Color theory, from light to meaning

Start with the band itself. Human eyes usually respond from about 380 nanometers at the violet end to about 700 at the red end, which is a frequency ratio of only 1.84, less than one musical octave. Newton's 1704 Opticks made the key distinction: a ray has no color in it, only a power to stir a sensation, the way a trembling string has no tone until someone hears it.

Beginners treat "the color of an object" as a fact about the object. A surface really has a habit of returning some wavelengths more than others, so what you see depends on what it was given. Newton placed paper, ashes, red lead and blue flowers in one separated color of light, and each took on that color at some brightness. Even in red light, red lead looked dimmer than white paper.

A second error is giving every color one explanation. A star emits, and hotter means bluer: about 3,000°C looks reddish and about 12,000°C looks blue. The sky scatters short blue waves more than long ones. The sea absorbs red and leaves blue. These are three separate routes.

Purple is a useful test case. Mixing the spectrum's two ends gave Newton purples unlike any single separated light, so the wheel holds hues the spectrum lacks.

Rule: when you explain a color, name the light, the material's route and the observer before you name the color.

You have L, M and S cones, peaking near 560, 530 and 425 nanometers. Calling them red, green and blue misleads: the L peak sits about 140 nanometers short of the red end, and all three respond across wide, overlapping ranges.

Each cone reports one number, which mixes wavelength with intensity. A 2019 review gives lights at 467 and 582 nanometers that produce the same M-cone response. One cone type therefore cannot see color at all. Color appears only when responses are compared.

Thomas Young reasoned his way to a few receptors in 1801, before anyone could measure one. His first three were red, yellow and blue, and later work moved to red, green and violet.

Trichromacy explains matching, but it leaves negative afterimages unexplained. Hering proposed opposed channels in 1878, and later physiology found relay cells excited by L and suppressed by M, and others setting S against L plus M. Asking which theory won misses the point, since current accounts keep both.

Keep two claims apart. Cone comparisons are not yet perceived hue, as a 2022 macaque study of visual cortex suggests. And the old idea of four pure "unique" hues remains contested: a 2017 study reported a brain marker, while its neural basis stays unsettled.

Rule: read a color signal as a comparison between cones, never as one cone's report.

Because each cone sends only one number, most spectral detail is discarded at the first step. A lemon's spectrum and a screen's three narrow primaries can both land on the same three cone responses, and then you cannot tell them apart. Every display depends on this. Munsell noted in 1907 how much it surprised painters: red and green light reach the eye as yellow, with nothing at yellow wavelengths required.

The common mistake is reading a match as sameness. A match is a fact about one observer in one setting, and two things can break it.

Change the light, and you get illuminant metamerism. Two fabric samples may match under shop lamps and split apart at home, because different reflectance spectra respond differently to a new source. Change the observer, and you get observer metamerism. Displays are generally tuned to a 1931 standard observer, a statistical stand-in that does not exactly describe any given person.

Technology can sharpen the second problem. Laser cinema projectors reach a wider gamut because each primary is very narrow, and that narrowness also makes small personal differences in cone sensitivity matter more. So two people with typical vision can disagree in front of a well-calibrated screen, and neither needs to be wrong.

Rule: if a match matters, test it under the light it will live in, and with more than one pair of eyes.

Combining beams adds energy. Red and green light make yellow, and spectral yellow and blue make white. Stirring paint mostly removes light. Blue and yellow paint make green because together they absorb most of the spectrum and leave what both return. Munsell lists stirred pairs such as red and green giving a yellow-gray, which is why a beginner's mix keeps coming out dull.

Trying to mix painters' powders into white, Newton got gray or dun brown, since each powder has already thrown away much of the light. Filters show the arithmetic cleanly: one passing 80% of a wavelength and another passing 50% leave 40%, because stacked filters multiply.

The third operation is mixing in the eye. Newton noticed that soap froth shows many colors up close and looks white from far enough away. Halftone printing depends on the same limit.

Numbers on a screen add one more trap, because code values are encoded. Decoding gives 0.216 for a value of 128, so #808080 carries about 22% of full light, and half the light needs a value near 188. Average pure red and pure green in code and you get rgb(128 128 0), luminance about 0.20. Average the light and you get a luminance of 0.46, close to rgb(188 188 0). The CSS specification warns that mixing in sRGB gives darker, grayer results.

Rule: ask whether light is being added, removed or blended by distance, then decide whether code values need decoding first.

Lovibond's 1914 table of past theories runs from seven primaries down to three. He also found the root of the old quarrel: scientists picked red, green and violet by mixing lights, and artists picked red, yellow and blue by mixing pigments. Munsell was already calling red, yellow and blue as true primaries a harmful teaching in 1907, while using five principal hues of his own. Red, yellow and blue remain a working painter's convention.

