Spectral theorem for normal operators on a finite-dim inner-product space (principal-axes theorem)

Intuition Beginner

A matrix tells you how a linear map stretches space. Most maps mix rotation and stretching together, and the rotation hides the simple part. The spectral theorem says that for a certain large class of maps — the normal ones — you can change to a special set of perpendicular axes in which the rotation disappears entirely, and the map becomes pure stretching along each axis.

A symmetric matrix on the plane is the cleanest example. Its level sets are ellipses, and the ellipses always have two perpendicular principal axes 01.01.15. Rotate to those axes and the matrix becomes diagonal, with the eigenvalues sitting on the diagonal as the stretch factors. The principal-axes theorem is the statement that this rotation-to-diagonal works in every dimension, not only on the plane.

The complex version covers more maps. A normal map is one that commutes with its mirror image (the adjoint). The class includes symmetric maps, rotations, and reflections. For every normal map there is an orthonormal basis of eigenvectors, and the map acts on each basis vector by a single scalar. The mixing of rotation and stretching always comes apart along perpendicular eigendirections.

Visual Beginner

The picture shows a real symmetric two-by-two matrix and the rotation that diagonalises it. On the left, a unit circle is mapped to an ellipse by the matrix; the principal axes of the ellipse point along the eigendirections, and the lengths of the semi-axes are the eigenvalues. On the right, an orthonormal basis rotated to match the eigendirections, with the matrix in the new basis written as a diagonal matrix carrying the two eigenvalues.

Two pieces of data are visible. The two eigendirections are perpendicular — that is what the spectral theorem guarantees for a symmetric matrix. The eigenvalues are the stretch factors along those directions, and the diagonal matrix in the rotated basis is the whole map.

Worked example Beginner

Take the two-by-two real symmetric matrix

$$S=\left(\begin{array}{cc}2& 1\\ 1& 2\end{array}\right).$$

Step 1. Compute the characteristic polynomial. Expand $\det (tI-S)=(t-2{)}^2-1=t^2-4t+3=(t-1)(t-3)$. The two eigenvalues are ${\lambda }_1=1$ and ${\lambda }_2=3$.

Step 2. Find an eigenvector for ${\lambda }_1=1$. Solve $(S-I)v=0$, where $S-I$ is the matrix with row entries $(1,1)$ and $(1,1)$. A solution is $v_1=(1,-1)$. Normalise to $u_1=(1/\sqrt{2},-1/\sqrt{2})$.

Step 3. Find an eigenvector for ${\lambda }_2=3$. Solve $(S-3I)v=0$, where $S-3I$ has row entries $(-1,1)$ and $(1,-1)$. A solution is $v_2=(1,1)$. Normalise to $u_2=(1/\sqrt{2},1/\sqrt{2})$.

Step 4. Check that $u_1$ and $u_2$ are perpendicular. The dot product is $u_1\cdot u_2=(1/\sqrt{2})(1/\sqrt{2})+(-1/\sqrt{2})(1/\sqrt{2})=1/2-1/2=0$. The eigenvectors are orthogonal as predicted by the spectral theorem.

Step 5. Assemble the diagonalisation. Form the orthogonal matrix $Q$ whose columns are $u_1$ and $u_2$:

$$Q=\frac{1}{\sqrt{2}}\left(\begin{array}{cc}1& 1\\ -1& 1\end{array}\right),Q^{\top }SQ=\left(\begin{array}{cc}1& 0\\ 0& 3\end{array}\right).$$

You can verify the right-hand equality by direct matrix multiplication. The original symmetric matrix becomes a diagonal matrix in the rotated basis, with the two eigenvalues on the diagonal.

What this tells us: every real symmetric matrix factors as $S=Q\Lambda Q^{\top }$ with $Q$ orthogonal and $\Lambda$ diagonal. The columns of $Q$ are an orthonormal eigenbasis, and the diagonal entries of $\Lambda$ are the eigenvalues. The factorisation is the principal-axes theorem in a single equation.

Check your understanding Beginner

Formal definition Intermediate+

Let $V$ be a finite-dimensional inner-product space over $\mathbb{F}\in \{\mathbb{R},\mathbb{C}\}$ with inner product $⟨\cdot ,\cdot ⟩$ — conjugate-linear in the second argument over $\mathbb{C}$, linear in both arguments over $\mathbb{R}$. Given a linear operator $T:V\to V$, the adjoint $T^{\ast }:V\to V$ is the unique linear operator satisfying $⟨Tv,w⟩=⟨v,T^{\ast }w⟩$ for all $v,w\in V$. Existence and uniqueness follow from the Riesz representation theorem for linear functionals on $V$ together with finite-dimensionality. In any orthonormal basis the matrix of $T^{\ast }$ is the conjugate transpose of the matrix of $T$; over $\mathbb{R}$ this is the ordinary transpose.

The operator $T$ is normal iff $T{T}^{\ast }=T^{\ast }T$. Three load-bearing special cases:

• Self-adjoint (or Hermitian over $\mathbb{C}$, symmetric over $\mathbb{R}$): $T=T^{\ast }$.
• Unitary (over $\mathbb{C}$) or orthogonal (over $\mathbb{R}$): $T{T}^{\ast }=T^{\ast }T=I$, equivalently $T$ preserves the inner product.
• Anti-self-adjoint (or skew-Hermitian / skew-symmetric): $T=-T^{\ast }$.

Each of the three classes is automatically normal because the defining identity implies commutation with $T^{\ast }$. The converse fails: the matrix $T=\mathrm{diag}(1,i)$ is normal but neither self-adjoint, unitary, nor skew-adjoint.

An orthogonal projection $P:V\to V$ is a linear operator satisfying $P^2=P$ and $P=P^{\ast }$. Its range is a subspace $W\subseteq V$, and $P$ is the orthogonal projection onto $W$ (along its orthogonal complement $W^{\perp }$). The spectral theorem produces a family of pairwise-orthogonal orthogonal projections summing to the identity.

The spectrum of $T$, written $\mathrm{spec}(T)$, is the set of eigenvalues, equivalently the roots of the characteristic polynomial ${\chi }_T(t)=\det (tI-T)$. Over $\mathbb{C}$ the spectrum is non-empty (fundamental theorem of algebra); over $\mathbb{R}$ the spectrum may be empty (e.g., the rotation matrix on ${\mathbb{R}}^2$).

