Spectral Theorem for Symmetric Matrices

⚡ Key Takeaways

One-line statement: Every real symmetric matrix can be diagonalized by an orthogonal matrix, with all-real eigenvalues. Key result: If $A = A^T$ , then $A = Q\Lambda Q^T$ where $Q$ is orthogonal and $\Lambda$ is diagonal real. Where it applies: quadratic forms, principal axes, covariance/PCA, vibrations, quantum observables, optimization (Hessians). Who/when: Cauchy (1829) for the reality of eigenvalues; the full orthogonal-diagonalization picture matured through the 19th century.

💡 The Core Insight

A symmetric matrix is exactly a self-adjoint operator with respect to the standard dot product: $\langle Ax, y\rangle = \langle x, Ay\rangle$ . Self-adjointness has two immediate, devastating consequences — eigenvalues cannot be complex, and eigenvectors for distinct eigenvalues are automatically orthogonal. Once you know orthogonality is built into the structure, the only remaining work is to peel off eigenvectors one at a time and show the process exhausts all of $\mathbb{R}^n$ . The result: $A$ acts as a pure scaling by $\lambda_i$ along each of $n$ mutually perpendicular axes.

DefinitionSymmetric Matrix and Orthogonal Matrix

A real matrix $A \in \mathbb{R}^{n\times n}$ is symmetric if it equals its transpose:

1

\quad A = A^T, \qquad \text{i.e.}\quad a_{ij} = a_{ji}\ \ \forall i,j.

Derivation hint: This is the definition; the transpose $A^T$ has entries $(A^T)_{ij} = a_{ji}$ , so equality means each off-diagonal pair matches. A real matrix $Q$ is orthogonal if its columns form an orthonormal basis of $\mathbb{R}^n$ :

2

\quad Q^TQ = QQ^T = I, \qquad \text{so}\quad Q^{-1} = Q^T.

Derivation hint: The $(i,j)$ entry of $Q^TQ$ is the dot product of columns $q_i$ and $q_j$ ; demanding it equal $\delta_{ij}$ is precisely orthonormality. Here $\langle x,y\rangle = x^Ty$ denotes the standard Euclidean inner product, and $\|x\| = \sqrt{x^Tx}$ .

TheoremSpectral Theorem (Real Symmetric Case)

Let $A \in \mathbb{R}^{n\times n}$ with $A = A^T$ . Then: 1. All eigenvalues of $A$ are real: $\lambda_1, \dots, \lambda_n \in \mathbb{R}$ . 2. $\mathbb{R}^n$ has an orthonormal basis $\{q_1, \dots, q_n\}$ of eigenvectors of $A$ . 3. There exists an orthogonal matrix $Q = [\,q_1\ \cdots\ q_n\,]$ and a real diagonal matrix $\Lambda = \operatorname{diag}(\lambda_1,\dots,\lambda_n)$ with

3

\quad A = Q\Lambda Q^T = \sum_{i=1}^{n} \lambda_i\, q_i q_i^T.

Derivation hint: Stack $Aq_i = \lambda_i q_i$ as the matrix equation $AQ = Q\Lambda$ , then right-multiply by $Q^T = Q^{-1}$ ; the rank-one expansion follows from $Q\Lambda Q^T = \sum_i \lambda_i q_i q_i^T$ since $Q\Lambda$ has columns $\lambda_i q_i$ . The matrices $P_i = q_iq_i^T$ are the orthogonal projections onto the eigenlines, satisfying

4

\quad P_i = P_i^T = P_i^2,\qquad P_iP_j = 0\ (i\neq j),\qquad \sum_{i=1}^n P_i = I.

Derivation hint: $P_i^2 = q_i(q_i^Tq_i)q_i^T = q_iq_i^T$ since $q_i^Tq_i=1$ ; $P_iP_j = q_i(q_i^Tq_j)q_j^T = 0$ since $q_i^Tq_j=0$ ; completeness $\sum P_i = QQ^T = I$ .

