Nonverbal arithmetic in humans: Light from noise

Sara Cordes, C. R. Gallistel, Rochel Gelman, Peter E. LathamPerception 2007 · DOI

Abstract

Humans possess an innate, nonverbal system for representing quantities as "mental magnitudes." This study investigates the internal representation of these magnitudes by analyzing the variability, or "noise," in human responses during nonverbal arithmetic tasks. By applying noise propagation analysis, the researchers provide evidence that the internal number line is linear rather than logarithmic, with variability that grows proportionally with the magnitude of the number.

Introduction

While it is well-established that humans and animals share a "number sense" for estimating quantities without counting, the underlying mechanics of how we perform calculations with these estimates remain a subject of debate. Two primary models compete to describe the internal number line: the logarithmic model, where distances between numbers are represented like decibels, and the linear model, which posits a linear scale where "noise" (uncertainty) grows proportionally with the number—a phenomenon known as scalar variability. This paper seeks to resolve this conflict by analyzing how noise propagates from initial estimates (operands) to final arithmetic results.

Methods

The researchers conducted psychophysical experiments with four adult subjects. To isolate nonverbal processes, two specific constraints were implemented: stimuli (dot flashes) were presented too rapidly (100ms) for counting, and participants performed "articulatory suppression" by repeating the word "the" throughout the trials to block internal verbalization.

The experimental design included three primary tasks:

  1. Baseline: Participants observed $N$ flashes and were asked to tap a button $N$ times.
  2. Addition: Participants observed two sequences ($N_1$ and $N_2$) and tapped the sum ($N_1 + N_2$).
  3. Subtraction: Participants observed two sequences and tapped the absolute difference ($|N_1 - N_2|$).
Schematic of the experimental tasks (Task 1: Baseline; Task 2: Arithmetic).

To analyze the data, the authors employed the Killeen variance equation to decompose response noise into three distinct sources: scalar noise (proportional to the number), binomial noise (errors in the physical tapping process), and additive noise (constant background variance).

Results and Key Insights

The study's findings strongly support a linear model of internal number representation.

Scalar Variability and Linearity

The data revealed that the standard deviation of responses was directly proportional to the mean response across all subjects. This hallmark of scalar variability is a defining characteristic of a linear system. As the number of flashes increased, the "blurriness" of the mental representation increased at a constant ratio.

Baseline response variability across four subjects, showing the linear relationship between the number of flashes and the number of responses.

Noise Propagation in Arithmetic

A key insight of the paper is "operand dominance": the variability in an arithmetic result is inherited from the uncertainty of the input numbers. The variance of the sum or difference was found to be approximately equal to the sum of the variances of the individual operands, plus an additional noise component related to the magnitude of the final answer. Interestingly, addition was consistently more variable than subtraction for the same pair of numbers, suggesting that the magnitude of the result itself contributes its own layer of scalar noise.

The Subtraction Evidence

The logarithmic model faces a significant hurdle with subtraction: $\log(0)$ is undefined. If the brain operated on a logarithmic scale, performance should degrade significantly as results approach zero. However, the researchers found that human performance remained smooth and consistent even for subtraction results of zero or near-zero, further supporting the linear representation hypothesis.

Discussion

The application of the Killeen variance equation provided a superior fit for the behavioral data compared to simpler models, confirming that human estimation is influenced by multiple robust noise factors, including setup and motor execution noise.

Limitations and Future Directions

The study's primary limitation is its small sample size (four subjects), though this is mitigated by the high volume of trials per participant (approx. 750). Additionally, while articulatory suppression is effective, it is difficult to entirely rule out the possibility that numerically fluent adults might occasionally use symbolic shortcuts.

Future research may apply this noise-propagation framework to other analog domains, such as the estimation of time or distance. Determining if the brain uses a universal combinatorial engine for all magnitudes could provide profound insights into the fundamental nature of human intuition and cognition.

Summary updated

BibTeX
@inproceedings{cordes2007nonverbal,
  author = {Sara Cordes and C. R. Gallistel and Rochel Gelman and Peter E. Latham},
  title = {{Nonverbal arithmetic in humans: Light from noise}},
  booktitle = {Perception},
  year = {2007},
  doi = {10.3758/bf03193955},
}
Figure