The Gold Standard of Visual Arts | Matching Colour Readings and the Reflectance Spectra They Hide

Imagine an artist combining painted paper and a printed element in one collage. Under the chosen studio light, the two areas seem well matched. In another display setting, a boundary becomes more apparent. The artist needs to distinguish a change in viewing conditions from a claim that the material itself has suddenly changed.

A colour record can answer a specific comparison question: how did these samples respond under this measurement arrangement? It does not automatically answer what each sample reflects at every wavelength or how every other viewer and light will compare them.

The Visual Arts root connects observation, material experiments and purposeful revision. The present lesson adds a quantitative reason to preserve the conditions of a colour comparison. The aim is not to make every studio decision mathematical, but to explain why apparently matching evidence can leave an important physical difference hidden.

Before discussing a mismatch, name the intended comparison. Is the task to match physical surfaces, reproduce a photograph on a screen, preserve a tonal relationship or interpret an artist’s colour choice? Those tasks need different evidence. A useful technical explanation helps the maker choose the right next test rather than impose one universal standard on every artistic intention.

Contents · Next chapter

For a simple nonfluorescent diffuse-surface model at a fixed geometry, the returned spectrum depends on illumination multiplied by reflectance. Illumination describes the light supplied across wavelengths. Reflectance describes the fraction returned under the model. Their product describes the light available to the receiving system, up to the chosen geometric and scaling conventions.

A surface that strongly reflects a band cannot return much light in that band if the source supplies little there. Conversely, increasing a source band changes the returned contribution without requiring a change in the surface’s reflectance.

Real materials can involve directional reflection, transmission, fluorescence and other effects. We initially exclude those so one relationship remains clear. A successful simple model explains the question it was designed for while retaining its boundary.

Scott Allen Burns’s research preprint on recovering reflectance from colour triplets introduces the forward relationship and the non-uniqueness of its inverse. The numerical examples here are independent constructions. They do not reproduce measured pigments or implement the preprint’s reconstruction methods. For a learner, the first task is simply to identify which quantity belongs to the lamp, which to the surface and which reaches the sensor. Losing those distinctions makes later explanations about colour correction unnecessarily confusing.

Contents · Previous chapter · Next chapter

A receiving channel can be sensitive to several parts of a spectrum. In a linear response model, its output is the sum of the returned contributions weighted by that channel’s sensitivity. Different channels use different weights.

With four simplified spectral bands, write a channel as C = ΣEᵢRᵢSᵢ. Here Eᵢ is illumination in band i, Rᵢ is surface reflectance and Sᵢ is that channel’s supplied sensitivity weight. Band-width factors are absorbed into the model’s weights. We use arbitrary compatible response units.

Three channel totals preserve three summaries. They do not retain the separate contribution from every band. If several different four-band reflectance patterns give those same summaries, the record cannot distinguish them under that illumination.

This is a forward-versus-inverse distinction. Calculating the readings from supplied illumination, reflectance and sensitivities can be straightforward. Recovering an unknown reflectance from the readings can remain non-unique. A precise calculator does not restore information discarded by summing. More digits in the three totals can improve their numerical representation while leaving the structural ambiguity intact. That is the central mechanism we will construct explicitly, rather than ask the reader to accept as a vague warning about imperfect colour.

Contents · Previous chapter · Next chapter

Our model has bands 1 to 4 and three channels. Channel 1 responds equally to bands 1 and 4. Channel 2 responds equally to bands 2 and 4. Channel 3 responds equally to bands 3 and 4. Other weights are zero.

Under illumination E = (1, 1, 1, 1), the channel formulas become C₁ = R₁ + R₄, C₂ = R₂ + R₄ and C₃ = R₃ + R₄. This arrangement is deliberately simple so the hidden degree of freedom can be followed by hand.

These are not standard human colour-matching functions and should not be labelled RGB, XYZ or cone measurements. A reading above one is possible in the arbitrary response scale because a channel sums contributions from more than one band. Reflectance itself remains between zero and one in each band.

