The spectral theorem

Intuition Beginner

A real symmetric matrix can always be diagonalised: rotate the axes to line up with a set of perpendicular "preferred directions," and the matrix becomes a diagonal list of real numbers called eigenvalues. That fact is the finite-dimensional spectral theorem, one of the most useful tools in linear algebra.

The spectral theorem of this unit extends that diagonalisation to spaces of functions and sequences, where the dimension is infinite. The role of "symmetric matrix" is played by a self-adjoint operator, and the finite list of eigenvalues may become a continuous spread of values across the real line.

The payoff is called functional calculus. Once the operator is resolved into its spectral pieces, you can apply any function to it — square roots, exponentials, heat kernels. That machinery is the engine behind quantum mechanics, vibration analysis, and partial differential equations.

Visual Beginner

A self-adjoint operator stretches space along mutually perpendicular axes. In finite dimensions the picture is an ellipsoid aligned with those axes. In infinite dimensions the axes form an orthonormal basis, or — when the spectrum spreads continuously — a continuous family of perpendicular directions indexed by the real line.

The operator's spectrum — the set of allowed eigenvalues together with any continuous spread of values — replaces the finite list of diagonal entries found in the matrix case.

Worked example Beginner

Take the symmetric matrix $A=\left(\begin{array}{cc}5& 2\\ 2& 5\end{array}\right)$ acting on the plane. Its eigenvalues are the values of $\lambda$ that make $A-\lambda I$ non-invertible. Solving $(5-\lambda {)}^2-4=0$ gives $\lambda =3$ and $\lambda =7$.

The value $3$ pairs with the direction $(1,-1)$, and the value $7$ pairs with $(1,1)$. Those two directions are perpendicular. After rotating the coordinate axes to align with them, the operator just stretches by $3$ along one axis and by $7$ along the other, with no mixing between them.

That diagonalisation, written $A=QD{Q}^{\ast }$ with $D$ the diagonal matrix of eigenvalues and $Q$ the rotation matrix whose columns are the perpendicular eigendirections, is the finite-dimensional spectral theorem in action.

An infinite-dimensional self-adjoint operator behaves the same way in spirit: perpendicular directions decompose it into stretching factors. When the operator is compact, the stretching factors form a sequence of eigenvalues tending to zero, and the diagonalisation generalises almost verbatim.

Check your understanding Beginner

Formal definition Intermediate+

Let $\mathcal{H}$ be a complex Hilbert space 02.11.08. A bounded operator $A\in \mathcal{B}(\mathcal{H})$ is self-adjoint when $A=A^{\ast }$, i.e. $⟨Ax,y⟩=⟨x,Ay⟩$ for all $x,y\in \mathcal{H}$ 02.11.01. The spectrum of $A$ is the compact set

$$\sigma (A)=\{\lambda \in \mathbb{C}:A-\lambda I\text{ is not invertible in }\mathcal{B}(\mathcal{H})\}.$$

For self-adjoint $A$, the spectrum is a non-empty compact subset of the real line.

A projection-valued measure (also called a resolution of the identity) on $\mathbb{R}$ is a map $E:{\mathcal{B}}_{\mathbb{R}}\to \mathcal{B}(\mathcal{H})$ from the Borel $\sigma$-algebra of $\mathbb{R}$ to the orthogonal projections on $\mathcal{H}$, satisfying:

1. $E(\varnothing )=0$, $E(\mathbb{R})=I$.
2. $E({\Delta }_1\cap {\Delta }_2)=E({\Delta }_1)E({\Delta }_2)$ for all Borel ${\Delta }_1,{\Delta }_2$.
3. $E\left({\bigcup }_{n=1}^{\infty }{\Delta }_n\right)={\sum }_{n=1}^{\infty }E({\Delta }_n)$ in the strong operator topology, for pairwise disjoint Borel ${\Delta }_n$.

For each $x,y\in \mathcal{H}$, the map $\Delta \mapsto ⟨E(\Delta )x,y⟩$ is a bounded complex Borel measure of total variation at most $\Vert x\Vert \Vert y\Vert$, denoted $d⟨E(\lambda )x,y⟩$. Integration against a bounded Borel function $f:\mathbb{R}\to \mathbb{C}$ is defined by

$$\int f(\lambda )d⟨E(\lambda )x,y⟩=⟨f(A)x,y⟩,$$

where $f(A):=\int fdE$ is the operator defined by the Riesz lemma (the sesquilinear form on the right is bounded). This is the Borel functional calculus.