ProofReality of Eigenvalues

Step 1. Let $\lambda \in \mathbb{C}$ be an eigenvalue with eigenvector $v \in \mathbb{C}^n$ , $v\neq 0$ : $Av = \lambda v$ . Write $\bar{v}$ for the complex conjugate and $v^* = \bar{v}^T$ for the conjugate transpose.

Step 2. Compute $v^*Av$ in two ways. On one hand,

5

\quad v^*Av = v^*(\lambda v) = \lambda\, v^*v = \lambda\,\|v\|^2.

Derivation hint: Substitute the eigenvalue equation $Av = \lambda v$ and factor the scalar $\lambda$ out of the inner product.

Step 3. On the other hand, since $A$ is real and symmetric ( $\bar A = A$ , $A^T = A$ ), $v^*A v$ is a scalar, hence equal to its own conjugate transpose:

6

\quad v^*Av = (v^*Av)^* = v^*A^* v = v^* \bar A^T v = v^*A v \quad\Rightarrow\quad \overline{\lambda\|v\|^2} = \lambda\|v\|^2.

Derivation hint: A $1\times1$ matrix equals its conjugate transpose; use $A^* = \bar A^T = A$ . Taking conjugate transpose of $\lambda\|v\|^2$ gives $\bar\lambda\|v\|^2$ , forcing $\bar\lambda = \lambda$ .

Step 4. Since $\|v\|^2 > 0$ , we get $\bar\lambda = \lambda$ , so $\lambda \in \mathbb{R}$ . A real eigenvalue of a real matrix then has a real eigenvector (the kernel of the real matrix $A - \lambda I$ is a real subspace). $\blacksquare$

∎

ProofOrthogonal Diagonalizability — Induction on Dimension

Step 1 (orthogonality for distinct eigenvalues). Suppose $Av = \lambda v$ , $Aw = \mu w$ with $\lambda \neq \mu$ . Then

7

\quad \lambda\langle v,w\rangle = \langle Av,w\rangle = \langle v, Aw\rangle = \mu\langle v,w\rangle \ \Rightarrow\ (\lambda-\mu)\langle v,w\rangle = 0.

Derivation hint: Self-adjointness $\langle Av,w\rangle = (Av)^Tw = v^TA^Tw = v^TAw = \langle v,Aw\rangle$ uses $A^T = A$ . Since $\lambda\neq\mu$ , conclude $\langle v,w\rangle = 0$ .

Step 2 (base case). For $n=1$ the matrix is already diagonal; take $Q = [1]$ .

Step 3 (inductive step). Assume the result for dimension $n-1$ . By reality, $A$ has a real eigenvalue $\lambda_1$ with a unit eigenvector $q_1$ ( $\|q_1\|=1$ ). Let $W = q_1^\perp = \{x : \langle x, q_1\rangle = 0\}$ , an $(n-1)$ -dimensional subspace.

Step 4 (invariance of $W$ ). $W$ is $A$ -invariant: if $x \in W$ then

8

\quad \langle Ax, q_1\rangle = \langle x, Aq_1\rangle = \langle x, \lambda_1 q_1\rangle = \lambda_1\langle x,q_1\rangle = 0,

Derivation hint: Move $A$ across the inner product via symmetry, then use $Aq_1 = \lambda_1 q_1$ and $x \perp q_1$ . Thus $Ax \in W$ . so $Ax \in W$ . The restriction $A|_W$ is a symmetric operator on the $(n-1)$ -dimensional inner-product space $W$ .

