[css-color-4] Linear Bradford matrices are inexact

[svgeesus]

From :

Conversion from colors specified using other white points is called a chromatic adaptation transform, which models the changes in the human visual system as we adapt to a new lighting condition. The Bradford algorithm [Bradford-CAT] is the industry standard chromatic adaptation transform, and is easy to calculate as it is a simple matrix multiplication.

As pointed out here the inverse matrix in color.js uses rounded off values, and in fact the matrices in the specification were calculated using those rounded-off values

// Chromatic adaptation

function D65_to_D50(XYZ) {
	// Bradford chromatic adaptation from D65 to D50
	// The matrix below is the result of three operations:
	// - convert from XYZ to retinal cone domain
	// - scale components from one reference white to another
	// - convert back to XYZ
	// http://www.brucelindbloom.com/index.html?Eqn_ChromAdapt.html
	var M =  [
		[  1.0479298208405488,    0.022946793341019088,  -0.05019222954313557 ],
		[  0.029627815688159344,  0.990434484573249,     -0.01707382502938514 ],
		[ -0.009243058152591178,  0.015055144896577895,   0.7518742899580008  ]
	];

	return multiplyMatrices(M, XYZ);
}

function D50_to_D65(XYZ) {
	// Bradford chromatic adaptation from D50 to D65
	var M = [
		[  0.9554734527042182,   -0.023098536874261423,  0.0632593086610217   ],
		[ -0.028369706963208136,  1.0099954580058226,    0.021041398966943008 ],
		[  0.012314001688319899, -0.020507696433477912,  1.3303659366080753   ]
	];

	return multiplyMatrices(M, XYZ);
}

which produce small inaccuracies when round-tripping. These can accumulate to the point they introduce small, but visible, artifacts.

Also there are also two Bradford transforms: the original, given in the references section, and the linear simplification introduced by the ICC which is a simple 3x3 matrix.

The spec says "linear Bradford" in 11. Converting Colors and elsewhere just uses "Bradford"; that should be corrected for clarity.

[svgeesus]

Related: linear Bradford slightly better than CAT16

Correct reference for linear Bradford would be E.3 Linearized Bradford transformation in ICC.1-2022-05

[svgeesus]

Matrices fixed by 9c4db75

See https://github.com/LeaVerou/color.js/pull/354/files

[svgeesus]

linear vs. original Bradford disambiguation fixed by 66a09e4

[romainmenke]

Is it possible that there was a copy/paste mistake here?
When using the values from 9c4db75 we get very different outputs.

- rgba(51, 102, 153, 0.8)
+ rgba(44, 111, 115, 0.8)

This is more than what could be explained by slightly different rounding.

[svgeesus]

Yes, see color-js/color.js#360 so I need to update again with these values.

(I admit I was surprised by the magnitude of change in that matrix)

[svgeesus]

Fixed by 5ae7930