Concept

Identifiability — where it appears

Whether a measurement's data can separate the parameters a model asks it for, as against how precisely it measures whatever it can separate. A pair that enters an observable only as a product, or that a similarity leaves untouched, is unidentifiable however carefully the measurement is made.

Named by 29 essays across 12 fields — each of them below, with the objects they name alongside it.

head0.668profile — the outline that names a faceeye0.669frontal — an eye in profile is a wedgeshoulders0.669frontal — the width that says two armslegs0.668profile — a stride is a side viewpond0.326plan — a rectangle of water is a rectangle1.000 — what its own aspect keepsshare of each part that reaches the picture, from the best single direction|d · n| for each part3.000 of 5, swept and in closed form

Assembled from several views

An Egyptian relief takes each part of a figure from the direction that identifies it — head in profile, eye and shoulders frontal, a pond in plan. What that buys is exactly measurable: any single viewing direction keeps at most √k of k perpendicular aspects, so the best compromise view retains 58% of each.

conventions · Aspective
481632641282560.3131030100how far away the scene is (m, log scale)pixels of parallax, and degrees of pose error (log scale)parallax, pxpose error, degreesparallax goes as distance to the -0.977 distances

Far enough away, a pair is one eye

Hold the baseline and walk the scene away, and the parallax a single homography cannot explain falls as the distance to the power −0.968 — one over the distance, which says the ratio of baseline to depth is the whole of it. The recovered translation direction follows it down, from 3.3° at four metres to 74.5° at two hundred and fifty-six.

manyviews · Degenerate
the seat the picture came fromthe screen, in planthe room sampledcurved monitor, 1 px of tolerance0.0% of the room inside

The screen that names the seat

A flat screen shows a homography of the intended picture from every seat in the room, and an observer's own framing is free to be a homography too — so a flat screen's picture is consistent with every seat there is. A curved one is not, and the seat comes back out of the picture in all three directions, in units of the screen's own radius.

screen · Curvedscreen
the pupilan opaque edge120 mmf/1.4, focused at 6 m3 of 5 rays get past

A pupil sees around an edge

Two backgrounds identical everywhere a pinhole can see, differing only in the strip an occluder hides from it, produce identical pinhole pictures and pupil pictures 42 per cent apart. So no function of the sharp image — no kernel, no depth-dependent kernel, nothing — produces the picture a real lens makes, and the reach behind the edge is R(Z₂/Z₁ − 1), which is 120 mm here.

sensor · Aperture
0102030400123how wide the audience is (m)the worst seat, in pixels of departure14.1 px per metrea flat panel of the same sizecurved television, the best single pre-warpthe flat panel's line is the axis

One picture and three people

A curved screen can be pre-warped for one seat, and the search over which seat to choose returns the middle one to a quarter of a per cent — there is nothing to be clever about. What the correction buys the sofa as a whole is six per cent, and the worst seat grows at fourteen pixels for every metre of audience, with no width at which it is zero except one person.

screen · Curvedscreen
00.50011.50210203040how many marks the fit was givenerror in the recovered translation direction, in degrees (powers of ten)flatwith depththe same reading error in both5 counts

An ambiguity is not an uncertainty

Eight marks to forty cuts a solid scene's pose error from 19.8° to 0.7° and leaves a flat one at 48°. The two failures look identical from inside — a confident answer, a residual at the floor — and they respond to opposite remedies, so telling them apart is worth more than either measurement.

twoviews · Critical
00.2500.5000.7501-10-50510across the edge, in pixels of the picturehow much of the background reaches the sensorwhere a pinhole puts the edgethe pupilone depth per samplethe same edge, two ways of blurring it70% apart

One depth per sample is not enough

A depth buffer keeps a single distance at each sample, so a post-process blur can only ask how far away the thing at this pixel is. Across an occluding edge that answer is two depths and an occlusion, and the gather it produces differs from the pupil's own integral by 70 per cent of full scale over a band eleven pixels wide.

pipeline · Zbuffer
1234correct from 15 cm, at 160 mm widefour moments · 8.5e-14 px

