The corner that answers every eye

Every picture on this site has a place it is correct from. That is the premise, and the point to stand at is where the collection makes it a number rather than a remark. This essay is about the one arrangement here that does not have such a place — not because it is badly made, but because every place is the right one.

The arrangement is three mirrors meeting at right angles, and it is on the back of every bicycle.

Three reflections make an inversion #

Two mirrors make one turn establishes that two reflections compose into a rotation about the line where the mirrors meet, by twice the angle between them. Add a third plane, perpendicular to both, and the composition changes character.

Take the three coordinate planes. The first reflection negates $x$, the second negates $y$, the third negates $z$, and each leaves the other two coordinates alone — so the composition sends $(x, y, z)$ to $(-x, -y, -z)$. That is the point inversion through the corner: every point taken to its opposite, every direction reversed.

The consequence for a ray is immediate and does not depend on which direction the ray came from. The inversion sends the direction $\mathbf{d}$ to $-\mathbf{d}$, so a ray that bounces off all three faces leaves antiparallel to the one that arrived, offset sideways but pointing exactly back.

Measured over five directions in section, and over three hundred and twenty-four directions in the full three-dimensional fan, the departure from antiparallel is under 10⁻¹⁰ degrees — arithmetic, because nothing is fitted.

A picture with no station point #

The three reflections commute. Negating $x$ and then $y$ is the same as negating $y$ and then $x$, so the order in which a ray meets the faces does not matter, and every ray in the octant is treated the same way.

That is what makes the arrangement unusual on this collection. Everything else here has a viewpoint the picture is correct from, and the collection spends a great deal of length measuring what standing somewhere else costs — standing in the wrong place for a perspective picture, an anamorph has one eye for the extreme case, and the whole of the viewing field in between. A corner reflector inverts the question. Light sent from anywhere comes back to the sender, so a lamp anywhere in the octant sees a bright corner, and there is no place from which the arrangement fails.

It is worth saying what a corner reflector is not. It is not a picture of anything: it returns light without forming an image, and a viewer looking into a perfect one sees their own eye and nothing else. So the arrangement belongs in this collection as the boundary of the viewing field rather than as a member of it — the case where the viewing-tolerance question has the answer “everywhere”, which is what the rest of the field’s numbers are measured against.

Half a degree costs a whole one #

A real corner is made to a tolerance, and the question with a number in it is what an error in one face costs the returning ray.

Tilt the third mirror by half a degree and try the same fan of three hundred and twenty-four incoming directions. The worst returning ray is out of antiparallel by 0.99999 degrees — exactly twice the mirror’s error — and the relation holds at every tilt tried, from a twentieth of a degree to one and six tenths.

The factor of two is the third appearance of the same fact in this field. A reflection’s effect is measured from the mirror, so reflecting a point moves it twice its distance to the plane; a mirror pair’s baseline is twice the camera’s distance to the glass; and a mirror turned by an angle turns the reflected ray by twice that angle. All three are one statement seen from three sides.

The best ray and the worst ray are not the same ray #

The number that matters for an instrument is the worst case, and quoting only the worst case hides something a reader needs.

The same fan of directions, at the same half degree of error, produces returning rays out by anything from 0.085 degrees to 0.999 — a spread of nearly twelve to one. Which value a particular ray gets depends on the direction it arrived from: a ray nearly along the axis of the corner is barely affected, and one arriving obliquely is affected by the whole of it.

So a tolerance measured on one ray is a statement about that ray. This is the same failure the collection has recorded twice under other names — a cross-ratio test evaluated at four consecutive divisions, where the value cannot distinguish anything, and a conformality test evaluated along the one tangent basis where the cylinder cannot fail. Both are cases of a necessary condition checked at the one input that makes it vacuous, and what a projection destroys carries the first.

Here the shape is milder — one ray gives a true number rather than a vacuous one — but the trap is the same: a single sample of a quantity that varies over its domain, reported as the quantity.

Where the factor of two comes from #

The demonstration above is numerical and it deserves a sentence of argument, because “twice” invites the question “twice what, exactly”.