The simplification is useful precisely because it makes the arithmetic visible. Once the principle is understood, the reader can recognise why real overlapping spectral sensitivities also need a carefully defined model rather than assuming three recorded numbers are a complete physical description of a surface.

Contents · Previous chapter · Next chapter

Let Surface A have reflectance (0.2, 0.4, 0.6, 0.4). Let Surface B have (0.4, 0.6, 0.8, 0.2). Every component lies within the allowed zero-to-one range, and the patterns differ in every band.

Under the uniform reference illumination, A gives channel readings (0.6, 0.8, 1.0). B gives exactly the same: 0.4 + 0.2 = 0.6, 0.6 + 0.2 = 0.8 and 0.8 + 0.2 = 1.0.

This is an exact match within the invented receiving system. It is not evidence that physical samples with these arbitrary band values would look identical to real viewers. Its logical job is to demonstrate that different reflectance inputs can produce identical recorded summaries. One valid pair is enough to show that the inverse mapping is not unique over this model’s permitted surfaces.

Contents · Previous chapter · Next chapter

Now use illumination E = (2, 1, 1, 1), doubling only band 1. For A, Channel 1 becomes 2 × 0.2 + 0.4 = 0.8. For B, it becomes 2 × 0.4 + 0.2 = 1.0. Channels 2 and 3 remain 0.8 and 1.0 for both.

The original match has broken in the model’s first channel. The surfaces have not changed. Their different reflectance patterns interact differently with the revised illumination.

The CIE's explanation of metameric matches describes the real colour-science distinction: different spectral properties can match for a specified illuminant and reference observer, with the match affected by a change in illuminant or observer. Our artificial example isolates the information-loss mechanism without claiming a standard colour-difference measurement.

An artist’s practical conclusion should remain proportionate. A match established under one light may need checking under the intended display conditions. It does not follow that every pair of paints will visibly separate, that the separation has one universal direction or that a numerical channel difference predicts a particular emotional effect. The model establishes why a single match cannot settle those later questions by itself.

Contents · Previous chapter · Next chapter

Suppose a perfect model white has reflectance (1, 1, 1, 1). Under the reference illumination, its channel readings are (2, 2, 2). Dividing each sample channel by the matching white channel gives both surfaces (0.3, 0.4, 0.5).

Under the changed illumination, white gives (3, 2, 2). Normalising A now gives approximately (0.2667, 0.4, 0.5), while B gives approximately (0.3333, 0.4, 0.5). The first-channel mismatch remains.

This calculation illustrates one simple per-channel reference operation. It is not a complete camera white-balance method or a model of human adaptation. Its purpose is to show that correcting against a shared white does not magically reconstruct four unknown reflectance contributions from three original summaries.

Reference measurements can be valuable. The mistake would be to promise that one reference operation removes every material-dependent response to changed light. In the constructed example, both samples share the same white normalisation yet retain different hidden spectra. A learner should explain which influence the operation addresses and which information remains unavailable. That distinction helps interpret colour-management tools as methods with defined purposes rather than universal buttons for recovering physical truth.

Contents · Previous chapter · Next chapter

Starting with A, add t to its first three reflectance components and subtract t from the fourth. The family is R(t) = (0.2 + t, 0.4 + t, 0.6 + t, 0.4 − t).

Under the reference illumination, every channel receives one +t and one −t contribution, which cancel. All permitted members therefore give (0.6, 0.8, 1.0). Surface B is the member with t = 0.2.

Requiring every reflectance component to remain between zero and one gives −0.2 ≤ t ≤ 0.4. For example, the first component requires t ≥ −0.2, while the third and fourth each require t ≤ 0.4. The other constraints are no tighter.