The wheel has the same history. Newton closed the spectrum into a circle, divided in the proportions of a musical scale, and used it as a calculator for mixtures. Goethe arranged his around the colors the eye seems to call up, and his 1810 book is valuable for that perception work. His account of the prism, set against Newton's, is wrong physics.

The sharpest trap is the word "complementary." Ask what partners red, and you get cyan by adding light, green on a painter's wheel or in Hering's pairs, and blue-green from Munsell. A 2020 study of afterimages found that no circle places those pairs as true diameters.

Even hue spacing is a choice. HSL (hue, saturation and lightness) puts red, yellow and green at 0°, 60° and 120°. Computed in the Oklab space, the hue angles are about 29°, 110° and 142°, so those equal steps are not equal to the eye.

Rule: before using "primary" or "complementary," name the system, because the answer changes with it.

Munsell's book opens on a letter from Samoa, in which Stevenson tries to order wallpaper using words like "topazy yellow." Names like that vary with the speaker. His answer was three separate scales.

Hue is a circle of ten named steps written with initials such as R, YR and BG. Value is a neutral scale from black at 0 to white at 10. Chroma is distance from gray, and it has no fixed top, because the strongest available pigment sets the limit for each hue. So R 5/9 reads as red at value 5 and chroma 9, his notation for vermilion.

The usual mistake is letting one word do several jobs. "Make it brighter" could mean a lighter value or a stronger chroma.

Measurement mattered to him because eyes disagree. He found school gray scales varying by over 10%, and measured one person's chosen middle gray at only 42% of the light, which he counted as nearly a full step too low. That implies a middle near 50% of the light. The modern Lab midpoint is about 18%, by the published formula. So "middle gray" needs the question: middle in energy, or in appearance?

His complements also use all three dimensions. A violet looks opposite a buttercup, yet it falls far short of the buttercup's chroma, so in his terms it cannot balance it.

Rule: restate any color request as a change in hue, value or chroma, and name the one you mean.

RGB (red, green and blue) is a model, while sRGB is a space with fixed primaries, a set white and an encoding curve. Browsers assume sRGB for colors with no tag, and must convert an image tagged for another space. Reuse the raw numbers and the colors come out wrong.

HSL is friendly and misleading. Blue and yellow both read 50% lightness there, while Lab puts them near 30 and 98. Lab, from 1976, was built on human vision data so that lightness steps look even. It bends saturated blues toward purple as chroma drops. Oklab was fitted to data on similar-looking colors, handles that better, and lets you measure difference as a plain straight-line distance.

A gamut is a volume, so always ask "in gamut for what?" A color outside it is legal and cannot be shown, and clipping it can change the hue, by 69° in one CSS specification example. Better mapping reduces chroma while holding hue. The space you mix in matters too: linear spaces model combined lights, and Oklab gives even steps.

Print raises the stakes, because cyan, magenta, yellow and black values are ink amounts with no fixed color. The CSS Color 5 draft works one value through two print conditions and gets a difference of 8.17 on the delta E 2000 scale, and 14.3 with a naive formula. Add that a print returns room light while a screen makes its own.

Rule: never pass color numbers along without the name of their space or print condition.

Complementary, triadic and analogous schemes are rules about angles. Since wheels differ, a scheme's output depends on an arbitrary starting choice, so it cannot be a law of perception on its own. That tells you what kind of tool they are.

The belief that opposites harmonize has been traced to Goethe. Munsell was already skeptical in 1907, writing that attempts to state laws of harmony had not succeeded. His example was a recipe of three parts yellow, five red and eight blue, which never says which yellow, red or blue. Chevreul's six harmonies, from analogy and from contrast, record a dye chemist's long craft experience. The book reports them as practice, not as tested effects.

Later measurement complicates the picture. A 2015 paper reports Schloss and Palmer's distinction between pair preference, pair harmony and figural preference, and their finding that both preference and harmony ratings fall as hue contrast rises, while preference rises with lightness contrast. You are reading that second-hand, so hold it with some care. The same paper's own imaging data comes from 18 people. A 2022 study found observers preferring a painting's original colors to hue-rotated versions, even when the image was scrambled, so color distribution seems to matter in ways a wheel does not capture.

Beginners treat a scheme as a guarantee and blame themselves when it looks wrong.

Rule: use schemes to generate candidates, then judge lightness, chroma and area in the actual arrangement.

The founding story is about black wool. Running the Gobelins dye works, Chevreul heard complaints that the blacks in blue and violet draperies looked weak. Black wool from other dye works was no better, because the fault lay in the neighboring colors. His law says two adjacent colors look as different as they can, in hue and in depth of tone. Staring at a color and then looking away gives a related effect, afterimages, which recordings suggest start with rebound responses in retinal ganglion cells.