The three settings of the spectral theorem are:

• (Compact case) $A\in \mathcal{K}(\mathcal{H})$ self-adjoint 02.11.05: orthonormal eigenbasis on $(\ker A{)}^{\perp }$, eigenvalues real, tending to $0$.
• (Bounded case) $A\in \mathcal{B}(\mathcal{H})$ self-adjoint: a projection-valued measure $E_A$ supported on $\sigma (A)\subset \mathbb{R}$ with $A={\int }_{\sigma (A)}\lambda d{E}_A(\lambda )$.
• (Unbounded case) $A:D(A)\subset \mathcal{H}\to \mathcal{H}$ self-adjoint 02.11.03: a projection-valued measure $E_A$ on $\mathbb{R}$ with $A={\int }_{\mathbb{R}}\lambda d{E}_A(\lambda )$ on the natural domain $D(A)=\{x:\int {\lambda }^{2}d⟨E(\lambda )x,x⟩<\infty \}$.

Key theorem with proof Intermediate+

Theorem 1 (compact self-adjoint spectral theorem). Let $T\in \mathcal{K}(\mathcal{H})$ be a compact self-adjoint operator on a Hilbert space $\mathcal{H}$. Then there exists a finite or countably infinite orthonormal set $\{e_n\}$ of eigenvectors of $T$, with corresponding real eigenvalues $\{{\lambda }_n\}$ satisfying ${\lambda }_n\to 0$ whenever the list is infinite, such that

$$Tx=\sum _n{\lambda }_n⟨e_n,x⟩e_n(x\in \mathcal{H}),$$

with convergence in operator norm. The orthogonal complement of $\overline{\mathrm{span}}\{e_n\}$ is $\ker T$. Equivalently, $\mathcal{H}=\overline{\mathrm{span}}\{e_n\}\oplus \ker T$, and $T$ acts as zero on $\ker T$ and as multiplication by ${\lambda }_n$ on each coordinate $⟨e_n,x⟩$.

Proof. If $T=0$ the claim is vacuous (take the empty list). Assume $T\ne 0$. We use the eigenvalue-existence lemma (proved in full in the Full proof set below): a non-zero compact self-adjoint operator has a non-zero real eigenvalue ${\lambda }_1$ with $|{\lambda }_1|=\Vert T\Vert$, witnessed by a unit eigenvector $e_1$.

Decompose $\mathcal{H}=\mathbb{C}{e}_1\oplus e_1^{\perp }$. The subspace $e_1^{\perp }$ is $T$-invariant: if $y\perp e_1$ then $⟨Ty,e_1⟩=⟨y,T{e}_1⟩={\lambda }_1⟨y,e_1⟩=0$, using self-adjointness. The restriction $T_1:=T{|}_{e_1^{\perp }}$ is again compact and self-adjoint, with $\Vert T_1\Vert \le \Vert T\Vert$.

Iterate: if $T_1\ne 0$, apply the lemma to obtain a unit eigenvector $e_2\perp e_1$ with eigenvalue ${\lambda }_2$ satisfying $|{\lambda }_2|=\Vert T_1\Vert \le |{\lambda }_1|$. Continue. Either the process terminates at step $N$ (in which case $T_N=0$, so $T$ has finite rank and the list $\{e_1,\dots ,e_N\}$ captures all of $T$ on $(\ker T{)}^{\perp }$), or it produces an infinite orthonormal sequence $\{e_n\}$ with $|{\lambda }_n|$ monotone non-increasing.

In the infinite case, $|{\lambda }_n|\to 0$: compactness of $T$ implies $\{T{e}_n\}=\{{\lambda }_n{e}_n\}$ has a convergent subsequence, but

$$\Vert {\lambda }_n{e}_n-{\lambda }_m{e}_m{\Vert }^2=|{\lambda }_n{|}^2+|{\lambda }_m{|}^2$$

for $n\ne m$ (orthonormality of $\{e_n\}$), so convergence of a subsequence forces $|{\lambda }_n|\to 0$.

Define $T^{\prime }x:={\sum }_n{\lambda }_n⟨e_n,x⟩e_n$. The partial sums of $T^{\prime }$ converge in operator norm (since $|{\lambda }_n|\to 0$ and the tail is bounded by ${\sup }_{n>N}|{\lambda }_n|$), so $T^{\prime }$ is compact self-adjoint with the same eigenvectors and eigenvalues as $T$ on $\overline{\mathrm{span}}\{e_n\}$. The difference $S:=T-T^{\prime }$ is compact self-adjoint, vanishes on each $e_n$, and hence descends to $(\overline{\mathrm{span}}\{e_n\}{)}^{\perp }$.

