Normal Operators and the Spectral Theorem

The whole theory of diagonalization splits cleanly into two worlds, and normality is the wall between them. On one side sit operators that can be brought to diagonal form by a unitary change of basis — a rotation of orthonormal axes — so that the eigenvectors form a genuine orthonormal frame. On the other side sit everything else: operators that are merely diagonalizable by some skewed, non-orthogonal basis, and operators that are defective and cannot be diagonalized at all. The spectral theorem names the wall exactly. An operator $N$ on a finite-dimensional complex space has an orthonormal eigenbasis if and only if it commutes with its adjoint, $NN^\dagger=N^\dagger N$ . Nothing weaker will do, and nothing stronger is needed. This one algebraic condition is the load-bearing beam under all of quantum mechanics and most of applied linear algebra. Observables are Hermitian ( $N=N^\dagger$ ) with real spectra; symmetries and time evolution are unitary ( $N^\dagger N=I$ ) with spectra on the unit circle; the generators of those symmetries are anti-Hermitian with imaginary spectra. All three are special cases of normal, differing only in where their eigenvalues sit in the complex plane, and all three inherit the same prize: a complete orthonormal set of eigenvectors that serves as a measurement basis, a mode basis, or a coordinate frame. Once the theorem is in hand, the spectral decomposition $N=\sum_k\lambda_k P_k$ resolves the operator into orthogonal projectors, and the functional calculus $f(N)=\sum_k f(\lambda_k)P_k$ takes square roots, logarithms, and exponentials of operators by acting on eigenvalues alone. The finite-dimensional statement here is the prototype for the Hilbert-space spectral theorem that von Neumann built quantum theory upon.

Key takeaways

An operator is normal when it commutes with its adjoint, $NN^\dagger=N^\dagger N$ ; equivalently $\lVert Nv\rVert=\lVert N^\dagger v\rVert$ for every vector $v$ .
The spectral theorem: on a finite-dimensional complex space $N$ is normal if and only if it has an orthonormal basis of eigenvectors — i.e. is unitarily diagonalizable, $N=U\Lambda U^\dagger$ .
Hermitian (real spectrum), anti-Hermitian (imaginary spectrum), and unitary (unit-circle spectrum) operators are the three physically central normal families, separated only by where their eigenvalues lie.
Normality gives the spectral decomposition $N=\sum_k\lambda_k P_k$ with orthogonal projectors $P_kP_l=\delta_{kl}P_k$ , $\sum_k P_k=I$ , which powers the functional calculus $f(N)=\sum_k f(\lambda_k)P_k$ .
The key structural fact: $N$ and $N^\dagger$ share eigenvectors with conjugate eigenvalues, which forces eigenvectors of distinct eigenvalues to be orthogonal.
Diagonalizable does not imply normal: a non-normal matrix can have a full eigenbasis, but never an orthonormal one — the worst cases are defective and land in Jordan form.
Key idea: normality ( $NN^\dagger=N^\dagger N$ ) is exactly the condition for an operator to possess an orthonormal eigenbasis; the spectral decomposition $N=\sum_k\lambda_k P_k$ into orthogonal projectors then follows, enabling the functional calculus $f(N)=\sum_k f(\lambda_k)P_k$ .

Where this lives today

The spectral theorem is executed millions of times a second in modern computation and measurement. Every quantum-mechanics simulation diagonalizes a Hermitian Hamiltonian for the energy levels and propagates states by $e^{-iHt/\hbar}=\sum_k e^{-iE_kt/\hbar}P_k$ — the functional calculus made literal; in quantum computing a measurement is a spectral decomposition of the observable, and gate synthesis takes logarithms and square roots of unitary operators through their eigenvalues. Data science leans on the theorem twice: principal component analysis is the spectral decomposition of a real symmetric covariance matrix, and the singular value decomposition — behind recommender systems, image compression, and least-squares fitting — is built by applying the spectral theorem to the Hermitian positive-semidefinite matrices $A^\dagger A$ and $AA^\dagger$ . Sparse variants (Lanczos, Davidson) extract a few extremal eigenpairs of enormous Hermitian matrices, and the normality guarantee is what lets those algorithms trust an orthonormal Krylov basis. An open frontier is non-Hermitian and $\mathcal{PT}$ -symmetric physics, where operators are deliberately non-normal: their eigenvectors coalesce at "exceptional points", the spectral theorem fails, and new sensor and laser designs exploit that failure.

