The Riesz Representation Theorem: Every Functional Is Secretly an Integral

The precise statement (two theorems under one name)

Riesz representation (Hilbert space). Let H be a Hilbert space over 𝔽 = ℝ or ℂ with inner product ⟨·,·⟩. For every bounded linear functional L: H → 𝔽 there exists a unique vector y ∈ H such that

• L(x) = ⟨x, y⟩ for all x ∈ H, and
• ‖L‖ = ‖y‖ (the operator norm equals the vector norm).

The map L ↦ y is a bijection H* → H; it is conjugate-linear in the complex case. So H is (conjugate-)isometrically isomorphic to its own dual.

Riesz–Markov–Kakutani (C(X)). Let X be a locally compact Hausdorff space and L: Cc(X) → ℝ a positive linear functional (f ≥ 0 ⇒ L(f) ≥ 0). Then there is a unique regular Borel measure μ on X with L(f) = ∫X f dμ for all f ∈ Cc(X). For compact X, the dual of C(X) is the signed (or complex) regular Borel measures with the total-variation norm.

The picture: a functional is a direction, a measure is a weighting

In finite dimensions this is just linear algebra you already know. Every linear functional on ℝⁿ is x ↦ x · y = ∑ᵢ xᵢ yᵢ for a fixed vector y. A functional is a direction; feeding it a vector reports the component along that direction. Riesz says this picture survives in infinite dimensions — provided you have a genuine inner product and completeness. The single vector y that represents L is the direction perpendicular to the level sets {L = 0}, scaled so the arithmetic comes out right.

For the C(X) version the picture shifts from geometry to weighting. A positive functional assigns a nonnegative number to every nonnegative function, monotonically and additively — exactly what an integral against a mass distribution does. Point evaluation f ↦ f(p) is integration against a Dirac mass δp; averaging is integration against Lebesgue measure. Riesz–Markov says these are all the examples: every positive functional is some measure telling you where the mass lives.

The key idea of the proof (the mechanism)

Hilbert case — orthogonal projection. If L ≡ 0 take y = 0. Otherwise N = ker L is a closed proper subspace (closed because L is continuous). Completeness gives the orthogonal decomposition H = N ⊕ N⊥, and since L ≠ 0 with codimension-1 kernel, N⊥ is one-dimensional; pick a unit vector z ∈ N⊥. Now set y = L(z) z. A one-line check shows x − (L(x)/L(z)) z ∈ ker L, hence is ⊥ to z, which forces ⟨x, y⟩ = L(x). Uniqueness: if ⟨x, y−y′⟩ = 0 for all x, plug x = y−y′.

Riesz–Markov — build the measure from the functional. You don't have a measure; you manufacture one. For an open set U define μ(U) = sup{ L(f) : 0 ≤ f ≤ 1, supp f ⊂ U }, extend to all sets by outer regularity, and prove — using Urysohn's lemma and partitions of unity to splice bump functions — that this is a genuine countably additive regular Borel measure whose integral reproduces L.

Worked example: functionals on ℓ² and on C[0,1]

ℓ² (the Hilbert case, concretely). Fix a sequence y = (yₙ) ∈ ℓ². Then L(x) = ∑ₙ xₙ ȳₙ is bounded with ‖L‖ = ‖y‖₂ by Cauchy–Schwarz. Riesz says the converse is airtight: every bounded functional on ℓ² arises this way from one square-summable y. Recover yₙ by testing on the standard basis: yₙ = L(eₙ). Square-summability of (L(eₙ)) is forced by boundedness.

C[0,1] (the measure case). Point evaluation L(f) = f(½) is positive and linear; its measure is the Dirac mass, ∫ f dδ_½ = f(½). The functional f ↦ ∫₀¹ f(t) dt is Lebesgue measure. A finite sum ∑ cᵢ f(tᵢ) with cᵢ > 0 (a quadrature rule!) is the measure ∑ cᵢ δ_tᵢ. Riesz–Markov guarantees these atomic-plus-continuous measures exhaust all positive functionals — nothing exotic hides in the dual.

Why the hypotheses matter — and what breaks

Completeness is not optional. Take the incomplete inner-product space V of finitely-supported sequences, dense in ℓ². The functional L(x) = ∑ₙ xₙ/n is bounded on V (the coefficients (1/n) are in ℓ²). Its would-be representer is y = (1/n)ₙ, which is not in V. So no y ∈ V represents L; without completeness the theorem fails. The proof needed N = ker L to be closed and to admit an orthogonal complement — both flow from completeness.

Positivity / p < ∞ matter too. In the Lᵖ duality, (L¹)* ≅ L∞ for σ-finite μ, but (L∞)* is strictly larger than L¹ — it contains purely finitely-additive "measures" (Banach limits), so the pattern stops at p = ∞. For C(X) you need regularity to pin down μ uniquely, and Hausdorff + local compactness to run Urysohn's lemma. Drop them and either existence or uniqueness collapses.

Closely related: this is the engine behind the Lax–Milgram theorem and the projection theorem, and it is weaker than the Hahn–Banach theorem (which needs no inner product but gives no uniqueness).

Why it matters: what the theorem unlocks

Riesz representation is the hinge on which functional analysis turns from abstract to computable.

• Weak solutions of PDEs. The Lax–Milgram theorem — the backbone of the finite element method and elliptic PDE theory — is Riesz representation plus a coercive bilinear form. "Find u with a(u,v) = L(v) for all v" is solvable precisely because functionals are inner products.
• Quantum mechanics. Dirac's bra–ket formalism is Riesz: every bra ⟨φ| is the functional ⟨φ, ·⟩ dual to the ket |φ⟩; the theorem legitimizes the identification of states with their duals.
• Adjoints and the spectral theorem. Existence of the Hilbert-space adjoint T* is proved by applying Riesz to x ↦ ⟨Tx, y⟩; self-adjointness underlies the spectral theorem.
• Probability and integration. Riesz–Markov constructs measures (including Lebesgue and Haar measure, and Wiener measure for Brownian motion) from the integral you want them to produce — turning "which measure?" into "which positive functional?"

The three faces of Riesz representation: same slogan (functionals = something concrete), different spaces and hypotheses.

Setting	Functionals represented	Represented by	Essential hypothesis
Hilbert space H (real or complex)	All bounded linear L: H → 𝔽	Unique y ∈ H, L(x) = ⟨x, y⟩, ‖L‖ = ‖y‖	Completeness of H (need closed subspace to have orthogonal complement)
C₀(X), X locally compact Hausdorff	All bounded linear functionals (dual space)	Unique finite signed/complex regular Borel measure μ, L(f) = ∫ f dμ, ‖L‖ = |μ|(X)	X locally compact Hausdorff; regularity of μ
C_c(X) or C(K), K compact	Positive linear functionals (f ≥ 0 ⇒ L(f) ≥ 0)	Unique positive regular Borel measure μ, L(f) = ∫ f dμ	Positivity of L; local compactness + Hausdorff
Lᵖ(μ), 1 ≤ p < ∞, μ σ-finite	All bounded linear functionals	Unique g ∈ Lᵍ, 1/p + 1/q = 1, L(f) = ∫ fg dμ	σ-finiteness (for p = 1); p < ∞ (fails for L∞)