Hilbert Space

Why Hilbert space matters

• Quantum mechanics. The state space of a quantum system is a complex Hilbert space. Pure states are unit vectors; observables are self-adjoint operators; measurements yield eigenvalues with Born-rule probabilities. The whole formalism — Schrödinger evolution, Heisenberg uncertainty, Bell inequalities — lives in Hilbert space. Von Neumann's 1932 axiomatization made this precise and remains the operating manual.
• Fourier analysis. The Fourier transform is a unitary isomorphism L²(ℝⁿ) → L²(ℝⁿ), preserving inner product (Plancherel). Discrete Fourier transforms diagonalize convolution on ℓ²(ℤ/N). Wavelets, splines, and time-frequency representations all live in Hilbert spaces and exploit orthonormal expansions for compression and analysis.
• Signal processing. Filters are operators on L² or ℓ²; matched filtering is inner product against a template. The energy of a signal is its L² norm; orthogonality of signals at different frequencies underpins channel allocation. MP3, JPEG2000, and FLAC all rely on Hilbert-space orthonormal bases (DCT, wavelets) for redundancy removal.
• Spectral theory. Self-adjoint operators on Hilbert space have spectral decompositions analogous to diagonalization of Hermitian matrices. The Laplacian on a domain is diagonalized by its eigenfunctions; the energy spectrum of a hydrogen atom is the spectrum of its Hamiltonian. Spectral theorems are the rigorous foundation of "diagonalizing observables."
• PDE theory. Weak solutions of elliptic, parabolic, and hyperbolic PDEs are sought in Sobolev spaces, which are Hilbert spaces. Lax-Milgram, Stampacchia, and Galerkin methods all run on Hilbert space inner products. Finite element methods discretize PDE problems by projecting onto finite-dimensional subspaces and solving the resulting linear systems.
• Reproducing kernel Hilbert spaces (RKHS). Kernel methods in machine learning (SVMs, Gaussian processes) implicitly construct an RKHS where data are embedded; the kernel K(x, y) = ⟨φ(x), φ(y)⟩ is an inner product in feature space. The "kernel trick" exploits Hilbert-space structure without explicit computation of the embedding.
• Statistical inference and stochastic processes. The space L² of a probability measure is the natural home of mean-square integrable random variables; conditional expectation is orthogonal projection. Karhunen-Loève expansions decompose random fields into eigenfunctions of the covariance operator — a Hilbert-space spectral problem.

Common misconceptions

• "Always finite-dimensional." ℝⁿ and ℂⁿ are Hilbert spaces, but finite-dim Hilbert space coincides with elementary linear algebra. The interesting cases are infinite-dimensional: ℓ², L², Sobolev spaces. Most theorems (Riesz, spectral theorem, completeness of ONB) become non-trivial only in infinite dimensions.
• "All separable Hilbert spaces are different." No — they are all isomorphic to ℓ². Separability + infinite-dimensional + Hilbert pins down the space up to isometric isomorphism. Physicists often refer to "the Hilbert space" because in this categorical sense, there is essentially one.
• "Norm = Euclidean." Only in finite dimensions and only with the canonical inner product. In L²(ℝ), ‖f‖ = √∫ |f|², which is Pythagorean in spirit but bears no resemblance to the Euclidean √(x² + y²) outside the formal analogy.
• "Inner product is symmetric." In a real Hilbert space, yes: ⟨x, y⟩ = ⟨y, x⟩. In a complex Hilbert space, the inner product is conjugate-symmetric: ⟨x, y⟩ = ⟨y, x⟩̄ (overline). Linearity is in one argument (usually the first by physicists' convention, second by mathematicians'); the other is conjugate-linear.
• "All linear maps are bounded." In infinite dimensions, no. A linear map T: H → H may fail to be bounded (continuous), in which case it is defined only on a dense subspace D(T) ⊂ H. Differential operators are typically unbounded; rigorous QM requires careful treatment of self-adjoint extensions, defect indices, and operator domains.
• "Compactness works the same." The closed unit ball in an infinite-dimensional Hilbert space is not compact (a fundamental difference from ℝⁿ). It is weakly compact (Banach-Alaoglu), and compact operators have well-behaved spectra (Hilbert-Schmidt theorem), but norm-compactness fails in infinite dimensions.