The physical picture

Read the adjoint $N^\dagger$ as the mirror image of $N$ across the inner product — the operator that moves $N$ from one slot of $\langle\cdot,\cdot\rangle$ to the other. Normality says $N$ and its mirror image commute: they never fight over directions, so they agree on a single common frame of axes, and those shared axes are precisely the orthonormal eigenvectors the spectral theorem promises. Geometrically a normal operator only stretches and rotates space along mutually perpendicular directions; it never shears. Hermitian operators pure-stretch along real eigenvalues, unitary operators pure-rotate with unit-modulus eigenvalues, and a general normal operator does both — but always along an orthogonal frame. The first figure shows where each family's eigenvalues may live in the complex plane, the cleanest way to see they are one genus with three species.

Figure

Where the spectrum lives: every normal operator has eigenvalues somewhere in the complex plane, and the three physical families are carved out by location — Hermitian on the real axis (λ real), anti-Hermitian on the imaginary axis (λ imaginary), unitary on the unit circle (modulus λ = 1). A general normal operator sits anywhere off the axes.

The second picture is the theorem working on real numbers. Take the Hermitian matrix $A=\begin{pmatrix}2&i\\-i&2\end{pmatrix}$ . Because it equals its own conjugate transpose it is normal, so it must have an orthonormal eigenbasis and a real spectrum. Solving the characteristic equation gives eigenvalues $\lambda=1$ and $\lambda=3$ (real, as promised), with orthonormal eigenvectors along perpendicular axes, and the operator reassembles as $A=1\cdot P_1+3\cdot P_3$ from two orthogonal projectors that sum to the identity — every claim of the spectral theorem visible in one $2\times2$ example.

Figure

The spectral theorem on numbers: the Hermitian matrix A with entries 2, i, −i, 2 has real eigenvalues λ = 1 and λ = 3 and orthonormal eigenvectors, and reassembles as A = 1·P₁ + 3·P₃ with orthogonal projectors summing to the identity. Trace = 1 + 3 = 4 and determinant = 1 × 3 = 3 both read straight off the spectrum.

The core equations

The definition is a single commutator identity, and everything downstream is a consequence of it. An operator is normal when it commutes with its own adjoint; because the adjoint is defined through the inner product, this is a statement about how $N$ sits inside the geometry of the space.

N N^\dagger = N^\dagger N \qquad (1)

where: $N$ — a (bounded) linear operator on a complex inner-product space; $N^\dagger$ — its adjoint, defined by $\langle N u, v\rangle=\langle u, N^\dagger v\rangle$ for all $u,v$ . The commutator definition has an equivalent purely analytic form that is often easier to check and more transparent physically: $N$ and $N^\dagger$ move every vector by the same length. This norm criterion is what makes the eigenvector orthogonality proof go through.

\lVert N v\rVert = \lVert N^\dagger v\rVert \quad \forall v \qquad (2)

where: $\lVert\cdot\rVert$ — the norm induced by the inner product, $\lVert v\rVert^2=\langle v,v\rangle$ . Reading: normality is exactly the condition that $N$ and $N^\dagger$ act with equal length everywhere. The payoff of the theorem is the spectral decomposition: a normal operator is a weighted sum of orthogonal projectors, one per distinct eigenvalue, and those projectors partition the identity. This single line replaces the operator by its spectrum plus a bookkeeping of eigenspaces.

N = \sum_{k=1}^{m} \lambda_k P_k, \qquad P_k P_l = \delta_{kl} P_k, \qquad \sum_k P_k = I \qquad (3)

where: $\lambda_k$ — the $m$ distinct eigenvalues of $N$ ; $P_k$ — the orthogonal projector onto the eigenspace of $\lambda_k$ , self-adjoint ( $P_k=P_k^\dagger$ ) and idempotent ( $P_k^2=P_k$ ); $m$ — number of distinct eigenvalues. The same content, written with an explicit change of basis, is the spectral theorem in matrix form. The columns of a single unitary matrix are the orthonormal eigenvectors, and conjugating by it diagonalizes $N$ .