Counterexamples to common slips

• The condition $T{T}^{\ast }=T^{\ast }T$ is genuinely weaker than self-adjointness. The diagonal matrix $\mathrm{diag}(1,i,-1,-i)\in M_4(\mathbb{C})$ is normal — every diagonal matrix is — but is not self-adjoint, since its eigenvalues are not real.
• The real-spectral-theorem direction "orthogonally diagonalisable implies symmetric" needs care: if $S=Q\Lambda Q^{\top }$ with $Q$ orthogonal and $\Lambda$ real diagonal, then $S^{\top }=Q{\Lambda }^{\top }{Q}^{\top }=Q\Lambda Q^{\top }=S$. The converse direction "symmetric implies orthogonally diagonalisable" is the substantive content of the theorem and uses the complex case as a black box.
• Real matrices satisfying $T{T}^{\top }=T^{\top }T$ (real-normal) are not in general orthogonally diagonalisable over $\mathbb{R}$. The rotation matrix $R=\left(\begin{array}{cc}0& -1\\ 1& 0\end{array}\right)$ is real-normal (in fact orthogonal) but has no real eigenvalues; the diagonalisation lives over $\mathbb{C}$. The real-normal canonical form replaces complex-conjugate eigenvalue pairs by real two-by-two rotation blocks.

Key theorem with proof Intermediate+

Theorem (complex spectral theorem; Shilov Ch. 10 §10.1 ^{[source pending]}; Halmos §79 ^{[source pending]}). Let $V$ be a finite-dimensional complex inner-product space and $T:V\to V$ a linear operator. The following are equivalent:

(i) $T$ is normal: $T T^ = T^* T$.*

(ii) $V$ has an orthonormal basis of eigenvectors of $T$.

(iii) There is a spectral decomposition $$ T = \sum_{i=1}^k \lambda_i P_i, $$ where ${\lambda }_1,\dots ,{\lambda }_k\in \mathbb{C}$ are the distinct eigenvalues of $T$ and $P_1,\dots ,P_k$ are pairwise-orthogonal orthogonal projections summing to the identity: ${\sum }_i{P}_i=I$ and $P_i{P}_j={\delta }_{ij}{P}_i$.

Proof. The implications $\text{(ii)}\Rightarrow \text{(iii)}\Rightarrow \text{(i)}$ are direct computations; the content is $\text{(i)}\Rightarrow \text{(ii)}$, which is what the five-step argument below establishes.

Step 1 (existence of an eigenvector). Over $\mathbb{C}$ the characteristic polynomial ${\chi }_T(t)\in \mathbb{C}[t]$ has degree $\dim V\ge 1$ and is non-constant, so by the fundamental theorem of algebra it has a root $\lambda \in \mathbb{C}$. The operator $T-\lambda I$ is then singular, hence has a non-zero element $v$ in its kernel: $Tv=\lambda v$. So $T$ has at least one eigenvector. This step is the only place complex coefficients enter; the real spectral theorem reduces to this step via complexification.

Step 2 (same eigenvector for $T$ and $T^{\ast }$, with conjugate eigenvalue). Suppose $Tv=\lambda v$ with $v\ne 0$ and $T$ normal. Then $(T-\lambda I)(T^{\ast }-\bar{\lambda }I)=T{T}^{\ast }-\lambda T^{\ast }-\bar{\lambda }T+|\lambda {|}^{2}I$, and the same expression equals $(T^{\ast }-\bar{\lambda }I)(T-\lambda I)$ since $T{T}^{\ast }=T^{\ast }T$. So $T-\lambda I$ is itself normal. Compute the norm: $$ |(T - \lambda I) v|^2 = \langle (T - \lambda I) v, (T - \lambda I) v \rangle = \langle v, (T - \lambda I)^* (T - \lambda I) v \rangle. $$ By normality $(T-\lambda I{)}^{\ast }(T-\lambda I)=(T-\lambda I)(T-\lambda I{)}^{\ast }$, and the same calculation gives $\Vert (T-\lambda I{)}^{\ast }v{\Vert }^2=\Vert (T-\lambda I)v{\Vert }^2$. Setting $(T-\lambda I)v=0$ then forces $(T^{\ast }-\bar{\lambda }I)v=0$, i.e., $T^{\ast }v=\bar{\lambda }v$. The eigenvector of $T$ at $\lambda$ is also an eigenvector of $T^{\ast }$ at $\bar{\lambda }$.

Step 3 (eigenspaces at distinct eigenvalues are orthogonal). Let $T{v}_1={\lambda }_1{v}_1$ and $T{v}_2={\lambda }_2{v}_2$ with ${\lambda }_1\ne {\lambda }_2$. Using Step 2, $T^{\ast }{v}_2={\bar{\lambda }}_2{v}_2$. Then $$ \lambda_1 \langle v_1, v_2 \rangle = \langle \lambda_1 v_1, v_2 \rangle = \langle T v_1, v_2 \rangle = \langle v_1, T^* v_2 \rangle = \langle v_1, \bar\lambda_2 v_2 \rangle = \lambda_2 \langle v_1, v_2 \rangle. $$ The conjugate-linearity of the inner product in the second slot turned ${\bar{\lambda }}_2$ into ${\lambda }_2$ on the right-hand side. Subtracting gives $({\lambda }_1-{\lambda }_2)⟨v_1,v_2⟩=0$, and since ${\lambda }_1\ne {\lambda }_2$ we conclude $⟨v_1,v_2⟩=0$.

Step 4 (orthogonal complement of an eigenspace is invariant). Let $E_{\lambda }=\ker (T-\lambda I)$ be an eigenspace and $W=E_{\lambda }^{\perp }$ its orthogonal complement. For $w\in W$ and $v\in E_{\lambda }$, compute $$ \langle T w, v \rangle = \langle w, T^* v \rangle = \langle w, \bar\lambda v \rangle = \lambda \langle w, v \rangle = 0, $$ where Step 2 supplied $T^{\ast }v=\bar{\lambda }v$, and conjugate-linearity in the second slot turned $\bar{\lambda }$ into $\lambda$ when pulled out. So $Tw\in W$, meaning $W$ is $T$-invariant. By the same argument applied to $T^{\ast }$, $W$ is also $T^{\ast }$-invariant, so the restriction $T{|}_W$ is again normal as an operator on $W$ with its inherited inner product.

Step 5 (induction on dimension). Argue by induction on $n=\dim V$. The base case $n=1$ is immediate: any unit vector is an orthonormal eigenbasis. For $n>1$, Step 1 supplies an eigenvalue ${\lambda }_1$ and eigenspace $E_{{\lambda }_1}$ of positive dimension. The orthogonal complement $W=E_{{\lambda }_1}^{\perp }$ has $\dim W<n$, the restriction $T{|}_W$ is normal by Step 4, and by inductive hypothesis $W$ admits an orthonormal eigenbasis for $T{|}_W$. Concatenate an orthonormal basis of $E_{{\lambda }_1}$ (taking each eigenvector to be a unit vector and applying Gram-Schmidt within the eigenspace) with the orthonormal eigenbasis of $W$. The result is an orthonormal eigenbasis of $V$ for $T$. This completes the induction and proves (i) $\Rightarrow$ (ii).

