Numeral Systems as Evolving Cognitive Technologies: Error Propagation, Compactness, and the Non-Optimality of Positional Notation

In the paper's words

Positional (Arabic-Indic) notation is conventionally treated as the endpoint of numeral-system evolution, but this paper argues that it is a historically contingent survivor whose dominance reflects fitness for particular task environments rather than global optimality. We formalize the tradeoff between additive systems (Roman-style symbol sums, modeled here as pure commutative sums without subtractive rules) and positional systems under explicit transcription-noise channels. Under a symbol-deletion channel we prove that additive error is bounded by the largest symbol value (a constant, $1000$), while positional error is (9H + d_k)· 10^k — proportional to the number's higher-order digits and unbounded in magnitude; under uniform-digit and uniform-deletion-position assumptions we derive an expected positional deletion error of $2187.5$ for four-digit numbers in [1000,3999] and an expected additive deletion error of approximately $449.4$, a ratio of about $4.87$.

The abstract, verbatim, and its opening sentence is the claim. Every line below is quoted from the text.

Numbers it reports

1. Under a symbol-deletion channel we prove that additive error is bounded by the largest symbol value (a constant, $1000$), while positional error is $(9H + d_k)\cdot 10^{k}$ — proportional to the number's higher-order digits and unbounded in magnitude; under uniform-digit and uniform-deletion-position assumptions we derive an expected positional deletion error of $2187.5$ for four-digit numbers in $[1000,3999]$ and an expected additive deletion error of approximately $449.4$, a ratio of about $4.87$.

   from Abstract

2. Positional notation wins on compactness (four digits versus an average of $4.45$ greedy-additive symbols over $1$–$3999$, ratio $4.45/4 \approx 1.11$) and is argued qualitatively to aid arithmetic tractability.

   from Abstract

3. A two-term cost model yields an explicit crossover: additive notation would be optimal whenever the error-aversion weight exceeds roughly $0.26$ times the per-symbol cost.

   from Abstract

4. Our answers, derived in full in Section 4, are: (1) additive systems have strictly bounded error propagation under deletion and exactly zero error under transposition, while positional error is unbounded in the number's magnitude; (2) additive notation costs a computable factor of $4.45/4 \approx 1.11$ in symbol count for the range examined; (3) no — the axes of the landscape (error resistance, compactness, arithmetic tractability) are satisfied by different systems in different environments, and a two-parameter cost model makes the crossover explicit.

   from 1. Introduction

5. A task environment assigns a per-symbol cost $\mu$ (time, ink, cognitive load) and an error-aversion weight $\lambda$ per $10^{3}$ units of expected error.

   from 3.3 Cost model

6. The cost of system $s$ is $$\kappa_s = \mu \, \bar{L}_s + \lambda \, \frac{\bar{E}_s}{10^{3}},$$ where $\bar{L}_s$ is expected string length and $\bar{E}_s$ expected corruption error.

   from 3.3 Cost model

How it is done

1. Under a symbol-deletion channel we prove that additive error is bounded by the largest symbol value (a constant, $1000$), while positional error is $(9H + d_k)\cdot 10^{k}$ — proportional to the number's higher-order digits and unbounded in magnitude; under uniform-digit and uniform-deletion-position assumptions we derive an expected positional deletion error of $2187.5$ for four-digit numbers in $[1000,3999]$ and an expected additive deletion error of approximately $449.4$, a ratio of about $4.87$.

   from Abstract

2. The supplied summary gives no numeral-specific content; we use it only as a parallel case of a technology whose adoption is gated by task-environment fit rather than intrinsic merit — the same gating we claim for positional notation. [3] A cognitive diversity framework for radar target classification.

   from 2. Background and Related Work

3. For concrete computation we use the greedy additive system $\mathcal{A}$ with the seven-symbol inventory $V=\{1,5,10,50,100,500,1000\}$, without subtractive rules, so that representation is a pure commutative sum and the greedy algorithm (take as many $v_m$ as possible, then $v_{m-1}$, and so on) is canonical.

   from 3.1 Notation families

4. Following the error-model template of robust design under parameterized channel uncertainty [8], we define two channels: Deletion channel $\mathcal{D}$: one symbol of the written form is deleted uniformly at random; the remaining symbols are read in written order (for positional strings, adjacent digit blocks merge).

   from 3.2 Noise channels

What would falsify it

1. We borrow the structural move — utility maximization under constraint, with an intermediate representational layer — and apply it to cultural artifacts: historical communities selecting notations satisficed under local task pressures, which predicts exactly the contingent, non-optimal outcomes we formalize. [6] Offloading Cognition onto Cognitive Technology.

   from 2. Background and Related Work

2. Falsification would occur if measured human transcription errors for positional notation consistently fell below the projected additive error across comparable tasks.

   from 6. Discussion

Where the claim stops

1. The projections rely on simplifying assumptions: uniform digit distributions, uniform deletion positions, and the greedy additive representation without subtractive rules.

   from 6. Discussion

2. Moreover, the tractability claim remains qualitative; a rigorous metric (e.g., average algorithmic step count) is left for future work.

   from 6. Discussion

3. Under a symbol-deletion channel we prove that additive error is bounded by the largest symbol value (a constant, $1000$), while positional error is $(9H + d_k)\cdot 10^{k}$ — proportional to the number's higher-order digits and unbounded in magnitude; under uniform-digit and uniform-deletion-position assumptions we derive an expected positional deletion error of $2187.5$ for four-digit numbers in $[1000,3999]$ and an expected additive deletion error of approximately $449.4$, a ratio of about $4.87$.

   from Abstract

Where to read

1. Abstract222 words · 1 min
2. 1. Introduction373 words · 2 min
3. 2. Background and Related Work871 words · 4 min
4. 3. Methods311 words · 1 min
5. 4. Analysis759 words · 3 min
6. 5. Results57 words · 1 min
7. 6. Discussion93 words · 1 min
8. 7. Conclusion74 words · 1 min