The same person, twice on one panel

A panel showing one figure at four moments is geometrically the least strange thing in this field — one camera, one floor, and every pair of copies meeting the horizon to 8.5 × 10⁻¹⁴ px. What the picture withholds is the order, and four copies admit twenty-four readings, and a reading convention supplies 4.6 bits from outside the marks. Enlarge one figure by six per cent and the horizon test that passed the panel catches it at 36 px.

conventions · Narrative
correct from 15 cm, at 160 mm widefive figures · 100% rank

Size that means rank

In a great many pictures the drawn height records importance rather than distance. That is a decision rather than a mistake, and it can be caught with a straightedge by carrying one figure's height across the room by the taught construction and see where it lands. A tenth of rank in the picture already misses the drawn head by 21 px, the miss is exactly linear in how much rank is there, and the whole test needs two references and no arithmetic.

conventions · Diminution
band 1band 2band 3every figure the same height, by constructionno horizon

A picture in bands

A register picture stacks its scene in horizontal bands, each with its own ground line and every figure drawn at one height. The feet line and the heads line of any pair are then parallel to the arithmetic floor — 0° against 15.4° in a photograph of the same figures — so the picture has no horizon anywhere in it, and what a reader recovers is an ordering with no metre attached.

conventions · Register
the plan, with its four walls rotated outward about their base linesthe planwalls rotated about their own base lines16 assemblies

The pond with its trees laid flat

An Egyptian garden pond is drawn in plan with its trees rotated outward about the bank they stand on. A rotation is an isometry, so every length in that drawing is exactly the length it is in the garden — zero error, not a small one. What is spent is the angle between any two faces, which reads 180° across every hinge and is 90° in the garden, and the four walls admit sixteen assemblies, so a reader supplies four bits to fold it back up.

conventions · Aspective
012204060how far off the axis the read directions reach, in degreeshow far the recovered wedge angle is out, in degreesthe wedge's own angle — nothing recovered60 trials at each spread, half a pixel of reading error0.068° at 55°

The wedge recovered with the camera

Admit the glass into the model and the fit finally has something left over. A two-degree wedge is invisible to a reading confined inside eighteen degrees of the axis — the fit calls the whole displacement a rotation of the camera and is right to — and by fifty-five degrees it comes back to 0.07 of a degree. What the picture does not separate is the ordinary glasses — assuming an index of 1.50 for a true 1.52 costs one per cent in the angle and 0.07 pixels of residual.

refraction · Slab
a single angle, 60°, for every partfits to 6.7e-13 px

No solid casts an aspective figure

Fitting the best single rigid view to an aspective figure — head and legs in profile, eye and shoulders turned square — misses its own marks by 2.6% of the drawn height, and no yaw does better than 3.0% in a full sweep. A genuine single-view drawing of the same body fits to 7.6e-13 pixels, and the five rotations recovered from the marks alone match the convention's own list to 0.0e+0°.

conventions · Aspective
horizonfaces the eye, gets nothingfaces the eye, gets nothingcorrect from 21 cm, at 160 mm wide4 facing faces unreached · 3.24 m²

Facing the reader is not being reachable

A face turns toward the eye or it does not, and that is a dot product any reader can compute. Whether the eye’s rays actually land on it is a different question with a different answer — on a seating rake, three faces of ten that face the reader receive nothing, and they are 27 per cent of the facing area. On a corner the same test loses nothing at all, which is what makes the gap occlusion rather than arithmetic.

viewing · Anamorph
00.2500.5000.75012345how high the eye is (m)share of the object, and share of the design that landsof the object's surfaceof the design that landsa corridor with a doorway, 7 facesbest surface share 100% at 2.5 m

The eye that reaches the most

A higher eye buys the faces occlusion was hiding and loses design off the far end of the object, so "the best eye" is not a question with an answer until somebody says which of the two they are paying for. On three objects the answer is as high as possible; on a corridor with a doorway in it the two quantities cross and the best height is two and a bit metres.