For (ii) $\Rightarrow$ (iii): given the orthonormal eigenbasis, group eigenvectors by eigenvalue. Let $P_i$ be the orthogonal projection onto the ${\lambda }_i$-eigenspace $E_{{\lambda }_i}$. Since the eigenspaces are pairwise orthogonal (Step 3) and span $V$ (eigenbasis exhausts $V$), the projections satisfy ${\sum }_i{P}_i=I$ and $P_i{P}_j={\delta }_{ij}{P}_i$. On any basis vector $v\in E_{{\lambda }_j}$, $Tv={\lambda }_{j}v={\sum }_i{\lambda }_i{P}_{i}v$ since $P_{i}v={\delta }_{ij}v$. By linearity $T={\sum }_i{\lambda }_i{P}_i$.

For (iii) $\Rightarrow$ (i): given the spectral decomposition, compute $T^{\ast }={\sum }_i{\bar{\lambda }}_i{P}_i$ using $P_i^{\ast }=P_i$. Then $T{T}^{\ast }={\sum }_{i,j}{\lambda }_i{\bar{\lambda }}_j{P}_i{P}_j={\sum }_i|{\lambda }_i{|}^2{P}_i=T^{\ast }T$, where the cross-terms with $i\ne j$ vanish by $P_i{P}_j=0$. So $T$ is normal. $\square$

Bridge. The complex spectral theorem builds toward every operator-theoretic structural result on a finite-dimensional inner-product space, and it appears again in 01.01.12 (singular value decomposition), where the spectral theorem applied to $T^{\ast }T$ produces the singular values and right singular vectors of an arbitrary linear map. The foundational reason the theorem holds is exactly the normality identity $T{T}^{\ast }=T^{\ast }T$: it forces $T$ and $T^{\ast }$ to share eigenvectors with conjugate eigenvalues (Step 2), and this is what makes the orthogonal complement of an eigenspace $T$-invariant (Step 4) instead of merely $T$-stable in some weaker sense. The central insight is that normality identifies the operator $T$ with a sum ${\sum }_i{\lambda }_i{P}_i$ over its spectrum, with the orthogonal projections $P_i$ resolving the identity. Putting these together, the spectral decomposition generalises to the real symmetric case (real eigenvalues + orthogonal eigenbasis), to the principal-axes theorem for quadratic forms, to the singular value decomposition for arbitrary operators, and is dual to the Schur decomposition that triangularises every operator over $\mathbb{C}$. The bridge is that the cap of the finite-dimensional spectral theorem — a finite sum $T={\sum }_i{\lambda }_i{P}_i$ — generalises to an integral $T={\int }_{\mathrm{spec}(T)}\lambda dE(\lambda )$ against a projection-valued measure $E$ in the unbounded self-adjoint case treated by von Neumann.

Exercises Intermediate+

Lean formalization Intermediate+

Mathlib's Mathlib.Analysis.InnerProductSpace.Spectrum provides the finite-dimensional spectral theorem for self-adjoint operators via LinearMap.IsSymmetric.eigenvectorBasis, which constructs an orthonormal basis of eigenvectors and supplies the diagonalisation. The normal-operator generalisation is partially packaged through IsStarNormal together with the spectral decomposition for elements of a star algebra, but no single theorem unifies (i) the normality condition, (ii) the orthonormal eigenbasis, and (iii) the spectral decomposition $T={\sum }_i{\lambda }_i{P}_i$ with the orthogonal-projection-family identities. The intended formalisation reads schematically:

import Mathlib.Analysis.InnerProductSpace.Adjoint
import Mathlib.Analysis.InnerProductSpace.Spectrum
import Mathlib.Analysis.NormedSpace.Star.Basic

variable {𝕜 V : Type*} [RCLike 𝕜]
  [NormedAddCommGroup V] [InnerProductSpace 𝕜 V] [FiniteDimensional 𝕜 V]
  (T : V →L[𝕜] V)

/-- Complex spectral theorem for normal operators on a finite-dim
inner-product space: a normal operator has an orthonormal basis of
eigenvectors. -/
theorem normal_spectralTheorem (hT : IsStarNormal T) :
    ∃ (n : ℕ) (b : Fin n → V) (λ : Fin n → 𝕜),
      Orthonormal 𝕜 b ∧ (∀ i, T (b i) = λ i • b i) ∧
      Submodule.span 𝕜 (Set.range b) = ⊤ :=
  sorry  -- Step 1 (FTA-driven eigenvector), Steps 2-4 (normality lemmas),
         -- Step 5 (induction on dim)

/-- Spectral decomposition: T = ∑ λᵢ • Pᵢ with orthogonal projections
Pᵢ summing to the identity and pairwise-orthogonal. -/
theorem normal_spectralDecomposition (hT : IsStarNormal T) :
    ∃ (k : ℕ) (λ : Fin k → 𝕜) (P : Fin k → V →L[𝕜] V),
      (∀ i, IsOrthogonalProjection (P i)) ∧
      (∀ i j, i ≠ j → P i * P j = 0) ∧
      (∑ i, P i = 1) ∧
      (T = ∑ i, (λ i : 𝕜) • P i) :=
  sorry  -- repackage normal_spectralTheorem by grouping eigenvectors

/-- Real spectral theorem: a symmetric operator on a finite-dim
real inner-product space is orthogonally diagonalisable. -/
theorem symmetric_spectralTheorem {V : Type*}
    [NormedAddCommGroup V] [InnerProductSpace ℝ V] [FiniteDimensional ℝ V]
    (T : V →L[ℝ] V) (hT : LinearMap.IsSymmetric T.toLinearMap) :
    ∃ (n : ℕ) (b : Fin n → V) (λ : Fin n → ℝ),
      Orthonormal ℝ b ∧ (∀ i, T (b i) = λ i • b i) :=
  sorry  -- complexify, apply normal_spectralTheorem, descend to ℝ

The proof gap to a clean Mathlib contribution: (i) define IsOrthogonalProjection if absent; (ii) prove the normality-implies-shared-eigenvector lemma T.IsStarNormal → T x = λ • x → T.adjoint x = (starRingEnd 𝕜 λ) • x; (iii) prove the orthogonal-complement-of-eigenspace-is-invariant lemma; (iv) package the induction-on-dimension step that combines these. Each piece is formalisable from existing Mathlib infrastructure, but no single named theorem currently presents the spectral decomposition in the textbook form $T={\sum }_i{\lambda }_i{P}_i$ with the projection-family identities as a single result.