N = U\,\Lambda\,U^\dagger, \qquad U^\dagger U = I,\; \Lambda=\mathrm{diag}(\lambda_1,\dots,\lambda_n) \qquad (4)

where: $U$ — unitary matrix whose columns are orthonormal eigenvectors of $N$ ; $\Lambda$ — diagonal matrix of eigenvalues, each repeated according to its multiplicity. Because the decomposition isolates each eigenvalue on its own orthogonal subspace, any function of the operator is computed by applying that function to the eigenvalues and leaving the projectors alone. This functional calculus is how $\sqrt{N}$ , $\log N$ , and $e^{N}$ acquire meaning.

f(N) = \sum_{k} f(\lambda_k)\, P_k \qquad (5)

where: $f$ — any function defined on the spectrum $\{\lambda_k\}$ ; for $f(x)=x$ this returns $N$ , for $f(x)=1$ it returns $I$ . Reading: functions of a normal operator act on eigenvalues only — the eigenspaces are frozen. The projectors are not extra data; they are polynomials in $N$ itself, recovered by Lagrange interpolation over the distinct eigenvalues. This is what makes the whole construction intrinsic to $N$ .

P_k = \prod_{l\neq k}\frac{N-\lambda_l I}{\lambda_k-\lambda_l} \qquad (6)

where: $\prod_{l\neq k}$ — the product over all other distinct eigenvalues; the polynomial equals $1$ on the $\lambda_k$ -eigenspace and $0$ on every other. Two structural identities underpin the orthogonal eigenbasis: $N$ and its adjoint share every eigenvector with conjugated eigenvalues, so distinct eigenvalues have orthogonal eigenvectors. Trace and determinant then read straight off the spectrum.

N v = \lambda v \;\Longleftrightarrow\; N^\dagger v = \bar\lambda\, v, \qquad \langle u,v\rangle=0 \ \text{ if } \ \lambda\neq\mu \qquad (7)

\operatorname{tr} N = \sum_k m_k\lambda_k, \qquad \det N = \prod_k \lambda_k^{\,m_k} \qquad (8)

where: $\bar\lambda$ — complex conjugate of $\lambda$ ; $u,v$ — eigenvectors for distinct eigenvalues $\lambda,\mu$ ; $m_k$ — multiplicity of $\lambda_k$ (algebraic equals geometric for a normal operator).

Derivation ASpectral theorem via Schur triangularization

The goal is to prove the hard direction of the spectral theorem: that normality alone forces a unitary eigenbasis. The cleanest finite-dimensional route does not touch eigenvectors one at a time; it first triangularizes $N$ by a unitary — which any complex matrix admits by Schur's theorem — and then shows normality is incompatible with any surviving off-diagonal entry, collapsing the triangle to a diagonal. We assume only that $\mathbb{C}$ is algebraically closed and Gram-Schmidt supplies orthonormal bases, and produce a unitary $U$ with $U^\dagger N U=\Lambda$ diagonal.

Step 1. By Schur's theorem there is a unitary $U$ with $U^\dagger N U=T$ upper-triangular, whose diagonal entries are the eigenvalues $\lambda_k$ of $N$ . Why: algebraic closure guarantees at least one eigenvector; peeling it off and orthonormalizing the remainder by Gram-Schmidt recursively builds the unitary that triangularizes any complex matrix, with no hypothesis on $N$ yet.

Step 2. Normality survives unitary conjugation. Compute $TT^\dagger=U^\dagger N U\,U^\dagger N^\dagger U=U^\dagger NN^\dagger U=U^\dagger N^\dagger N U=T^\dagger T$ , so $T$ is a normal upper-triangular matrix. Why: this is where the hypothesis $NN^\dagger=N^\dagger N$ enters — it passes intact to $T$ , and a normal triangular matrix is about to be forced diagonal.

Step 3. Compare the $(1,1)$ entries of $TT^\dagger$ and $T^\dagger T$ . Since $T$ is upper-triangular, $(TT^\dagger)_{11}=\sum_j|T_{1j}|^2=|T_{11}|^2+|T_{12}|^2+\cdots+|T_{1n}|^2$ , while $(T^\dagger T)_{11}=|T_{11}|^2$ . Equality forces $|T_{12}|^2+\cdots+|T_{1n}|^2=0$ . Why: the whole first row of a triangular matrix contributes to one product but not the other, so normality can only hold if that row has no off-diagonal content.

Step 4. Therefore the entire first row above the diagonal vanishes. Repeating the $(j,j)$ comparison for $j=2,3,\dots$ — each time the already-cleared rows drop out — forces every $T_{jl}=0$ for $l\neq j$ . Hence $T=\Lambda$ is diagonal. Why: the argument cascades down the rows, each step killing one row of off-diagonal entries, until nothing off-diagonal remains.