viewing · Anamorphreach
2 transversalsnoneunfalsifiable3 transversalsnoneunfalsifiable4 transversals15 transversals26 transversals38 transversals510 transversals713 transversals1017 transversals14independent statements a pavement makesthree unknowns, one equation per mark

Four marks before anything is said

A reader fitting a correct perspective to a row of transversals has three numbers to choose, so three transversals fit whatever they are and the fourth is the first that can disagree. Below that count a pavement is unfalsifiable, and a great many painted pavements are below it.

construction · Attribution
10121416-1000100how far the distance point was misplaced, in pixelsthe distance the finished drawing is correct from, in cmas intendedevery point passes the reader's own test4/3 to 8e-15

The slip that leaves no trace

A distance point put twenty-four pixels wrong moves the pavement by two and a half and leaves the reader's projective test reading exactly four thirds. The same slip on Alberti's section moves the drawing by the same amount and is caught, so the difference is not the size of the error — it is that one of them lands back on the set of correct drawings.

construction · Attribution
horizoncorrect from 19 cm, at 160 mm wide2 centres · 2.12 px

How many lamps a drawing has

The shadow field recovers a lamp by intersecting drawn lines. Two lamps make that a partition rather than an intersection — and two centres fit any bundle better than one, on a one-lamp drawing as readily as on a two-lamp one, so a count is a decision that needs a noise level before it exists. A criterion built on a penalty instead of a noise level returns four.

light · Twolamps
02.5057.5010020406080how far the floor is folded from flat (°)what a 10 cm error in the height leaves (mm)two degrees: 0.58 mma design 1.4 m wide, the eye 1.62 m up, the crease at 1.2 mflat: 1.5e-13 mm · 90°: 9.3 mm

A fold names the height

A pavement anamorph’s marks fix where the reader must stand and leave how tall they are entirely free — every height explains the marks exactly, to the last bit. Put one crease in the floor and the freedom is gone, because two degrees of fold makes a ten-centimetre error in the height leave six tenths of a millimetre, and a right angle makes it nine.

viewing · Anamorphrecovery
00.2500.5000.7501-1-0.50000.500how far apart the lamps are, log₁₀ metreshow often the drawing is read as two lampshalf the drawings1 px of clicking, 12 drawings each, separated across the viewbelow the crossing the second lamp is not in the picture

The distance at which two lamps part

Two lamps five centimetres apart are one lamp, and the drawing is right to say so. The separation at which they become two is proportional to how carelessly the picture was clicked — 3.6 centimetres at half a pixel, 7.2 at one, 28 at four — with no floor anywhere, so nothing but care stands between a reader and any separation at all.

light · Twolamps
0.4000.6000.80011.201.40braccia in the pavement (log₁₀)the worst departure of a mark, in pixels (log₁₀)steppedmeasured from the zerothe same hand, laid off two ways√n against n^0.21

Stepped, or measured from the zero

The same hand at the same precision, laying the same braccia off two ways — dividers walked from the last mark accumulate and grow as the square root of the count, while marks set from a common origin do not accumulate at all. The difference is not which method was used — it is where the zero is, and only the second is recorded in the drawing.

construction · Handerror
05101520123how many decades of the control were sweptlargest floor the data cannot rule out (% of the first sample)everything in here is still possiblea floor-free law at 1% noise21.5% at 0.3 decades · 0.00% at 3

What a null result is worth in decades

The first draft of this expected a short sweep to invent a floor, on the reasoning that least squares always spends a free parameter. It does not — on exact data the fitted floor of a floor-free law comes back at three parts in a quadrillion. The failure is the other one and it is worse because it looks like a result. Over a third of a decade at one per cent noise, floors of a fifth of the first sample are still consistent with the data, and the fit reports none while telling the truth.

