The Cross Product Is a Coincidence

In 1844, a Gymnasium teacher in Stettin published a book of mathematics so far ahead of its readers that the publisher eventually recycled the unsold copies as waste paper.

The teacher was Hermann Grassmann, the book was Die lineale Ausdehnungslehre, and the idea inside it was that geometry deserved an algebra of its own — not coordinates standing in for points, but directed lengths, swept areas, and oriented volumes you could add and multiply directly. Möbius, the nearest mathematician who might have understood it, confessed he could not get through the philosophical prose. Gauss replied politely that the ideas seemed to run along paths he had walked privately for half a century, and engaged no further. Grassmann spent the rest of his career teaching secondary school, gave his spare hours to Sanskrit, and produced a Rigveda dictionary that Vedic scholars still cite. He died in 1877, one year before William Kingdon Clifford published the paper that finally took his product seriously.

Eleven months before Grassmann's book appeared, on October 16, 1843, William Rowan Hamilton was walking his wife along the Royal Canal in Dublin when the solution to a problem he had chased for thirteen years arrived all at once, and he scratched it into the stone of Broome Bridge: i² = j² = k² = ijk = −1. Two algebras of space, born a year apart. Hamilton's quaternions got the professorships, the societies, and a devoted priesthood; Grassmann's extension theory got the pulping mill. And then, in the 1880s and 1890s, Josiah Willard Gibbs at Yale and Oliver Heaviside in London quietly strip-mined the quaternion for parts. They kept the dot product and the cross product of its vector piece, threw away the rest, and when Edwin Bidwell Wilson turned Gibbs's lectures into the 1901 textbook Vector Analysis, the notation war was over. Every physics and engineering curriculum since has taught the settlement as if it were the mathematics itself.

The settlement was a good one. Heaviside used it to compress Maxwell's twenty equations into the four you learned. But it smuggled one artifact into the curriculum with three quiet defects, and every student meets all three and is taught not to ask about any of them.

The cross product takes two vectors and returns a third vector perpendicular to both, with magnitude equal to the area of the parallelogram the two span, and direction chosen by the right-hand rule. Defect one: it does not generalize. In two dimensions there is nothing perpendicular to hand the answer to; in four dimensions there are too many perpendicular directions to choose from. The construction works in exactly one number of dimensions, and the curriculum treats this the way a family treats an eccentric uncle — present at every dinner, never discussed.

Defect two is stranger. Hold a current-carrying loop of wire up to a mirror. The reflected loop runs its current the other way around, which is physically honest — mirrors do that. But the magnetic-field arrow you assigned with your right hand does not reflect like the arrows of velocity or force do. Compute B for the mirrored loop and it points opposite to the mirror image of the original B. Physics students learn to call these strange arrows axial vectors, as opposed to the well-behaved polar kind, memorize which quantities are which — torque, angular momentum, the magnetic field, anything a curl produced — and move on. The vocabulary names the symptom and buries the cause.

Defect three is hiding in the phrase “right-hand rule” itself. Nothing about a spinning flywheel or a current loop knows which of your hands is which. Outside the weak nuclear force, physics is ambidextrous; a hand should never be an input to a law of nature. Yet there it is, in every definition of the cross product, a small anatomical convention doing load-bearing work. All three defects have the same explanation, and Grassmann had already published it: the thing the cross product returns was never a vector in the first place.

Grassmann's product — he called it the outer product, written u ∧ v and read “u wedge v” — answers a different question. Not “what direction is perpendicular to u and v?” but “what did u and v sweep?” Slide the tail of v along u and you sweep a parallelogram. The wedge product is that parallelogram, taken seriously as an algebraic object: an oriented patch of area called a bivector, with a magnitude (the area) and a circulation (first along u, then along v). It does not point anywhere. It lies in the plane the two vectors span, the way a patch of oil lies on water.