The record therefore hides a continuous family, not merely one exceptional alternative. This strengthens the inverse-problem lesson. Choosing one plausible reconstruction from a colour triplet does not establish that the chosen spectrum is the actual surface spectrum. It may be useful under additional assumptions, but those assumptions need to remain visible. An algorithm can return one neat curve because it imposes a preference or constraint; the existence of that output does not erase the alternatives compatible with the original readings.

Contents · Previous chapter · Next chapter

Under E = (2, 1, 1, 1), the first-channel response of the family is 2(0.2 + t) + (0.4 − t) = 0.8 + t. In this restricted family, an exact first-channel reading of 1.0 would identify t = 0.2.

The added condition supplies information that the original reference reading lacked. Repeating the original illumination with identical ideal conditions would repeat the same cancellation, whereas the changed band weighting reveals the hidden parameter.

This does not establish that two lights identify every possible real reflectance spectrum. Our family has one remaining parameter within a finite four-band construction. Real spectra can contain far more degrees of freedom, and practical measurements carry uncertainty and other conditions.

The transferable question is whether the added measurement distinguishes the alternatives that matter. A second photograph with unchanged relevant spectral conditions may improve precision or documentation without resolving the same physical ambiguity. An appropriately different measurement can add new information, but its scope still depends on the model. More measurements are useful when they answer a missing question, rather than simply make the evidence collection look larger.

An added measurement can also constrain a parameter without identifying it exactly. Keep the original reference triplet exact within the toy model, but suppose the changed-light first channel is known only to lie between 0.98 and 1.02 response units. Since that channel is 0.8 + t, the compatible parameter range is 0.18 to 0.22.

The corresponding reflectance ranges are 0.38–0.42, 0.58–0.62, 0.78–0.82 and 0.18–0.22 in the four bands. Those ranges are linked by the same t. They should not be treated as four independently selectable intervals. Selecting the first band’s upper endpoint 0.42 and the fourth band’s upper endpoint 0.22 together would give a reference Channel 1 of 0.64, contradicting the stipulated 0.6. Both values lie in their separate ranges, but that combination does not belong to the family.

This is an important evidence habit beyond colour. Uncertain reconstructed quantities can remain related even when reported separately. A useful result preserves the parameter relationship or an equivalent joint description. The added light has narrowed a specific ambiguity under explicit assumptions; it has not made the original record a direct spectral measurement or established a universal uncertainty allowance for real colour instruments.

A restricted inverse problem can sometimes have a unique answer, but only because the restriction supplies information. Suppose an independently justified model requires the same reflectance r in all four bands. Under the uniform illumination, all three readings must then equal 2r. A supplied triplet (0.8, 0.8, 0.8) would identify r = 0.4 within that restricted model.

Without the equal-band restriction, the same triplet also fits reflectance (0.2, 0.2, 0.2, 0.6). The unique result was conditional on the independently supplied structure, not a new power of the three readings themselves. A reconstruction method should therefore explain which constraints or preferences select its answer.

The restriction can also fail a model check. Our original triplet (0.6, 0.8, 1.0) cannot come from equal reflectance in all four bands under this receiver and reference illumination, because its three outputs are unequal. Forcing one grey-model parameter to fit that record exactly would conceal the disagreement rather than solve it.

This is a helpful extension for artists reading technical tools: a plausible-looking reconstructed curve is useful when its assumptions fit the intended task. A neat output alone does not establish that the assumed material family actually describes the object.

Contents · Previous chapter · Next chapter

Let the illumination vary as E(s) = (1 + s, 1, 1, 1), with s between zero and one. At s = 0 it is the reference light; at s = 1 it is the changed light. The raw first-channel difference B minus A is 0.2s.

At s = 0.5, A’s first channel is 0.7 and B’s is 0.8. The white first channel is 2.5, giving normalised values 0.28 and 0.32. Their difference is 0.04. The general normalised difference is 0.2s/(2 + s).

The raw channel relationship is linear in this supplied illumination parameter. The normalised relationship is not the same straight line because its denominator also changes. This is a useful reminder to preserve the stage of the calculation when comparing numbers.

