The room a divergent picture is a photograph of

An inverse perspective is a leaning plane settled what a single divergent construction depicts: not a trapezium widening with depth, but an ordinary rectangle on a plane tilted toward the camera, with the splay alone fixing how far — 36.3° toward the lens at a splay of 1.32, 53.5° away at 0.62. That answer comes from one quadrilateral, drawn on its own. A real room has more than one plane meeting at a corner, so the natural next question is what happens when a second one-point construction — a wall, standing on the floor’s own far edge — is drawn beside the first, each with its own, independently chosen splay.

The question this essay was built to answer was where the required lean passes 90° and the depicted scene stops existing outright, on the reasoning that a plane cannot lean further than edge-on. That turns out to be the wrong plane to ask it of. The floor’s own splay-to-tilt map never gets near edge-on at all — it is bounded well under 89° for every spread the recovery machinery sweeps, and it flattens out near 77° long before that — while the companion wall’s map crosses exactly 90° at a wall spread of 3.000, independently of which floor it is standing on. The scene does not stop existing because a floor tips over; a specific, computable wall does.

Two independent constructions rarely agree #

The recovery behind every number in this essay is planeOfQuad, which takes any four corners, finds the two vanishing points its opposite sides imply, and returns the plane whose normal those two directions determine — the same computation the essay above runs on a single quadrilateral, generalised so a wall’s four corners can be handed to it exactly as a floor’s can.

Run it once on the floor and once on the wall in the figure above and it returns two tilts, 36.3° and 29.3°, computed from nothing shared between them except one assumed camera — the same principal point and the same focal length, standing in for an eye neither drawing states. There is nothing in that computation that requires the two recovered planes to meet at a right angle, because nothing about drawing two one-point constructions on a shared sheet enforces it. A draughtsman choosing a floor’s splay and a wall’s splay independently — which is exactly what a splay is, a free choice rather than a measurement — has no reason built into the construction to land on a buildable corner, and 6.9° is what “no reason” looks like as a number.

The assumed camera is itself a free choice, and it is worth being explicit that this essay fixes it throughout rather than exploring it. Every figure here recovers both planes against one stated principal point and one stated focal length, chosen once and reused for the floor and the wall alike. A different assumed camera would in general read different individual tilts off the identical marks — the recovery is only ever a statement about these marks, under that camera, never about the marks alone. What survives a change of assumed camera and what does not is not this essay’s question; holding it fixed is what lets the floor and the wall be compared to each other at all; it is not offered as the one true camera the panel was drawn against.

The dihedral crosses ninety without being asked to #

If the wall’s splay is free, the dihedral it produces alongside a fixed floor ought to move as that splay moves, and it is worth checking that the 6.9° above is not simply where the pair happens to sit rather than a floor for the family.

Those two readings already say the dihedral moves a long way for a small change in wall spread — 87.7° of arc between two spreads only 0.55 apart, one of them 83.1° short of square and the other 4.6° past it. That is worth stating plainly before searching for an exact right angle: a corner that happens to be square is not a special property of divergent constructions, it is a coincidence available to the same family that produces every other angle, reachable by moving one free parameter a modest distance, and the interesting question is not whether square is reachable but which wall spread reaches it for a stated floor.

Solving for a buildable corner #

solveSquareWallSpread is a bisection on the dihedral itself, run against wallSpread alone with the floor held fixed, and it lands on 0.511 for this floor — a specific, unremarkable-looking number with nothing about it that a draughtsman drawing by eye would be likely to land on unassisted. That is the honest content of “buildable” here: consistency with one right-angled corner, recovered after the fact, rather than a property either drawing states about itself. A panel drawn with wall spread 0.511 and one drawn with wall spread 1.15 differ from each other by an amount well within what a hand-drawn divergent construction varies by in practice, and only one of the two is a room.

A solved wall spread also exists for other floors, which is worth confirming rather than assuming from the one case above. Floors of spread 0.62, 0.8, 1.6 and 2.5 each return their own square-corner wall spread from the identical bisection, none of them equal to 0.511 and none of them equal to each other — the solved value is a joint property of the pair, not a constant the family carries. That a solution exists at all, for floors this different from one another, is closer to the interesting fact than any one of the individual numbers: the family is rich enough that a buildable corner is never far away, which is exactly why an arbitrary pairing so reliably misses one.