Two properties fall straight out of the sweeping picture. Swap the vectors and you sweep the same patch with the opposite circulation,

and sweep a vector along itself and you enclose nothing at all,

which is the algebra's way of saying that only the independent parts of two vectors make area. The product is linear in each argument, so it plays well with coordinates. In ℝ³ the expansion works out to

where e₁₂ is the unit patch in the 1–2 plane and so on. Look at those three coefficients for a moment. They are, up to a sign and a shuffle, exactly the three components of the cross product you memorized. The numbers were always right. What the numbers are attached to is the issue: the cross product files them under three axes, while the wedge files them under three planes. Three numbers filed under planes, in a space that happens to have three planes through the origin's coordinate pairs — that “happens to” is the whole story.

In n dimensions, a vector has n components — one per axis. A bivector has one component per coordinate plane, and the number of coordinate planes is the number of ways to choose two axes from n:

Run the numbers. In two dimensions there is one plane, so a bivector is a single signed number — an area with a circulation. In three dimensions, three planes, three components. In four dimensions, six. A bivector in ℝ⁴ has six components, and no four-component vector can impersonate it; in ℝ² the impersonation fails in the other direction. For a space with any directions in it at all, the equation n(n−1)/2 = n has exactly one solution, and it is three. The cross product exists because three is the only number of dimensions where planes and directions are counted by the same number. That is the coincidence in the title. (Seven dimensions admits a different binary product, inherited from the octonions, but it is not this construction and it surrenders most of the properties that make the three-dimensional one useful.)

A matching count means a translation is possible. It does not perform the translation. For that you need a machine that takes each plane to the one direction it leaves out, and that machine needs two pieces of information the wedge product never asked for.

The machine is written ⋆, named for William Vallance Douglas Hodge, who built it into his theory of harmonic integrals at Cambridge in the 1930s. The Hodge star takes any k-dimensional blade in an n-dimensional space and returns its orthogonal complement — the (n−k)-dimensional blade made of everything the original left out, carrying the same magnitude. To know what “orthogonal” means, the star consults a metric: lengths and angles are an input. And to decide between the two ways of orienting the complement, it consults a chosen ordering of the axes — an orientation. That second input is the right hand. It was never anatomy; it was a bit, a single sign convention, fed to an operator the curriculum declined to name.

In ℝ² the star of a vector is that vector turned a quarter to the left: everything perpendicular to a direction, in a plane, is another direction. Apply it twice and you are facing backwards, ⋆⋆ = −1, which is the plane's way of admitting it contains a hidden rotation. In ℝ³ the star trades planes for their normals and volumes for numbers, and composing it with the wedge gives the formula the whole settlement rests on:

The cross product is two operations wearing one trench coat. The wedge does the geometry: it sweeps the parallelogram, records the area, fixes the circulation, and does all of it identically in every dimension, with no metric and no hands. The star does the packaging: it exchanges the honest bivector for the one arrow that fits in the leftover direction — a swap that is only available because of the counting coincidence, and only well-defined because you fed it a metric and a handedness.

Every mystery from the earlier list now dissolves on contact. The mirror problem: a reflection reverses orientation, so the star's sign convention flips, so every starred quantity — torque, angular momentum, B — comes out backwards relative to unstarred ones. The bivector underneath behaved perfectly all along; it is the arrow costume that fails the mirror test. The hand in the definition: that is the orientation input, made flesh. And the refusal to generalize is just the count from the last section — in four dimensions the star of a plane is another plane, and there is no arrow to exchange it for.

It is worth sitting for a moment with what ℝ⁴ actually does to these objects, because four dimensions is where the shorthand's debts come due, and it is also where the physics lives. The star maps grade k to grade 4−k, so planes map to planes: the complement of e₁₂ is e₃₄, a plane that shares nothing with the first but a single point. Two full two-dimensional sheets, meeting only at the origin — there is no way to picture that in three dimensions, and no way to avoid it in four. Rotation changes character for the same reason. A rotation happens in a plane, not around an axis; the axis you learned is just ⋆ applied to the rotation plane, one more three-only artifact. In four dimensions a rigid body can rotate in e₁₂ and e₃₄ simultaneously, at independent rates — a double rotation with no fixed axis anywhere, every point moving except one.