Advanced results Master

Theorem (real spectral theorem; Hoffman-Kunze §8.4 ^{[source pending]}). Let $V$ be a finite-dimensional real inner-product space and $T:V\to V$ a linear operator. Then $T$ is orthogonally diagonalisable iff $T = T^$(symmetric),inwhichcase$V$admitsanorthonormaleigenbasisofrealeigenvectorsandthespectrumliesin$\mathbb{R}$.*

The proof complexifies $V$ to $V_{\mathbb{C}}$, applies the complex spectral theorem to $T_{\mathbb{C}}$, uses the self-adjoint reality-of-eigenvalues argument to confine the spectrum to $\mathbb{R}$, and descends an orthonormal real eigenbasis by averaging each complex eigenvector with its conjugate. The full descent argument is given in the proof set below.

Theorem (principal-axes theorem for quadratic forms; Shilov Ch. 10 §10.1 ^{[source pending]}). Every quadratic form $Q:{\mathbb{R}}^n\to \mathbb{R}$, $Q(x)=x^{\top }Sx$ with $S$ real symmetric, is congruent under orthogonal change of basis to a diagonal form $Q(y)={\sum }_i{\lambda }_i{y}_i^2$, where ${\lambda }_i$ are the eigenvalues of $S$. The directions of the new basis are the principal axes of $Q$.

Specifically, the orthogonal matrix $Q$ in the diagonalisation $S=Q\Lambda Q^{\top }$ is the change-of-basis matrix whose columns are the orthonormal eigenvectors of $S$. Setting $y=Q^{\top }x$ rewrites $x^{\top }Sx=(Qy{)}^{\top }S(Qy)=y^{\top }\Lambda y={\sum }_i{\lambda }_i{y}_i^2$. Sylvester's law of inertia 01.01.15 further refines this: the signature $(p,q,z)$ — the number of positive, negative, and zero eigenvalues — is invariant under arbitrary congruence by ${\mathrm{GL}}_n(\mathbb{R})$, not only orthogonal congruence.

Theorem (Schur decomposition; Schur 1909 Math. Ann. 66, 488–510 ^{[source pending]}). Every linear operator $T:V\to V$ on a finite-dimensional complex inner-product space admits a unitary basis change taking $T$ to upper-triangular form: $T = U R U^$with$U$unitaryand$R$upper-triangular.Thediagonalentriesof$R$aretheeigenvaluesof$T$, in any chosen order.*

The Schur form holds for every operator, not only normal ones. The spectral theorem is the special case where normality forces the strictly upper-triangular part of $R$ to vanish, so $R$ is diagonal. The Schur form is the strongest universal canonical form available in the unitary-similarity category, and it is the input to QR-style numerical eigenvalue algorithms.

Theorem (real-normal canonical form). A real linear operator $T:V\to V$ satisfying $T{T}^{\top }=T^{\top }T$ is orthogonally similar to a block-diagonal matrix with real one-by-one blocks $({\lambda }_i)$ for real eigenvalues and real two-by-two blocks $\left(\begin{array}{cc}a& -b\\ b& a\end{array}\right)$ for complex-conjugate eigenvalue pairs $a\pm bi$ ($b>0$).

Real-orthogonal matrices (e.g., rotation matrices) are real-normal but in general have no real eigenvalues; the two-by-two rotation blocks encode the complex-conjugate eigenvalue pairs on the unit circle.

Theorem (functional calculus for normal operators). Let $T$ be normal with spectral decomposition $T={\sum }_i{\lambda }_i{P}_i$. For any function $f:\mathrm{spec}(T)\to \mathbb{C}$, define $$ f(T) = \sum_i f(\lambda_i) P_i. $$ The map $f\mapsto f(T)$ is a $$-algebrahomomorphismfromthealgebraoffunctionson$\mathrm{spec}(T)$totheoperatoralgebra$\mathrm{End}(V)$.Ittakestheidentityfunction$f(\lambda) = \lambda$to$T$,theconstant$1$totheidentityoperator,$f g$to$f(T) g(T)$,andcomplexconjugation$\bar f$totheadjoint$f(T)^$.Forpolynomial$f \in \mathbb{C}[t]$,$f(T)$agreeswiththepolynomialevaluation;for$f = \exp$,$f(T) = \exp(T) = \sum_i e^{\lambda_i} P_i$;for$f(\lambda) = \sqrt{\lambda}$onapositiveoperator,$f(T)$ is the unique positive square root.

The finite-dimensional functional calculus is the model for the continuous functional calculus on a C*-algebra (Gelfand-Naimark) and for the Borel functional calculus on a von Neumann algebra; the spectral-measure formulation extends $f(T)$ from continuous to Borel functions and from ${\sum }_i$ to ${\int }_{\mathrm{spec}(T)}fdE$.

Theorem (compact self-adjoint operators on Hilbert space; Riesz-Schauder, Hilbert 1906 + Riesz 1907 ^{[source pending]}). Let $H$ be a separable Hilbert space and $T:H\to H$ a compact self-adjoint operator. The spectrum of $T$ is a countable subset of $\mathbb{R}$ with $0$ as the only possible accumulation point. Each non-zero eigenvalue has finite-dimensional eigenspace, and $H$ admits an orthonormal basis of eigenvectors of $T$. Equivalently, $$ T = \sum_{n=1}^\infty \lambda_n P_n, $$ with the sum converging in the operator norm and ${\lambda }_n\to 0$.

This is the infinite-dimensional generalisation that motivated the development of functional analysis: Hilbert's 1906 study of symmetric integral kernels on $L^2[a,b]$ established the spectral theorem for the compact case, and Riesz 1907/1910 abstracted it to compact operators on an arbitrary Hilbert space, founding the Riesz-Schauder theory. The hypothesis "compact" replaces "finite-dimensional"; the conclusion is the same diagonalisation, now indexed by a countable spectrum accumulating only at $0$.

Theorem (bounded self-adjoint operators on Hilbert space; multiplication-operator form, Hahn-Hellinger). Let $H$ be a separable Hilbert space and $T:H\to H$ a bounded self-adjoint operator. There exist a $\sigma$-finite measure space $(X,\mu )$, a bounded measurable function $f:X\to \mathbb{R}$, and a unitary isomorphism $U:H\to L^2(X,\mu )$ such that $UT{U}^{-1}$ is multiplication by $f$: $$ (U T U^{-1} \psi)(x) = f(x) \psi(x). $$ Equivalently (and more invariantly), there is a projection-valued measure $E$ on the Borel $\sigma$-algebra of $\mathrm{spec}(T)\subseteq \mathbb{R}$ such that $T={\int }_{\mathrm{spec}(T)}\lambda dE(\lambda )$.