foundations · Buyable
two lamps: what one centre leaves, floor by floora flat floor138.33 pxa dished floor138.33 pxa dished floor, k = 0.02138.33 pxa dished floor, k = 0.06138.33 pxa dished floor, k = 0.12138.33 pxa ridged floor138.33 pxa ridged floor, k = 0.02138.33 pxa ridged floor, k = 0.06138.33 pxa ridged floor, k = 0.12138.33 pxa floor with a step138.33 pxa floor with a step, k = 0.02138.33 pxa floor with a step, k = 0.06138.33 pxa floor with a step, k = 0.12138.33 pxtwo lamps, four floors, four curvaturesspread over all thirteen: 2.8e-14 px

A floor cannot fake a second lamp

Cast the same two lamps onto four floors at four curvatures and ask how well one centre explains the drawing. Every one of the thirteen answers is 138.3277 pixels — the same to fifteen digits, because a floor decides where along a ray the shadow's tip landed, and a line through a point and another point that has slid along it is the same line.

light · Curvedreceiver
wrong hzcorrectno shapefrom a zerosteppeda rulethe distance pointAlberti's sectionthe measuring pointa photographthe constant ratio60144157203036060read with the panel's own horizon60 drawings each

What a panel says about its maker

The reading assembled over this row, run against every procedure sixty times and scored — with the failures reported as carefully as the successes, because three of the five rows are refusals. A drawing names the class of error in it, not the recipe that produced it, and one procedure it never names at all.

construction · Attribution
how clearly each line belongs to one pencil2 lamps, 5 posts251.5 px100% assigned right3 lamps, 5 posts12.1 px73% assigned right4 lamps, 5 posts38.6 px40% assigned right0.5 px of clicking, lamps spread over 2 mthe counting of unknowns cannot see this

The drawing does not run out of lines

Every post supplies a line to every lamp, so a drawing of five posts offers ten lines to two lamps and twenty to four — the unknowns and the constraints grow together and two posts fix any number of lights. What runs out is the partition, whose margin falls from 251 pixels to six as the share of lines assigned correctly falls from all to just over half.

light · Lightrecovery
00.2500.5000.7501020406080angle off the optical axis (degrees)image radius, against the radius the same lens gives at 90°equidistantequal-areastereographicorthographicflat planeeach curve divided by its own radius at 90°54.0% apart at 45°

Which rule a fisheye obeys, from straightness alone

Four candidate rules for a fisheye lens part by 54.0 per cent at 45° off axis, and a plumb-line fit shown no scene, no camera and no calibration target can still name which one took a photograph — reliably from about 45° of half-field. Below that the four are indistinguishable in the marks, and naming one collapses to guessing.

curved · Fitlaw
lamp 1lamp 2two lamps, close together, in plan1 dark region

Counting shadows is not counting lamps

Two lamps close together cast one connected dark patch and the drawn lines say two. One lamp behind two cards casts two patches and the lines say one, exactly. And the arrangement where the patch count is right — two lamps far apart — has no fully dark region at all, so the same floor answers one, two or zero depending on which darkness is being counted.

light · Shadow
under the lampa scatter, in planground conditioning 2.1e-1

The arrangement the count cannot see

Five posts laid out five different ways give the same leverage to a sixth and the same separation limit to a quarter, and the sixth arrangement — posts strung out along their own shadows, built to be exactly degenerate — is no worse than the rest. The degeneracy belongs to the family the count does not use, and its conditioning there is exactly zero.

light · Pointcount
five points · worst 2.10 px5 straight lines, drawn as arcs between their vanishing pointsworst 2.10 px at 40° of tilt

The arcs the five-point construction actually draws

The taught five-point construction draws circular arcs between five vanishing points and instructs a draughtsman to graduate the radius evenly. Read that way, the arcs miss a straight line's true image by up to 3.75 pixels on a 300-pixel disc. Read at the stereographic scale instead, the same arcs are exact to 4.3e-13 pixels — the construction was always drawing one projection, and the taught scale was never it.

curved · Curvilinear

Named alongside it

The objects these essays reach for when they reach for this one.

ResidualConditioningdegrees of freedomLight recoveryOcclusionPoint lightFree parameterHorizoninstrument limitSamplingAnamorphosisDrawing convention

All concepts