Red, green, blue, and yellow feel fundamental — unmixed, irreducible, uniquely themselves. Red doesn't seem like it contains other colors the way orange contains red and yellow. This phenomenological fact, the existence of four “unique hues,” has no established neural correlate. We know the retina has three cone types. We know opponent processing creates red-green and blue-yellow channels. But why these four hues appear unique — why they feel like primitives rather than mixtures — remains unexplained by physiology alone.
Belsten, Frady, and Olshausen (arXiv:2603.24293) show that the four unique hues emerge from sparse coding of color statistics in natural scenes. Simulating cone responses across 503 calibrated natural images, they find a strongly non-Gaussian distribution in 3D color space with heavy tails in distinct, asymmetric directions. When a sparse coding model is trained on this data — minimizing the total sum of coefficients needed to represent the distribution — a six-basis-vector solution converges to the four unique hues plus black and white.
Note (2026-11-12): checked against the paper's full text. The pipeline and the result match: cone responses to 503 calibrated natural images, a sphering step that yields cone-opponent axes from the data, and a nonnegative sparse coding model whose six-vector solution aligns with the four unique hues plus black and white, with excitatory and inhibitory interactions from the nonlinear inference. One scope point: the six-vector solution is one of a series of models with four to eight basis vectors. In the paper the other sizes give different bases (at seven, the black vector splits into two with slight hue components; at eight, a violet vector appears), and the model is constrained to nonnegative coefficients; the paper argues that the six-vector alignment would be unlikely by chance. 'The efficient alphabet' therefore refers to that six-vector model, and the link between irreducibility and a basis vector is the essay's gloss on the paper's 'linking principle'.
The key mechanism is the nonlinearity of sparse inference. Excitatory interactions between adjacent basis vectors enable intermediate hues: the representation can combine, say, red and yellow to encode orange. But inhibitory interactions between opposite pairs enforce mutual exclusivity: red suppresses green, blue suppresses yellow. The opponent structure isn't imposed by the model architecture — it emerges from the optimization.
The result connects subjective color experience to environmental statistics through a single computational principle. The world's surfaces, illuminants, and reflectances create a specific distribution of cone activations. Efficient coding of that distribution — representing it with minimal total activation — naturally produces basis vectors at the positions we experience as unique hues. The feeling of irreducibility is the perceptual signature of a basis vector: something that participates in representing other colors but is not itself decomposed.
This doesn't explain consciousness. It explains why, given that we perceive color at all, we perceive these particular colors as fundamental. The unique hues aren't arbitrary neural conventions. They're the efficient alphabet for the world's actual color statistics.