The degenerate spread, avoided #

Every wall spread used above sits comfortably away from one value that has to be avoided for the same reason divergent’s own floor construction avoids it: a spread of exactly 1 draws a wall whose two long sides are parallel on the page, which is a parallel projection with no vanishing point at all rather than a one-point construction with an unhelpful one. planeOfQuad cannot recover a tilt from a pair of parallel sides any more than it could from the floor at its own splay of 1, and the wall spreads used in every figure above — 0.6, 0.511, 1.15 — are each chosen clear of it.

The neighbourhood of that gap is worth naming even though no figure here is placed inside it, because it is where the dihedral does something no smooth reading of the figures above would predict. Swept directly rather than placed, the dihedral runs to roughly 125° just below a wall spread of 1 and reappears at roughly 36° just above it — a jump belonging to the construction’s own degenerate point rather than to any property of the room, the same discontinuity the picture whose lines spread already reports for a single one-point construction crossing its own splay of 1. Two constructions sharing a sheet do not smooth that discontinuity away; each carries its own, independently.

The control: a photograph is not ambiguous #

Both figures so far reconstruct planes from marks that were never a projection of anything solid — divergent is a construction, built directly in the picture plane rather than projected from a scene. The control has to remove that: build an actual floor and wall at a genuine right angle in space, project them with one ordinary pinhole, and run the identical recovery.

90.000000000° is not “solved for,” the way the previous figure’s 0.511 was; it falls out of the recovery unassisted, because the marks genuinely came from a right-angled scene and the reconstruction has nothing to reconcile. The contrast between this figure and squared is the whole difference between consistency and evidence: a solved wall spread makes a drawing consistent with a room, and a photograph is one, and the recovery machinery cannot tell the difference from the marks alone — which is exactly why 0.511 had to be searched for while 90.000000000° simply appeared.

Where a floor’s own lean tops out #

The original question behind this essay assumed a plane’s lean could run away to edge-on and take the whole scene with it. Checked directly rather than assumed, the floor’s own splay-to-tilt map does no such thing.

Sweeping divergent’s own splay from just above 1 out to 200 — the same sweep the recovery machinery already runs to confirm it before relying on it — the recovered tilt tops out at 77.40°, at a splay near 159, and it barely moves from there: at a splay of a thousand it reads 77.51°, at a million, 77.53°. The floor’s own map is not approaching 90° at all; it is approaching a different, lower ceiling and has essentially reached it long before any spread a draughtsman would plausibly use. Over that whole range it stays under 89°, which is the bound the recovery machinery checks rather than merely expects. A plane that levels off 12 or 13 degrees short of edge-on is not a plane whose lean is going to make a scene stop existing, however far its splay is pushed.

The reason the floor levels off rather than running away is where its own sides’ vanishing point is allowed to sit. Checked directly across the same range, that point’s picture height falls from several thousand pixels at a splay just above 1 to about 330 px as the splay grows without bound — and it never crosses the assumed camera’s own principal-point height of 215 px anywhere in between. planeOfQuad’s recovered tilt is built from that one moving point together with the near and far edges’ fixed horizontal direction, and a vanishing point that stays on one side of the principal point’s height, however far it travels, cannot swing the recovered plane past the tilt that height corresponds to. The wall’s own sides’ vanishing point is not confined the same way, and that is exactly where the next section’s crossing comes from.

Where a wall’s own lean does not top out #

The wall is a different quadrilateral from the floor — its own bottom and top edges sit at different picture heights, and as wallSpread grows those edges’ own vanishing point sweeps upward through the picture rather than settling toward a ceiling the way the floor’s does.

3.000 is where the wall’s two sides’ vanishing point crosses the assumed camera’s own principal-point height — a genuine, checkable geometric event rather than an approach to a limit, which is why the tilt reaches 90° exactly rather than merely nearing it. And running the identical sweep against a different floor, spread 0.62 rather than 1.32, returns the same 3.000 to six decimal places: the wall’s own critical spread depends only on the wall’s two edge heights and its own spread, not on which floor it happens to be standing on. That independence is worth pausing on, because it is the reverse of the naive expectation that a “critical divergence” for the whole room would depend on the whole room.