The mistake is assuming the surround always pushes away. In assimilation, a target moves toward the color of its surroundings instead. A 2019 model treats the two as acting together, near neighbors pulling and remote ones pushing.

Constancy is the larger achievement. The light reaching your eye blends illumination with surface, and no unique answer can be pulled from that alone. How well people manage is still debated. Lab tasks report 15% to 80%, while a 2024 study of remembered familiar objects under five office lights found people mostly picked the same chip. Scenes with real shadows show clear failures: a 2014 study found the same paint shifting in lightness, hue and chroma.

The 2015 dress photo split viewers between blue and black, its true colors, and white and gold. The leading explanation is differing assumptions about the light, and one small study's added causes are not settled.

Rule: judge a color in its final surroundings and lighting, never from an isolated sample.

Start by dropping "sees in gray." Complete loss of color is rare, about 1 in 30,000 by one review's estimate. Most cases are red-green, split into altered forms, protanomaly and deuteranomaly, and absent forms, protanopia and deuteranopia. Deuteranomaly, the most common, is mild.

The genetics follow from one fact: L and M pigment genes sit on the X chromosome. Males need one altered copy, and females usually need two. Squaring 8% gives 0.64%, which roughly fits the 0.5% often reported for women of that ancestry.

Beginners often quote 8% of men as a world figure. It describes Northern European ancestry, and a 2024 meta-analysis of 16 studies in Africa pooled 2.71% overall. Many people find out late, since screening is limited and mild cases go unnoticed.

John Dalton offers a good case of testing a guess. He confused scarlet with green and suspected a blue tint in his eye fluid. Examination after death found it clear, and DNA much later showed the middle-wave pigment was the one he lacked.

Filter glasses are contested. In a 2022 study of 16 people, nobody moved category on a clinical test with them on, while a study the National Eye Institute publicized reported gains after two weeks of use. Orientation for study, not clinical advice: a change in color vision is a reason to see an eye doctor.

Rule: never make color the only carrier of a message, since some readers cannot rely on it and may not know why.

WCAG, the Web Content Accessibility Guidelines, defines relative luminance by decoding each sRGB channel and then weighting it: 0.2126 red, 0.7152 green, 0.0722 blue. Green carries about ten times the weight of blue. For the contrast ratio, add 0.05 to each luminance and divide the lighter by the darker, so white on black gives 21:1.

Level AA sets a floor of 4.5:1 for ordinary text. Large text, meaning 18 point or 14 point bold, needs 3:1. Interface parts and meaningful graphics need 3:1. The boundary is invisible to the eye: on white, #767676 computes to 4.54 and #777777 to 4.48.

Beginners assume strong hue difference means strong contrast. Pure red on #008000 green is a classic opposite pair on a painter's wheel, and it computes to 1.28:1. Pure yellow on white is 1.07:1, and pure blue on black is 2.44:1.

Every solid background clears 4.5:1 with black or with white text, because the worst case sits at the square root of 21, about 4.58. The pivot is a luminance near 0.179, close to gray code value 118, below the middle of the code range.

Passing is a floor. The formula ignores font and surroundings, and color alone must not carry meaning. In figures, a simulation over-warns: of images in one journal, 85.6% held a risky pair, and hand review of a sample judged 12.8% hard to read. Rainbow data maps are a real problem, since uneven steps add artificial boundaries.

Rule: compute the ratio, then add a second cue besides color.

Start by separating association from effect. Associations are well documented: a classifier given 711 people's ratings picked which of 12 color terms they were rating 38.7% of the time, against 8.33% by chance, and picked their country 80.2% of the time, against 25%. Words and patches give nearly the same patterns, which suggests much of this lives in concepts. Red alone has been linked to love, anger and power, so no one-to-one chart holds.

Effects are another matter. A 2007 series reported lower scores after a little red before a test, with Cohen's d between about -0.51 and -1.14. A 2020 meta-analysis of 67 effects found about -0.06 for anagrams, -0.04 for knowledge tests and -0.34 for reasoning. Once publication bias was corrected for, nothing reliable remained. Early studies ran much larger than later ones, the usual warning sign.

Newer claims often rest on single experiments: red speeding a walk to a dating interview and slowing a walk to an intelligence interview, or red reversing its effect for Chinese stockbrokers, whose market shows rises in red. A 2015 review called the field nascent, citing poor blinding and colors not controlled for lightness or chroma.

Preferences vary too. Pink for girls was absent in three small-scale societies and large in one global city. Warm and cool are perceptual labels that reverse physics, since a hotter star is bluer.

Rule: when you read "red makes people," ask what was manipulated, in how many people, and whether anyone replicated it.