Special Classes of Operators
Four families defined purely in terms of $T$ and $T^*$ — self-adjoint, unitary, normal, and projection operators — that account for most operators encountered in practice.
Self-adjoint operators
Self-adjoint operators are the infinite-dimensional analogue of real symmetric (or Hermitian) matrices. Two immediate consequences:
- $\langle Tx, x \rangle \in \mathbb{R}$ for all $x \in H$;
- all eigenvalues of $T$ (if any) are real — proved in Chapter 6.
Unitary operators
Equivalently, $T$ is unitary iff $T$ is a surjective isometry: $\|Tx\| = \|x\|$ for all $x$, and $T$ is onto. Unitary operators are exactly the isomorphisms of Hilbert space that preserve the inner product, $\langle Tx, Ty \rangle = \langle x, y \rangle$, so they are the natural notion of "change of orthonormal basis" or symmetry of $H$.
Normal operators
Both self-adjoint and unitary operators are normal, but normal operators form a strictly larger class. Normality is exactly the condition needed for the spectral theorem: normal operators are precisely those that are "diagonalizable" with respect to a suitable spectral measure. A useful equivalent characterization on a Hilbert space:
Sketch. $\|Tx\|^2 - \|T^*x\|^2 = \langle T^*Tx, x\rangle - \langle TT^*x, x \rangle = \langle (T^*T - TT^*)x, x\rangle$, which vanishes for all $x$ exactly when $T^*T - TT^* = 0$ (using that a self-adjoint operator $A$ with $\langle Ax,x\rangle = 0$ for all $x$ must be $0$).
Orthogonal projections
Every orthogonal projection $P$ corresponds to a closed subspace $M = \operatorname{ran}(P)$, with $P$ acting as the identity on $M$ and as $0$ on $M^{\perp}$, and every closed subspace arises this way. Projections satisfy $0 \le \langle Px, x \rangle \le \|x\|^2$ and $\|P\| \in \{0, 1\}$ for $P \ne 0$.