Step 5 (apply induction). By the inductive hypothesis $A|_W$ has an orthonormal basis of eigenvectors $q_2,\dots,q_n$ of $W$ . Adjoining $q_1$ gives an orthonormal basis $\{q_1,\dots,q_n\}$ of $\mathbb{R}^n$ of eigenvectors. Forming $Q = [q_1\cdots q_n]$ gives $AQ = Q\Lambda$ , i.e. $A = Q\Lambda Q^T$ . $\blacksquare$

∎

⚠️ The induction does NOT require distinct eigenvalues

A common misreading is that orthogonality of eigenvectors only follows when eigenvalues are distinct (Step 1). For repeated eigenvalues the eigenspace is multi-dimensional, and orthogonality within it must be imposed by hand (e.g. Gram–Schmidt). The induction above sidesteps this: it never assumes simple eigenvalues — it only ever needs one eigenvector at a time and the invariance of the orthogonal complement.

AC

Augustin-Louis Cauchy

1789–1857

Cauchy, working on the principal axes of quadric surfaces and of moments of inertia, proved in 1829 that the eigenvalues of a symmetric matrix (he worked with the "characteristic equation" of a real symmetric system) are always real. He recognized that the principal-axis problem — finding directions in which a quadratic form decouples into independent squares — was equivalent to diagonalizing a symmetric array. His result removed the apparent paradox that physical, real-valued problems might produce complex resonant frequencies, and it anchored the later development of the spectral theory of operators by Hilbert.

🧠 Intuition

Picture a symmetric matrix as a stretching field. The unit sphere $\{\|x\|=1\}$ is mapped by $A$ to an ellipsoid. The Spectral Theorem says this ellipsoid's principal axes are mutually perpendicular, that they point along the eigenvectors $q_i$ , and that the semi-axis lengths are $|\lambda_i|$ . No shearing, no twisting — pure stretching along orthogonal directions, possibly with a flip ( $\lambda_i < 0$ ). The orthogonal change of basis $Q^T$ rotates your viewpoint so these axes become the coordinate axes; in that frame $A$ is simply $\Lambda$ , and $Ax$ scales coordinate $i$ by $\lambda_i$ . The quadratic form $x^TAx = \sum_i \lambda_i\, y_i^2$ where $y = Q^Tx$ , so symmetry $\Leftrightarrow$ "diagonalizable by rotation" $\Leftrightarrow$ "a quadratic form is a weighted sum of squares of orthogonal coordinates."

Figure

A symmetric matrix maps the unit circle to an ellipse whose axes lie along the orthonormal eigenvectors q₁, q₂, scaled by eigenvalues λ₁, λ₂.

ExampleA 2×2 Symmetric Matrix

Given. Diagonalize $A = \begin{pmatrix} 2 & 1 \\ 1 & 2 \end{pmatrix}$ (symmetric).

Step 1 — eigenvalues. $\det(A-\lambda I) = (2-\lambda)^2 - 1 = \lambda^2 - 4\lambda + 3 = (\lambda-1)(\lambda-3)$ . So $\lambda_1 = 3$ , $\lambda_2 = 1$ (both real ✓).

Step 2 — eigenvectors. For $\lambda_1=3$ : $(A-3I)v = \begin{pmatrix}-1&1\\1&-1\end{pmatrix}v = 0 \Rightarrow v_1 = (1,1)^T$ . For $\lambda_2=1$ : $\begin{pmatrix}1&1\\1&1\end{pmatrix}v=0 \Rightarrow v_2 = (1,-1)^T$ .

Step 3 — check orthogonality. $v_1^Tv_2 = (1)(1)+(1)(-1) = 0$ ✓ (automatic, since $\lambda_1\neq\lambda_2$ ).

Step 4 — normalize. $q_1 = \tfrac{1}{\sqrt2}(1,1)^T$ , $q_2 = \tfrac{1}{\sqrt2}(1,-1)^T$ .

Step 5 — assemble.

9

\quad Q = \frac{1}{\sqrt2}\begin{pmatrix}1&1\\1&-1\end{pmatrix},\quad \Lambda = \begin{pmatrix}3&0\\0&1\end{pmatrix},\quad A = Q\Lambda Q^T.

Derivation hint: Place normalized eigenvectors as columns of $Q$ in the same order as their eigenvalues in $\Lambda$ ; verify $Q^TQ=I$ and recompute $Q\Lambda Q^T$ .