Step 5. Reassemble: $N=U\Lambda U^\dagger$ with $U$ unitary, so the columns of $U$ are an orthonormal eigenbasis of $N$ with eigenvalues $\lambda_k$ . This is the spectral theorem. Why: a unitary similarity to a diagonal matrix is exactly the statement that the eigenvectors can be chosen orthonormal.

Go deeper · Derivation B · 1 minOrthogonal eigenvectors and the converse

Derivation BOrthogonal eigenvectors and the converse

The goal is to close the iff by proving the two remaining facts that make the eigenbasis orthonormal and the condition necessary: first, that a normal operator and its adjoint share eigenvectors with conjugate eigenvalues, so distinct eigenvalues are automatically orthogonal; second, that any operator with an orthonormal eigenbasis is normal, so nothing outside the normal class can be unitarily diagonalized. Together with the previous derivation this establishes: normal if and only if unitarily diagonalizable.

Step 1. For any scalar $\lambda$ , the shifted operator $N-\lambda I$ is normal whenever $N$ is, since it also commutes with its adjoint $N^\dagger-\bar\lambda I$ . Apply the norm criterion (2): $\lVert(N-\lambda I)v\rVert=\lVert(N^\dagger-\bar\lambda I)v\rVert$ . Why: this single identity says $Nv=\lambda v$ holds exactly when $N^\dagger v=\bar\lambda v$ , so $N$ and $N^\dagger$ have the very same eigenvectors, with conjugated eigenvalues — the structural heart of normality.

Step 2. Let $Nv=\lambda v$ and $Nw=\mu w$ with $\lambda\neq\mu$ . Evaluate $\langle v,w\rangle$ two ways using the lemma: $\lambda\langle v,w\rangle=\langle\bar\lambda v,w\rangle=\langle N^\dagger v,w\rangle=\langle v,Nw\rangle=\mu\langle v,w\rangle$ . Why: pushing $N^\dagger$ off the first slot and $N$ onto $w$ turns one inner product into two different scalar multiples of itself, which can only be consistent if that inner product is zero.

Step 3. Subtract to get $(\lambda-\mu)\langle v,w\rangle=0$ ; since $\lambda\neq\mu$ , we conclude $\langle v,w\rangle=0$ . Within a single eigenspace Gram-Schmidt supplies an orthonormal basis, so stacking the eigenspaces yields a full orthonormal eigenbasis. Why: distinct eigenspaces are forced perpendicular, and the only freedom left — inside a degenerate eigenspace — is removed by orthonormalizing there.

Step 4. Converse. Suppose $N=U\Lambda U^\dagger$ with $U$ unitary. Then $NN^\dagger=U\Lambda U^\dagger U\bar\Lambda U^\dagger=U\Lambda\bar\Lambda U^\dagger=U\bar\Lambda\Lambda U^\dagger=N^\dagger N$ , because diagonal matrices commute. Why: an orthonormal eigenbasis makes $N$ and $N^\dagger$ simultaneously diagonal, and diagonal matrices always commute — so a unitary eigenbasis cannot exist without normality.

Go deeper · Derivation C · 1 minSpectral projectors and the functional calculus

Derivation CSpectral projectors and the functional calculus

The goal is to show that the projectors in the spectral decomposition are genuine, computable objects — polynomials in $N$ — and that once they exist, every function of $N$ is fixed by its values on the spectrum. This is what promotes the abstract theorem into a working tool: it is why $e^{-iHt/\hbar}$ , $\sqrt{\rho}$ , and $\log N$ are well-defined and easy to evaluate. We start from $N=\sum_k\lambda_k P_k$ and build both the projectors and the calculus.

Step 1. Powers are trivial because the projectors are orthogonal: $N^p=\big(\sum_k\lambda_kP_k\big)^p=\sum_k\lambda_k^p P_k$ , since every cross term $P_kP_l$ with $k\neq l$ vanishes. Why: orthogonality of the eigenspaces means the operator never mixes them, so raising $N$ to a power just raises each eigenvalue to that power while the frame stands still.

Step 2. Extend by linearity to any polynomial or convergent power series $f(x)=\sum_p c_p x^p$ : $f(N)=\sum_p c_p N^p=\sum_k\big(\sum_p c_p\lambda_k^p\big)P_k=\sum_k f(\lambda_k)P_k$ . On a finite spectrum any two functions agreeing on the $\lambda_k$ give the same operator. Why: the operator only ever "sees" $f$ through the finite list of eigenvalues, so the calculus reduces an operator problem to $m$ scalar evaluations.