If $S\ne 0$, the eigenvalue-existence lemma supplies a unit eigenvector $e_{\ast }\perp \{e_n\}$ with $S{e}_{\ast }=\mu e_{\ast }$ and $|\mu |=\Vert S\Vert >0$. But then $T{e}_{\ast }=T^{\prime }{e}_{\ast }+S{e}_{\ast }=\mu e_{\ast }$ (since $e_{\ast }\perp \{e_n\}$ means $T^{\prime }{e}_{\ast }=0$), contradicting the maximality of the construction. So $S=0$ and $T=T^{\prime }$. Finally, on $(\overline{\mathrm{span}}\{e_n\}{)}^{\perp }$ the operator $T$ vanishes, so that complement is $\ker T$. $\square$

Theorem 2 (bounded self-adjoint spectral theorem; statement, sketched). Let $A\in \mathcal{B}(\mathcal{H})$ be self-adjoint. There exists a unique projection-valued measure $E_A$ on the Borel $\sigma$-algebra of $\mathbb{R}$, supported on $\sigma (A)$, such that

$$A={\int }_{\sigma (A)}\lambda d{E}_A(\lambda ).$$

The map $f\mapsto f(A)=\int fd{E}_A$ is an isometric $\ast$-homomorphism from the bounded Borel functions on $\sigma (A)$ into $\mathcal{B}(\mathcal{H})$, extending the continuous functional calculus $C(\sigma (A))\to C^{\ast }(A)\subset \mathcal{B}(\mathcal{H})$ ^{[Reed-Simon §VII]}.

The construction proceeds in three stages. First, the continuous functional calculus: the commutative unital C*-algebra $C^{\ast }(A,I)$ generated by $A$ is isometrically $\ast$-isomorphic to $C(\sigma (A))$ via the Gelfand transform, with $A$ corresponding to the identity function $\lambda \mapsto \lambda$. Second, for each pair $x,y\in \mathcal{H}$, the functional $f\mapsto ⟨f(A)x,y⟩$ is a bounded linear functional on $C(\sigma (A))$ of norm at most $\Vert x\Vert \Vert y\Vert$; by the Riesz-Markov representation theorem, it is represented by a unique complex Borel measure ${\mu }_{x,y}$ of total variation at most $\Vert x\Vert \Vert y\Vert$. Third, the measures $\{{\mu }_{x,y}\}$ are assembled into the projection-valued measure $E_A$ via $⟨E_A(\Delta )x,y⟩={\mu }_{x,y}(\Delta )$, and the calculus extends from continuous to bounded Borel functions. Uniqueness of $E_A$ follows from the uniqueness of each ${\mu }_{x,y}$ on continuous test functions, plus a monotone-class argument that promotes equality on $C(\sigma (A))$ to equality on all Borel sets.

Theorem 3 (unbounded self-adjoint spectral theorem; statement). Let $A:D(A)\subset \mathcal{H}\to \mathcal{H}$ be a self-adjoint operator in the sense of 02.11.03 (densely defined, closed, $A=A^{\ast }$ as operators including domains). There exists a unique projection-valued measure $E_A$ on $\mathbb{R}$ such that

$$Ax={\int }_{\mathbb{R}}\lambda d{E}_A(\lambda )x,D(A)=\left\{x\in \mathcal{H}:{\int }_{\mathbb{R}}{\lambda }^{2}d⟨E_A(\lambda )x,x⟩<\infty \right\}.$$

The map $f\mapsto \int fd{E}_A$ extends to unbounded Borel functions, defining operators on the maximal domains where the relevant quadratic integral is finite ^{[Reed-Simon §VIII]}. The proof and the role of the resolution of the identity $E_A$ are developed in 02.11.03; this unit states the result and connects it to the bounded setting via the Cayley transform $U=(A-i)(A+i{)}^{-1}$, which is unitary and reduces the unbounded theorem to the bounded spectral theorem for unitary operators.

