the geometry of numbers
Number theory asks about integers and their combinations; geometry asks about shapes and volumes. The geometry of numbers is Minkowski's discovery that you can answer hard arithmetic questions by drawing the integers as a regular grid of points and reasoning about which convex shapes are forced to contain grid points. It converts statements like 'this equation has an integer solution' into statements like 'this convex body is too big to miss the lattice,' trading algebra for the visual, volume-based reasoning of convex geometry.
The setting is a lattice L in R^n — all integer combinations of a fixed basis, a perfectly periodic array of points with a covolume d(L) measuring the volume per point. The two pillars are Minkowski's theorems. The first: a centrally symmetric convex body of volume exceeding 2^n d(L) must contain a nonzero lattice point. The second governs the successive minima lambda_1 <= ... <= lambda_n of a symmetric convex body K — the smallest dilations at which K first captures 1, 2, ..., n linearly independent lattice points — bounding their product: (2^n/n!) d(L) <= lambda_1 ... lambda_n vol(K) <= 2^n d(L). Reduction theory (Hermite, Minkowski, and the algorithmic LLL) supplies bases that are 'short and nearly orthogonal,' making the lattice's hidden short vectors visible and computable.
The pay-off spans pure and applied mathematics. Classically it proves Lagrange's four-square theorem, Dirichlet's approximation theorem, and the finiteness of class numbers and unit groups in algebraic number theory (Minkowski's bound, Dirichlet's unit theorem). Modern lattices underlie sphere packing (the optimal packings in dimensions 8 and 24 are the E8 and Leech lattices), integer programming, and lattice-based cryptography, whose security rests on the conjectured hardness of finding short lattice vectors. An honest caveat: 'geometry of numbers' names a method and viewpoint, not a single theorem; its power is real but its bounds are often not sharp, and finding the shortest vector in a general high-dimensional lattice is believed to be computationally intractable — which is precisely why cryptographers like it.
Dirichlet's approximation theorem, geometrized: to approximate a real number alpha by a fraction p/q, consider in R^2 the symmetric convex region |q| <= Q and |q*alpha - p| <= 1/Q, a thin parallelogram of area 4. By Minkowski's first theorem (area 4 = 2^2 times covolume 1) it contains a nonzero integer point (p, q), giving |alpha - p/q| <= 1/(q*Q) <= 1/q^2 — a good rational approximation, conjured purely from a volume count.
A thin area-4 parallelogram must contain a lattice point, yielding |alpha - p/q| <= 1/q^2.
It is a viewpoint and toolkit, not one theorem, and its volume bounds are generally not optimal. Finding the genuinely shortest vector of a high-dimensional lattice is conjectured to be hard (NP-hard under randomized reductions) — the foundation of post-quantum lattice cryptography, not a shortcoming to be fixed.