What the peak means for the picture is worth stating without softening it. A wall spread of exactly 3.000 depicts a plane with no width left to draw four corners on — genuinely edge-on, seen exactly along its own surface — and a wall spread on the far side of 3.000 is read by the identical machinery as a different, shallower plane, because the vanishing point that determines the tilt has swept past the principal point and the construction cannot distinguish “further divergent” from “coming back the other way.” That is not the scene ceasing to exist in the manner originally proposed; it is the recovery becoming unable to tell two different walls apart, which is a narrower and more specific kind of limit.

A single construction, for comparison #

It is worth setting the two-plane panel beside the one-plane construction it is built from, because the contrast says exactly what the second plane adds.

A single such construction has exactly one recovered tilt and nothing to disagree with — the picture whose lines spread and the essay built on it are both about that one number and what it does and does not encode. The moment a second one-point construction is drawn beside the first and asked to share an edge, there are two tilts where there was one, and a dihedral where there was nothing to compare. Divergence itself is not the source of the disagreement measured above — a single divergent panel is perfectly self-consistent, as consistent as any parallel or convergent one — the disagreement is entirely a property of the pair, which is the same shape of finding two stations in one picture reaches from a different construction: one rule, applied once, cannot disagree with itself; the interesting numbers start when a second, independently chosen instance of the same rule is asked to share a sheet with the first.

The honest limit #

Nothing in this essay’s recovery tests whether a divergent panel was drawn as a room. It tests whether the marks are consistent with one, which a solved wall spread guarantees and an arbitrary one does not, and consistency is a weaker claim than intention. Counting the eyes needs the room makes the identical distinction from the opposite direction — a fitted centre that lands exactly where a room’s own measurements say it should is consistency with a real room, not proof that a real room was photographed, and the same caution applies here in reverse: a solved 0.511 is consistency with a right angle, not evidence a right angle was intended by whoever chose it.

The critical wall spread is also not the kind of limit the opening question imagined. It is a peak in a map that keeps producing answers past it, not a wall the construction refuses to draw — the picture whose lines spread meets the same shape at its own splay of exactly 1, where the construction does not fail so much as change what kind of object it is describing. And the floor’s own bounded tilt is a fact about divergent’s particular proportions rather than a law that every one-point construction must obey; a floor built with a different depth-to-width ratio would level off at a different ceiling, and nothing here measures where that ceiling moves.

Nor does anything here say what a reader looking at such a panel would actually notice. A dihedral of 6.9° instead of 90° is a large number on the page this essay draws it on, but the panel itself shows two flat shapes meeting along one edge, and nothing about the drawing marks that edge as a corner that ought to be square rather than one that is simply however it is. A picture with two eyes in it makes the equivalent point about a genuinely two-eyed composite: the seam is there in the geometry whether or not a viewer’s eye catches it, and this essay’s dihedral is the same kind of fact — real, computed, and silent about how conspicuous it is.

What this joins #

Recovering the camera from a picture only works when the picture is a projection of one scene from one point, and everything in this essay is what happens to that recovery when it is quietly handed two independent constructions and not told there are two. Where parallel lines meet is the foundational fact both planeOfQuad calls lean on — a vanishing point is where a direction’s own rays converge, and a plane’s tilt is read off two such points at once. Brunelleschi drilled a hole in his panel and the point a picture has to be seen from both insist that a single picture states a single place to stand; this essay’s panel, read as two rooms sharing an edge, has no such place at all, only two planes and an angle between them that a real corner would fix and a drawn one leaves free.

The shape of the whole essay is worth stating once more, plainly, because it is easy to lose under the specific numbers. A single divergent construction is exactly as consistent as any other drawing system this collection measures — it depicts one rectangle, on one tilted plane, with nothing left over. Consistency only becomes a live question the moment a second such construction is asked to share an edge with the first, and the question it raises is not whether either half is a good picture but whether the pair, together, could be a photograph of anything at all. That the answer is usually no, and occasionally yes at one specific, computable wall spread, is a statement about pairs of constructions rather than about divergence itself — the same lesson two stations in one picture draws from a rule glued to another rule rather than a plane glued to another plane.