spectral theorem (self-adjoint)

In Volume I you learned that some matrices diagonalize and some do not. The first triumph of spectral theory is to name a large, natural class that always diagonalizes, and diagonalizes in the nicest possible way: self-adjoint operators. An operator T on a finite-dimensional inner product space is self-adjoint when T = T^*, meaning <Tu, v> = <u, Tv> for all vectors u and v. For a real matrix this says A = A^T (symmetric); for a complex matrix it says A = A^* (Hermitian, the conjugate transpose).

The theorem states: every self-adjoint operator has an orthonormal basis of eigenvectors, and every eigenvalue is real. So you can write the whole space as an orthogonal direct sum of eigenspaces, and the operator just scales each one by a real number. There are no defective directions, no Jordan blocks, no complex eigenvalues hiding — the geometry is fully transparent. The realness of the eigenvalues follows from a one-line computation: if Tv = lambda v with v nonzero, then lambda <v,v> = <Tv,v> = <v,Tv> = conj(lambda) <v,v>, forcing lambda = conj(lambda).

Why does orthogonality come for free here, when a general diagonalizable matrix only gives you a (possibly skewed) eigenbasis? Because eigenvectors for distinct eigenvalues are automatically orthogonal: if Tu = lambda u and Tv = mu v with lambda != mu, then lambda <u,v> = <Tu,v> = <u,Tv> = mu <u,v>, so (lambda - mu)<u,v> = 0 forces <u,v> = 0. Within a repeated eigenvalue you can run Gram-Schmidt to orthonormalize. The payoff is enormous: self-adjoint operators are the matrices of physics (observables, covariance, stress) precisely because their eigenvalues are real and their eigendirections are perpendicular.