An aperture is a filter

An aperture sounds like a question about precision — a smaller one measures a smaller area, and that is all. On a translucent sample it is a question about colour.

The claim #

A finite aperture is a spectrally selective attenuation, and the spectrum it selects is decided by the sample rather than by the instrument.

• The light that escapes the aperture is the light that travelled furthest, and how far light travels is set by how weakly the material absorbs at that wavelength.
• So the bands reflecting most lose most. The attenuation is largest where the reflectance is largest, which flattens the curve.
• Every material in the table loses chroma, from 10 per cent on coated paper to 65 on marble, through a two-millimetre radius.
• The loss is not a calibration error and cannot be calibrated away, because the correction would need the sample’s own diffusion length at every wavelength.
• And the effect is invisible in the one number an instrument reports, because a reading that is 4 per cent low is a reading, and nothing in it says which 4 per cent.

Why the loss is not flat #

The kernel of a scattering surface has a width, and the width is the diffusion length — how far light travels sideways before absorption ends it. Two things set that length: how strongly the material scatters, and how strongly it absorbs.

Absorption is where the colour is. A blue pigment absorbs in the long wavelengths and not in the short ones, which is what makes it blue. So at the wavelengths it reflects, absorption is weak and the diffusion length is long; at the wavelengths it absorbs, the length is short.

An aperture collects what comes back inside it and loses the rest. A short kernel fits inside; a long one does not. So the loss falls hardest at exactly the wavelengths the sample was reflecting most strongly, which is the definition of an attenuation that flattens a curve.

The consequence is a desaturation, and it is the same shape on every material in the table. Through a two-millimetre radius the chroma of a coated paper falls from 3.36 to about 3.0; a wall emulsion from 5.56 to 4.7; a pigmented plastic from 7.32 to 5.9; skin from 13.21 to 9.5; marble from 7.15 to 4.8; wax from 5.67 to 2.0.

What that costs in the unit a specification uses #

The colour difference between what an instrument reads through a four-millimetre radius and what the sample’s own reflectance would give runs, across the six materials:

• coated paper, 0.53 ΔE₀₀ — below almost every published tolerance, and not zero;
• wall emulsion, 1.16 — at the tolerance a paint batch would be judged on;
• pigmented plastic, 1.96 — twice it;
• skin, 5.83;
• pale marble, 12.65;
• candle wax, 17.55.

Halving the aperture roughly doubles all six, and doubling it roughly halves them: at eight millimetres the plastic is at 0.96 and at two it is at 4.04.

The first row is the one to keep. A coated printing paper is the most opaque material in the table and about as far from translucent as an ordinary sample gets, and it still reads 0.53 ΔE₀₀ from its own colour. A delivery tolerance is three tolerances and the tightest of them is around one; this is half of that, on the material the industry considers solved, from a departure nobody in that industry’s specifications names.

The correction that cannot be written #

The obvious response is that an instrument could correct for this, the way it corrects for the sphere lighting its own sample or for a lamp that has drifted.

It cannot, and the reason is worth stating exactly. The correction factor is the share of the kernel the aperture recovers, and that share depends on the diffusion length, which depends on the absorption and the scattering — which is to say, on the answer. A calibration tile has its own kernel and it is not the sample’s, so calibrating against it corrects for the tile’s loss and leaves the sample’s.

This is the same shape of problem as the sphere’s substitution error, where the correction needs the reflectance the instrument is trying to find, and there the answer is to iterate: take the raw reading as a first estimate, compute the correction it implies, repeat. Here that route is closed, because the reading is one number per wavelength and the correction needs two coefficients per wavelength. One measurement cannot recover two unknowns, which is the ordinary shape of an inverse problem in this collection and is the reason the practical answer is a second aperture rather than a cleverer calculation.

What was computed, and how #

Six materials, each stated as an absorption spectrum and a scattering spectrum rather than as a reflectance, and each turned into eighty-one slabs — one per band of the site’s wavelength grid.

For each band the kernel is integrated twice: once over the whole plane, which gives the reflectance the model wants, and once against the aperture’s weight, which gives the reading. The two integrals share a quadrature, so their ratio is a clean statement about the aperture rather than a difference of two numerical errors.

The colours are then computed the way every colour on this site is: the reflectance multiplied by D65, integrated against the 1931 observer, converted to CIELAB against the light’s own white. Nothing about the colour arithmetic changes; the only thing that changes is which reflectance is handed to it.

One number in the sweep is worth its own line. The share recovered at a four-metre radius is 99.95 per cent rather than 100, on a sample whose diffusion length is four millimetres — a thousand diffusion lengths out, and the last twentieth of a per cent is still outside. That is not a rounding artefact; it is the kernel’s tail, which falls as a power rather than exponentially once the exponential term has died. The limit is approached and not reached, which is why the argument for a wide aperture is quantitative rather than categorical.