A single reflection in a plane with normal $\mathbf{n}$ sends a direction to $\mathbf{d} - 2(\mathbf{n}\cdot\mathbf{d})\mathbf{n}$. Rotate the plane’s normal by a small angle $\epsilon$ about some axis, and the reflected direction rotates by $2\epsilon$ about that same axis — the derivative of the map with respect to the normal’s orientation is exactly two, uniformly. A ray meeting three faces meets the tilted one once, so it picks up $2\epsilon$ once, and the other two faces carry that error through unchanged because they are isometries.

The reason the worst case is exactly $2\epsilon$ and the typical case is less is that the error is a rotation about a particular axis, and its effect on a ray depends on the angle between the ray and that axis. A ray parallel to the axis is unaffected; a ray perpendicular to it takes the whole $2\epsilon$.

The spread has a law, and it makes the worst case the honest number #

The section above reports a spread of nearly twelve to one and leaves it as a caution. The law behind it is one line, and having it turns the caution round.

The tilt is a rotation of one mirror’s normal by $\epsilon$ about some axis, so the returning ray is rotated by $2\epsilon$ about that same axis — and rotating a direction $\mathbf{d}$ by an angle $\delta$ about an axis $\hat{\mathbf{a}}$ moves it by $\delta\sin\vartheta$, with $\vartheta$ the angle between the two. So the departure from antiparallel is

$$2\epsilon\,\sin\vartheta.$$

At half a degree of tilt that gives 1.000° for a ray perpendicular to the tilt axis — the measured worst case — and 0.085° needs $\sin\vartheta = 0.085$, a ray within 4.9° of the axis. The twelve-to-one spread is the sine running over the fan, and its lower end is one narrow direction rather than a typical one.

Which is the useful correction. Averaging $\sin\vartheta$ over directions gives $\pi/4 = 0.785$, so the typical returning ray is out by about four fifths of the worst case, not by a twelfth of it. The distribution is bunched near the maximum and has a thin tail toward zero, because a sine is flat near its peak and steep near its root.

So the tolerance quoted as $2\epsilon$ is very nearly what a randomly chosen ray receives, and the best-case figure is the misleading one. That reverses the instinct the previous section leaves — the worst case is not a pessimistic bound here, it is a good estimate — and it is worth having explicitly because a designer reading “0.085 to 0.999 degrees” would reasonably budget somewhere in the middle and be wrong by a factor of two.

The same expression says which orientation of a manufacturing error is expensive. An error whose axis lies along the corner’s own diagonal — the direction most incoming rays come from — is the cheap one, because those rays have a small $\sin\vartheta$; an error about an axis perpendicular to it is the expensive one. A moulded reflector’s faces are not out of square in a random direction, so a maker who knows which way their tooling errs can spend the tolerance where it costs least, and the factor available is the same twelve.

The eye that sees itself #

There is a consequence worth drawing out, because it is where the arrangement touches the rest of the viewing field rather than sitting beside it.

A viewer looking into a perfect corner reflector sees their own pupil at the centre of every face, and nothing else. Light from any other object bounces back to that object rather than to the viewer, so the only thing the viewer can see is the only thing sending light along the path that returns to the viewer — which is the viewer. The three faces show three copies of it and the copies are inverted through the corner, which is a parity a reader can check on an asymmetric eye.

That is the exact opposite of the anamorph. An anamorph has one eye is about a design that is correct from one place and incoherent everywhere else; a corner reflector is coherent everywhere and shows nothing but the observer. Between the two lie the ordinary pictures, which are approximately right over a region and whose region where the anamorph still works measures.

Both extremes have the same underlying cause and it is worth naming. A picture’s tolerance is set by how fast the picture changes as the eye moves, and the corner reflector’s picture changes with the eye — it is the eye — so moving the eye cannot break it.

What this is for #

Three uses, and they are the reason the arrangement is worth a rung rather than a footnote.

A retroreflector on a road sign or a bicycle sends a car’s headlights back to the car, which is why it appears bright from the driver’s seat and dim from the pavement. The brightness is not a property of the material; it is a property of the geometry, and the tolerance measured above is why a moulded plastic reflector with faces a degree out of square still works at the distances involved.