The multiplication-operator form replaces the finite sum ${\sum }_i{\lambda }_i{P}_i$ by integration of the identity function against the spectral measure $E$. Hahn-Hellinger established the multiplicity-theoretic refinement: the unitary equivalence class of $T$ is determined by a multiplicity function on $\mathrm{spec}(T)$.

Theorem (unbounded self-adjoint operators on Hilbert space; von Neumann 1929 Math. Ann. 102, 49–131 ^{[source pending]}). Let $T:D(T)\to H$ be an unbounded self-adjoint operator on a separable Hilbert space, with dense domain $D(T)\subseteq H$ and $T = T^$asunboundedoperators.Thereexistsauniqueprojection-valuedmeasure$E$ontheBorel$\sigma$-algebraof$\mathbb{R}$ such that* $$ T = \int_{\mathbb{R}} \lambda , dE(\lambda), \qquad D(T) = \left{ \psi \in H : \int_{\mathbb{R}} \lambda^2 , d|E_\lambda \psi|^2 < \infty \right}. $$

The unbounded case is the technically deepest version of the spectral theorem and the one that underwrites quantum mechanics: observables are unbounded self-adjoint operators on a Hilbert space of state vectors, and the projection-valued measure $E$ is the measurement postulate. Stone's theorem identifies the one-parameter unitary group $U_t=e^{itT}=\int e^{it\lambda }dE(\lambda )$ as the dynamical evolution generated by the observable $T$.

Theorem (normal operators on Hilbert space). A bounded normal operator $T$ on a separable Hilbert space admits a projection-valued measure $E$ on the Borel $\sigma$-algebra of $\mathrm{spec}(T)\subseteq \mathbb{C}$ such that $T={\int }_{\mathrm{spec}(T)}\lambda dE(\lambda )$. Equivalently, $T$ is unitarily equivalent to multiplication by a bounded measurable complex-valued function on a $\sigma$-finite measure space.

The complex spectrum supports the projection-valued measure; the proof passes through the C*-algebra generated by $T$ and applies Gelfand-Naimark to identify it with continuous functions on $\mathrm{spec}(T)$. The unbounded normal case requires an additional commutativity hypothesis between $T$ and $T^{\ast }$ on domains (Putnam-Fuglede).

Synthesis. The finite-dimensional spectral theorem for normal operators is the foundational reason the inner-product geometry of $V$ admits an operator-adapted orthonormal basis: the operator $T$ and its adjoint $T^{\ast }$ commute exactly when their joint eigenspaces resolve the identity, and this is exactly the content of the orthogonal-eigenbasis statement. The central insight is that normality identifies $T$ with a finite sum ${\sum }_i{\lambda }_i{P}_i$ of orthogonal projections weighted by complex scalars, and the projections form a partition of unity in the inner-product geometry of $V$. Putting these together, one spectral framework produces the real symmetric case (eigenvalues confined to $\mathbb{R}$), the unitary case (eigenvalues on the unit circle), the anti-self-adjoint case (eigenvalues imaginary), the principal-axes theorem for quadratic forms, the singular value decomposition via the spectral theorem applied to $T^{\ast }T$, and the functional calculus $f(T)={\sum }_{i}f({\lambda }_i)P_i$ for arbitrary spectral functions. The bridge is that the finite sum generalises to an operator-valued integral $T={\int }_{\mathrm{spec}(T)}\lambda dE(\lambda )$ against a projection-valued measure, and this is exactly the form taken by the spectral theorem for unbounded self-adjoint operators on Hilbert space, builds toward 02.11.03 (unbounded self-adjoint operators) where the integral form supplies the dynamical-evolution generator of quantum mechanics, and appears again in 01.01.12 (singular value decomposition) as the diagonalisation of $T^{\ast }T$ that produces singular values for arbitrary linear maps.

The spectral framework identifies several pairings that look distinct at first inspection. The real-symmetric diagonalisation is exactly the orthogonal-congruence classification of quadratic forms via the principal-axes theorem. The unitary diagonalisation is exactly the Fourier-style decomposition of a unitary operator into eigenspaces on the unit circle, the finite-dimensional model for Stone's theorem on one-parameter unitary groups. The Schur decomposition supplies an upper-triangular form for every complex operator, and normality is the precise additional condition that collapses Schur to diagonal. This pattern recurs at every level of generality: in the Hilbert-space case the spectral theorem for compact self-adjoint operators (Hilbert-Riesz-Schauder) gives the same orthonormal-eigenbasis statement with a countable spectrum, the bounded self-adjoint case replaces the eigenbasis by a multiplication operator on an $L^2$-space (Hahn-Hellinger), and the unbounded case packages everything in a projection-valued measure (von Neumann). At each level, the finite-dimensional formula $T={\sum }_i{\lambda }_i{P}_i$ is the model, and the upgrade is to a measure-theoretic integral. The foundational identification — operator decomposes against its spectrum — is preserved across the whole hierarchy.

Full proof set Master

Theorem (complex spectral theorem), proof. Given in the Intermediate-tier section: existence of an eigenvalue from the fundamental theorem of algebra (Step 1), shared eigenvector for $T$ and $T^{\ast }$ with conjugate eigenvalue using normality of $T-\lambda I$ (Step 2), orthogonality of eigenspaces at distinct eigenvalues via the adjoint-pairing argument (Step 3), $T$-invariance of the orthogonal complement of an eigenspace (Step 4), and induction on $\dim V$ (Step 5). The implication (ii) $\Rightarrow$ (iii) groups eigenvectors by eigenvalue and reads off the projection-family decomposition; (iii) $\Rightarrow$ (i) computes $T{T}^{\ast }$ and $T^{\ast }T$ both as ${\sum }_i|{\lambda }_i{|}^2{P}_i$. $\square$

Proposition (real spectral theorem), proof. Let $T:V\to V$ be a symmetric operator on a finite-dim real inner-product space. Complexify: $V_{\mathbb{C}}=V{\otimes }_{\mathbb{R}}\mathbb{C}\cong V\oplus iV$, with the Hermitian inner product $⟨v_1+i{w}_1,v_2+i{w}_2{⟩}_{\mathbb{C}}=⟨v_1,v_2⟩+⟨w_1,w_2⟩+i(⟨w_1,v_2⟩-⟨v_1,w_2⟩)$ restricting to the real inner product on $V$. Extend $T$ to a $\mathbb{C}$-linear operator $T_{\mathbb{C}}(v+iw)=Tv+iTw$. Verify that $T_{\mathbb{C}}$ is Hermitian-self-adjoint: $$ \langle T_\mathbb{C}(v_1 + i w_1), v_2 + i w_2 \rangle_\mathbb{C} = \langle T v_1, v_2 \rangle + \langle T w_1, w_2 \rangle + i(\langle T w_1, v_2 \rangle - \langle T v_1, w_2 \rangle), $$ and on the other side $$ \langle v_1 + i w_1, T_\mathbb{C}(v_2 + i w_2) \rangle_\mathbb{C} = \langle v_1, T v_2 \rangle + \langle w_1, T w_2 \rangle + i(\langle w_1, T v_2 \rangle - \langle v_1, T w_2 \rangle). $$ By $T=T^{\top }$ in the real inner product, the two sides match. So $T_{\mathbb{C}}$ is self-adjoint, hence normal.