Bridge. The spectral theorem builds toward the rigorous mathematical framework of quantum mechanics 12.02.02, where every observable is a self-adjoint operator resolved into its projection-valued measure, and appears again in the Dirac operator 03.09.08 and elliptic operators 03.09.09, whose spectra encode geometric and topological data of the underlying manifold. This is exactly the mechanism by which the Laplacian's eigenvalues reveal the shape of a drum or a Riemannian manifold, and the central insight linking functional analysis to both physics and geometry: unbounded self-adjoint operators 02.11.03 become tractable only through their spectral resolutions. Putting these together, the spectral theorem is the diagonalisation engine by which infinite-dimensional operator theory inherits the clarity of finite-dimensional linear algebra 01.01.13, and the bridge is that every self-adjoint operator is, up to a unitary change of variables, a multiplication operator.

Exercises Intermediate+

Lean formalization Intermediate+

lean_status: none — Mathlib provides the continuous functional calculus for bounded self-adjoint and normal operators (via ContinuousMap and the Gelfand transform on the commutative C*-subalgebra CStarAlgebra.adjoin), and the spectral theorem for finite-dimensional Hermitian endomorphisms via eigenvector decomposition. What is absent in packaged curriculum form is the infinite-dimensional content of this unit:

• Projection-valued measures as countably-additive Boolean-algebra homomorphisms from the Borel $\sigma$-algebra of $\sigma (A)$ to orthogonal projections, with their integration theory.
• The equivalence between the projection-valued measure form and the multiplication-operator form (unitary equivalence to multiplication by $\lambda$ on some $L^2(\sigma (A),\mu )$).
• The Borel functional calculus (extending the continuous calculus to bounded Borel functions, not just continuous ones).
• The unbounded self-adjoint spectral resolution, including the domain characterisation via square-integrability against the resolution of the identity.

Each of these requires new Mathlib infrastructure: a type of projection-valued measures, a Bochner-style integration theory for them, and the Riesz-Markov representation argument that lifts the continuous calculus on $C(\sigma (A))$ to a Borel calculus. The expected formalization would build on MeasureTheory.Measure and ComplexInnerProduct to define SpectralMeasure A : Measure ℝ valued in orthogonal projections, prove uniqueness, and derive the functional calculus as the associated integration operator.

Advanced results Master

Multiplication-operator form. The spectral theorem admits an equivalent and often more concrete statement. A bounded self-adjoint operator $A\in \mathcal{B}(\mathcal{H})$ is unitarily equivalent to a multiplication operator: there exists a measure space $(X,\mu )$, a unitary $U:\mathcal{H}\to L^2(X,\mu )$, and a real-valued essentially bounded function $\phi \in L^{\infty }(X,\mu ;\mathbb{R})$ such that

$$UA{U}^{-1}f=\phi \cdot f(f\in L^2(X,\mu )).$$

The function $\phi$ plays the role of the diagonal of the matrix: the spectrum $\sigma (A)$ is the essential range of $\phi$. For a cyclic vector $x_0\in \mathcal{H}$ (one for which $\{E_A(\Delta )x_0\}$ spans a dense subspace), the measure space can be taken as $\sigma (A)$ with measure ${\mu }_{x_0,x_0}$, and $\phi (\lambda )=\lambda$. The general case decomposes $\mathcal{H}$ into at most countably many cyclic subspaces and takes the disjoint union ^{[Reed-Simon §VII]}.

Continuous functional calculus via Gelfand duality. For self-adjoint $A$, the closed unital subalgebra $C^{\ast }(A,I)\subset \mathcal{B}(\mathcal{H})$ generated by $A$ is a commutative unital C*-algebra. Gelfand duality 39.01.01 identifies commutative C*-algebras with spaces of continuous functions on compact Hausdorff spaces: $C^{\ast }(A,I)\cong C(\sigma (A))$, with $A$ mapping to the identity function ${\mathrm{id}}_{\lambda }$. This is the continuous functional calculus: for $f\in C(\sigma (A))$, the operator $f(A)$ is the inverse image of $f$ under this isomorphism. The map $f\mapsto f(A)$ is an isometric $\ast$-isomorphism preserving composition: $(g\circ f)(A)=g(f(A))$ whenever $f$ maps $\sigma (A)$ into the domain of $g$ 39.01.02.

Spectral measures and the Riesz-Markov lift. The continuous calculus extends to bounded Borel functions by a measure-theoretic lifting. For each pair $x,y\in \mathcal{H}$, the functional $f\mapsto ⟨f(A)x,y⟩$ on $C(\sigma (A))$ is bounded of norm at most $\Vert x\Vert \Vert y\Vert$; the Riesz-Markov representation theorem 02.11.02 supplies a unique regular complex Borel measure ${\mu }_{x,y}$ on $\sigma (A)$ realising it. The family $\{{\mu }_{x,y}\}$ is sesquilinear in $(x,y)$ and satisfies the positivity condition ${\mu }_{x,x}(\sigma (A))=\Vert x{\Vert }^2$, ${\mu }_{x,x}(\Delta )\ge 0$ for Borel $\Delta$. This positivity is what allows the measures to be assembled into a single projection-valued measure $E_A$ via $⟨E_A(\Delta )x,y⟩={\mu }_{x,y}(\Delta )$ ^{[Reed-Simon §VII]}.