A corner cube in a surveying prism returns a laser to its source over a kilometre without alignment, which is the whole reason such instruments can be operated by one person. The tolerance matters more here: at a kilometre, two degrees of return error is thirty-five metres of miss, so the faces are made to arcseconds.

And the lunar retroreflectors left on the Moon return a pulse to the Earth over three hundred and eighty thousand kilometres. The same factor of two sets their tolerance and the same argument applies unchanged, which is the sort of scale invariance this collection likes: the geometry does not know how far away anything is.

Two mirrors are not enough, and four are no better #

A pair of perpendicular mirrors sends a ray back in the plane perpendicular to their common line and does nothing at all to the component along it. So a two-mirror corner retroreflects in one dimension: it is a strip that works for rays arriving in one plane, which is why a reflective strip on the side of a lorry is a row of two-mirror grooves and why it fails when the headlights are above it.

Three mirrors close the remaining dimension because three perpendicular planes span the space, and there is nothing left over. A fourth mirror adds no dimension to close, so it adds nothing — and in fact takes something away, because a ray meeting four faces has been reflected an even number of times and comes out with its handedness preserved and its direction not reversed. A four-bounce path is a rotation.

This is the same accounting the previous rung makes for the dihedral group, and it is worth stating in the same words: what an arrangement of mirrors supplies is a composition, and the composition’s parity decides what it does. An odd number of reflections reverses handedness; three perpendicular ones reverse every direction; anything else does something else.

The boundary, stated #

Three exactly-flat mirrors, exactly perpendicular, and a ray that meets all three. Each of those can fail.

A ray entering near an edge may meet only two faces, in which case the composition is a rotation rather than an inversion and it does not come back — which is why a real corner cube’s effective aperture is smaller than its opening, and why the returned light falls off toward the rim rather than stopping at it. Curved faces have no fixed plane and the whole argument goes; a curved mirror has no eye is the general statement. And nothing here is about the wavelength of the light: a corner small enough for diffraction to matter is an optics problem rather than a geometry one, and this collection computes geometry.

The arrangement read as a picture surface #

This collection files everything under what it does to a projection, and it is worth asking what a corner reflector is under that heading.

A picture surface takes a direction from the eye and puts a mark somewhere. A corner reflector takes a direction from a source and returns it, unmarked — so it is not a picture surface at all, and its entry in the table no picture surface keeps everything draws up would be a row of blanks. It preserves no straightness because it draws no lines; it preserves no angle because it draws no angles; and it has no area scale because it has no area to scale.

What it does preserve is the one thing every other surface in the collection destroys: the identity of the direction. Every other surface here maps a bundle of directions onto a plane or a curve and loses something in the mapping, and the whole curved field is an accounting of what. A corner reflector maps every direction to itself, negated, exactly, and loses nothing at all — because it never leaves the space of directions.

That is why it sits at the boundary of the viewing field rather than inside it, and why the answer to “where must a reader stand” is “anywhere”. A surface that keeps everything can afford to.

What is measured here #

Three numbers and a rejection.

A right-angled corner returns three hundred and twenty-four incoming directions antiparallel to under 10⁻¹⁰ degrees. A corner with one face half a degree out of square returns the worst of them 0.99999 degrees off, which is twice the mirror’s error to four parts in a hundred thousand, and the best of them 0.085 degrees off — a spread of nearly twelve to one across the same fan. And the relation between the mirror’s error and the worst return is a straight line of slope two across a thirtyfold change in the error, which is what says the factor is a law rather than a coincidence at one setting.

The short version #

Three reflections in mutually perpendicular planes compose into the point inversion, so a corner reflector sends every ray back the way it came whatever direction it arrived from. It is the one arrangement in this collection with no station point, because every station point is correct.

An error of $\epsilon$ in one face costs the worst returning ray exactly $2\epsilon$ and the best about a twelfth of that, so the arrangement degrades gracefully and its tolerance has to be quoted as a worst case over directions rather than measured on one ray. The factor of two is the same one that makes a mirror pair’s baseline twice the distance to the glass.