JOVANA
Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
All guides

Cauchy-Schwarz and the geometry it unlocks

One inequality — |<u, v>| <= ||u|| ||v|| — turns three dry axioms into real geometry: the triangle inequality, well-defined angles, and the identities that recover the inner product from the norm alone.

The master inequality

The single most important consequence of the axioms is the Cauchy-Schwarz inequality: for all u, v, |<u, v>| <= ||u|| ||v||, with equality exactly when u and v are parallel. Without it the word "angle" would be meaningless in an abstract space, because we could not guarantee that <u, v> / (||u|| ||v||) lands in [-1, 1].

The proof is a one-line trick that you should keep in your pocket: pick the scalar t that makes ||u - t v|| smallest, then expand 0 <= <u - t v, u - t v> and read off the bound. The minimizing t = <u, v> / <v, v> is the coefficient of the best projection of u onto v — geometry and algebra meeting in one move.

Recovering the product from the norm

Does the norm remember the inner product? Yes — and there is a clean test for when a norm came from an inner product at all: the parallelogram law, ||u + v||^2 + ||u - v||^2 = 2 ||u||^2 + 2 ||v||^2. The sum of the squared diagonals of a parallelogram equals the sum of the squared sides. A norm satisfies this if and only if it arises from an inner product.

When the law holds, the polarization identity reconstructs the product explicitly. Over R: <u, v> = ( ||u + v||^2 - ||u - v||^2 ) / 4. Over C an extra imaginary piece appears to recover the phase. So norm and inner product carry exactly the same information — knowing all lengths is the same as knowing all angles.

Is the max-norm ||x||_inf = max(|x1|, |x2|) from an inner product?
Test the parallelogram law on u = (1, 0), v = (0, 1):

  ||u + v||_inf = max(1,1) = 1   -> squared 1
  ||u - v||_inf = max(1,1) = 1   -> squared 1
  LHS = 1 + 1 = 2
  RHS = 2*1 + 2*1 = 4

  2 != 4   ->  law FAILS  ->  no inner product induces the max-norm.

Now the Euclidean norm on the same u, v:
  LHS = (sqrt2)^2 + (sqrt2)^2 = 4 = RHS   ->  law holds.
The parallelogram law is a litmus test: not every norm is an inner-product norm.

Best approximation and Bessel

Expand a vector in an orthonormal set e_1, ..., e_k. The coefficients c_i = <v, e_i> are its Fourier coefficients, and Bessel's inequality says sum |c_i|^2 <= ||v||^2. Geometrically, projecting onto the span of the e_i can only shorten v; the gap is the squared distance to that subspace.

When the orthonormal set is a full basis, Bessel becomes an equality — Parseval's identity, ||v||^2 = sum |c_i|^2. This is the abstract heart of energy conservation in Fourier analysis: the total energy of a signal equals the sum of the energies of its frequency components.