Functional calculus as the payoff. The Borel functional calculus turns the self-adjoint operator $A$ into a variable that can be substituted into any Borel function. Three flagship consequences:

• Square roots and absolute values. $|A|=(A^2{)}^{1/2}$ and the polar decomposition $A=U|A|$ are defined via the calculus; this is the input to the Schatten ideal theory and the singular-value decomposition for compact operators 02.11.05.
• Exponentials and groups. For self-adjoint $A$, the operator $e^{itA}$ (defined by the calculus applied to $\lambda \mapsto e^{it\lambda }$) is unitary, and $t\mapsto e^{itA}$ is a strongly continuous one-parameter unitary group. Stone's theorem ^{[Stone 1932]} identifies strongly continuous one-parameter unitary groups with self-adjoint generators: $U(t)=e^{itA}$ for a unique self-adjoint $A$. This is the mathematical statement that self-adjoint operators are the infinitesimal generators of quantum-mechanical time evolution.
• Resolvents and the Laplace transform. The resolvent $(A-z{)}^{-1}$ for $\mathrm{Im}z\ne 0$ is recovered from the spectral measure by $(A-z{)}^{-1}=\int (\lambda -z{)}^{-1}d{E}_A(\lambda )$. This identity underlies scattering theory, the spectral shift function, and the Birman-Krein formula.

Unbounded theorem and quantum mechanics. The unbounded spectral theorem is the rigorous form of the Born rule and the spectral measurement postulate: the possible outcomes of measuring the observable $A$ on a state $\psi$ are the points of $\sigma (A)$, and the probability of finding the outcome in a Borel set $\Delta$ is $⟨E_A(\Delta )\psi ,\psi ⟩$. The expectation value is $⟨A\psi ,\psi ⟩=\int \lambda d⟨E_A(\lambda )\psi ,\psi ⟩$. This is why observables must be self-adjoint: only self-adjoint operators admit projection-valued measures, hence well-defined measurement statistics ^{[von Neumann 1932]}.

Sturm-Liouville and elliptic PDE. A regular Sturm-Liouville operator on $[a,b]$ (such as $-d^2/d{x}^2+V(x)$ with suitable boundary conditions) is self-adjoint on $L^2([a,b])$ and has compact resolvent; hence its spectrum is a discrete list of real eigenvalues tending to $+\infty$, with an orthonormal eigenbasis. This is the content of Sturm-Liouville theory, and it generalises to elliptic self-adjoint operators on compact manifolds 03.09.09: the Laplacian on a compact Riemannian manifold has discrete spectrum $0={\lambda }_0<{\lambda }_1\le {\lambda }_2\le \cdots \to \infty$, and Weyl's law gives the asymptotic counting function $N(\Lambda )\sim {\Lambda }^{n/2}/(2\pi {)}^n\mathrm{Vol}(M)$ ^{[Reed-Simon §XIII]}.

Synthesis. The spectral theorem is the foundational reason that self-adjoint operators — bounded or unbounded — admit a complete structural description as multiplication operators, or equivalently as operator-valued integrals against projection-valued measures. The compact case generalises finite-dimensional diagonalisation 01.01.13 to operators whose eigenvalues accumulate only at zero; the bounded case resolves every self-adjoint operator into a direct integral over its spectrum; and the unbounded case supplies the functional calculus that quantum mechanics 12.02.02 and the theory of elliptic partial differential equations 02.18.03 absolutely require. This is exactly the bridge between algebra and measure: the C*-algebra generated by a self-adjoint operator is isomorphic to continuous functions on its spectrum, the central insight of Gelfand duality 39.01.01. Putting these together, the spectral theorem unifies matrix theory, integral equations, quantum observables, vibration analysis, and elliptic operator theory 03.09.09 into a single diagonalisation framework, and the bridge is that every self-adjoint operator, no matter how abstractly presented, is ultimately a multiplication operator in disguise.

