We now study operators with special relationships to their adjoints. Throughout this section, is a finite-dimensional inner product space.
9.4.1 Self-Adjoint Operators
The most important case is when .
Definition 9.4.1 (Self-adjoint operator). An operator is self-adjoint (or Hermitian) if , i.e., for all .
Note. In an orthonormal basis, is self-adjoint iff its matrix is symmetric (real case, ) or Hermitian (complex case, ), by Theorem 9.3.4. Self-adjoint operators need not be invertible: for example, the zero operator is self-adjoint.
Proposition 9.4.2 (Basic properties of self-adjoint operators). Let be self-adjoint.
- is real for all .
- Every eigenvalue of is real.
- Eigenvectors of for distinct eigenvalues are orthogonal.
- If is a -invariant subspace, then is also -invariant.
Proof. (i) .
(ii) If with , then , so . (Over there is nothing to prove; the content is in the complex case.)
(iii) If and with , then is real by (ii), and So , and .
(iv) This is Lemma 9.3.7(i) with . ∎
Note. Property (iv) is what makes self-adjoint operators diagonalizable: once we find an eigenvector, its orthogonal complement is invariant, and we can proceed by induction.
9.4.2 The Spectral Theorem for Real Self-Adjoint Operators
Over , every operator has an eigenvector. Over , the main work is to show that a self-adjoint operator has one. The key is that quadratic polynomials without real roots cannot kill anything.
Lemma 9.4.3 (Positivity lemma). Let be self-adjoint on a finite-dimensional real inner product space. For any real , the operator is invertible.
Proof. For , using , So for . Thus is injective, hence invertible since is finite-dimensional. ∎
Corollary 9.4.4 (Quadratics without real roots). If has no real roots (i.e., ), and is self-adjoint on a finite-dimensional real inner product space, then is invertible.
Proof. Complete the square: .
Since , we have . The operator is self-adjoint ( by Proposition 9.3.3). By the lemma, is invertible. ∎
Theorem 9.4.5 (Spectral theorem (real self-adjoint operators)). Let be a self-adjoint operator on a finite-dimensional real inner product space. Then has an orthonormal basis consisting of eigenvectors of . In particular, is diagonalizable, and its characteristic polynomial splits into linear factors over .
Proof. Step 1: has an eigenvector if . Let and choose any nonzero . The vectors are linearly dependent, so there is a nonzero polynomial with . It is not constant, since a nonzero multiple of is nonzero.
Factor over : with , , and . (Such a factorization exists by induction on . By the Fundamental Theorem of Algebra (Chapter 4), has a root . If is real, then with by the factor theorem (Lemma 4.3.14). Otherwise because has real coefficients, and since , applying the factor theorem twice in gives with , a real quadratic with no real roots. Conjugating all coefficients gives , so , and has real coefficients because has no zero divisors (Lemma 4.3.3).) Since is a ring homomorphism , Each is invertible by the corollary. If every were injective, then would be a composite of injective operators, hence injective; but with . So some has a nonzero kernel, and has an eigenvector with real eigenvalue .
Step 2: orthonormal eigenbasis, by induction on . If , the empty list works. Otherwise, by Step 1 let be an eigenvector with , normalized so . Let . Since , the orthogonal complement is -invariant (Proposition 9.4.2(iv)), and .
Since is invariant under , the restriction is self-adjoint for the restricted inner product (Lemma 9.3.7(ii)). By induction, has an orthonormal basis of eigenvectors of . Then is an orthonormal list of length , hence an orthonormal basis of , consisting of eigenvectors.
Step 3: the characteristic polynomial. In this basis has the diagonal matrix with , so its characteristic polynomial is (Definition 7.4.7). ∎
Note. The spectral theorem says that self-adjoint operators on real inner product spaces are “as nice as possible”: they are diagonalizable, with real eigenvalues, and the eigenspaces are mutually orthogonal (Proposition 9.4.2(iii)). The matrix in an orthonormal eigenbasis is . In matrix terms: for every real symmetric matrix there is a matrix with (its columns are the orthonormal eigenvectors) such that is diagonal.
Skew-Adjoint Operators
The “opposite” of self-adjoint is skew-adjoint.