Physical light mixing in the stipulated linear model also differs from averaging encoded image colours in editing software. The model’s inputs are spectral-band contributions, not arbitrary displayed colour values. A learner can compute the intermediate case and then explain which quantities were mixed. That explanation prevents a correct mathematical interpolation from being transferred into a different colour representation without checking its meaning.

Contents · Previous chapter · Next chapter

If all four illumination components double together, the raw channel outputs of both A and B double to (1.2, 1.6, 2.0). Their match remains intact in the linear model. The reference-white channels also double, so the simple normalised triplets remain unchanged.

The previous mismatch arose from changing the relative spectral contributions, not merely multiplying the whole illumination by one positive constant. This is a valuable control case because it identifies which relationship caused the constructed separation.

Real appearance can depend on brightness and viewing conditions in ways our simple channel model does not describe. The calculation should not become a promise that changing brightness never affects perception. It establishes an invariance within the stated linear physical model.

For an independent explanation, ask the learner to compare E = (2, 2, 2, 2) with E = (2, 1, 1, 1). Both contain a larger first-band contribution than the original. Only one preserves all relative band proportions. The distinction is more precise than saying that a brighter lamp changed the colour. Identify what actually changed before assigning a mechanism to a visible difference.

Contents · Previous chapter · Next chapter

So far, the receiving channel weights have stayed fixed. Change Channel 1 to weights (2, 0, 0, 1) under the original uniform illumination. A then gives 0.8 in that channel and B gives 1.0, while the other unchanged channels still match.

This is a mathematical change in receiver sensitivity, not a model of one named person’s vision or a particular camera. It shows why the observer or measuring system belongs in a match statement alongside the illuminant.

CIE distinguishes observer-related metamerism from illuminant-related metamerism. A match for one reference observer does not settle every real observer’s comparison. That scientific distinction should encourage precise documentation rather than unsupported claims about who sees colour correctly.

In visual practice, separate a camera record, an instrument comparison and an in-person viewing judgement. They can inform one another while using different receiving systems and contexts. A photograph that looks well matched on one display does not independently establish that two physical materials have equal spectra or will match for every viewer in the room. The useful next step follows the intended outcome: document a reproduction, compare the actual surfaces or investigate a physical measurement question.

Contents · Previous chapter · Next chapter

A digital image stores a representation created through a camera and processing chain. Its three-channel values can depend on illumination, sensor sensitivities, processing, encoding and other choices. They are not automatically three direct samples of a surface’s reflectance at three wavelengths.

Even the ideal linear triplet in our model failed to identify the four-band reflectance uniquely. A processed photograph does not remove that structural difficulty merely by offering a colour picker with many decimal places or an apparently precise hexadecimal code.

The existing painting guide on colour, materials and viewing conditions develops the broader artistic and perceptual context. The present model explains one specific reason a matching record cannot promise a matching physical response under new illumination.

An image can still be an excellent record for its defined purpose. Preserve source files, relevant capture information and the intended reproduction context where the task requires them. Avoid changing the claim from these displayed regions have matching recorded values to these materials are physically identical. The first may be directly checkable in a file. The second requires different evidence, and many distinct materials can share a colour relationship under one observation condition.

A later processing stage can discard additional information even when the original channel responses differ. For a separate invented example, let two linear triplets be (0.8, 1.2, 1.5) and (0.8, 1.8, 1.9). Suppose a display stage clips every value above 1 to 1. Both displayed triplets become (0.8, 1, 1).

Their match at that later stage does not show that the original receiving system produced equal responses. Information was lost through clipping. The example should not be labelled a proof of metamerism for the original receiver, because its unprocessed readings did distinguish the inputs.

This gives an image reviewer a concrete next question: at which stage did the equality appear? A matching pixel value could reflect the physical receiving summaries, a shared reference transformation, clipping or another part of the representation. Keeping an appropriate original record makes those alternatives easier to investigate.