Full proof set Master

Proposition (eigenvalue existence for compact self-adjoint operators). Let $T\in \mathcal{K}(\mathcal{H})$ be self-adjoint with $T\ne 0$. Then either $\Vert T\Vert$ or $-\Vert T\Vert$ is a non-zero eigenvalue of $T$.

Proof. For self-adjoint $T$, the operator norm is computed by the quadratic form:

$$\Vert T\Vert =\underset{\Vert x\Vert =1}{\sup }|⟨Tx,x⟩|.$$

To see this, set $M:={\sup }_{\Vert x\Vert =1}⟨Tx,x⟩$ and $m:={\inf }_{\Vert x\Vert =1}⟨Tx,x⟩$. The bound $\Vert T\Vert \ge \max (|M|,|m|)$ is immediate. For the reverse, use the polarisation-like identity (valid for self-adjoint $T$):

$$4\mathrm{Re}⟨Tx,y⟩=⟨T(x+y),x+y⟩-⟨T(x-y),x-y⟩,$$

which yields, after Cauchy-Schwarz applied to each quadratic form on the right, $|\mathrm{Re}⟨Tx,y⟩|\le \max (|M|,|m|)\cdot \Vert x\Vert \Vert y\Vert$. Replacing $x$ by a phase-rotated $e^{i\theta }x$ controls $|⟨Tx,y⟩|$, and taking $\sup$ over unit $x,y$ gives $\Vert T\Vert \le \max (|M|,|m|)$. Hence $\Vert T\Vert =\max (|M|,|m|)$.

Assume $\Vert T\Vert =M$ (the case $\Vert T\Vert =-m$ is analogous, replacing $T$ by $-T$). Take a sequence $\{x_n\}$ of unit vectors with $⟨T{x}_n,x_n⟩\to M$. Expand:

$$\Vert T{x}_n-M{x}_n{\Vert }^2=\Vert T{x}_n{\Vert }^2-2M⟨T{x}_n,x_n⟩+M^2\Vert x_n{\Vert }^2.$$

Using $\Vert T{x}_n\Vert \le \Vert T\Vert =M$ and $\Vert x_n\Vert =1$, the right side is bounded above by $M^2-2M⟨T{x}_n,x_n⟩+M^2\to 2M^2-2M^2=0$. So $\Vert T{x}_n-M{x}_n\Vert \to 0$.

By compactness of $T$, the sequence $\{T{x}_n\}$ has a convergent subsequence $T{x}_{n_k}\to y\in \mathcal{H}$. Then $M{x}_{n_k}\to y$ as well (since $\Vert T{x}_n-M{x}_n\Vert \to 0$), so $x_{n_k}\to y/M$; call this limit $e$. By continuity of $T$ and the convergence, $Te=\lim T{x}_{n_k}=y=Me$. Finally, $\Vert e\Vert =\lim \Vert x_{n_k}\Vert =1$, and $⟨Te,e⟩=M=\Vert T\Vert \ne 0$, so $e\ne 0$. Hence $M$ is a non-zero eigenvalue with unit eigenvector $e$. $\square$

Proposition (uniqueness of the projection-valued measure). Let $A\in \mathcal{B}(\mathcal{H})$ be self-adjoint. If $E$ and $E^{\prime }$ are two projection-valued measures on $\sigma (A)$ both satisfying $A=\int \lambda dE(\lambda )=\int \lambda d{E}^{\prime }(\lambda )$, then $E=E^{\prime }$.

Proof. By hypothesis, the two measures give the same operator when integrated against the identity function $\lambda \mapsto \lambda$. The polynomial approximation argument (Weierstrass) shows they also agree on every polynomial in $\lambda$, and by norm-limits, on every continuous function $f\in C(\sigma (A))$. Hence for every $x,y\in \mathcal{H}$ and every $f\in C(\sigma (A))$,

$$\int fd⟨E(\lambda )x,y⟩=\int fd⟨E^{\prime }(\lambda )x,y⟩.$$

Two finite regular Borel measures that integrate every continuous function identically are equal (Riesz-Markov uniqueness). So $⟨E(\Delta )x,y⟩=⟨E^{\prime }(\Delta )x,y⟩$ for every Borel $\Delta$ and every $x,y$, giving $E(\Delta )=E^{\prime }(\Delta )$ as operators. $\square$

Proposition (spectral mapping for the functional calculus). Let $A$ be self-adjoint and $f:\sigma (A)\to \mathbb{C}$ be a bounded Borel function. Then $\sigma (f(A))\subset \overline{f(\sigma (A))}$, with equality when $f$ is continuous.

