A new preprint posted to arXiv suggests that the four so-called 'unique hues' — red, green, blue, and yellow — emerge naturally from the statistical structure of the visual world, rather than from hardwired biological quirks of the human eye or brain.

Researchers analyzed simulated cone responses across a dataset of 503 calibrated natural images and found that the distribution of color signals in three-dimensional color space is strongly non-Gaussian, featuring heavy tails that extend in specific, asymmetrically arranged directions. These directional biases, they argue, are not random.

The team then applied a sparse coding model — a mathematical framework that seeks the most efficient representation of data using as few active components as possible — to this color distribution. When configured with six basis vectors, the model reliably converged on representations corresponding to the four unique hues (red, green, blue, and yellow), plus black and white. No manual tuning toward these outcomes was required; they emerged from optimization alone.

The model also reproduced two hallmark features of human color perception: the ability to combine adjacent unique hues to perceive intermediate colors (such as orange between red and yellow), and the mutual exclusivity of opposite pairs, meaning observers cannot perceive a color as simultaneously red-green or blue-yellow. These opponent relationships, long documented in psychophysics, fell out of the model's nonlinear inference dynamics as excitatory interactions between adjacent hues and inhibitory interactions between opposing ones.

The study offers a 'linking principle' between the statistical regularities of natural scenes and the phenomenology of color appearance — a longstanding puzzle in vision science. Prior theories have variously attributed unique hues to cone photoreceptor tuning, postreceptoral opponent channels, or cultural learning, but none has provided a fully satisfying computational account grounded in environmental statistics.

Important caveats apply. The work is a preprint posted at https://arxiv.org/abs/2603.24293 and has not yet undergone peer review. The image dataset, while calibrated, comprises 503 images — a relatively modest sample that may not capture the full diversity of natural lighting and surface conditions across ecological environments. The sparse coding framework is a computational abstraction and may not map directly onto specific neural circuits. The study also does not address cross-cultural variation in unique hue perception, which has been documented in the psychophysical literature and could complicate a purely statistical environmental account.