The rate is the practical fact. Going from a four-millimetre radius to forty on marble takes the recovery from 63 per cent to 95; the next factor of ten takes it to 99.5. Each decade of aperture buys a decade of the remaining error, which is the signature of a tail falling as one over the radius, and it means there is no aperture at which the problem stops rather than merely becoming small.

What an instrument would have to report instead #

An honest reading of a translucent sample is two numbers and a statement, not one number.

The two numbers are readings at two apertures. Their ratio is a measurement of the sample’s transport, independent of its colour to first order, and it is the quantity that says how far from the limit the wider of the two is. An instrument with a four-millimetre and an eight-millimetre port already has the hardware; what it lacks is a field in the data format for the second reading and an agreed way to report the pair.

The statement is which of the two was used for the colour. That sounds like bookkeeping and it is the whole of the problem: two laboratories reporting one number each, from the two ports of the same instrument, will disagree by 1.00 ΔE₀₀ on a pigmented plastic and by 5.90 on marble, with both instruments in calibration and both operators correct.

This collection has met that shape before and named it: a tolerance cannot cross a condition. A number agreed between two parties who measured under different conditions is not a tighter agreement than the conditions allow; it is an agreement about something neither of them measured.

Sixty-five per cent is the wax, not the marble #

The chroma sweep is summarised as running from 10 per cent on coated paper to 65 on marble, and the six figures beside it put the 65 on a different material.

Marble loses a third and wax loses two thirds. The claim’s range is right and its upper label is on the wrong row, which matters because marble is the material the ΔE₀₀ table makes the villain — 12.65 against wax’s 17.55 — and on chroma the ordering is the other way round by a wide margin.

The two rankings differ for a reason worth having. Marble’s ΔE is large because it is light and loses a great deal of lightness; wax’s chroma loss is large because its chroma was small to begin with, so losing 3.67 units of it is two thirds. A percentage of a small chroma and an absolute ΔE₀₀ are not the same ranking, and the table supports both.

The desaturation is a quarter of the effect #

The essay’s title claim is that an aperture reads a sample not just dark but desaturated, and the two halves are not the same size.

Take the pigmented plastic at a two-millimetre radius, where both numbers are given: the chroma falls from 7.32 to 5.90 and the colour difference is 4.04 ΔE₀₀. Putting the chroma change through CIEDE2000’s own weighting — a mean chroma of 6.6 gives S_C = 1.297 — makes the chroma term 1.09.

That is 27 per cent of the total difference and 7 per cent of its square. The other 93 per cent of the squared difference is lightness and hue, and on this mechanism it is overwhelmingly lightness: the aperture throws light away.

So the finding is real and secondary. An aperture reads a translucent sample dark, and desaturated by about a quarter as much. The desaturation is the interesting half because it is the half nobody expects and the half no level correction removes — a reading scaled back up to the right lightness is still the wrong colour — but it is not where the ΔE₀₀ comes from, and an essay whose title is the smaller term should say which term is which.

The exponent is not quite one, and it varies #

Halving the aperture roughly doubles all six is stated as a rule of thumb and the two materials with figures give it a number.

The plastic goes 1.96 at four millimetres to 0.96 at eight and 4.04 at two, which are exponents of 1.030 and 1.044 — inverse-linear to within five per cent. Marble goes 12.65 at four to 6.75 at eight, from the stated 5.90 disagreement, which is an exponent of 0.906.

So the rule is inverse-linear on shallow materials and slightly shallower on deep ones, running from about 1.04 down to 0.91 as the diffusion length grows. That is the same relation the kernel essay fits at 0.87 on marble’s recovered share, reached from the other end, and the two agree to within four per cent on the one material they share.

The consequence for the two-aperture remedy is that a doubling buys slightly less on the materials that need it most. On the plastic a four-to-eight step removes 51 per cent of the departure; on marble it removes 47. Both are close enough to half that the rule of thumb survives, and the direction is the unhelpful one.

The tail’s own rate checks too. Marble’s missing share runs 37 per cent at four millimetres, 5 at forty, 0.5 at four hundred and 0.05 at four metres — each decade dividing it by 7.4, then 10, then 10. So the tail settles into an exact one-over-radius fall after the first decade, which is the slower of the two regimes and the reason the last twentieth of a per cent is still outside a thousand diffusion lengths away.

The three ways a reading can be low #

It is worth separating this departure from the two the collection already has, because all three make a reading low and only one of them is fixable by the instrument.

The sphere lights its own sample, so a dark specimen sits in a darker sphere than the white tile it was calibrated against. The correction needs the sample’s reflectance, which is the unknown, and the instrument iterates to it in one or two passes because the dependence is weak. Fixable, and fixed.