Proof. For continuous $f$: if $\zeta \notin f(\sigma (A))$, then $g(\lambda )=(\zeta -f(\lambda ){)}^{-1}$ is a well-defined continuous function on $\sigma (A)$, and $g(A)$ is a two-sided inverse of $\zeta I-f(A)$ by the multiplicativity of the continuous functional calculus. So $\zeta \notin \sigma (f(A))$, giving $\sigma (f(A))\subset f(\sigma (A))$. Conversely, if $\zeta =f({\lambda }_0)$ for some ${\lambda }_0\in \sigma (A)$, take a sequence of unit vectors $x_n$ "concentrated near ${\lambda }_0$" via the spectral measure: $x_n:=E_A(({\lambda }_0-1/n,{\lambda }_0+1/n))y/\Vert E_A(\cdots )y\Vert$ for suitable $y$ with $E_A(\{{\lambda }_0\})=0$ (or an eigenvector when $E_A(\{{\lambda }_0\})\ne 0$). Then $\Vert (\zeta I-f(A))x_n\Vert \to 0$ by continuity of $f$ at ${\lambda }_0$, so $\zeta I-f(A)$ is not bounded below, hence not invertible: $\zeta \in \sigma (f(A))$. For bounded Borel $f$, only the inclusion survives: $\sigma (f(A))\subset \overline{f(\sigma (A))}$, since the essential range of $f$ replaces the pointwise image. $\square$

Connections Master

• Bounded linear operators 02.11.01 — the spectral theorem is a statement about elements of $\mathcal{B}(\mathcal{H})$ (in the bounded case); the operator norm, spectrum, and adjoint operation are the inputs that the theorem reorganises into a diagonal form.

• Unbounded self-adjoint operators 02.11.03 — the unbounded case of the spectral theorem lives here; the resolution of the identity $E_A$ and the natural domain $D(A)$ defined by square-integrability are the load-bearing output for quantum mechanics and differential operators.

• Hahn-Banach theorem 02.11.02 — the Riesz-Markov representation theorem, used to construct the spectral measure from the continuous functional calculus, is a direct descendant of the Hahn-Banach extension machinery in its measure-theoretic form.

• Compact operators 02.11.05 — the compact self-adjoint case is the cleanest entry point to the spectral theorem, producing an honest eigenvector basis; the Riesz-Schauder theory of compact operator spectra is the finite-rank-limit backbone of the general theorem.

• Hilbert space 02.11.08 — the ambient structure; the inner product, the adjoint, and the notion of orthonormal basis all live here, and without the Hilbert structure (in mere Banach spaces) the spectral theorem fails.

• Spectral theorem for normal operators on a finite-dimensional inner product space 01.01.13 — the finite-dimensional prototype; this unit is the infinite-dimensional generalisation that handles continuous spectrum and unbounded operators.

• Dirac operator 03.09.08 and elliptic operators 03.09.09 — geometric operators whose spectra (eigenvalues of the Laplacian, Dirac spectrum) encode topological and geometric data of the underlying manifold; the spectral theorem is what makes "eigenvalue counting" rigorous.

• C*-algebras 39.01.01 and Gelfand duality 39.01.02 — the continuous functional calculus is a direct application of Gelfand duality for commutative C*-algebras; the spectral theorem is the Hilbert-space-concrete realisation of the abstract duality.

Historical & philosophical context Master

The spectral theorem has its origin in David Hilbert's work on integral equations, communicated in five notes between 1904 and 1906 and collected in his 1912 book Grundzuge einer allgemeinen Theorie der linearen Integralgleichungen. Hilbert studied what are now called Hilbert-Schmidt integral operators with symmetric kernels and discovered that they admit a complete orthonormal eigenbasis with real eigenvalues — the first form of the spectral theorem, in the compact self-adjoint case. His "spectral decomposition" of bounded quadratic forms in infinitely many variables was the seed from which the entire theory of operator algebras grew ^{[Hilbert 1912]}.

Hellinger and Toeplitz, in their monumental 1927 Enzyklopadie article Integralgleichungen und Gleichungen mit unendlichvielen Unbekannten, systematised and extended Hilbert's theory to handle operators with continuous spectrum. They introduced the spectral measure viewpoint, in which the operator is resolved into a continuous family of perpendicular projections indexed by the real line, rather than a discrete list of eigenvectors. This was the conceptual leap that made the bounded spectral theorem possible as a statement about general self-adjoint operators, not just compact ones ^{[Hellinger-Toeplitz 1927]}.