And the six components of a bivector in ℝ⁴ are not a curiosity. Six is three plus three, and that arithmetic is physical: the electromagnetic field is a single bivector F on spacetime whose six components are the three of E and the three of B. What a boost mixes, what an observer splits into “electric” and “magnetic,” is one geometric object that never fit in any arrow. In vacuum, Maxwell's four equations collapse to a pair — dF = 0 and d⋆F = 0 — with the star (built there from spacetime's metric rather than Euclid's) rotating the electric face into the magnetic one. Readers of Seams and Signals have met d already, as the coboundary that turns local measurements into loop sums; the star is its missing partner, the operator that turns “around the loop” into “through what the loop misses.”

None of this needs to stay verbal. I built a small instrument for it: two to four dimensions, a pair or triple of draggable vectors, the wedge drawn as the patch it is, and the star drawn as whatever the complement actually looks like in that dimension. Open Wedge & Star.

Start in ℝ² and drag the tips: the parallelogram's area is the product's magnitude, the little circulation arrow is its sign, and dragging v across u flips the circulation the instant the sweep reverses. Pull the two vectors nearly parallel and watch the patch die — u ∧ v forgets everything the vectors share. Then move to ℝ³ and turn on the star: the amber arrow standing on the patch is your old friend u × v, reconstructed from parts, and swapping u with v flips patch and arrow together. Orbit the scene until the plane is edge-on and the arrow is pointing at you; that view is the right-hand rule, drawn instead of mimed.

Then switch to ℝ⁴, where the rose-colored fourth axis enters, and star the same parallelogram. The dual is a patch now, not an arrow, tilted so that it shares nothing with the first but the origin — drag a vector's e₄ slider and watch the two planes negotiate their perpendicularity in real time. Wedge three vectors and the star hands back the one direction all three left out, which is as close to a cross product as four dimensions will ever offer: it needs three arguments there, not two. And if you want the full tour, set the view spinning — a slow double rotation, the kind with no axis — and watch objects leave ordinary space and come back without ever moving discontinuously.

None of this says the cross product was a mistake. Gibbs and Heaviside were solving a real problem — quaternions were a poor fit for the physics of their century, and the vector settlement carried electromagnetism, fluid dynamics, and two generations of engineering on its back. Shorthands earn their keep. But a shorthand quietly becomes a worldview when nobody writes down what it abbreviated, and for a hundred and twenty years the standard curriculum has taught the abbreviation as the thing itself, axial vectors and hand rules and all, while the two clean operations underneath went missing from the syllabus.

So the next time you catch yourself curling your fingers around an axis — for a torque, a flux, a curl — notice the moment your hand enters the calculation, and ask what it is being asked to decide. It is choosing an orientation for a star you were never told about, converting a swept plane into a pointed arrow because in three dimensions the bookkeeping happens to permit it. The arrow was the shadow. The plane was the thing.

[1] Michael J. Crowe, A History of Vector Analysis (1967). The standard account of the whole affair — Grassmann's reception and the fate of the 1844 edition, Hamilton and the quaternion school, and the Gibbs–Heaviside settlement.

[2] Hermann Grassmann, Die lineale Ausdehnungslehre (1844). The outer product, oriented magnitudes of every grade, and most of exterior algebra, sixty years early.

[3] W. R. Hamilton's account of October 16, 1843 survives in his 1865 letter to his son Archibald; the Broome Bridge carving itself did not.

[4] Edwin Bidwell Wilson, Vector Analysis (1901). The textbook, built from Gibbs's Yale lectures, that fixed the dot and cross product as the standard language.

[5] W. V. D. Hodge, The Theory and Applications of Harmonic Integrals (1941). The star operator in its native habitat.

[6] Leo Dorst, Daniel Fontijne, and Stephen Mann, Geometric Algebra for Computer Science (2007). A modern computational treatment in which bivectors, not cross products, do the rotating.