Definition and basic theorems

A Hilbert space H is a vector space over 𝕂 = ℝ or ℂ equipped with an inner product ⟨·, ·⟩: H × H → 𝕂 satisfying:

• Conjugate symmetry: ⟨x, y⟩ = ⟨y, x⟩̄ (just symmetry in the real case).
• Linearity in the first argument: ⟨ax + by, z⟩ = a⟨x, z⟩ + b⟨y, z⟩.
• Positive-definite: ⟨x, x⟩ ≥ 0 with equality iff x = 0.
• Completeness: every Cauchy sequence in the induced norm ‖x‖ = √⟨x, x⟩ converges in H.

Foundational tools:

• Cauchy-Schwarz inequality. |⟨x, y⟩| ≤ ‖x‖ · ‖y‖, with equality iff x, y are linearly dependent. Proves the triangle inequality and continuity of inner product.
• Parallelogram law. ‖x + y‖² + ‖x − y‖² = 2‖x‖² + 2‖y‖². Characterizes Hilbert norms among Banach norms; spaces failing this are not Hilbert.
• Polarization identity. Recovers the inner product from the norm: ⟨x, y⟩ = (1/4)(‖x+y‖² − ‖x−y‖²) (real); ⟨x, y⟩ = (1/4) Σ_{k=0}^3 i^k ‖x + i^k y‖² (complex).

Canonical examples

• ℓ² (square-summable sequences). ℓ² = {(xₙ)_{n=1}^∞ : Σ|xₙ|² < ∞} with inner product ⟨x, y⟩ = Σ xₙ ȳₙ. The standard separable Hilbert space; the canonical orthonormal basis is e_k = (0, …, 0, 1, 0, …) with 1 in position k.
• L²(X, μ). Square-integrable measurable functions on a measure space (X, μ): L²(X) = {f: ∫ |f|² dμ < ∞} (modulo a.e. equality), with ⟨f, g⟩ = ∫ f ḡ dμ. Examples: L²(ℝⁿ), L²([0, 2π]) (with Fourier basis), L²(S²) (with spherical harmonics), L²(probability space) (mean-square integrable random variables).
• Sobolev spaces H^k(Ω). Functions f on a domain Ω ⊂ ℝⁿ with ‖f‖² = Σ_{|α| ≤ k} ‖∂^α f‖²_{L²} < ∞. The natural setting for elliptic PDE; H¹₀ is used in the Lax-Milgram theorem to give weak solutions of Δu = f.
• Reproducing kernel Hilbert spaces. Function spaces on a set X where evaluation at x is a bounded linear functional; by Riesz, evaluation is inner product with a kernel function K(·, x). Hardy spaces H²(𝔻), Bergman spaces, and Gaussian process regression spaces are all RKHS.
• Hardy space H²(𝔻). Holomorphic functions on the unit disc whose boundary values are L²(unit circle); the canonical orthonormal basis is {1, z, z², z³, …}, and the inner product comes from L²(unit circle). Foundation of Beurling's invariant subspace theorem and Hardy-Littlewood maximal inequalities.
• Bosonic Fock space. Symmetric tensor algebra over a single-particle Hilbert space H: F_+(H) = ⊕_{n=0}^∞ Sym^n H, with the obvious inner product. State space for non-relativistic many-boson systems; with H = L²(ℝ³), this is the second-quantized Fock space. Fermionic version uses antisymmetric tensor algebra.

Orthonormal bases and Parseval

An orthonormal set {eₙ} ⊂ H satisfies ⟨eₙ, eₘ⟩ = δₙₘ. It is an orthonormal basis (or "complete orthonormal set") if its linear span is dense in H.

Theorem (separable Hilbert space): every separable Hilbert space H has a countable orthonormal basis. Given any orthonormal basis {eₙ}, every x ∈ H can be expanded:

x = Σₙ ⟨x, eₙ⟩ eₙ    (convergence in norm).

Parseval's identity:    ‖x‖² = Σₙ |⟨x, eₙ⟩|².
Plancherel inner product:  ⟨x, y⟩ = Σₙ ⟨x, eₙ⟩ ⟨eₙ, y⟩.

The map x ↦ (⟨x, eₙ⟩) is an isometric isomorphism H → ℓ². So all separable infinite-dimensional Hilbert spaces are isomorphic to ℓ².