\boxed{A = \tfrac{1}{\sqrt2}\begin{pmatrix}1&1\\1&-1\end{pmatrix}\begin{pmatrix}3&0\\0&1\end{pmatrix}\tfrac{1}{\sqrt2}\begin{pmatrix}1&1\\1&-1\end{pmatrix}}

ExampleRepeated Eigenvalue — Gram–Schmidt Within an Eigenspace

Given. $A = \begin{pmatrix} 2 & 0 & 0\\ 0 & 3 & 1\\ 0 & 1 & 3\end{pmatrix}$ (symmetric).

Step 1 — eigenvalues. Block structure: the lower $2\times2$ block $\begin{pmatrix}3&1\\1&3\end{pmatrix}$ has eigenvalues $4, 2$ ; the top gives $2$ . So eigenvalues are $\lambda = 4$ (simple) and $\lambda = 2$ (multiplicity 2).

Step 2 — eigenspace for $\lambda=4$ . $(A-4I) = \begin{pmatrix}-2&0&0\\0&-1&1\\0&1&-1\end{pmatrix}$ , kernel spanned by $(0,1,1)^T$ . Normalize: $q_1 = \tfrac1{\sqrt2}(0,1,1)^T$ .

Step 3 — eigenspace for $\lambda=2$ . $(A-2I) = \begin{pmatrix}0&0&0\\0&1&1\\0&1&1\end{pmatrix}$ , kernel = solutions of $y+z=0$ , i.e. span of $(1,0,0)^T$ and $(0,1,-1)^T$ .

Step 4 — orthogonalize within the 2D eigenspace. The two vectors $(1,0,0)$ and $(0,1,-1)$ are already orthogonal (dot product $0$ ). Normalize: $q_2 = (1,0,0)^T$ , $q_3 = \tfrac1{\sqrt2}(0,1,-1)^T$ .

Step 5 — result.

10

\quad Q = \begin{pmatrix} 0 & 1 & 0\\ \tfrac1{\sqrt2} & 0 & \tfrac1{\sqrt2}\\ \tfrac1{\sqrt2} & 0 & -\tfrac1{\sqrt2}\end{pmatrix},\qquad \Lambda = \operatorname{diag}(4,2,2).

Derivation hint: Within a repeated eigenspace, choose any orthonormal basis (Gram–Schmidt if needed); eigenvectors from different eigenspaces are auto-orthogonal. Columns of $Q$ pair with $\Lambda$ entries in order.

\boxed{A = Q\,\operatorname{diag}(4,2,2)\,Q^T \quad\text{with orthonormal columns } q_1,q_2,q_3.}

⚠️ Symmetry is essential — diagonalizability alone is not enough

"Every diagonalizable matrix is orthogonally diagonalizable" is false. A non-symmetric matrix may have a full set of eigenvectors that are not orthogonal, so $A = P\Lambda P^{-1}$ with $P$ invertible but $P^{-1}\neq P^T$ . The Spectral Theorem's converse holds tightly: $A$ is orthogonally diagonalizable over $\mathbb{R}$ if and only if $A$ is symmetric. (Indeed if $A = Q\Lambda Q^T$ then $A^T = Q\Lambda^TQ^T = Q\Lambda Q^T = A$ .)

CounterexampleA non-symmetric matrix lacking real spectral structure

Take $B = \begin{pmatrix} 0 & -1\\ 1 & 0\end{pmatrix}$ (rotation by $90^\circ$ ). It is real but not symmetric ( $B^T = -B$ ). Its eigenvalues are $\pm i$ — not real — and it has no real eigenvectors at all. This shows that the reality and orthogonal-diagonalizability conclusions genuinely require symmetry; they fail for general real matrices. A second instructive case: $C = \begin{pmatrix} 1 & 1\\ 0 & 1\end{pmatrix}$ is non-symmetric, has the real repeated eigenvalue $1$ , yet is defective (only a one-dimensional eigenspace), so it is not diagonalizable at all. Symmetry rules out both pathologies simultaneously.