Step 3. Construct the projectors explicitly by Lagrange interpolation. Define $P_k=\prod_{l\neq k}(N-\lambda_l I)/(\lambda_k-\lambda_l)$ . On the $\lambda_j$ -eigenspace each factor acts as the scalar $(\lambda_j-\lambda_l)/(\lambda_k-\lambda_l)$ , so the product is $1$ when $j=k$ and $0$ when $j\neq k$ (one factor kills it). Why: this is the interpolating polynomial that equals $1$ at $\lambda_k$ and $0$ at every other eigenvalue, which is exactly the defining action of the projector $P_k$ .

Step 4. Verify closure and self-adjointness. Summing the interpolants gives $\sum_k P_k=I$ (the constant polynomial $1$ ); orthogonality $P_kP_l=\delta_{kl}P_k$ follows from the product structure; and since $N$ normal makes each eigenspace orthogonal to the rest, $P_k=P_k^\dagger$ . Why: these are the three defining properties in (3), now derived rather than assumed, so the spectral decomposition is self-consistent and unique.

Limiting cases

Normal operators form one family; the physically important members are singled out by where their eigenvalues sit in $\mathbb{C}$ , with the extreme case being the loss of normality altogether.

Case	Result	Meaning
Hermitian (self-adjoint)	$N=N^\dagger \Rightarrow \lambda_k\in\mathbb{R}$	Real spectrum on the real axis; quantum observables. Orthonormal eigenbasis is the measurement basis.
Anti-Hermitian	$N=-N^\dagger \Rightarrow \lambda_k\in i\mathbb{R}$	Purely imaginary spectrum; generators of unitary one-parameter groups, e.g. $-iH/\hbar$ .
Unitary	$N^\dagger N=I \Rightarrow \lvert\lambda_k\rvert=1$	Spectrum on the unit circle; preserves norms and inner products — symmetries and time evolution.
Positive (semi)definite	$N=N^\dagger,\ \lambda_k\ge 0$	Real non-negative spectrum; arises as $A^\dagger A$ , density matrices, metric and inertia tensors.
Non-normal	$NN^\dagger\neq N^\dagger N$	No orthonormal eigenbasis. May still be diagonalizable by a skewed basis; at worst defective (Jordan).

Common trapDiagonalizable does not mean normal — a full eigenbasis need not be an orthonormal one.

The single most common error here is to treat "has $n$ independent eigenvectors" and "is normal" as the same statement. They are not. An operator is diagonalizable if some invertible $S$ gives $S^{-1}NS$ diagonal — the eigenvectors are independent but may be badly skewed. An operator is normal only if a unitary $U$ gives $U^\dagger N U$ diagonal — the eigenvectors are orthonormal. The matrix $\begin{pmatrix}1&1\\0&2\end{pmatrix}$ is diagonalizable (eigenvalues $1$ and $2$ ) yet not normal, and its eigenvectors $(1,0)$ and $(1,1)$ are not orthogonal. So the presence of a complete eigenbasis tells you nothing about orthogonality; only the commutator test $NN^\dagger=N^\dagger N$ does. The failure is worst for defective matrices like $\begin{pmatrix}1&1\\0&1\end{pmatrix}$ , which have too few eigenvectors to diagonalize at all and belong to the Jordan-form world.

Worked examples

1 · Full spectral decomposition of a Hermitian matrix (fully worked)

Take $A=\begin{pmatrix}2&i\\-i&2\end{pmatrix}$ . Confirm it is normal, find its spectrum and orthonormal eigenbasis, build the spectral projectors, verify $A=\sum_k\lambda_kP_k$ , and compute $\sqrt{A}$ by the functional calculus.

Step 1. Confirm normality. Here $A^\dagger=\begin{pmatrix}2&i\\-i&2\end{pmatrix}=A$ , so $A$ is Hermitian and hence trivially normal ( $AA^\dagger=A^\dagger A=A^2$ ). Being Hermitian, its eigenvalues must be real — a prediction we can check.

Step 2. Solve the characteristic equation. $\det(A-\lambda I)=(2-\lambda)^2-(i)(-i)=(2-\lambda)^2-1=0$ , so $2-\lambda=\pm1$ and the eigenvalues are $\boldsymbol{\lambda=3}$ and $\boldsymbol{\lambda=1}$ — real, as required. Consistency checks: $\operatorname{tr}A=4=3+1$ and $\det A=3=3\times1$ , matching (8).