Apply the complex spectral theorem: $V_{\mathbb{C}}$ admits an orthonormal eigenbasis $\{e_1,\dots ,e_n\}$ of $T_{\mathbb{C}}$ with eigenvalues ${\lambda }_i$. Each ${\lambda }_i$ is real, by the self-adjoint reality argument: ${\lambda }_i⟨e_i,e_i{⟩}_{\mathbb{C}}=⟨T_{\mathbb{C}}{e}_i,e_i{⟩}_{\mathbb{C}}=⟨e_i,T_{\mathbb{C}}{e}_i{⟩}_{\mathbb{C}}={\bar{\lambda }}_i⟨e_i,e_i{⟩}_{\mathbb{C}}$, forcing ${\lambda }_i={\bar{\lambda }}_i\in \mathbb{R}$.

Complex conjugation on $V_{\mathbb{C}}$ (sending $v+iw\mapsto v-iw$) commutes with $T_{\mathbb{C}}$ because $T$ is real. For each real $\lambda$, the eigenspace $E_{\lambda }(T_{\mathbb{C}})$ is conjugation-invariant, so it has a real basis: pick any $z\in E_{\lambda }(T_{\mathbb{C}})$, decompose $z=v+iw$ with $v,w\in V$, then $T_{\mathbb{C}}z=Tv+iTw=\lambda (v+iw)$ gives $Tv=\lambda v$ and $Tw=\lambda w$, so $v,w\in E_{\lambda }(T):=\ker (T-\lambda I){|}_V$. Choosing $z,\bar{z}$ as a real-spanning pair and applying Gram-Schmidt in the real eigenspace gives a real orthonormal basis of $E_{\lambda }(T)$.

Distinct real eigenspaces of $T$ are orthogonal in $V$ by the same calculation as Step 3 of the complex case (specialised to real ${\lambda }_1\ne {\lambda }_2$, where the conjugate-linear distinction vanishes). Concatenating the real orthonormal bases of the distinct real eigenspaces produces a real orthonormal eigenbasis of $V$, and the matrix of $T$ in this basis is real diagonal. $\square$

Proposition (principal-axes theorem), proof. Let $S$ be a real symmetric $n\times n$ matrix and $Q(x)=x^{\top }Sx$ the associated quadratic form on ${\mathbb{R}}^n$ 01.01.15. By the real spectral theorem there is an orthogonal matrix $P$ and a real diagonal matrix $\Lambda$ with $S=P\Lambda P^{\top }$. Change variable: set $y=P^{\top }x$, so $x=Py$. Substitute into $Q$: $$ Q(x) = x^\top S x = (P y)^\top (P \Lambda P^\top)(P y) = y^\top P^\top P \Lambda P^\top P y = y^\top \Lambda y = \sum_{i=1}^n \lambda_i y_i^2, $$ using $P^{\top }P=I$. The new basis vectors (columns of $P$) are the principal axes of $Q$, the eigenvalues ${\lambda }_i$ are the diagonal entries of $\Lambda$, and $Q$ is diagonalised as a sum of weighted squares. $\square$

Proposition (Schur decomposition), proof. Let $T:V\to V$ be a linear operator on a finite-dim complex inner-product space, $\dim V=n$. Argue by induction on $n$. The base case $n=1$ is immediate. For $n>1$, $T$ has an eigenvalue ${\lambda }_1\in \mathbb{C}$ with eigenvector $v_1$; normalise to $u_1=v_1/\Vert v_1\Vert$. Let $W=(\mathrm{span}\{u_1\}{)}^{\perp }\subseteq V$, a subspace of dimension $n-1$. Note: $W$ is not in general $T$-invariant (it is $T^{\ast }$-invariant, by the same argument as Step 4 of the spectral theorem but with $T$ and $T^{\ast }$ swapped). However, the composite ${\pi }_W\circ T{|}_W:W\to W$ is a well-defined linear operator on $W$, where ${\pi }_W$ is orthogonal projection onto $W$.

By inductive hypothesis there is an orthonormal basis $\{u_2,\dots ,u_n\}$ of $W$ in which ${\pi }_W\circ T{|}_W$ is upper triangular. Concatenate to form the orthonormal basis $\{u_1,u_2,\dots ,u_n\}$ of $V$. The matrix of $T$ in this basis has first column $({\lambda }_1,0,\dots ,0{)}^{\top }$ (since $T{u}_1={\lambda }_1{u}_1$) and the lower-right $(n-1)\times (n-1)$ block agrees with the matrix of ${\pi }_W\circ T{|}_W$, which is upper-triangular by inductive hypothesis. Off-diagonal entries above the diagonal in the first row are the inner products $⟨T{u}_j,u_1⟩$ for $j>1$, which need not vanish — but they sit in the upper triangle, so the resulting matrix is upper triangular. Setting $U$ to be the unitary matrix whose columns are $\{u_1,\dots ,u_n\}$ gives $T=UR{U}^{\ast }$ with $R$ upper triangular. $\square$

Proposition (functional calculus is a $\ast$-homomorphism), proof. Let $T$ be normal with spectral decomposition $T={\sum }_i{\lambda }_i{P}_i$, and for $f:\mathrm{spec}(T)\to \mathbb{C}$ define $f(T)={\sum }_{i}f({\lambda }_i)P_i$. Use $P_i{P}_j={\delta }_{ij}{P}_i$ and $P_i^{\ast }=P_i$:

(i) Linearity: $(\alpha f+\beta g)(T)={\sum }_i(\alpha f({\lambda }_i)+\beta g({\lambda }_i))P_i=\alpha f(T)+\beta g(T)$.

(ii) Multiplicativity: $f(T)g(T)=({\sum }_{i}f({\lambda }_i)P_i)({\sum }_{j}g({\lambda }_j)P_j)={\sum }_{i,j}f({\lambda }_i)g({\lambda }_j)P_i{P}_j={\sum }_{i}f({\lambda }_i)g({\lambda }_i)P_i=(fg)(T)$.

(iii) Identity: the constant function $1$ gives $1(T)={\sum }_i{P}_i=I$.

(iv) Star: $\overline{f}(T)={\sum }_i\overline{f({\lambda }_i)}{P}_i=({\sum }_{i}f({\lambda }_i)P_i{)}^{\ast }=f(T{)}^{\ast }$, since each $P_i$ is self-adjoint.