Famous orthonormal bases:

• L²([0, 2π]): {e^{inx}/√(2π)} for n ∈ ℤ — Fourier basis.
• L²(ℝ) with weight e^{−x²}: Hermite polynomials (after weighting); equivalently L²(ℝ) has Hermite functions H_n(x) e^{−x²/2} as an ONB. These are eigenfunctions of the harmonic oscillator Hamiltonian.
• L²([0, 1]): Legendre polynomials P_n(x) (after normalization).
• L²(S²): spherical harmonics Y_l^m.
• L²(ℝ): wavelet bases (Haar, Daubechies, Meyer) — useful in numerics and signal processing where multi-resolution structure is needed.

Riesz representation theorem

Statement: let H be a Hilbert space and φ: H → 𝕂 a continuous (equivalently, bounded) linear functional. Then there exists a unique y ∈ H such that

φ(x) = ⟨x, y⟩    for all x ∈ H,
‖φ‖_{H*} = ‖y‖_H.

The map y ↦ ⟨·, y⟩ is an antilinear (in complex case) isometry H → H* (a linear isometry in the real case). So the dual H* is identified with H itself.

Sketch of proof: if φ ≡ 0, take y = 0. Else N = ker φ is a proper closed subspace; pick z ∈ N^⊥ with ‖z‖ = 1. For any x ∈ H, x − (φ(x)/φ(z)) z ∈ N, so ⟨x − (φ(x)/φ(z)) z, z⟩ = 0, giving φ(x) = φ(z) · ⟨x, z⟩ = ⟨x, φ̄(z) z⟩. Take y = φ̄(z) z.

Riesz representation is the workhorse of functional analysis. It transforms statements about linear functionals (often arising from variational problems, weak solutions, dual measures) into statements about elements of the original space. Lax-Milgram for elliptic PDEs, dual problems in optimization, the construction of unbounded self-adjoint operators — all run on Riesz.

Quantum mechanics in Hilbert space

The standard formulation of quantum mechanics, due to von Neumann (1929-1932):

• The state space of a system is a complex Hilbert space H. Pure states are unit vectors |ψ⟩ ∈ H, with phase identification |ψ⟩ ∼ e^{iθ}|ψ⟩.
• Observables are self-adjoint operators A on H; the spectrum σ(A) ⊂ ℝ is the set of possible measurement outcomes.
• Measurement of A in state |ψ⟩ yields outcome λ with probability |⟨ψ_λ | ψ⟩|², where |ψ_λ⟩ is the eigenvector. Generalization to continuous spectrum uses spectral measures.
• Time evolution is governed by the Schrödinger equation iℏ ∂_t |ψ⟩ = Ĥ |ψ⟩, where Ĥ is the Hamiltonian. Equivalent to |ψ(t)⟩ = e^{−iĤt/ℏ} |ψ(0)⟩, with the right-hand side defined via the spectral theorem for self-adjoint operators.
• Composite systems live in tensor products H_A ⊗ H_B; entanglement is the failure of states to factor.

For a single non-relativistic particle in 3D, H = L²(ℝ³). Position is the multiplication operator (X̂ψ)(x) = x ψ(x); momentum is P̂ = −iℏ ∇. The canonical commutation [X̂, P̂] = iℏ Î holds on a dense domain. The Heisenberg uncertainty principle ‖ψ‖ Δx · Δp ≥ ℏ/2 follows from Cauchy-Schwarz applied to commutators.

Bounded and unbounded operators

A linear operator T: H → H is bounded if there exists C such that ‖Tx‖ ≤ C ‖x‖ for all x; the smallest such C is ‖T‖. The space B(H) of bounded operators is a Banach algebra under composition.

Classes of operators:

• Self-adjoint: T* = T (where T* is defined by ⟨Tx, y⟩ = ⟨x, T*y⟩). Spectrum is real; spectral theorem applies.
• Unitary: U*U = UU* = I. Preserves inner product: ⟨Ux, Uy⟩ = ⟨x, y⟩. Time evolution operators e^{−iĤt/ℏ} are unitary.
• Compact: images bounded sets to relatively compact sets. Compact self-adjoint operators have a discrete spectrum {λₙ} → 0 with eigenvectors forming an ONB (Hilbert-Schmidt theorem). Schmidt decomposition for compact operators between Hilbert spaces.
• Trace-class: compact with Σ |λₙ| < ∞. Density matrices in QM are trace-class self-adjoint with trace 1.
• Hilbert-Schmidt: Σ |λₙ|² < ∞. Form a Hilbert space themselves under ⟨A, B⟩_HS = tr(A*B). Includes integral operators with L² kernels.

