The Gradient, the Curl,
and the Rest

Give me any smooth vector field on the plane — arrows of wind, of water, of “how much does this beat that.” The Helmholtz–Hodge theorem (Helmholtz, 1858) promises it splits, uniquely and orthogonally, into:

the gradient part — the steepest-ascent field of some invisible hill, a potential φ. It has sources and sinks and no rotation anywhere; its streamlines run downhill and never close. This is the part a single number per point captures perfectly.
the curl part — pure circulation, described by a stream function ψ. It has whirls and no sources anywhere; its streamlines close into loops. This is the part that circulates: no hill has these as its slope, so no potential can hold it.

Below, that split is not described — it is computed, live, on whatever field you make. The engine is a spectral Hodge decomposition: at every wave, the flow parallel to the wave is gradient, the flow across it is curl, and “parallel ⟂ across” is why the two halves are exactly orthogonal. Drag on the canvas to paint; the halves re-derive on release.

Load pure gradient and the bar goes all gold: a hill’s slope, zero curl, streamlines that run and never return. Load pure curl and it goes all blue: a whirlpool, zero divergence, every streamline a closed loop. A source is almost pure gradient; a vortex almost pure curl; a saddle is a hill (all gradient) even though it looks like it spins. Messy is a real field — some of each — and the bar reads off exactly how much of it a potential could ever have captured, and how much it could not. That number is not an estimate. Switch to the curl part and you are looking at the residue: the essay’s “part no potential can hold,” drawn.

Why the two halves can’t leak into each other

The bar always sums to 100%, and the check above always reads orthogonal, because the split is a projection. Think in waves. Any field is a sum of plane waves; for each wave with direction k, the flow pointing along k is the only flow a potential can make (a hill’s slope points along the wave), and the flow pointing across k is the only flow a stream function can make. Along and across are perpendicular. So the decomposition just resolves each wave into its along-component and its across-component — the same move as dropping a vector onto two perpendicular axes. Nothing is lost, nothing is shared, and the energies add by Pythagoras. The gradient half is provably curl-free; the curl half is provably divergence-free; and the checkable proof of both is running above, on your own field, every time you let go of the mouse.

The rest — a ghost that only topology can see

Helmholtz on the open plane says two parts. But this canvas is secretly a torus: leave the right edge and you re-enter on the left. On a doughnut a new thing becomes possible — a flow that is both curl-free and divergence-free at once, with no sources and no whirls, that still goes somewhere: a uniform drift that wraps clean around the hole. It belongs to neither half. It is harmonic, and on a torus the space of such flows is exactly two-dimensional — the doughnut’s first Betti number is 2, one winding each way. Press + add drift and watch the violet slice appear: not an error, not roundoff, but the shape of the space itself showing up in the arithmetic. Change the topology and you change how many ways a flow can refuse to be a gradient or a curl. That is Hodge theory — the count of harmonic pieces reads off the holes.

The same theorem, three ways down

This is the continuous face of a decomposition the Wasteland has already built twice, on discrete things. Replace the grid with a graph of who beat whom and the very same Hodge split appears — a “gradient” that is a consistent ranking, and a “curl” that is rock-paper-scissors, the cyclic fraction a leaderboard throws away. That is the argument the prose coda makes: a leaderboard is a potential; the curl is what it cannot hold.