(v) Identity function: $\mathrm{id}(T)={\sum }_i{\lambda }_i{P}_i=T$.

The map $f\mapsto f(T)$ is therefore a $\ast$-algebra homomorphism from the algebra of functions on $\mathrm{spec}(T)$ (with pointwise operations and complex conjugation) to $\mathrm{End}(V)$. For $f\in \mathbb{C}[t]$ a polynomial, $f(T)={\sum }_k{a}_k{T}^k$ by linearity and multiplicativity, agreeing with the standard polynomial evaluation. For $f=\exp$, $\exp (T)={\sum }_i{e}^{{\lambda }_i}{P}_i$ is a finite sum that equals the operator-norm exponential ${\sum }_{k\ge 0}{T}^k/k!$ by uniform-convergence comparison on the finite spectrum. $\square$

Theorem (compact self-adjoint operators), stated without independent proof — see Hilbert 1906 Nachr. Gött. IV ^{[source pending]} and Riesz 1907 C. R. Acad. Sci. Paris 144, 615–619 ^{[source pending]}. The Riesz-Schauder argument adapts the five-step proof of the finite-dimensional case to compact operators: the spectrum is a countable set with $0$ as the only possible accumulation point (Riesz lemma), each non-zero eigenvalue has finite multiplicity (compactness of the unit ball in the eigenspace), and the eigenvectors span the closure of the range (by induction on the largest eigenvalue, using that the orthogonal complement of an eigenspace is invariant and that the restriction is still compact and self-adjoint). The full development is in Reed-Simon Vol. I §VI.

Theorem (unbounded self-adjoint operators, von Neumann's spectral theorem), stated without independent proof — see von Neumann 1929 Math. Ann. 102, 49–131 ^{[source pending]}. The unbounded case requires the Cayley transform $C=(T-iI)(T+iI{)}^{-1}$, which converts an unbounded self-adjoint $T$ into a bounded unitary $C$, then applies the spectral theorem for unitary operators (a normal-operator special case with spectrum on the unit circle), and inverts the Cayley transform to recover the projection-valued measure of $T$. The detailed argument is given in Reed-Simon Vol. I §VIII; the finite-dimensional theorem is the model that orients the entire construction.

Connections Master

• Eigenvalue, eigenvector, characteristic polynomial 01.01.08. The spectral theorem is the structural refinement of the eigenvalue-eigenvector theory for normal operators on an inner-product space. The eigenvalues ${\lambda }_i$ of $T$ are the roots of ${\chi }_T(t)$, and normality is the precise condition that promotes "every operator has eigenvectors over $\mathbb{C}$" to "every normal operator has an orthonormal eigenbasis". Without the inner-product structure and the adjoint $T^{\ast }$, the orthonormality clause is meaningless; without normality, eigenspaces at distinct eigenvalues need not be orthogonal and the orthogonal-complement-of-eigenspace need not be $T$-invariant.

• Bilinear / quadratic form 01.01.15. The principal-axes theorem is the orthogonal-congruence diagonalisation of a real quadratic form, and it is exactly the real spectral theorem applied to the symmetric matrix representing the form. Sylvester's law of inertia further classifies real quadratic forms up to arbitrary ${\mathrm{GL}}_n(\mathbb{R})$-congruence by the signature $(p,q,z)$, recovering the eigenvalue-sign information from the spectral decomposition. The bridge between the inner-product-geometric statement (orthogonal congruence) and the bilinear-form classification (Sylvester's law) is the spectral theorem.

• Singular value decomposition 01.01.12. For an arbitrary linear map $A:V\to W$ between finite-dim inner-product spaces, the operator $A^{\ast }A$ on $V$ is self-adjoint and positive semi-definite, so the spectral theorem gives an orthonormal eigenbasis $\{v_i\}$ with non-negative eigenvalues ${\sigma }_i^2\ge 0$. The numbers ${\sigma }_i$ are the singular values of $A$; setting $u_i=A{v}_i/{\sigma }_i$ for ${\sigma }_i>0$ produces the orthonormal basis on the $W$-side, and $A={\sum }_i{\sigma }_i{u}_i{v}_i^{\ast }$ is the singular value decomposition. The SVD is the spectral theorem on $A^{\ast }A$ paired with the recipe for transporting the eigenbasis to the codomain.

• Jordan canonical form 01.01.11. Over $\mathbb{C}$ every linear operator $T$ admits a Jordan canonical form, the strongest similarity-classification statement available without an inner-product structure. The Schur decomposition is the unitary-similarity refinement (every operator is unitarily upper-triangularisable), and the spectral theorem is the further refinement available when $T$ is normal (the upper-triangular part collapses to diagonal). Jordan and spectral are dual viewpoints: Jordan classifies up to $\mathrm{GL}(V)$-conjugacy without inner-product structure, spectral classifies up to $U(V)$-conjugacy with inner-product structure, and normal operators are exactly the ones whose Jordan form is already diagonal.

• Unbounded self-adjoint operators 02.11.03. The infinite-dimensional generalisation of the spectral theorem to unbounded self-adjoint operators on a separable Hilbert space (von Neumann 1929) is the foundation of quantum-mechanical observable theory. The finite-sum spectral decomposition $T={\sum }_i{\lambda }_i{P}_i$ generalises to an operator-valued integral $T=\int \lambda dE(\lambda )$ against a projection-valued measure $E$ on the Borel $\sigma$-algebra of $\mathrm{spec}(T)\subseteq \mathbb{R}$, and Stone's theorem identifies the one-parameter unitary group $e^{itT}$ as the dynamical evolution. The finite-dim theorem proved here is the model that orients the entire functional-analytic theory.

Historical & philosophical context Master

The principal-axes theorem in its real symmetric form is due to Cauchy, who established the orthogonal diagonalisation of real symmetric matrices in his 1829 Exercices de Mathématiques IV (140–160) ^{[source pending]} in the context of celestial mechanics: the secular perturbations of planetary orbits were governed by a real symmetric matrix arising from the gravitational potential, and Cauchy needed the principal axes of the corresponding quadratic form to decouple the perturbation equations. The real eigenvalues recorded the frequencies of the linearised motion; the orthogonal eigenvectors recorded the normal modes. The theorem was first applied as a problem-solving tool in mathematical astronomy before being recognised as a structural result in linear algebra. Hermite extended the result to complex Hermitian matrices in 1855 (Comptes Rendus 41, 181–183, and J. Reine Angew. Math. 50, 273–283) ^{[source pending]}, establishing the reality of eigenvalues for the Hermitian case and the corresponding orthogonal diagonalisation.

