Rayleigh quotient

Eigenvalues are usually found by solving the characteristic equation, but for self-adjoint operators there is a far more flexible handle on them: a single number you can compute for any nonzero vector. The Rayleigh quotient of a vector v is R(v) = <Tv, v> / <v, v>. Geometrically it measures how much T stretches v along v's own direction, normalized so that scaling v does not change the answer.

Here is the key fact for self-adjoint T. The Rayleigh quotient is always real, and as v ranges over all nonzero vectors, R(v) ranges over exactly the interval [lambda_min, lambda_max], from the smallest to the largest eigenvalue. The endpoints are achieved precisely at the corresponding eigenvectors: R(v) is maximized at the top eigenvector and minimized at the bottom one. So eigenvalues stop being roots of a polynomial and become the extreme values of an optimization problem.

This variational viewpoint is enormously useful. It turns eigenvalue estimation into maximizing or minimizing a smooth function, which numerical methods can attack directly (power iteration and its refinements live here). It also makes bounds easy: plug in any test vector v and you instantly know lambda_min <= R(v) <= lambda_max, no diagonalization required. And it is the seed of the full min-max theorem, which extends this idea from the two extreme eigenvalues to every eigenvalue in between.