Definition 9.4.6 (Skew-adjoint operator). An operator is skew-adjoint (or anti-self-adjoint) if , i.e., for all .
Note. In an orthonormal basis, is skew-adjoint iff its matrix satisfies (real case) or (complex case). Such matrices are called skew-symmetric and skew-Hermitian, respectively.
Proposition 9.4.7 (Properties of skew-adjoint operators). Let be skew-adjoint on a finite-dimensional real inner product space.
- for all .
- The only possible eigenvalue of is .
- is invertible.
Proof. (i) Taking in the definition gives . Over , , so .
(ii) If with , then by (i), so .
(iii) If , then ; by (ii), is not an eigenvalue, so . Thus is injective, hence invertible since is finite-dimensional. ∎
Note. Over , a skew-adjoint operator can have no eigenvalues at all: rotation by in , with matrix , is skew-adjoint. Its eigenvalues as a complex matrix are ; Corollary 9.4 below shows that the complex eigenvalues of a real skew-symmetric matrix are always purely imaginary.
9.4.3 Normal Operators and the Complex Spectral Theorem
Over , the spectral theorem extends to a larger class of operators: those that commute with their adjoint.
Definition 9.4.8 (Normal operator). An operator is normal if
Example 9.4.9 (Classes of normal operators).
- Self-adjoint operators () are normal: .
- Skew-adjoint operators () are normal: .
- Operators with are normal: . These are the orthogonal and unitary operators of Section 9.5.
Lemma 9.4.10 (Characterization of normality). An operator is normal if and only if for all .
Proof. By the defining property of the adjoint and , So for all iff for all , where . If is normal, . Conversely, is self-adjoint (Proposition 9.3.3), and we claim that a self-adjoint with for all is zero. For any , using . So . Over this says . Over , replacing by gives . In both cases for all ; taking gives . ∎
Proposition 9.4.11 (Normal operators and eigenspaces). Let be normal.
- If , then .
- Eigenvectors for distinct eigenvalues are orthogonal.
- .
Proof. (i) The operator is normal: its adjoint is , and these two commute since and commute. By the lemma, , so .
(ii) If and with , then by (i) So , giving .
(iii) The inclusion is (i). Since is also normal, with , applying (i) to and the eigenvalue gives the reverse inclusion. ∎
Theorem 9.4.12 (Complex spectral theorem). Let be a normal operator on a finite-dimensional complex inner product space. Then has an orthonormal basis consisting of eigenvectors of . Moreover:
- if is self-adjoint, its eigenvalues are real;
- if is skew-adjoint, its eigenvalues are purely imaginary;
- if (i.e., is unitary, Section 9.5), its eigenvalues lie on the unit circle .
Proof. We induct on ; the case is trivial. Let . Since is algebraically closed, has an eigenvector (Theorem 7.5.1); let be a unit eigenvector with .
By Proposition 9.4.11(i), . So is invariant under both and . By Lemma 9.3.7(i), applied to and to (using ), the subspace is invariant under both and . By Lemma 9.3.7(ii), , so and is normal. Since , by induction has an orthonormal basis of eigenvectors of . Then is an orthonormal eigenbasis of .
For the eigenvalues: if with , then . If , this gives . If , it gives . If , then , so . ∎
Note. The complex spectral theorem is stronger than the real one: every normal operator (not just every self-adjoint one) is diagonalizable in an orthonormal basis. The price is working over with Hermitian inner products. Normality is also necessary: if has an orthonormal eigenbasis, its matrix in that basis is diagonal, has the diagonal matrix , and diagonal matrices commute, so is normal.
Corollary 9.4.13 (Eigenvalues of real symmetric and skew-symmetric matrices). Let be a real matrix, and consider its complex eigenvalues (the roots in of its characteristic polynomial).
- If , all complex eigenvalues of are real.
- If , all complex eigenvalues of are purely imaginary.
- If , all complex eigenvalues of have absolute value .
Proof. Let , , with the standard Hermitian inner product. The standard basis is orthonormal, so has matrix (the entries of are real). Thus is self-adjoint in case (i), skew-adjoint in case (ii), and satisfies in case (iii). The roots of the characteristic polynomial of are the eigenvalues of , and the theorem applies. ∎