Step 3. Find the eigenvectors. For $\lambda=3$ : $(2-3)v_1+iv_2=0\Rightarrow v_1=iv_2$ , giving the normalized $v_3=\tfrac1{\sqrt2}(i,1)$ . For $\lambda=1$ : $(2-1)v_1+iv_2=0\Rightarrow v_1=-iv_2$ , giving $v_1=\tfrac1{\sqrt2}(-i,1)$ . Their overlap is $\langle v_1,v_3\rangle=\overline{(-i)}\,(i)+\overline{(1)}\,(1)=(i)(i)+1=-1+1=\mathbf{0}$ — orthogonal, exactly as the spectral theorem demands.

Step 4. Build the projectors $P_k=v_kv_k^\dagger$ . From $v_3=\tfrac1{\sqrt2}(i,1)$ , $P_3=\tfrac12\begin{pmatrix}1&i\\-i&1\end{pmatrix}$ ; from $v_1=\tfrac1{\sqrt2}(-i,1)$ , $P_1=\tfrac12\begin{pmatrix}1&-i\\i&1\end{pmatrix}$ . Check the resolution of identity: $P_1+P_3=\tfrac12\begin{pmatrix}2&0\\0&2\end{pmatrix}=I$ , and idempotency $P_3^2=P_3$ , orthogonality $P_3P_1=0$ both hold.

Step 5. Verify the spectral decomposition. $3P_3+1P_1=\tfrac12\begin{pmatrix}3+1&3i-i\\-3i+i&3+1\end{pmatrix}=\tfrac12\begin{pmatrix}4&2i\\-2i&4\end{pmatrix}=\begin{pmatrix}2&i\\-i&2\end{pmatrix}=A.$ The operator is reassembled from its spectrum.

Step 6. Functional calculus. $\sqrt{A}=\sqrt3\,P_3+\sqrt1\,P_1=\tfrac12\begin{pmatrix}\sqrt3+1&(\sqrt3-1)i\\-(\sqrt3-1)i&\sqrt3+1\end{pmatrix}\approx\begin{pmatrix}1.366&0.366\,i\\-0.366\,i&1.366\end{pmatrix}.$ Squaring returns $A$ exactly: the $(1,1)$ entry is $1.366^2+(0.366)^2=1.866+0.134=\mathbf{2}$ and the $(1,2)$ entry is $2(1.366)(0.366)i=\mathbf{i}$ . The square root of an operator is just the square root of its eigenvalues.

2 · A non-normal operator has no orthonormal eigenbasis

Take $M=\begin{pmatrix}1&1\\0&2\end{pmatrix}$ . Show it is diagonalizable but not normal, and that its eigenvectors are not orthogonal.

Step 1. Test normality by computing both products. $MM^\dagger=\begin{pmatrix}1&1\\0&2\end{pmatrix}\begin{pmatrix}1&0\\1&2\end{pmatrix}=\begin{pmatrix}2&2\\2&4\end{pmatrix}$ , while $M^\dagger M=\begin{pmatrix}1&0\\1&2\end{pmatrix}\begin{pmatrix}1&1\\0&2\end{pmatrix}=\begin{pmatrix}1&1\\1&5\end{pmatrix}$ . Since $\begin{pmatrix}2&2\\2&4\end{pmatrix}\neq\begin{pmatrix}1&1\\1&5\end{pmatrix}$ , we have $MM^\dagger\neq M^\dagger M$ — not normal.

Step 2. It is nonetheless diagonalizable. The eigenvalues are $1$ and $2$ (triangular, read off the diagonal), which are distinct, so two independent eigenvectors exist: $u_1=(1,0)$ for $\lambda=1$ and $u_2=\tfrac1{\sqrt2}(1,1)$ for $\lambda=2$ .

Step 3. Check orthogonality. $\langle u_1,u_2\rangle=\tfrac1{\sqrt2}(1\cdot1+0\cdot1)=\tfrac1{\sqrt2}\neq0$ . The eigenvectors are not orthogonal, so $M$ can be written $S^{-1}\!\Lambda S$ with an invertible but non-unitary $S$ , and never as $U^\dagger\Lambda U$ — precisely the distinction the spectral theorem draws: a full eigenbasis exists, but the orthonormal one does not, because $M$ fails the normality test.

Physics of the portrait

The spectral theorem is a result a mathematician proves and a physicist cannot live without. Its finite-dimensional form is nineteenth-century linear algebra; its infinite-dimensional form, and its installation as the grammar of quantum mechanics, is the work of one person above all.

