Color naming needs a non-uniform perceptual prior: Individual differences in hue scaling and the vision–language interface
Abstract
How a language divides the color spectrum into named categories is increasingly explained as efficient compression: a vocabulary is a code that names colors as accurately as possible with as few words as possible. Every such model needs a description of how similar colors look, and nearly all use the same one: a standard, evenly spaced color space (CIE L*a*b*, or CIELAB) in which each color carries a fixed, symmetric blur of perceptual uncertainty. Here I argue that this assumption is wrong in a way that matters. When people with normal color vision judge hues they disagree in systematic ways, and the pattern of that disagreement, summarized by factor analysis, points to roughly eight narrow, largely independent regions that tile the color circle and lie near category boundaries rather than their best examples. I show how to build this measured structure into the standard model by stretching the circle at the category boundaries. A full test substitutes the reshaped space into the information-bottleneck model and rescores all 110 World Color Survey languages. Under the model’s near-uniform default prior the warp is close to a wash; but under a realistic, measured communicative need a modest warp improves fit for most languages, a small but systematic benefit the uniform prior masked, although the predicted optimal category systems themselves barely change. The effect is small and conditional, marking when a non-uniform perceptual space matters. The proposal yields a prediction testable with public data and code, informative whether it succeeds or fails.
Related
- Individual differences reveal shared computational structure in perceptual decision-making
- Individual differences in Game Transfer Phenomena are associated with creativity and positive schizotypy but not with perceptual priming
- A brain-behavior generative model links individual brain dynamics to perceptual decisions
- Alignment of visual brain encoding models with individual perceptual similarity judgments
- Respiration and individual differences shape behavioral and inter-brain synchronization
BibTeX
@misc{peterzell2026color,
title = {{Color naming needs a non-uniform perceptual prior: Individual differences in hue scaling and the vision–language interface}},
author = {David Henry Peterzell},
howpublished = {PsyArXiv},
year = {2026},
month = oct,
doi = {10.31234/osf.io/u69mg_v3},
url = {https://osf.io/preprints/psyarxiv/u69mg_v3/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Color naming needs a non-uniform perceptual prior: Individual differences in hue scaling and the vision–language interface
AU - David Henry Peterzell
PY - 2026
DA - 2026/10/19
PB - PsyArXiv
DO - 10.31234/osf.io/u69mg_v3
UR - https://osf.io/preprints/psyarxiv/u69mg_v3/
AB - How a language divides the color spectrum into named categories is increasingly explained as efficient compression: a vocabulary is a code that names colors as accurately as possible with as few words as possible. Every such model needs a description of how similar colors look, and nearly all use the same one: a standard, evenly spaced color space (CIE L*a*b*, or CIELAB) in which each color carries a fixed, symmetric blur of perceptual uncertainty. Here I argue that this assumption is wrong in a way that matters. When people with normal color vision judge hues they disagree in systematic ways, and the pattern of that disagreement, summarized by factor analysis, points to roughly eight narrow, largely independent regions that tile the color circle and lie near category boundaries rather than their best examples. I show how to build this measured structure into the standard model by stretching the circle at the category boundaries. A full test substitutes the reshaped space into the information-bottleneck model and rescores all 110 World Color Survey languages. Under the model’s near-uniform default prior the warp is close to a wash; but under a realistic, measured communicative need a modest warp improves fit for most languages, a small but systematic benefit the uniform prior masked, although the predicted optimal category systems themselves barely change. The effect is small and conditional, marking when a non-uniform perceptual space matters. The proposal yields a prediction testable with public data and code, informative whether it succeeds or fails.
ER -