The geometry throws away the interface reflection — or keeps it — so 45°/0° and a sphere with the port open report two different quantities. Not an error at all: two answers to two questions, and the fix is to name the question, which is what the geometry designation in a specification is for.

The aperture throws away the tail of the kernel. Not fixable by iteration, because the correction needs two coefficients where the reading gives one; not fixable by naming, because naming the aperture does not make two apertures comparable. The only route is a second measurement.

Three departures, three different remedies, and the only thing they have in common is that each one makes the number smaller than the sample’s own.

Where the model stops #

Three limits, and the third is the one that would change the numbers most.

The slab is semi-infinite. A real sheet has a back, and light reaching it either returns or leaves. A thin translucent sheet on a dark backing reads darker than the same sheet on a white one, and that is a second departure with its own two-parameter arithmetic.

The sample is uniform. A textured or fibrous material has a kernel that depends on direction, so its reading would depend on how the sample was placed in the instrument.

And the reading is monochromatic band by band. Real instruments have a bandpass, and a bandpass is a choice this collection has measured elsewhere; the two departures would interact, since a wide bandpass averages over bands whose diffusion lengths differ.

A one-millimetre aperture is smaller than any hand-held instrument offers and is what a microscope objective effectively has.

What a reader can check without an instrument #

The arithmetic is easier to believe with a demonstration that needs nothing but a hand.

Hold a finger in front of a bright lamp. The light coming through it is red, and it is red for the reason this essay is about: at the long wavelengths haemoglobin has stopped absorbing, so the diffusion length is centimetres, and at the short ones it is under a millimetre. What comes out at any distance from where the light went in is therefore the long-wavelength part, and the further from the entry point the redder it is.

An instrument with a small aperture is looking at the near part of that same distribution. It sees the light that came back quickly, which is the light that was absorbed quickly, which is the light at the wavelengths the sample is worst at reflecting. The finger and the aperture are the same measurement with the sign of the distance reversed.

That is also why skin is the material in this table that costs a photographer most and a colourist least. A camera photographing a face is lighting an area vastly larger than any pixel, so the field’s factor in the pairing is zero and the picture’s colour is right; a spectrophotometer on the same cheek is lighting four millimetres and reads 5.83 ΔE₀₀ away from it.

The generalisation #

The useful shape here is that a loss which depends on the quantity being measured is not a calibration, and the sign of that is easy to check: if the correction factor has the answer in it, no reference sample can supply it.

Colour measurement has three departures of that kind and this collection now has all three. The sphere’s substitution error depends on the sample’s reflectance and is fixed by iterating, because the dependence is weak and contractive. A fluorescent sample’s reading depends on the lamp’s ultraviolet, and that one is fixed by specifying the lamp — which is what the M-conditions are — because the dependence is on the instrument rather than on the sample. The aperture’s loss depends on the sample’s transport, and neither route works: it does not iterate and it cannot be specified away.

What is left is to measure it, which needs a second reading at a second aperture, and to report it, which needs a field in the specification that no specification has. The industries where it bites most — plastics, textiles, food, dentistry — have all arrived at the same practical rule from experience, which is to light a wider disc than is looked at.

And the same kernel taken to its limit: as the aperture closes on a sample, the colour it returns walks somewhere, and where it walks depends on how the diffusion length varies across the band.

Who found it, and when #

The effect is old in practice and has been standardised as advice rather than as a model. ASTM and CIE documents on translucent samples say to use the largest aperture the specimen allows, to light an area larger than the one measured, and to back the specimen with a stated material — three instructions that between them name all three departures above without naming any of them as a quantity.

The plastics industry has had the sharpest version of the problem since the 1960s, because a pigmented polymer is translucent by nature and its colour is a contractual matter. The dental literature came to it independently in the 1990s, since a tooth is a translucent object measured against an unstated backing, and its papers are the ones that report inter-instrument disagreements of several ΔE on the same specimen.

What none of those literatures does is treat the aperture as a filter with a spectrum. The advice is about magnitude — the reading is low — and the arithmetic here says the reading is also the wrong colour, in a direction the sample chooses.

Where the ladder goes next #

The sharpest unfinished question is whether the loss can be inverted from two readings rather than merely bounded. Two apertures give two numbers per band, and the kernel has two parameters per band, so the counting works. Whether the inversion is stable is a different matter, and it is exactly the sort of question this collection has learned to ask about how many measurements a recovery needs before assuming that a matched count is enough.

The other open item is the interaction with gloss. An aperture removes body light and leaves interface light untouched, so the ratio between them rises as the aperture narrows — which means the two departures do not add, and what they do together is smaller than either measured alone suggests.