Origins · 1 minJohn von Neumann and how this was discovered

John von Neumann

1903–1957 · mathematical foundations of quantum mechanics

John von Neumann took a scatter of nineteenth-century results — Cauchy and Hermite on real eigenvalues, Hilbert on integral operators, Schur on triangularization — and forged the spectral theorem into the backbone of quantum theory. In papers around 1929 and in his 1932 Mathematische Grundlagen der Quantenmechanik, he proved the spectral theorem for unbounded self-adjoint operators on an abstract Hilbert space, replacing the finite sum $N=\sum_k\lambda_kP_k$ by an integral $N=\int\lambda\,\mathrm{d}E(\lambda)$ against a projection-valued measure. This was exactly what quantum mechanics needed: position and momentum are unbounded with continuous spectra, and Born's rule, state collapse, and the split between observables (self-adjoint) and symmetries (unitary) all rest on it. He also settled the delicate gap between merely symmetric and genuinely self-adjoint operators — invisible in finite dimensions but decisive for whether an observable even has a well-defined spectrum. He went on to found operator algebras, game theory, and the stored-program computer, but it is the spectral theorem that made him the person who wrote quantum mechanics in its final mathematical language.

History

1829

— Cauchy proves the eigenvalues of a real symmetric matrix are all real

1855

— Hermite extends the real-eigenvalue result to complex Hermitian forms

1909

— Schur proves every complex matrix is unitarily triangularizable

1918

— Toeplitz singles out the normal matrices satisfying NN† = N†N

1929

— von Neumann casts the spectral theorem for unbounded self-adjoint operators in Hilbert space

Going further

Texts. Stone & Goldbart, Mathematics for Physics, ch. on linear operators and the spectral theorem · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., ch. 5–6 · Hassani, Mathematical Physics, ch. 4–6 (operator theory and spectral decomposition) · Byron & Fuller, Mathematics of Classical and Quantum Physics, ch. 5 · Nakahara, Geometry, Topology and Physics, §1.3 (linear operators) · Halmos, Finite-Dimensional Vector Spaces, §§78–84 · Reed & Simon, Methods of Modern Mathematical Physics I, ch. VII (the Hilbert-space spectral theorem).

Provenance. All results are standard linear algebra and are recomputable. A verifier should confirm: (1) the Schur-triangularization argument that normality forces zero off-diagonal entries, via the row-norm comparison, against Stone & Goldbart and Halmos §78–79; (2) the adjoint-eigenvalue lemma $Nv=\lambda v\Leftrightarrow N^\dagger v=\bar\lambda v$ and the resulting eigenspace orthogonality against Arfken–Weber ch. 6 and Hassani ch. 4; (3) the functional-calculus and Lagrange-projector identities by direct substitution using $P_kP_l=\delta_{kl}P_k$ ; (4) the worked $2\times2$ example $A=\begin{pmatrix}2&i\\-i&2\end{pmatrix}$ with spectrum $\{1,3\}$ , projectors $P_{1,3}=\tfrac12\begin{pmatrix}1&\mp i\\\pm i&1\end{pmatrix}$ , and $\sqrt A=\sqrt3P_3+P_1$ . The spectral classifications (real / imaginary / unit-modulus eigenvalues) follow directly from $N=N^\dagger$ , $N=-N^\dagger$ , $N^\dagger N=I$ .

Prerequisites. Adjoint and Hermitian Operators · Inner-Product Spaces and Norms · Eigenvalues, Eigenvectors, and the Characteristic Equation · Diagonalization and the Spectral Decomposition · Orthogonality and the Gram-Schmidt Process. Leads to. Jordan Canonical Form and Defective Matrices · Change of Basis and Similarity Transformations · Determinants and Their Properties.

Ask the tutor: Ask me to walk through why $NN^\dagger=N^\dagger N$ — and nothing weaker — is exactly the price of an orthonormal eigenbasis, or to diagonalize a normal operator of your choice step by step.

What is the difference between a normal, a Hermitian, and a unitary operator?

Normal is the umbrella; Hermitian and unitary are two special cases under it. An operator is normal if it commutes with its adjoint, $NN^\dagger=N^\dagger N$ — that is the only requirement, and it is exactly what guarantees an orthonormal eigenbasis. It is Hermitian (self-adjoint) if additionally $N=N^\dagger$ , which pins its eigenvalues to the real axis. It is unitary if instead $N^\dagger N=I$ , which pins its eigenvalues to the unit circle. Both special cases inherit the orthonormal eigenbasis from normality; they differ only in where the spectrum lands. Anti-Hermitian operators ( $N=-N^\dagger$ , imaginary spectrum) are a third such family. So every Hermitian or unitary operator is normal, but most normal operators are neither.

