Delta-E: The Number and Its Limits
A single averaged colour-difference figure tells you less than the calibrator who sold it to you wants you to believe.
What ΔE Measures — and Which Formula to Demand
Delta-E (ΔE) quantifies the distance between two colours in a perceptually uniform colour space. The original 1976 formulation, ΔE76, computed simple Euclidean distance in CIELAB — fast, intuitive, and wrong enough to matter. CIELAB is not actually perceptually uniform: a ΔE76 of 3 in a saturated region looks noticeably smaller to a human observer than a ΔE76 of 3 in a neutral grey. The CIE knew this by the 1990s and progressively corrected it, culminating in ΔE2000 — a formula with separate weighting terms for lightness, chroma, and hue, plus an interaction term that handles the blue-violet region where CIELAB is most distorted.
The practical consequence: a display that scores ΔE76 < 2 on average can still show ΔE2000 values above 3 on saturated primaries. Any review or calibration report quoting ΔE without specifying the formula version is incomplete. ΔE2000 is the version that correlates reliably with what a trained observer will actually notice, and it is the version that matters when calibrating a display.
The perceptibility threshold commonly cited is ΔE2000 ≈ 1. Below that, differences are generally indistinguishable under typical viewing conditions. A well-calibrated professional monitor targets ΔE2000 < 1 across the test patch set; consumer panels considered colour-accurate typically land between 1 and 3. Above 3 is visible to most people; above 5 is obvious.
Why a Low Average Hides Real Problems
Averaged ΔE is where the metric becomes dangerous. A panel that scores an average ΔE2000 of 1.2 across one hundred test patches can still produce a ΔE2000 of 6 on a specific saturated red or a near-white with a strong colour cast. The average absorbs the outlier. Skin tones, memory colours — the blue of a clear sky, the green of foliage — fall in regions of colour space where the eye is most discriminating, and a spot error there is perceptually far more damaging than the averaged number suggests.
The correct practice is to examine the full distribution: minimum, maximum, and the 95th-percentile value alongside the mean. A panel with a mean of 0.8 and a maximum of 1.4 is a different instrument from one with a mean of 0.9 and a maximum of 5.1, even though both might be marketed as "under 1 ΔE average." Reporting the test patch set used matters equally — fifty neutral-biased patches will flatter a display that falls apart on saturated colours.
ΔE is a useful engineering yardstick, not a certification of accuracy. Read the distribution, name the formula, and check which patches were measured before trusting the number.