Many operators of physical interest are unbounded (defined only on dense subspaces). Position, momentum, and the Hamiltonian of any system with continuous spectrum are unbounded. Self-adjointness for unbounded operators requires care: the operator and its adjoint must have the same domain. von Neumann's theory of self-adjoint extensions handles this; defect indices classify symmetric operators by their possible self-adjoint extensions.

Where Hilbert spaces show up

• Quantum mechanics — entire formalism. Every textbook QM calculation runs on Hilbert space arithmetic. The hydrogen atom's energy levels are eigenvalues of the Hamiltonian; spin-1/2 systems live in ℂ²; spin-J systems in ℂ^{2J+1}; harmonic oscillators in L²(ℝ); free particles in L²(ℝ³). Tensor products handle composite systems and entanglement.
• Quantum chemistry — Hartree-Fock and DFT. Electron orbitals are vectors in L²(ℝ³); the Slater-determinant ansatz lives in antisymmetric tensor products. Variational methods minimize ⟨ψ|Ĥ|ψ⟩ over restricted subspaces of L². Density functional theory replaces the wave function with the electron density but still relies on Hilbert-space variational structure.
• Signal processing — wavelets and frames. Signal compression and analysis use orthonormal bases (DCT, JPEG) or overcomplete frames (Gabor, wavelets) of L²(ℝ). MRA (multiresolution analysis) provides hierarchical orthonormal bases that adapt to local signal features.
• Machine learning — kernel methods. SVMs, Gaussian processes, kernel ridge regression all map data into a reproducing kernel Hilbert space and apply linear algorithms there. The "kernel trick" computes inner products without explicit feature extraction. Mercer's theorem provides the spectral decomposition of kernels.
• Statistical mechanics — equilibrium states. Quantum statistical mechanics works in operator algebras on a Hilbert space: states are positive normalized linear functionals (or density matrices), dynamics is implemented by time evolution operators. Tomita-Takesaki theory and KMS conditions characterize equilibrium states.
• PDE theory — weak solutions. Sobolev spaces H^k are Hilbert spaces; weak solutions of elliptic PDEs are sought there via Lax-Milgram or variational methods. Finite element methods discretize by projecting onto finite-dimensional subspaces of H^k. Spectral methods use the eigenfunctions of the Laplacian as a basis.
• Quantum information — qubits and channels. A qubit is a unit vector in ℂ²; n qubits in ℂ^{2ⁿ}. Quantum gates are unitary operators; measurements are projections. Quantum channels are completely positive trace-preserving maps on density operators. Stinespring's dilation theorem expresses any channel as a unitary on a larger Hilbert space.

History

David Hilbert, in a series of papers from 1904-1910, developed the theory of integral equations with kernels K(x, y) on bounded intervals. Solutions Σ aₙ φₙ to integral equations led him to consider the space ℓ² of square-summable sequences (an, n ≥ 1) — what we now call an "ℓ²-space," though Hilbert thought of it concretely as Fourier coefficient sequences.

Fréchet, F. Riesz, and E. Schmidt extended this to L²-spaces of functions and abstract inner-product structures in the 1900s-1920s. Banach's 1932 "Théorie des opérations linéaires" gave the first systematic abstract theory of normed spaces.

John von Neumann, in his 1929 paper "Allgemeine Eigenwerttheorie Hermitescher Funktionaloperatoren" and in the 1932 book "Mathematische Grundlagen der Quantenmechanik," abstracted the Hilbert space concept (giving it the name) and used it as the foundational framework for quantum mechanics. The spectral theorem for unbounded self-adjoint operators, the rigorous treatment of canonical commutation relations, and the proof of the Stone-von Neumann uniqueness theorem all came from this period.

Modern Hilbert-space theory continues to be central: in mathematical physics (rigorous QFT, axiomatic and algebraic QFT), in the Connes program for noncommutative geometry, in machine learning (RKHS), and in quantum information (qubits, entanglement entropies, quantum error correction). The "infinite-dimensional Euclidean space" remains one of the most productive abstractions in mathematics.