Why does normality, and not just diagonalizability, matter physically?

Because physics needs the eigenvectors to be orthonormal, not merely independent. In quantum mechanics the eigenvectors of an observable are the mutually exclusive measurement outcomes, and mutual exclusivity is orthogonality — a state in one eigenspace must have zero overlap with another, or Born's probabilities would not sum to one. Only a normal (here Hermitian) operator supplies that orthonormal set. A merely diagonalizable observable with skewed eigenvectors would give overlapping, non-orthogonal "outcomes" with no consistent probability interpretation. The same need appears in normal-mode analysis, where orthogonality of modes is what decouples the equations of motion, and in numerical linear algebra, where an orthonormal eigenbasis is numerically stable while a nearly-parallel one is catastrophically ill-conditioned.

How is the functional calculus

f(N)=\sum_k f(\lambda_k)P_k

actually used?

It is how every function of an operator is computed in practice. To evolve a quantum state you need $e^{-iHt/\hbar}$ ; diagonalize the Hermitian $H=\sum_kE_kP_k$ once and then $e^{-iHt/\hbar}=\sum_k e^{-iE_kt/\hbar}P_k$ for all times, turning an operator exponential into scalar exponentials. Thermal physics needs $e^{-\beta H}$ for the partition function by the same recipe. Quantum information needs $\sqrt{\rho}$ and $\log\rho$ (for fidelity and entropy), each obtained by applying $\sqrt{\cdot}$ or $\log$ to the eigenvalues of the density matrix. The rule is always the same: diagonalize, transform the eigenvalues, leave the projectors alone. It works precisely because the projectors are orthogonal, so the eigenspaces never mix under any function.

Does the spectral theorem survive in infinite dimensions?

The idea survives, but the machinery gets subtler and the sum usually becomes an integral. For a bounded self-adjoint or normal operator on a Hilbert space, $N=\int_\sigma\lambda\,\mathrm{d}E(\lambda)$ , where $E$ is a projection-valued measure — the continuous analogue of the $P_k$ . The point spectrum (honest eigenvalues) can now be joined by a continuous spectrum, as for the position operator, whose "eigenstates" are not normalizable functions. Unbounded operators such as momentum and energy require care about their domain, and one must distinguish merely symmetric operators from truly self-adjoint ones; only the latter get a spectral theorem. This is exactly the territory von Neumann charted, and it is why the finite-dimensional version is the right first thing to master.

Can a real matrix be normal, and what does it look like?

Yes. Over the reals the adjoint is the transpose, so normal means $NN^{\mathsf T}=N^{\mathsf T}N$ . Real symmetric matrices ( $N=N^{\mathsf T}$ ) are the real Hermitian case, with real eigenvalues and orthonormal eigenvectors — the principal-axis theorem for inertia, stress, and quadratic forms. Real orthogonal matrices ( $N^{\mathsf T}N=I$ ) are the real unitary case: rotations and reflections, eigenvalues on the unit circle. A real rotation by an angle other than $0$ or $\pi$ has complex conjugate eigenvalues $e^{\pm i\theta}$ and no real eigenvectors — normal's way of saying a genuine rotation has no fixed axis in the plane it turns. The clean statement lives over $\mathbb{C}$ , where every normal operator diagonalizes.

What goes wrong for a non-normal operator, and where does it lead?

A non-normal operator cannot be diagonalized by any orthonormal basis. If it still has a full set of (skewed) eigenvectors it is diagonalizable by a non-unitary $S$ , and the trouble is mostly numerical: nearly-parallel eigenvectors make the eigenvalues hypersensitive to perturbations. If it lacks a full eigenbasis it is defective — the shear $\begin{pmatrix}1&1\\0&1\end{pmatrix}$ is the canonical example — and no similarity can diagonalize it. The right normal form there is the Jordan canonical form, which replaces diagonal entries by Jordan blocks encoding the missing eigenvectors. Non-normality is also physically meaningful: it signals dissipation, gain, or open-system dynamics, and in $\mathcal{PT}$ -symmetric optics the coalescence of eigenvectors at exceptional points is engineered on purpose.