the separating hyperplane theorem
If two crowds of people stand apart with a gap between them, you can run a single straight rope down the gap that keeps one crowd strictly on the left and the other on the right. The separating hyperplane theorem makes this precise for convex sets: two convex sets that do not overlap can be split by a single flat wall, with one set on each side. It is the geometric reason linear classifiers exist, the cornerstone of convex duality, and the finite-dimensional shadow of the Hahn-Banach theorem.
The cleanest version: let A and B be nonempty disjoint convex sets in R^n. Then there is a nonzero vector u and a number c with <u, x> <= c for all x in A and <u, y> >= c for all y in B — a hyperplane <u, x> = c with A on one side, B on the other. If in addition A is compact and B is closed (and they are disjoint), the separation is strict: there is a gap, <u, x> <= c1 < c2 <= <u, y>, so the wall has breathing room on both sides. The proof idea for a point and a closed convex set: take the unique nearest point q in the set to the outside point p; the perpendicular bisector hyperplane through the midpoint, with normal p - q, separates them — uniqueness of the nearest point is where convexity does the work.
This one theorem powers an astonishing amount: the duality of linear programming, the existence of supporting prices in economics (the second welfare theorem), the margin in support vector machines, Farkas' lemma, and minimax theorems in game theory. Two honest caveats. First, mere disjointness gives only weak (non-strict) separation in general — two disjoint convex sets can be 'asymptotically tangent' (like the region under y = e^x and the region y <= 0) with no strict gap, so you must assume compactness somewhere for strict separation. Second, in infinite dimensions one needs the sets to have nonempty interior or to invoke Hahn-Banach with care; closedness alone is not enough.
In R^2 let A be the closed unit disk centered at the origin and B the single point (3, 0). They are disjoint convex sets, and the vertical line x = 2 separates them strictly: every point of A has x <= 1 < 2 < 3, the x-coordinate of B. The nearest point of A to (3,0) is (1,0); the separating direction u = (1,0) is exactly (3,0) - (1,0) normalized, illustrating the nearest-point construction.
A disk and an outside point are split by a vertical line; the normal points along nearest-to-far.
Convexity of BOTH sets is essential. Two non-convex sets that interlock like jigsaw pieces — say two nested C-shapes — cannot be separated by any hyperplane even when disjoint. The theorem buys separation only for convex sets, which is why so many problems are first 'convexified'.