The definitive treatment of the unbounded case is due to John von Neumann and Marshall Stone, working independently in the early 1930s. Von Neumann's 1932 Mathematische Grundlagen der Quantenmechanik gave the spectral theorem for unbounded self-adjoint operators its modern form, with the resolution of the identity as the central object, and used it to put the newly-discovered matrix mechanics of Heisenberg and the wave mechanics of Schrodinger on a common rigorous footing. Stone's 1932 Linear Transformations in Hilbert Space developed the same theory from the angle of one-parameter unitary groups (Stone's theorem, identifying strongly continuous unitary groups with self-adjoint generators) ^{[Stone 1932]}. The equivalence of the von Neumann and Stone approaches — via the Cayley transform relating self-adjoint operators to unitary operators — is now the standard pedagogical route.

Philosophically, the spectral theorem marks the moment when infinite-dimensional analysis ceased to be a collection of clever tricks for specific integral and differential equations and became a unified structural theory. The theorem says that self-adjointness is the exact condition under which an operator admits a "diagonal form," and that this diagonal form — whether a discrete eigenbasis or a continuous spectral measure — is unique. It is the conceptual hinge on which quantum mechanics, the theory of elliptic PDE, and the modern theory of operator algebras all turn.

Bibliography Master

@book{reedSimon1980,
  author    = {Reed, Michael and Simon, Barry},
  title     = {Methods of Modern Mathematical Physics, {V}ol.~{I}: Functional Analysis},
  publisher = {Academic Press},
  address   = {New York},
  year      = {1980},
  note      = {Revised and enlarged edition; \S\S VII--VIII cover the spectral theorem}
}

@book{kreyszig1978,
  author    = {Kreyszig, Erwin},
  title     = {Introductory Functional Analysis with Applications},
  publisher = {John Wiley \& Sons},
  address   = {New York},
  year      = {1978},
  note      = {\S 9.4--9.5 give the compact self-adjoint eigenbasis theorem at intermediate level}
}

@book{kato1976,
  author    = {Kato, Tosio},
  title     = {Perturbation Theory for Linear Operators},
  edition   = {2nd},
  publisher = {Springer-Verlag},
  address   = {Berlin},
  year      = {1976},
  note      = {\S VI covers spectral representation and functional calculus of self-adjoint operators}
}

@book{dunfordSchwartz1963,
  author    = {Dunford, Nelson and Schwartz, Jacob T.},
  title     = {Linear Operators, {P}art {II}: Spectral Theory},
  publisher = {Interscience Publishers},
  address   = {New York},
  year      = {1963},
  note      = {\S\S X--XII develop the general spectral theory of bounded and unbounded operators}
}

@book{hilbert1912,
  author    = {Hilbert, David},
  title     = {Grundz{\"u}ge einer allgemeinen {T}heorie der linearen {I}ntegralgleichungen},
  publisher = {B.~G.~Teubner},
  address   = {Leipzig},
  year      = {1912},
  note      = {Expansion of the 1904--1906 communications; founding text of spectral theory}
}

@incollection{hellingerToeplitz1927,
  author    = {Hellinger, Ernst and Toeplitz, Otto},
  title     = {Integralgleichungen und Gleichungen mit unendlichvielen Unbekannten},
  booktitle = {Encyklop{\"a}die der mathematischen Wissenschaften},
  series    = {Band {II}, Teil {C}, Heft 13},
  publisher = {B.~G.~Teubner},
  address   = {Leipzig},
  year      = {1927}
}

@book{vonneumann1932,
  author    = {von Neumann, John},
  title     = {Mathematische {G}rundlagen der {Q}uantenmechanik},
  publisher = {Springer-Verlag},
  address   = {Berlin},
  year      = {1932},
  note      = {English translation by R.~T.~Beyer, Princeton University Press, 1955}
}

@book{stone1932,
  author    = {Stone, Marshall Harvey},
  title     = {Linear Transformations in {H}ilbert Space and their Applications to Analysis},
  series    = {American Mathematical Society Colloquium Publications},
  volume    = {15},
  publisher = {American Mathematical Society},
  address   = {New York},
  year      = {1932}
}

---

Missing-unit backfill. Produced as the spectral theorem target cited by the functional-analysis chain (Hahn-Banach 02.11.02, bounded operators 02.11.01, unbounded self-adjoint operators 02.11.03) and by the geometry/QM/PDE spines downstream.