The same reasoning also explains why an apparently stronger contrast adjustment cannot recover differences already clipped away. It may change the visible representation of what remains, but it does not identify the discarded original values. Treat a processed match as evidence about that processing stage unless suitable additional records support a stronger statement about the surfaces.

Contents · Previous chapter · Next chapter

The artificial model treats each sample as an isolated uniform response. It does not represent surrounding colours, edges, spatial arrangement, adaptation or the viewer’s interpretation of a scene. Those relationships matter in visual arts and should not be collapsed into the word metamerism.

For example, moving an unchanged grey patch from a pale surround to a dark surround can change its appearance without changing its reflectance or the illumination in the simplified comparison. That is a different question from two different reflectances producing matching channel readings under one light.

An artist investigating a mismatch should therefore identify what changed. Was it the illuminant, material, viewing geometry, surrounding image, reproduction chain or more than one of these? The right comparison isolates a useful relationship while retaining the work’s actual context.

This is where controlled studies and whole-work viewing complement one another. A small swatch comparison can reveal a material-light relationship. Seeing that swatch within the collage can reveal a different compositional problem. Neither result replaces the other. Good visual reasoning gives each test a clear job, so a successful isolated colour match does not become an unsupported claim that the passage will serve the whole artwork’s intention.

Contents · Previous chapter · Next chapter

Gloss and directional reflection can change what reaches an observer at different angles. Fluorescent materials can emit light at wavelengths different from those absorbed, requiring a richer model than a simple same-band illumination-times-reflectance product.

NIST’s surface-colour measurement research describes measuring spectral reflectance across illumination and viewing geometries. It supports the practical point that geometry can be part of a surface-colour measurement rather than an irrelevant detail.

A changed appearance also need not imply permanent material alteration. Conversely, actual fading or another material change can alter reflectance over time. A quick return to a previous light may help frame a comparison, but a real conservation claim requires suitable evidence and expertise, not a casual studio observation.

For this article, keep the boundary narrow. The four-band examples exclude fluorescence, directional effects and material ageing. Their purpose is to demonstrate non-unique spectral information under a specified receiver and light. If the actual question involves another mechanism, extend or replace the model appropriately. A model becomes more useful when its limits help select the next question, rather than being presented as a complete explanation of every colour difference.

Contents · Previous chapter · Next chapter

Suppose a collage maker wants a printed repair patch to remain visually unobtrusive beside painted paper in a planned display setting. A useful record identifies the materials, relevant light, viewing arrangement and the purpose of the comparison. A match under a convenient studio lamp is a starting observation, not the whole acceptance question.

Use suitable ordinary materials and the arrangements appropriate to the work. The learning task can simply compare supplied swatches or documented images under stated conditions; it does not require modifying valuable artwork, using unknown chemicals or improvising intense lighting.

Record what was observed without inventing an instrument result. The boundary became more visible under the second condition is a useful visual observation. A numerical standard colour difference needs an appropriate method. A photograph should be identified as a reproduction rather than substituted for the actual surface comparison.

The maker may choose to embrace a changing boundary instead of eliminating it. The artistic decision belongs to the work’s intention. Scientific understanding explains why a match can depend on conditions; it does not dictate that every artwork must maintain one appearance everywhere. The practical gain is a more deliberate decision, supported by an account the maker can revisit when presentation conditions change.

Contents · Previous chapter · Next chapter

Keep A, B and the original three artificial channels, but use E = (1, 1, 1, 2), doubling only band 4. Calculate the two response triplets before reading the answer. Then ask whether normalising against the same model white restores their original match.

A gives (1.0, 1.2, 1.4): each of its first three reflectances receives an added contribution of 2 × 0.4. B gives (0.8, 1.0, 1.2), with a band-4 contribution of 2 × 0.2. The mismatch now appears in every channel.

White gives (3, 3, 3). Normalising A gives approximately (0.3333, 0.4000, 0.4667), while B gives approximately (0.2667, 0.3333, 0.4000). Their match is not restored by this reference operation.

