sequence F(2n−1)
(OEIS A001519: 1, 2, 5, 13, 34, 89...).
Satellite Node (1 Core Letter):
Letter: ፈ (fa). Generates the unshifted Odd Fibonacci Sequence F(2n−1) from the first term (OEIS A001519:
1, 1, 2, 5, 13, 34…). Because its initial index includes the leading 1, it forms its own distinct sequence row while sharing the exact same x² − 3x + 1 = 0 polynomial

The 14 Transient Letters:
The remaining 14 letters (ሀ, ሐ, መ, ዘ, ቀ, ኀ, ነ, ከ, ዐ, የ, ደ, ገ, ጸ, ፀ) function as transient states.
• The 2-Letter Pair: መ, ዘ (ma, za) form a shared sequence group (1,3,10,32,96…). Accounts for the 2 in the multiset.
• The 12 Individual Letters: ሀ, ሐ, ቀ, ኀ, ነ, ከ, ዐ, የ, ደ, ገ, ጸ, ፀ each generate a unique sequence trajectory, accounting for the remaining 12 single-letter rows in the multiset (1+12=13).
• Spectral Convergence: The minimal linear recurrences of these 14 transient letters contain x2 −3x+1
as an algebraic factor, with all other characteristic roots having strictly smaller modulus. This ensures that their asymptotic growth converges to ϕ2 . Full sequence terms and recurrence polynomials for these letters are documented in the Zenodo census data.

How common is the Golden Ratio pattern?
In our broader census of 36 writing systems, this exact φ² algebraic hub recurs across 4 scripts: Ge'ez, Hebrew, Phoenician, and Classical Thai. That represents an observed base rate of approximately 11% across the corpus.
These occurrences are not fully independent: Hebrew and Phoenician share a direct linguistic lineage, and Thai belongs to a script family with historical connections to ancient Semitic scripts. The number of completely independent historical lineages is therefore smaller than four. Whether φ² reflects a general mathematical attractor in letter naming remains an open question to be tested against randomized baseline models.

5. Script-by-Script Spectral Audit
A. Caucasian Albanian (52 Letters): The Low-Collapse Network
• Algebraic Collapse: The 52 letters collapse into 49 distinct sequences because 3 letter pairs (dat/rat, cha/sha, and choy/shoy) produce identical sequences despite having distinct literal names. Under staged disclosure, the exact operator rules responsible for this insensitivity remain to be formally detailed.
• Spectral Convergence: Across all 49 sequence paths, 100% of the alphabet converges toward the dominant growth ratio: λ ≈ 2.956080. The fact that almost every letter generates its own trajectory (46 single-letter rows) suggests a dispersed communicating network consistent with an irreducible transition matrix under Perron–Frobenius theory (Lind & Marcus, 1995).
B. Georgian (33 Letters): The High-Collapse Core
• Cluster Structure: Georgian partitions into 11 distinct sequence groups, displaying strong internal clustering: six multi-letter groups and five single-letter groups.
• Spectral Coupling: Across all 11 sequence groups, including Row 1 (the 5 vowels: an, en, in, on, un, matching OEIS A120925) and Rows 2–11 (the 28 consonants), all 33 letters converge toward the dominant growth rate λ ≈ 2.521380, indicating a coupled internal core.
C. Classical Armenian (38 Letters): The Decomposed Regime
Recent convolutional neural network models identified Classical Armenian as sharing the highest visual similarity with Ethiopic among the tested scripts (Zemene et al., 2026). However, our spectral audit demonstrates that their sequence patterns tell a very different story:
• Structural Shattering: Armenian does not possess a unified dominant eigenvalue. It fragments into 25 distinct sequence families.
• Bimodal Growth Regime:
o Polynomial Regime (λ → 1.0): 22 sequence groups (covering 33 letters) exhibit slow polynomial growth, generating combinatorial numbers (e.g., OEIS A000124, A000125, A000217, A050407, A055795, A180415; Sloane, 2026).
o Base-2 Exponential Regime (λ = 2.0): Letters ի (ini), գ (gim), and ղ (ghad) decouple into powers of two and Mersenne sequences (OEIS A000225; Sloane, 2026).
• Scientific Synthesis: Armenian and Ge’ez represent a clear divergence between morphological shape and recurrence dynamics. This divergence indicates that visual glyph design and letter naming conventions can evolve under distinct rules.
D. Old South Arabian (29 Letters): The Unrecursive Baseline
All 29 letters collapse into a single trivial sequence: 1, 1, 1, 1, 1... (OEIS A000012; Sloane, 2026).
Significance: For Old South Arabian in this specific census, we used atomic inputs: single Unicode consonant signs, not multi-character names. Each sign refers only to itself, so there are no mathematical links between letters, and the operator correctly returns a constant sequence (λ=1.000). We chose this atomic representation on purpose to show how the operator behaves when there are absolutely no recursive links.
Other extinct scripts in our census, such as Caucasian Albanian and Phoenician, give non-trivial growth. This is because they were evaluated through multi-character names or historical values, which create links between letters. Because the input type is fundamentally different, Old South Arabian serves as a strict operational baseline, not a direct algebraic comparison with these scripts.
Future Direction: In a forthcoming paper, we will run the Old South Arabian script using its reconstructed multi-character Proto-Semitic roots (e.g., Alef → 𐩰𐩡𐩱). This will allow us to compare the atomic Unicode baseline directly against the recursive linguistic network, demonstrating how the matrix connects when native multi-character names are applied.
E. Comparative Mechanics: Closure vs. Subordination (Ge’ez vs. Caucasian Albanian)
Although Ge’ez and Caucasian Albanian both report a single dominant growth rate, they reach it in two completely different ways:
• Caucasian Albanian reaches it through closure: the system behaves as if its matrix is strongly connected (irreducible), meaning every letter participates in feedback without a separate transient periphery.
• Ge’ez reaches it through subordination: the system behaves as if its matrix is reducible, featuring a
12-letter Fibonacci core surrounded by 14 transient peripheral letters that feed into it.
This structural difference is testable from sequence data: core letters follow exact recurrences from term one, while transient letters exhibit measurable numerical drift before settling on φ². Full sequence terms and recurrence data for these transient letters are available in the Zenodo census data.
6. Proposed Joint Validation Plan & Next Steps
Why this test matters:
We want to answer a simple question: Did ancient letters get their shapes from how their names are built, or are shapes and names totally separate?
• Possibility A (Connected): Scribes drew similar shapes for letters that share similar naming patterns.
• Possibility B (Independent): Scribes borrowed shapes from other cultures, but kept their own letter names.
• Possibility C (Convergent Evolution): Human handwriting naturally gravitates toward optimal stroke geometries for certain phonological sounds.
• Possibility D (Shared Substrate): Both scripts inherited their structures from an earlier, common ancestral writing system.
We propose a simple, three-step collaborative roadmap to cross-validate these findings alongside neural paleography models (such as Zemene et al., 2026):