The normal-operator framing is due to Toeplitz, whose 1918 paper Das algebraische Analogon zu einem Satze von Fejér (Math. Z. 2, 187–197) ^{[source pending]} identified normality $T{T}^{\ast }=T^{\ast }T$ as the precise condition for unitary diagonalisability in the finite-dimensional complex setting. Schur's 1909 paper Über die charakteristischen Wurzeln einer linearen Substitution (Math. Ann. 66, 488–510) ^{[source pending]} had already established the unitary upper-triangularisation of arbitrary complex matrices, and Toeplitz's contribution was the recognition that the strictly upper-triangular part of the Schur form vanishes exactly when the operator is normal — collapsing the Schur decomposition to a diagonal one and identifying normality as the algebraic signature of the spectral theorem.

The infinite-dimensional generalisations developed in parallel with the finite case. Hilbert's Grundzüge einer allgemeinen Theorie der linearen Integralgleichungen — published in six instalments in the Nachrichten der Gesellschaft der Wissenschaften zu Göttingen between 1904 and 1910 and collected as a monograph in 1912 — established the spectral theorem for symmetric integral operators on $L^2$, the first infinite-dimensional spectral theorem ^{[source pending]}. Riesz's 1907 Comptes Rendus note (144, 615–619) and 1910 Mathematische Annalen paper (69, 449–497) abstracted Hilbert's argument to compact symmetric operators on a separable Hilbert space, founding what is now called Riesz-Schauder theory ^{[source pending]}. The bounded self-adjoint case received its multiplication-operator form through Hahn and Hellinger in the 1910s and 1920s, and the unbounded case was settled by von Neumann's 1929 Allgemeine Eigenwerttheorie Hermitescher Funktionaloperatoren (Math. Ann. 102, 49–131) ^{[source pending]}, where the projection-valued measure formulation $T=\int \lambda dE(\lambda )$ first appeared. Von Neumann's 1932 monograph Mathematische Grundlagen der Quantenmechanik (Springer 1932) then identified self-adjoint operators with quantum-mechanical observables and the spectral measure with the measurement postulate, an identification that has structured both functional analysis and the foundations of physics ever since.

Bibliography Master

@article{Cauchy1829Exercices,
  author  = {Cauchy, Augustin-Louis},
  title   = {Sur l'{\'e}quation {\`a} l'aide de laquelle on d{\'e}termine les in{\'e}galit{\'e}s s{\'e}culaires des mouvements des plan{\`e}tes},
  journal = {Exercices de Math{\'e}matiques},
  volume  = {4},
  year    = {1829},
  pages   = {140--160}
}

@article{Hermite1855,
  author  = {Hermite, Charles},
  title   = {Remarques sur un th{\'e}or{\`e}me de M. Cauchy},
  journal = {Comptes Rendus de l'Acad{\'e}mie des Sciences},
  volume  = {41},
  year    = {1855},
  pages   = {181--183}
}

@article{Schur1909,
  author  = {Schur, Issai},
  title   = {{\"U}ber die charakteristischen Wurzeln einer linearen Substitution mit einer Anwendung auf die Theorie der Integralgleichungen},
  journal = {Mathematische Annalen},
  volume  = {66},
  year    = {1909},
  pages   = {488--510}
}

@article{Toeplitz1918,
  author  = {Toeplitz, Otto},
  title   = {Das algebraische Analogon zu einem Satze von Fej{\'e}r},
  journal = {Mathematische Zeitschrift},
  volume  = {2},
  year    = {1918},
  pages   = {187--197}
}

@article{Hilbert1906,
  author  = {Hilbert, David},
  title   = {Grundz{\"u}ge einer allgemeinen Theorie der linearen Integralgleichungen, Vierte Mitteilung},
  journal = {Nachrichten der K. Gesellschaft der Wissenschaften zu G{\"o}ttingen},
  year    = {1906},
  pages   = {157--227}
}

@book{Hilbert1912IntegralEquations,
  author    = {Hilbert, David},
  title     = {Grundz{\"u}ge einer allgemeinen Theorie der linearen Integralgleichungen},
  publisher = {Teubner},
  address   = {Leipzig},
  year      = {1912}
}

@article{Riesz1907,
  author  = {Riesz, Frigyes},
  title   = {Sur les syst{\`e}mes orthogonaux de fonctions},
  journal = {Comptes Rendus de l'Acad{\'e}mie des Sciences},
  volume  = {144},
  year    = {1907},
  pages   = {615--619}
}

@article{Riesz1910,
  author  = {Riesz, Frigyes},
  title   = {Untersuchungen {\"u}ber Systeme integrierbarer Funktionen},
  journal = {Mathematische Annalen},
  volume  = {69},
  year    = {1910},
  pages   = {449--497}
}

@article{vonNeumann1929,
  author  = {von Neumann, John},
  title   = {Allgemeine Eigenwerttheorie Hermitescher Funktionaloperatoren},
  journal = {Mathematische Annalen},
  volume  = {102},
  year    = {1929},
  pages   = {49--131}
}

@book{vonNeumann1932,
  author    = {von Neumann, John},
  title     = {Mathematische Grundlagen der Quantenmechanik},
  publisher = {Springer},
  address   = {Berlin},
  year      = {1932}
}

@book{HalmosFDVS,
  author    = {Halmos, Paul R.},
  title     = {Finite-Dimensional Vector Spaces},
  publisher = {Springer-Verlag},
  series    = {Undergraduate Texts in Mathematics},
  edition   = {2},
  year      = {1974}
}

@book{ShilovLinearAlgebra,
  author    = {Shilov, Georgi E.},
  title     = {Linear Algebra},
  publisher = {Dover},
  year      = {1977},
  note      = {English translation; Russian original 1969}
}

@book{HoffmanKunze,
  author    = {Hoffman, Kenneth and Kunze, Ray},
  title     = {Linear Algebra},
  publisher = {Prentice-Hall},
  edition   = {2},
  year      = {1971}
}

@book{AxlerLADR,
  author    = {Axler, Sheldon},
  title     = {Linear Algebra Done Right},
  publisher = {Springer},
  series    = {Undergraduate Texts in Mathematics},
  edition   = {3},
  year      = {2015}
}

@book{ReedSimonI,
  author    = {Reed, Michael and Simon, Barry},
  title     = {Methods of Modern Mathematical Physics, Vol. I: Functional Analysis},
  publisher = {Academic Press},
  edition   = {Revised and Enlarged},
  year      = {1980}
}

@book{KatoPerturbation,
  author    = {Kato, Tosio},
  title     = {Perturbation Theory for Linear Operators},
  publisher = {Springer-Verlag},
  series    = {Grundlehren der mathematischen Wissenschaften},
  volume    = {132},
  edition   = {2},
  year      = {1980}
}