For a second task, explain why doubling every band instead would preserve the raw match. The successful answer identifies the hidden reflectance differences and how the new illumination weights them. It does not rely on naming a lamp warm or cool, or guessing what hue the arbitrary sensors represent. This is independent transfer: the reader chooses the contribution account from changed conditions and keeps the physical and perceptual claims within the model’s scope.

Contents · Previous chapter · Next chapter

A dependable comparison record states whether its values describe reflectance, illumination, linear channel response, encoded image values or a perceptual judgement. It identifies the receiver, geometry, reference operation and relevant assumptions. Those labels prevent numbers from drifting between stages of the colour chain.

For the numerical model, check that reflectances stay within zero and one, weights and illumination are those supplied, and every channel includes the correct contributions. Confirm the reference match, then independently calculate the changed-light responses. A failure of the match under new conditions should follow from those contributions, not from an unexplained software transformation.

Repair begins with separating lamp, surface and returned light. Stabilisation adds the two-surface counterexample and white reference. Extension explores the matching family, added measurements and receiver changes. An artist can use the conceptual route without calculating every channel; a mathematically interested learner can defend the non-uniqueness explicitly.

The Education parent connects this habit with learning that travels. A useful final explanation names what the original match established, constructs or recognises an alternative consistent with it, and identifies which new evidence could distinguish the alternatives that matter to the next decision.

Contents · Previous chapter · Next chapter

Does a matching colour triplet identify one reflectance spectrum?

Not in general. The model constructs a continuous family of different reflectances giving the same triplet under one light and receiver. Additional justified constraints can narrow an inverse problem, but those constraints remain part of the result.

Are the model channels human cone measurements or RGB?

No. They are invented linear receiving channels chosen to make the arithmetic transparent. Their exact values should not be interpreted as standard colour coordinates or used to predict a particular human colour sensation.

Does a white reference guarantee matching under another light?

The simple normalisation example shows that it need not. A reference operation can address particular influences without recovering all missing spectral information. Use the actual colour-management method within its documented purpose and conditions.

Can two lights identify every possible surface spectrum?

That does not follow. The second light identifies one parameter in our restricted family. A richer real spectrum can retain other unresolved degrees of freedom, along with practical measurement uncertainty and material effects.

Is every appearance change metamerism?

No. Spatial context, adaptation, geometry, processing and actual material alteration can also matter. Identify the changed relationship before assigning a mechanism. A single word should not stand in for several different physical and perceptual questions.

Must an artist eliminate every condition-dependent colour change?

No. The work’s intention determines the decision. Understanding the mechanism supports deliberate material and presentation choices. The important habit is to describe the comparison honestly and avoid promising a universal match from limited evidence.

Contents · Previous chapter · Next chapter

Connect the model with artistic practice

Return to The Gold Standard of Visual Arts for purposeful observation and revision. Colour Is Not a Property of Paint Alone develops the broader painting context. Recovering Composition from Detail Images examines a different kind of information missing from an image record.

For a materials-production comparison, Textile Manufacturing connects shade checks with a defined material and use. It complements the information-loss model rather than replacing the artist’s intention with a factory acceptance rule.

For the image-making chain, Colour Ambiguity in Photography follows the camera’s observation conditions. Colour Reconstruction explains sampling and rendering; White Balance examines neutral references; Printing follows the return to a reflective object. These guides supply practical contexts for the narrower spectral ambiguity constructed here.

Check the primary explanations

CIE's metamerism publication description distinguishes reference illuminants and observers. Burns's reflectance-reconstruction preprint explains why colour triplets do not generally determine a unique spectrum. NIST's surface-colour research documents the role of measurement geometry.

All four-band tables, channel weights, matching families and transfer calculations here are original hypothetical models. Source routes checked on 22 October 2026. Use How X Works for wider subject connections, choosing the next guide according to the physical, artistic or evidence question that remains unresolved.

Contents · Previous chapter · Return to Visual Arts