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

Compact operators: the ones that still behave like matrices

Most operators in infinite dimensions are badly behaved. Compact operators are the precious exception — limits of finite-rank maps that inherit eigenvalue structure, the Fredholm alternative, and a spectral theorem nearly identical to Vol I's.

What goes wrong, and what compactness rescues

A general bounded operator can refuse to have a single eigenvalue (the shift is one such), so the whole eigenvalue/diagonalization machinery of Vol I has no foothold. We need a subclass that keeps enough finite-dimensional rigidity for eigenvalues to exist and behave. That subclass is the compact operators.

A compact operator T maps the (infinite, non-compact) unit ball to a set whose closure *is* compact — equivalently, every bounded sequence (x_n) has a subsequence with (T x_n) convergent. The cleanest mental model: T is a bounded operator that is a limit, in operator norm, of finite-rank operators. Compact operators are "almost finite-dimensional," and that is exactly why matrix intuition returns.

The spectral theorem for compact self-adjoint operators

Here is the payoff that makes compactness worth defining. The spectral theorem for compact self-adjoint operators says: a compact self-adjoint T has an orthonormal system of eigenvectors {e_k} with real eigenvalues {lambda_k}, and T x = sum lambda_k <x, e_k> e_k. This is the Vol I spectral theorem A = Q D Q^T, copied almost verbatim into infinite dimensions — only the finite sum became an infinite one.

Compact self-adjoint T:  the eigenvalue picture

  eigenvalues are real:        lambda_1, lambda_2, lambda_3, ...
  ordered by size:             |lambda_1| >= |lambda_2| >= ... 
  the ONLY accumulation point of {lambda_k} is 0:
        lambda_k -> 0  as k -> infinity.
  each nonzero lambda_k has FINITE-dimensional eigenspace.

Diagonal action (a Fourier-style expansion):
        T x = sum_k  lambda_k <x, e_k> e_k

Compare Vol I:   A x = sum_k  lambda_k <x, q_k> q_k   (finite sum)
Same formula -- the leap to infinity cost us only:
  (a) infinitely many eigenvalues, and
  (b) they must pile up at 0 (an infinite-dim necessity).
Compact self-adjoint operators diagonalize exactly like symmetric matrices — but eigenvalues must accumulate only at 0.

One structural detail is forced by infinite dimensions: the eigenvalues can only accumulate at 0. There can be infinitely many of them, but for any cutoff only finitely many exceed it in size. This is the same accounting that makes the SVD and low-rank approximation of Vol I work — truncating after the largest few eigenvalues gives the best low-rank approximation of T.

The Fredholm alternative: solving (T - lambda I) x = y

Compactness also rescues the existence-uniqueness theory of linear systems. The Fredholm alternative, for a compact T and lambda != 0, restores the exact finite-dimensional dichotomy you knew from linear systems: either (T - lambda I) x = y has a unique solution for every y, or the homogeneous equation (T - lambda I) x = 0 has nontrivial solutions. One or the other — never both, never neither.