Y′CbCr, In Depth

Two reasons, one biological, one historical. The eye resolves brightness detail far more finely than color detail: the retina has ~95 million luminance-capable rods and cones against ~6 million color-differencing cones, and the visual system builds edges almost entirely from the luminance channel. Send full-resolution color and most of those bits describe detail no one can see.

The historical reason: when color TV arrived in 1953, millions of black-and-white sets already existed. NTSC’s answer was to keep broadcasting the B&W picture, which is exactly what Y′ is, and hide the color in two narrow difference signals an old set would ignore. Compatibility invented the architecture; the eye’s anatomy made it nearly free. Every codec and cable since has kept it.

The apostrophe in Y′ is load-bearing. CIE luminance Y is computed from linear light. Video luma Y′ is computed from gamma-encoded R′G′B′: the weighted sum happens after the transfer function, because that is cheap in hardware and was the only option in 1953 vacuum-tube economics. The two are close but not equal, and the difference has a name: since some true luminance leaks into the chroma channels, heavily saturated colors reconstruct with small brightness errors after subsampling.

For measurement work the distinction matters constantly: a probe reads CIE luminance in cd/m², a waveform monitor shows luma in code values or IRE. They correspond through the EOTF (see Transfer Functions), and conflating them is one of the most common errors in display QC.

The weights are not taste. They are the middle row of the RGB→XYZ matrix for the standard’s primaries and white point: the exact contribution of each primary to CIE luminance. Change the primaries and the row changes: BT.601’s 0.299 / 0.587 / 0.114 comes from the 1953 NTSC phosphors, BT.709’s 0.2126 / 0.7152 / 0.0722 from the HD primaries, BT.2020’s 0.2627 / 0.6780 / 0.0593 from the wide-gamut set.

The chroma divisors, 2(1−Kb) and 2(1−Kr), just normalize: they scale B′−Y′ and R′−Y′ so both difference signals span exactly ±0.5 for any legal color. Everything in the lab card above is those two lines of arithmetic, live.

Encode with one standard’s coefficients, decode with another’s, and every saturated color shifts: greens go yellowish, reds drift orange, skin turns subtly wrong, while grays survive untouched (a neutral has Cb = Cr = 0 under any matrix, which is why the bug hides from casual viewing). Set ENCODE to BT.601 and DECODE AS to BT.709 above and study the neon.

This is a real, endemic failure: SD content flagged as HD, scalers that assume 709 for everything, cameras that tag nothing. It is also the calibrator’s argument for testing with color patches and not just a gray ramp: a grayscale-only check literally cannot see a matrix mismatch.

Read J:a:b over a J-pixel-wide, two-row block: J luma samples per row, a chroma samples in the first row, b additional chroma samples in the second. 4:4:4 is chroma everywhere. 4:2:2 is chroma every second pixel horizontally, full vertically (half the chroma data). 4:2:0 is every second pixel in both directions (a quarter). 4:1:1, DV tape’s choice, keeps every fourth pixel horizontally, full vertically. The “4” is a fossil: it stood for sampling at 4× the NTSC subcarrier, ≈13.5 MHz, frozen into notation by CCIR 601 in 1982.

The engineering ladder follows the math: 4:2:0 ships to consumers (every streaming service, every Blu-ray), 4:2:2 is the production format with enough chroma to survive keying and grading, 4:4:4 lives in VFX, graphics and calibration paths. Move the loupe over sign text above and step through them: edges hold, color edges smear.

In 8-bit limited range, luma spans 16–235 and each chroma channel spans 16–240, centered on 128: code 128 means “no color difference”, a neutral. That offset is why a Y′CbCr frame of pure gray reads (Y′, 128, 128), and why chroma clipping behaves symmetrically around the middle of the scale rather than at zero.

The footroom and headroom (1–15, 236–254) exist for filter overshoot and analog heritage. The full story, including what happens when a chain mixes limited and full range, is the Signal Range module. Toggle RANGE above and watch the code readout in the lab card re-scale.

One idea, many uniforms. YUV is the analog PAL form (and, loosely, everyone’s slang for the whole family), YIQ is NTSC’s rotated variant with its axes aligned to the colors the eye resolves best, YPbPr is the analog component version on three RCA/BNC cables carrying the same differences as continuous voltages, and Y′CbCr is the digital, quantized form this module runs on. Calling a digital file “YUV” is universal and technically wrong, which makes it a fine shibboleth.

The modern heir is ICtCp (BT.2100): the same luma-plus-two-differences shape, but built through the PQ curve and LMS cone space so that a step of one code means roughly the same visual change everywhere, the property Y′CbCr never had. It is the basis of ΔE-ITP verification (see that module), and the clearest sign that television still believes in the 1953 split; it just keeps re-deriving it with better math.

Y′CbCr is where three classic field problems live. Matrix mismatch: colors shift, grays don’t, so catch it with saturated patches, never with a ramp. Subsampling in the measurement path: a 4:2:2 link soft-filters single-pixel patterns, so test patterns must be designed for the link that carries them. Range mismatch riding on top (limited vs full) crushes or lifts the whole scale. Any of the three can masquerade as “the display is off.”

The discipline is to verify the signal path before touching the display’s controls: confirm what format and matrix the panel reports receiving, use patches that exercise chroma, and only then trust the probe. It isn’t the display until the pipe is proven.