Step 1: Compare Shapes Against Math (The Mantel Test)
Researchers can run this cross-validation test immediately using only the per-letter sequences and hub assignments already published on Zenodo; zero access to the proprietary EOA operator is required. We propose a rank-based Mantel permutation test (Mantel, 1967) between the visual similarity matrix derived from neural paleography models (such as Zemene et al., 2026) and a sequence-distance matrix built from the Zenodo census data, with the exact distance definition to be fixed jointly before any test is run.
(Note: Because some letters share the same mathematical sequence, we use a standard ranked comparison so the statistics stay accurate).
What if they don't match? A negative result is an informative finding either way. A null correlation is consistent with the hypothesis that visual glyph shapes and letter names evolved independently. Whether the data shows a correlation or a clear dissociation, the comparison provides a valuable methodological contribution.
Step 2: Check the 8 Main Ge’ez Letters
We test whether the neural paleography model (Zemene et al., 2026) places these 8 primary Ge’ez letters
(ለ, ሠ, ረ, ሰ, ተ, ወ, ጠ, ጰ) into a shared visual cluster.
Step 3: A Joint Publication
If the cross-validation demonstrates statistically significant correlation (or informative divergence), we welcome co-authoring a joint interdisciplinary paper combining these two perspectives: bringing together neural computer-vision paleography with discrete mathematical sequence dynamics.
7. Data Availability & Staged Disclosure Policy
• Open Census Data: The full 36-script census tables, OEIS verification suites, and BigInt audit scripts are registered under Creative Commons Attribution 4.0 International (CC BY 4.0) at Zenodo: DOI: 10.5281/zenodo.23858517 (BuDargham, 2026).

• Staged Disclosure Terms: To manage the ongoing development of the EOA β-operator while enabling full academic verification, the raw N×N incidence matrices and generative code are available under staged disclosure. Complete matrix datasets and collaborative co-authorship pipelines are provided under standard institutional research agreements.

References
Berlekamp, E. R. (1968). Algebraic Coding Theory. McGraw-Hill.
BuDargham, B. (2026). The Alphabet as a Sequence Engine: A Census of What Letter Names Generate, and What the OEIS Has Never Seen. Zenodo. DOI: 10.5281/zenodo.23858517.
Lind, D., & Marcus, B. (1995). An Introduction to Symbolic Dynamics and Coding. Cambridge University Press.
Mantel, N. (1967). The detection of disease clustering and a generalized regression approach. Cancer Research, 27(2), 209–220.
Massey, J. L. (1969). Shift-Register Synthesis and BCH Decoding. IEEE Transactions on Information Theory, 15(1), 122–127.
Sloane, N. J. A. (Ed.). (2026). The On-Line Encyclopedia of Integer Sequences. Published electronically at https://oeis.org.
Zemene, D., et al. (2026). Machine learning techniques for exploring influence, commonalities, and shared origin of scripts: Cases of Ethiopic, Armenian, Georgian, and Caucasian Albanian scripts. Digital Scholarship in the Humanities, 41(2), 1092. https://doi.org/10.1093/llc/fqag029

Top comments (0)