⚠️ Order matters: pair columns with eigenvalues

In $A = Q\Lambda Q^T$ the $i$ -th column of $Q$ must be the eigenvector for the $i$ -th diagonal entry of $\Lambda$ . Reordering eigenvalues without reordering columns yields a wrong matrix. Also, eigenvectors must be normalized ( $\|q_i\|=1$ ); using un-normalized eigenvectors gives $A = P\Lambda P^{-1}$ but $P^T \neq P^{-1}$ , breaking the orthogonal form.

connection

The Spectral Theorem is the finite-dimensional, real prototype of a tower of results. Over $\mathbb{C}$ , the natural class is Hermitian Matrices ( $A = A^*$ ), unitarily diagonalizable with real eigenvalues; the even broader class of Normal Matrices ( $AA^* = A^*A$ ) is exactly the set of unitarily diagonalizable matrices. It underlies the Singular Value Decomposition, since the SVD of any matrix $M$ comes from spectrally decomposing the symmetric matrices $M^TM$ and $MM^T$ . It governs Quadratic Forms through Sylvester's law of inertia and the principal-axis theorem, and it specializes to Positive Definite Matrices (all $\lambda_i > 0$ ). In infinite dimensions it generalizes to the Spectral Theorem for Compact Self-Adjoint Operators and ultimately to the spectral theory of bounded and unbounded operators on Hilbert Spaces. It is also the engine of Principal Component Analysis via the covariance matrix.

🔬 Where this appears

⏳ History

- 1773 — Lagrange. Studies the secular (characteristic) equation in celestial mechanics; encounters symmetric systems whose roots determine planetary perturbations and intuits their reality.

- 1820s — Cauchy. Develops the algebra of quadratic forms and the principal-axis transformation for quadrics and inertia.

- 1829 — Cauchy. Publishes the theorem that the roots of the characteristic equation of a real symmetric matrix are real — the first rigorous proof of reality of eigenvalues.

- 1852 / 1858 — Sylvester. Formulates the law of inertia: the number of positive, negative, and zero eigenvalues of a symmetric matrix is invariant under congruence — coining "matrix" and clarifying signature.

- 1855 — Cayley. Introduces systematic matrix notation, enabling clean statements like

A = Q\Lambda Q^T

.

- 1870s — Weierstrass & Jordan. Develop canonical forms; clarify why symmetry avoids defective (Jordan-block) behavior.

- 1904–1910 — Hilbert. Extends spectral ideas to infinite-dimensional symmetric (bounded) operators and integral equations, coining "spectrum."

- 1929–1932 — von Neumann. Establishes the spectral theorem for unbounded self-adjoint operators on Hilbert space, the rigorous foundation of quantum mechanics.

open-problem

The undergraduate theorem is complete and admits no open questions — but its boundaries are where the subtleties live. (1) Numerical stability: eigenvalue/eigenvector computation for symmetric matrices is far better-conditioned than for general matrices, but eigenvectors of nearly-degenerate eigenvalues are ill-conditioned (the eigenspace is stable, individual vectors are not). (2) Generalized eigenproblems: $Ax = \lambda Bx$ with $A,B$ symmetric and $B$ positive definite still diagonalizes simultaneously, but if $B$ is merely symmetric (indefinite) the clean theory breaks. (3) Infinite dimensions: for unbounded self-adjoint operators the eigenvector expansion must be replaced by a projection-valued measure (continuous spectrum), and not every self-adjoint operator has eigenvectors — the spectral theorem survives only in integral form. (4) Non-real fields: over fields without a notion of orthogonality (e.g. finite fields), "symmetric" loses its spectral force entirely.

Spectral Theorem for Symmetric Matrices

Formal Statement

Why It Holds

Building Intuition

Worked Examples

Common Misconceptions

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also