Contribution
No new result; a careful from-scratch exposition by an REU student (mentor Jake Fiedler), following Axler, Rudin and Stein–Shakarchi. It builds inner products, norms, Hilbert spaces and orthogonal projection, proves the Riesz representation theorem, proves it again constructively through orthonormal bases, and derives the Radon–Nikodym theorem from it.
Key insight
The first proof finds the representer as a multiple of a unit vector normal to ker φ, and the author says it “seems to fall from the sky”. The second proof removes the mystery: expand in any orthonormal basis and the representer’s coefficients are the conjugated values of φ on the basis, so Riesz is the infinite-dimensional version of the finite-dimensional dual-basis computation of §3.
Assumptions
- Convention. The inner product is linear in its second argument and conjugate-linear in its first (Def. 1.1), the opposite of LIT-230. So the theorem reads φ(v) = ⟨u, v⟩.
- Completeness. Used once, in the Hilbert projection theorem (Thm 2.2), where the minimizing sequence converges (“only here do we invoke the completeness of the space”).
- Choice. Orthonormal bases exist by Zorn’s lemma (§4); the construction is explicit only where a basis can be built (Gram–Schmidt in the separable case, footnote 11).
- Radon–Nikodym is proved for finite positive μ and ν with ν ≪ μ; the σ-finite, signed and complex cases are said to follow from monotone convergence and the Hahn decomposition, not proved.
Key results
- