The Spectral Theorem for Self-Adjoint Operators

Which linear operators can be diagonalized using an orthonormal basis? This is one where lengths and angles behave exactly as in Euclidean space. This is not the same question as "which operators are diagonalizable." An arbitrary diagonalizable operator may need a badly-skewed basis. That basis could be full of enormous or nearly-parallel vectors. The Spectral Theorem answers the orthonormal question completely on real inner product spaces. The operators that admit an orthonormal eigenbasis are exactly the self-adjoint ones. This matters because self-adjoint operators are everywhere real geometry and real physics touch linear algebra. Quadratic forms, second derivatives (Hessians), moments of inertia, covariance matrices, and the Hamiltonians of quantum mechanics with real symmetric matrices are all self-adjoint. Understanding them reduces entirely to understanding real numbers — their eigenvalues — sitting on orthogonal axes. The real case is subtler than the complex case treated in The Complex Spectral Theorem. This is because over R\mathbb{R} a matrix need not have any eigenvalues at all: rotations have none. Self-adjointness is precisely the extra structure that rules out rotation-like behavior. It forces real eigenvalues to exist. The big picture is a rigidity phenomenon. On a complex inner product space, the Complex Spectral Theorem shows something important. Normality (TT∗=T∗TTT^* = T^*T) is equivalent to the existence of an orthonormal eigenbasis. On a real inner product space, normality is not enough. A rotation matrix is normal but has no real eigenvectors at all. The correct hypothesis strengthens to self-adjointness (T=T∗T = T^*). This article proves both directions of the real theorem from scratch. It states the two results side by side as one unified diagonalization statement. It also shows exactly where the extra hypothesis is used. This article delivers a complete, hypothesis-explicit proof of the real Spectral Theorem. It is built on the machinery of Self-Adjoint Operators together with the complex case. We prove existence of at least one real eigenvalue for self-adjoint operators. This is the crux of the real argument. Then we induct using invariant orthogonal complements. We close with worked numerical examples. We also give a counterexample showing why self-adjointness cannot be weakened to normality in the real setting. Finally, we look at where this theorem resurfaces — most immediately in Positive Operators and The Singular Value Decomposition.

Key takeaways
  • On a finite-dimensional real inner product space, TT has an orthonormal eigenbasis iff TT is self-adjoint (T=T∗T = T^*).
  • On a finite-dimensional complex inner product space, TT has an orthonormal eigenbasis iff TT is normal (TT∗=T∗TTT^*=T^*T); self-adjoint operators are a special case with real eigenvalues.
  • The hard direction in R\mathbb{R} is existence: a self-adjoint operator on a nonzero real space must have a real eigenvalue — this fails for general normal real operators (rotations).
  • Proof strategy: find one eigenvalue (via the complexification/quadratic-form argument), split off its eigenvector, restrict to the orthogonal complement, and induct — an orthogonal-complement induction.
  • In matrix language: AA real symmetric   ⟺  \iff A=QDQTA = Q D Q^{T} for some orthogonal QQ and real diagonal DD; AA complex Hermitian or normal   ⟺  \iff A=UDU∗A = UDU^* for unitary UU.
  • Key idea: self-adjointness is exactly the algebraic condition that converts the geometric wish "diagonalize with an orthonormal basis" into the guaranteed existence of real eigenvalues and orthogonal eigenspaces.

Where this lives today

Quadratic forms and optimization. Every quadratic form q(x)=xTAxq(x) = x^T A x with AA symmetric is diagonalized by the Spectral Theorem into q(y)=λ1y12+⋯+λnyn2q(y) = \lambda_1 y_1^2 + \cdots + \lambda_n y_n^2 in the rotated coordinates y=QTxy = Q^Tx; the signs of the λi\lambda_i classify qq as positive definite, indefinite, etc., which is the entire content of the second-derivative test in multivariable calculus (the Hessian is symmetric) and underlies convex optimization.

Principal Component Analysis and statistics. The covariance matrix of a data set is symmetric positive semidefinite; its orthonormal eigenbasis (principal components) gives the directions of maximal variance, ordered by eigenvalue, foundational to dimensionality reduction and data compression.

Physics. Moments of inertia are encoded by a real symmetric tensor whose orthonormal eigenbasis gives the principal axes of rotation of a rigid body; in quantum mechanics, physical observables are modeled by self-adjoint (Hermitian) operators precisely because the Spectral Theorem guarantees real eigenvalues (measurement outcomes) and an orthonormal eigenbasis (measurement collapse states).

Graph theory and spectral graph theory. The Laplacian matrix of a graph is real symmetric (as seen in Example 1's K3K_3 Laplacian); its eigenvalues and orthonormal eigenvectors encode connectivity, expansion, and clustering structure, central to spectral clustering algorithms.

Numerical linear algebra. Algorithms such as the QR algorithm and Jacobi eigenvalue algorithm exploit symmetry to compute eigendecompositions stably and efficiently, since orthogonal similarity transformations preserve conditioning far better than general similarity transformations.

Onward. The theorem is the direct ancestor of Positive Operators (self-adjoint operators with nonnegative eigenvalues, characterizing ⟨Tv,v⟩≥0\langle Tv,v\rangle \ge 0) and The Singular Value Decomposition (applying the Spectral Theorem to the self-adjoint, positive semidefinite operator T∗TT^*T for an arbitrary — not necessarily self-adjoint — operator TT, to produce singular values and singular vector bases).

Ask the tutor: Can you show me how applying this theorem to T∗TT^*T for a non-self-adjoint TT leads directly to the singular value decomposition?

DefinitionSelf-adjoint operator and orthonormal eigenbasis

Let VV be a finite-dimensional inner product space over F\mathbb{F} (where F=R\mathbb{F} = \mathbb{R} or F=C\mathbb{F} = \mathbb{C}), and let T∈L(V)T \in \mathcal{L}(V) be a linear operator on VV.

Adjoint. The adjoint of TT is the unique operator T∗∈L(V)T^* \in \mathcal{L}(V) satisfying

⟨Tv,w⟩=⟨v,T∗w⟩for all v,w∈V.\langle Tv, w \rangle = \langle v, T^*w \rangle \quad \text{for all } v, w \in V.

(Existence and uniqueness of T∗T^* is standard and assumed from Self-Adjoint Operators.)

Self-adjoint. TT is self-adjoint if T=T∗T = T^*, i.e. ⟨Tv,w⟩=⟨v,Tw⟩\langle Tv, w\rangle = \langle v, Tw\rangle for all v,w∈Vv,w \in V. In matrix terms with respect to an orthonormal basis, this means the matrix of TT equals its conjugate transpose: A=A∗A = A^* (Hermitian when F=C\mathbb{F}=\mathbb{C}; symmetric, A=ATA = A^T, when F=R\mathbb{F}=\mathbb{R}).

Normal. TT is normal if TT∗=T∗TTT^* = T^*T. Every self-adjoint operator is normal (since T=T∗T=T^* trivially commutes with itself), but on real spaces the converse fails.

Orthonormal eigenbasis. TT admits an orthonormal eigenbasis if there is a basis e1,…,ene_1,\dots,e_n of VV that is orthonormal (⟨ei,ej⟩=δij\langle e_i,e_j\rangle = \delta_{ij}) and consists of eigenvectors of TT, i.e. Tei=λieiTe_i = \lambda_i e_i for scalars λi∈F\lambda_i \in \mathbb{F}.

Intuition

Think of a self-adjoint operator as a machine that stretches space along a family of mutually perpendicular directions, with no rotation, shear, or complex twisting mixed in. The symmetric-matrix condition A=ATA = A^T is exactly what prevents "off-diagonal antisymmetric" behavior — the part of a matrix that generates rotation. Decompose any real square matrix as A=S+KA = S + K, its symmetric and antisymmetric parts; the antisymmetric part KK is responsible for rotation-like eigenvalue-free behavior (think of (0−110)\begin{pmatrix}0&-1\\1&0\end{pmatrix}, a quarter turn with no real eigenvectors). Requiring AA to be symmetric (K=0K = 0) removes this obstruction entirely, and the theorem says that is exactly enough to guarantee real eigenvalues and orthogonal eigenspaces. Geometrically, a self-adjoint operator on Rn\mathbb{R}^n sends the unit sphere to an ellipsoid whose axes are orthogonal and whose axis lengths are ∣λi∣|\lambda_i| — this is the picture behind quadratic forms, principal axes of inertia, and (later) singular values. The proof strategy mirrors this picture: find one axis (an eigenvector), note the operator preserves the hyperplane orthogonal to it, and repeat inside that smaller space.

Figure
A self-adjoint operator finds an eigenvector, then acts on the orthogonal complement
x y e₁ (λ₁ = 3) e₂ (λ₂ = −1) Image of unit circle: ellipse with orthogonal axes e₁ ⊥ e₂
TheoremThe Spectral Theorem, unified statement

Let VV be a nonzero finite-dimensional inner product space over F\mathbb{F}, and let T∈L(V)T \in \mathcal{L}(V).

(Complex case, F=C\mathbb{F}=\mathbb{C}, cited from <a class="wiki-link" href="/wiki/math-linalg-complex-spectral-theorem" data-track="wiki_article.body.wiki_link">The Complex Spectral Theorem</a>). VV has an orthonormal basis consisting of eigenvectors of TT if and only if TT is normal, i.e. TT∗=T∗TTT^* = T^*T.

(Real case, F=R\mathbb{F}=\mathbb{R}). VV has an orthonormal basis consisting of eigenvectors of TT if and only if TT is self-adjoint, i.e. T=T∗T = T^*. Equivalently, in matrix form: a real matrix AA is orthogonally diagonalizable (A=QDQTA = QDQ^T, QQ orthogonal, DD real diagonal) if and only if AA is symmetric; a complex matrix AA is unitarily diagonalizable (A=UDU∗A = UDU^*, UU unitary, DD diagonal) if and only if AA is normal.

We prove the real case in full; the complex case is assumed as a prerequisite result, but we will use one direction of its proof technique (the "peel off an eigenvector, restrict to orthogonal complement" induction) explicitly, adapting it to the real setting where the existence of even one eigenvalue must be earned.

LemmaSelf-adjoint operators have real eigenvalues, and eigenvalues exist

Let V≠{0}V \neq \{0\} be a finite-dimensional real inner product space and let T∈L(V)T \in \mathcal{L}(V) be self-adjoint. Then TT has at least one eigenvalue, and every eigenvalue of TT is real (the latter is automatic in R\mathbb{R}, but the content is existence).

Proof

Technique: variational/quadratic-form argument for existence, reduction to the complex case for the "no complex eigenvalues" bookkeeping. Let n=dim⁡V≥1n = \dim V \ge 1. Fix an orthonormal basis of VV and let AA be the matrix of TT in this basis; since T=T∗T = T^*, the matrix AA satisfies AT=AA^T = A (real symmetric). Step 1: AA, viewed as a complex matrix, is Hermitian, hence has nn real eigenvalues (with multiplicity). View AA as an operator on Cn\mathbb{C}^n with the standard Hermitian inner product. Since AA is real symmetric, A∗=A‾T=AT=AA^* = \overline{A}^T = A^T = A, so AA is Hermitian as a complex matrix, in particular normal. By the Complex Spectral Theorem, AA (acting on Cn\mathbb{C}^n) has an orthonormal eigenbasis v1,…,vn∈Cnv_1,\dots,v_n \in \mathbb{C}^n with Avi=λiviAv_i = \lambda_i v_i, λi∈C\lambda_i \in \mathbb{C}. We claim each λi∈R\lambda_i \in \mathbb{R}. Indeed, for any eigenpair (λ,v)(\lambda, v) with v≠0v \neq 0 in Cn\mathbb{C}^n,

λ⟨v,v⟩=⟨Av,v⟩=⟨v,A∗v⟩=⟨v,Av⟩=⟨v,λv⟩=λ‾⟨v,v⟩,\lambda \langle v,v\rangle = \langle Av, v\rangle = \langle v, A^*v\rangle = \langle v, Av\rangle = \langle v, \lambda v \rangle = \overline{\lambda}\langle v,v\rangle,

using A∗=AA^* = A and sesquilinearity (⟨v,λv⟩=λ‾⟨v,v⟩\langle v,\lambda v\rangle = \overline\lambda \langle v,v\rangle in the convention linear in the first slot). Since ⟨v,v⟩>0\langle v,v\rangle > 0, we get λ=λ‾\lambda = \overline\lambda, so λ∈R\lambda \in \mathbb{R}. Step 2: at least one such λi\lambda_i yields a genuine real eigenvector for TT on VV. We now must produce an actual eigenvector inside the real space VV, not merely in Cn\mathbb{C}^n. Take any eigenvalue λ∈R\lambda \in \mathbb{R} of AA found in Step 1 (there is at least one since n≥1n\ge1), with complex eigenvector v=x+iy∈Cnv = x + iy \in \mathbb{C}^n, x,y∈Rnx,y \in \mathbb{R}^n, v≠0v \neq 0. Since AA is a real matrix and λ\lambda is real,

A(x+iy)=λ(x+iy)  ⟹  Ax+iAy=λx+iλy.A(x+iy) = \lambda(x+iy) \implies Ax + iAy = \lambda x + i\lambda y.

Comparing real and imaginary parts (both AA and λ\lambda are real, so this is valid termwise):

Ax=λx,Ay=λy.Ax = \lambda x, \qquad Ay = \lambda y.

Since v=x+iy≠0v = x+iy \neq 0, at least one of x,yx,y is a nonzero real vector. That vector, call it u∈Rnu \in \mathbb{R}^n, satisfies Au=λuAu = \lambda u with u≠0u \neq 0. Thus λ\lambda is a genuine real eigenvalue of AA acting on Rn\mathbb{R}^n, i.e. of TT acting on VV, and uu (identified with the corresponding vector in VV via the fixed orthonormal basis) is a real eigenvector. This proves TT has at least one (real) eigenvalue, completing the lemma. ■\blacksquare

∎
Proof

Proof of the Spectral Theorem, real case. Technique: induction on dim⁡V\dim V, using invariant orthogonal complements — the same architecture used in the complex case, with the Lemma supplying the needed existence step.

(⇐\Leftarrow) Self-adjoint   ⟹  \implies orthonormal eigenbasis exists. We induct on n=dim⁡Vn = \dim V. Base case n=1n=1. Any nonzero vector, normalized, is an orthonormal eigenbasis trivially (every operator on a 11-dimensional space is a scalar multiple of the identity). Inductive step. Suppose n≥2n \ge 2 and the result holds for all self-adjoint operators on real inner product spaces of dimension less than nn. Let T∈L(V)T \in \mathcal{L}(V) be self-adjoint with dim⁡V=n\dim V = n. By the Lemma, TT has a real eigenvalue λ1\lambda_1 with a nonzero eigenvector ww; let e1=w/∥w∥e_1 = w/\|w\|, so Te1=λ1e1Te_1 = \lambda_1 e_1 and ∥e1∥=1\|e_1\| = 1. Let U={v∈V:⟨v,e1⟩=0}U = \{v \in V : \langle v, e_1\rangle = 0\} be the orthogonal complement of span⁡(e1)\operatorname{span}(e_1). We know dim⁡U=n−1\dim U = n - 1 (standard orthogonal complement dimension count in a finite-dimensional inner product space) and V=span⁡(e1)⊕UV = \operatorname{span}(e_1) \oplus U. Claim: UU is invariant under TT, i.e. T(U)⊆UT(U) \subseteq U. Let u∈Uu \in U, so ⟨u,e1⟩=0\langle u, e_1\rangle = 0. Then

⟨Tu,e1⟩=⟨u,T∗e1⟩=⟨u,Te1⟩=⟨u,λ1e1⟩=λ1⟨u,e1⟩=0,\langle Tu, e_1 \rangle = \langle u, T^*e_1\rangle = \langle u, Te_1\rangle = \langle u, \lambda_1 e_1\rangle = \lambda_1 \langle u, e_1\rangle = 0,

using self-adjointness (T∗=TT^* = T) in the first equality. Hence Tu∈UTu \in U, proving the claim. Claim: T∣UT|_U, the restriction of TT to UU, is a self-adjoint operator on UU. First, T∣UT|_U is a well-defined operator on UU by the invariance just shown. For u1,u2∈U⊆Vu_1, u_2 \in U \subseteq V,

⟨T∣Uu1,u2⟩=⟨Tu1,u2⟩=⟨u1,Tu2⟩=⟨u1,T∣Uu2⟩,\langle T|_U u_1, u_2\rangle = \langle Tu_1, u_2 \rangle = \langle u_1, Tu_2\rangle = \langle u_1, T|_U u_2\rangle,

where the middle equality is self-adjointness of TT on VV, and the inner product on UU is simply the restriction of the inner product on VV. This shows T∣UT|_U satisfies the defining property of self-adjointness with respect to the inner product on UU. Since dim⁡U=n−1<n\dim U = n-1 < n, the inductive hypothesis applies to T∣UT|_U on UU: there is an orthonormal basis e2,…,ene_2,\dots,e_n of UU consisting of eigenvectors of T∣UT|_U, hence of TT (as TT agrees with T∣UT|_U on UU). Now consider e1,e2,…,ene_1, e_2, \dots, e_n. Each eie_i (i≥2i\ge2) lies in UU, hence is orthogonal to e1e_1 by definition of UU; the eie_i for i≥2i \ge 2 are mutually orthonormal by construction; and ∥e1∥=1\|e_1\|=1. So e1,…,ene_1,\dots,e_n is an orthonormal set of nn vectors in the nn-dimensional space VV, hence an orthonormal basis (an orthonormal set is automatically linearly independent: if ∑ciei=0\sum c_i e_i = 0, taking the inner product with eje_j gives cj=0c_j = 0 for each jj; a linearly independent set of size n=dim⁡Vn = \dim V is a basis). Each eie_i is an eigenvector of TT: e1e_1 by construction, and e2,…,ene_2,\dots,e_n by the inductive hypothesis applied to T∣U=TT|_U = T on UU. This completes the induction, proving that self-adjoint TT admits an orthonormal eigenbasis.

(⇒\Rightarrow) Orthonormal eigenbasis exists   ⟹  \implies TT is self-adjoint. Suppose e1,…,ene_1,\dots,e_n is an orthonormal basis of VV with Tei=λieiTe_i = \lambda_i e_i, λi∈R\lambda_i \in \mathbb{R}. We must show ⟨Tv,w⟩=⟨v,Tw⟩\langle Tv,w\rangle = \langle v, Tw\rangle for all v,w∈Vv, w \in V; by bilinearity it suffices to check this on basis vectors v=eiv = e_i, w=ejw = e_j. We compute both sides:

⟨Tei,ej⟩=⟨λiei,ej⟩=λi⟨ei,ej⟩=λiδij,\langle Te_i, e_j\rangle = \langle \lambda_i e_i, e_j \rangle = \lambda_i \langle e_i, e_j\rangle = \lambda_i \delta_{ij},
⟨ei,Tej⟩=⟨ei,λjej⟩=λj⟨ei,ej⟩=λjδij.\langle e_i, Te_j\rangle = \langle e_i, \lambda_j e_j\rangle = \lambda_j \langle e_i, e_j\rangle = \lambda_j \delta_{ij}.

Since λi,λj\lambda_i, \lambda_j are real, when i=ji = j both sides equal λi\lambda_i; when i≠ji \neq j both sides equal 00. In either case ⟨Tei,ej⟩=⟨ei,Tej⟩\langle Te_i, e_j\rangle = \langle e_i, Te_j\rangle. By bilinearity this extends to all v,w∈Vv,w \in V, so ⟨Tv,w⟩=⟨v,Tw⟩\langle Tv,w\rangle = \langle v,Tw\rangle for all v,wv,w, which is exactly the statement T=T∗T = T^*, i.e. TT is self-adjoint. Both directions are established, completing the proof of the real Spectral Theorem. The complex statement is The Complex Spectral Theorem, assumed as a prerequisite and invoked only in Step 1 of the Lemma above (to extract real eigenvalues), not in a circular way, since that invocation concerns the complex vector space Cn\mathbb{C}^n and its independently-established theory. ■\blacksquare

∎
CorollaryEigenspace orthogonality and orthogonal diagonalization

Let TT be self-adjoint on a real inner product space VV, with distinct eigenvalues λ1,…,λk\lambda_1,\dots,\lambda_k. Then eigenvectors for distinct eigenvalues are orthogonal: if Tv=λivTv = \lambda_i v and Tw=λjwTw = \lambda_j w with λi≠λj\lambda_i \neq \lambda_j, then ⟨v,w⟩=0\langle v,w\rangle = 0. Consequently V=Eλ1⊕⋯⊕EλkV = E_{\lambda_1} \oplus \cdots \oplus E_{\lambda_k} is an orthogonal direct sum of eigenspaces, and in matrix form A=QDQTA = QDQ^{T} where QQ is orthogonal (its columns are the orthonormal eigenbasis) and D=diag⁡(λ1,…,λn)D = \operatorname{diag}(\lambda_1,\dots,\lambda_n) (with repetitions according to multiplicity).

Proof

Using self-adjointness,

λi⟨v,w⟩=⟨λiv,w⟩=⟨Tv,w⟩=⟨v,Tw⟩=⟨v,λjw⟩=λj⟨v,w⟩,\lambda_i \langle v, w \rangle = \langle \lambda_i v, w\rangle = \langle Tv, w\rangle = \langle v, Tw\rangle = \langle v, \lambda_j w\rangle = \lambda_j \langle v,w\rangle,

so (λi−λj)⟨v,w⟩=0(\lambda_i - \lambda_j)\langle v,w\rangle = 0; since λi≠λj\lambda_i \neq \lambda_j, we conclude ⟨v,w⟩=0\langle v,w\rangle = 0. For the matrix statement: let e1,…,ene_1,\dots,e_n be the orthonormal eigenbasis from the theorem, QQ the matrix with these as columns. Orthonormality of the columns means QTQ=IQ^TQ = I, i.e. QQ is orthogonal (Q−1=QTQ^{-1}=Q^T). Since Tei=λieiTe_i = \lambda_i e_i, we have AQ=QDAQ = QD where D=diag⁡(λ1,…,λn)D = \operatorname{diag}(\lambda_1,\dots,\lambda_n), so A=QDQ−1=QDQTA = QDQ^{-1} = QDQ^T. ■\blacksquare

∎

Worked examples

Example 1: A 3×33\times3 real symmetric matrix with an asymmetric spread of eigenvalues. Let

A=(2−1−1−12−1−1−12).A = \begin{pmatrix} 2 & -1 & -1 \\ -1 & 2 & -1 \\ -1 & -1 & 2 \end{pmatrix}.

This is symmetric (A=ATA = A^T), so the theorem guarantees an orthogonal diagonalization. Compute the characteristic polynomial. Note A=3I−JA = 3I - J where JJ is the all-ones matrix. The matrix JJ has eigenvalues 33 (eigenvector (1,1,1)(1,1,1)) and 00 (multiplicity 22, on the plane orthogonal to (1,1,1)(1,1,1), since JJ has rank 11 and trace 33). Hence A=3I−JA = 3I - J has eigenvalues

3−3=0(eigenvector (1,1,1)),3−0=3(multiplicity 2, on (1,1,1)⊥).3 - 3 = 0 \quad \text{(eigenvector } (1,1,1)\text{)}, \qquad 3 - 0 = 3 \quad \text{(multiplicity 2, on } (1,1,1)^\perp\text{)}.

Check directly: A(1,1,1)T=(2−1−1, −1+2−1, −1−1+2)T=(0,0,0)TA(1,1,1)^T = (2-1-1,\,-1+2-1,\,-1-1+2)^T = (0,0,0)^T. ✓. For the eigenspace of λ=3\lambda = 3: solve (A−3I)v=0(A-3I)v = 0, i.e.

(−1−1−1−1−1−1−1−1−1)v=0  ⟺  v1+v2+v3=0,\begin{pmatrix} -1&-1&-1\\-1&-1&-1\\-1&-1&-1\end{pmatrix}v = 0 \iff v_1+v_2+v_3=0,

a 22-dimensional space, namely (1,1,1)⊥(1,1,1)^\perp, exactly as predicted by the Corollary (eigenvectors for λ=0\lambda=0 and λ=3\lambda=3 must be orthogonal, and indeed (1,1,1)⊥{v1+v2+v3=0}(1,1,1) \perp \{v_1+v_2+v_3=0\}). An orthonormal basis of (1,1,1)⊥(1,1,1)^\perp: take u2=12(1,−1,0)u_2 = \tfrac{1}{\sqrt2}(1,-1,0) and u3=16(1,1,−2)u_3 = \tfrac{1}{\sqrt6}(1,1,-2) (obtained via Gram–Schmidt on (1,−1,0)(1,-1,0) and (1,0,−1)(1,0,-1)). One checks ⟨u2,u3⟩=112(1−1+0)=0\langle u_2,u_3\rangle = \tfrac{1}{\sqrt{12}}(1-1+0) = 0. Together with u1=13(1,1,1)u_1 = \tfrac{1}{\sqrt3}(1,1,1), the matrix

Q=(13121613−1216130−26)Q = \begin{pmatrix} \tfrac{1}{\sqrt3} & \tfrac{1}{\sqrt2} & \tfrac{1}{\sqrt6} \\ \tfrac{1}{\sqrt3} & -\tfrac{1}{\sqrt2} & \tfrac{1}{\sqrt6} \\ \tfrac{1}{\sqrt3} & 0 & -\tfrac{2}{\sqrt6} \end{pmatrix}

is orthogonal, and A=Q diag⁡(0,3,3) QTA = Q\,\operatorname{diag}(0,3,3)\,Q^T, exactly as the Corollary predicts. This matrix AA is in fact the graph Laplacian of the triangle graph K3K_3 — a first hint of where symmetric matrices arise combinatorially.

Example 2: A non-symmetric self-adjoint check and explicit failure diagnosis. Let

B=(1231).B = \begin{pmatrix} 1 & 2 \\ 3 & 1\end{pmatrix}.

Here B≠BTB \neq B^T (BT=(1321)≠BB^T = \begin{pmatrix}1&3\\2&1\end{pmatrix} \neq B), so BB is not self-adjoint (w.r.t. the standard inner product) and the theorem does not apply directly — but it is instructive to see the diagonalization exist over R\mathbb{R} anyway (since tr⁡(B)2−4det⁡(B)=4−4(1−6)=4+20=24>0\operatorname{tr}(B)^2 - 4\det(B) = 4 - 4(1-6) = 4+20 = 24 > 0) while failing to be orthogonal. Eigenvalues: λ2−2λ+(1−6)=0  ⟹  λ2−2λ−5=0  ⟹  λ=1±6\lambda^2 - 2\lambda + (1-6) = 0 \implies \lambda^2 - 2\lambda - 5 = 0 \implies \lambda = 1 \pm \sqrt6. For λ1=1+6\lambda_1 = 1+\sqrt6: solve (B−λ1I)v=0(B - \lambda_1 I)v = 0:

(−623−6)v=0  ⟹  v=(2,6)T (up to scale),\begin{pmatrix} -\sqrt6 & 2 \\ 3 & -\sqrt6\end{pmatrix}v = 0 \implies v = (2, \sqrt6)^T \text{ (up to scale)},

checking −6(2)+2(6)=0-\sqrt6(2) + 2(\sqrt6) = 0. ✓ For λ2=1−6\lambda_2 = 1-\sqrt6: similarly v=(2,−6)Tv = (2,-\sqrt6)^T up to scale, checking 6(2)+2(−6)=0\sqrt6(2)+2(-\sqrt6) = 0. ✓ Now test orthogonality: ⟨(2,6),(2,−6)⟩=4−6=−2≠0\langle (2,\sqrt6), (2,-\sqrt6)\rangle = 4 - 6 = -2 \neq 0. The eigenvectors are not orthogonal, exactly as expected since BB is diagonalizable over R\mathbb{R} (real eigenvalues exist here by luck of the discriminant) but is not self-adjoint, so the Corollary's orthogonality conclusion need not — and does not — hold. This shows the theorem's hypothesis is doing real work even when eigenvalues happen to be real: self-adjointness is what forces orthogonality of eigenvectors, not merely diagonalizability.

CounterexampleWhy self-adjointness cannot be weakened to "normal" on real spaces

Consider the rotation matrix

R=(0−110)∈L(R2),R = \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix} \in \mathcal{L}(\mathbb{R}^2),

representing counterclockwise rotation by 90°90°, with the standard inner product on R2\mathbb{R}^2.

RR is normal. We compute RT=(01−10)=−RR^T = \begin{pmatrix}0&1\\-1&0\end{pmatrix} = -R. Then

RRT=R(−R)=−R2=−(−100−1)=(1001),RR^T = R(-R) = -R^2 = -\begin{pmatrix}-1&0\\0&-1\end{pmatrix} = \begin{pmatrix}1&0\\0&1\end{pmatrix},
RTR=(−R)R=−R2=(1001).R^TR = (-R)R = -R^2 = \begin{pmatrix}1&0\\0&1\end{pmatrix}.

So RRT=RTR=IRR^T = R^TR = I; RR is normal (in fact orthogonal).

RR is not self-adjoint. RT=−R≠RR^T = -R \neq R since R≠0R \neq 0.

RR has no real eigenvalues, hence no real eigenvectors, hence no orthonormal eigenbasis of R2\mathbb{R}^2. The characteristic polynomial is det⁡(R−λI)=λ2+1\det(R - \lambda I) = \lambda^2 + 1, whose roots ±i\pm i are not real. Geometrically, RR rotates every nonzero vector by 90°90°, so RvRv is never a real scalar multiple of vv for v≠0v \neq 0: no line through the origin is preserved. This shows that on a real inner product space, normality alone (the hypothesis that suffices in the complex Spectral Theorem) is not enough to guarantee an orthonormal eigenbasis — the conclusion of the theorem genuinely fails. The extra strength of self-adjointness (T=T∗T=T^*, not merely TT∗=T∗TTT^*=T^*T) is exactly what the Lemma used to force a real eigenvalue to exist in the first place; dropping it collapses the proof at Step 2 of the Lemma, where the real matrix RR, viewed over C\mathbb{C}, does have eigenvalues ±i\pm i, but these are non-real, so the "compare real and imaginary parts" argument produces no real eigenvector at all.

Scope, edge cases & common traps

SituationWhat happensWhy
F=R\mathbb{F} = \mathbb{R}, TT self-adjointOrthonormal eigenbasis exists, all eigenvalues realFull theorem applies
F=R\mathbb{F}=\mathbb{R}, TT normal but not self-adjointMay have no real eigenvalues at allRotation counterexample above
F=C\mathbb{F}=\mathbb{C}, TT normal (includes self-adjoint, unitary, skew-Hermitian)Orthonormal eigenbasis existsThe Complex Spectral Theorem
F=C\mathbb{F}=\mathbb{C}, TT self-adjoint (Hermitian)Orthonormal eigenbasis exists and all eigenvalues realSpecial case of complex theorem plus the real-eigenvalue computation in Step 1 of the Lemma
TT diagonalizable but eigenvectors not orthogonalDiagonalizable ⇏\not\Rightarrow orthogonally diagonalizableExample 2 above: real eigenvalues existed but eigenvectors were skew
dim⁡V=0\dim V = 0Statement is vacuousNo content; excluded by "nonzero" hypothesis for cleanliness

The mathematician behind it

Origins · 1 minAugustin-Louis Cauchy and how this was discovered
AC
Augustin-Louis Cauchy
1789–1857 · Early 19th-century France, Ecole Polytechnique and the rigorization of analysis

Cauchy is best known for founding rigorous analysis — the epsilon-delta definitions of limits, continuity, and convergence that underlie the entire discipline — but his work on quadratic forms and what we would now call symmetric matrices, developed in the study of the principal axes of surfaces defined by quadratic equations (conics and quadrics), anticipates the real Spectral Theorem in geometric form. Cauchy proved, in effect, that quadratic forms in several variables can be brought to a sum-of-squares diagonal form by an orthogonal change of coordinates, work motivated by celestial mechanics and the theory of elasticity, where symmetric matrices arise as moment-of-inertia tensors and stress tensors. The abstract operator formulation used today, with inner product spaces and adjoints, was assembled later by Hilbert, Schmidt, and others extending these ideas to infinite dimensions, but the finite-dimensional real symmetric case traces its geometric heart to Cauchy's principal axis theorem.

History
- 1826 — Cauchy studies principal axes of quadric surfaces, effectively diagonalizing quadratic forms via orthogonal transformations.
- 1829 — Cauchy proves that the eigenvalues of a real symmetric matrix are real, using a variational argument on quadratic forms.
- 1858 — Cayley's memoir on matrix algebra provides the algebraic language (matrix multiplication, matrix polynomials) needed to phrase the theorem cleanly.
- 1904–1910 — Hilbert and Schmidt extend spectral decomposition to self-adjoint operators on infinite-dimensional Hilbert spaces, laying groundwork for spectral theory used in quantum mechanics.
- 1925–1930 — von Neumann formalizes the spectral theorem for unbounded self-adjoint operators, the mathematical backbone of quantum observables.
Why does the complex Spectral Theorem only need "normal" while the real one needs "self-adjoint"?

Normality (TT∗=T∗TTT^*=T^*T) is the correct necessary-and-sufficient condition for a unitary diagonalization whenever the field is algebraically closed, because over C\mathbb{C} every operator has a full set of nn eigenvalues (with multiplicity), and normality is exactly what makes the corresponding generalized eigenvectors resolve into genuine orthogonal eigenvectors. Over R\mathbb{R}, the field is not algebraically closed, so an operator may have no real eigenvalues at all (e.g., a rotation). Self-adjointness is strictly stronger than normality on a real space, and it is precisely strong enough to force the existence of a real eigenvalue in the first place, which is the missing ingredient normality alone cannot supply.

Is every symmetric matrix diagonalizable, and is that the same as this theorem?

Every real symmetric matrix is diagonalizable — this is implied by the Spectral Theorem, since an orthonormal eigenbasis is in particular a basis of eigenvectors. But the Spectral Theorem asserts strictly more: not just that a full eigenbasis exists, but that it can be chosen orthonormal. There exist non-symmetric real matrices that are diagonalizable (Example 2's BB has distinct real eigenvalues, hence is diagonalizable) but whose eigenvectors are not orthogonal. Diagonalizability is a weaker, more common property; orthogonal diagonalizability is special to self-adjoint operators.

Does the theorem require finite dimension?

Yes, as stated here. The proof relies on dimension induction and on the existence of a characteristic polynomial with finitely many roots, both finite-dimensional notions. Infinite-dimensional generalizations exist (spectral theory of self-adjoint operators on Hilbert spaces), but they require substantially more machinery — spectral measures rather than finite eigenbases — and are a separate, more advanced theory built by Hilbert, von Neumann, and others, not a direct corollary of the finite-dimensional statement.

How does this theorem relate to quadratic forms and the second derivative test?

A quadratic form q(x)=xTAxq(x)=x^TAx with AA symmetric corresponds bijectively to a self-adjoint operator. The Spectral Theorem diagonalizes AA via an orthogonal change of variables y=QTxy=Q^Tx, turning qq into λ1y12+⋯+λnyn2\lambda_1y_1^2+\cdots+\lambda_ny_n^2. The signs of the λi\lambda_i determine whether qq is positive definite, negative definite, or indefinite, which is exactly the content of the second-derivative (Hessian) test for local extrema in multivariable calculus, since the Hessian of a smooth function is always symmetric by equality of mixed partials.

What goes wrong in the proof if we only assume TT is diagonalizable, not self-adjoint?

The induction step in the proof crucially uses self-adjointness twice: once to guarantee an eigenvalue exists at all (via the Lemma, which needs T=T∗T=T^* for the reality argument), and again to show the orthogonal complement UU of an eigenvector is TT-invariant, which used ⟨Tu,e1⟩=⟨u,T∗e1⟩=⟨u,Te1⟩\langle Tu,e_1\rangle = \langle u, T^*e_1\rangle = \langle u,Te_1\rangle. If TT is merely diagonalizable without being self-adjoint, the eigenspaces need not be mutually orthogonal (Example 2), so the orthogonal complement of one eigenvector is not guaranteed to be invariant, and the induction cannot proceed.

Why is the complexification trick in the Lemma not circular, given the theorem is about real spaces?

The Lemma invokes the Complex Spectral Theorem applied to the matrix AA reinterpreted as acting on Cn\mathbb{C}^n with the standard Hermitian inner product — a genuinely different, independently established result about complex normal operators. Nothing about the real Spectral Theorem is assumed in that invocation; it is a one-directional import of a prerequisite fact, used only to extract candidate eigenvalues, after which the argument returns entirely to real linear algebra (comparing real and imaginary parts) to build the real eigenvector. This is a legitimate, non-circular reduction.

Does the theorem say anything about the eigenvalues' signs or magnitudes?

Not directly — the theorem guarantees the eigenvalues are real and the eigenbasis is orthonormal, but says nothing about their sign or size in general. Refinements exist: a self-adjoint operator is called positive semidefinite if all eigenvalues are ≥0\ge 0 (equivalently ⟨Tv,v⟩≥0\langle Tv,v\rangle\ge0 for all vv), which is the subject of Positive Operators, and this refinement is exactly what powers constructions like matrix square roots and the Singular Value Decomposition via T∗TT^*T.

Before this

Where this leads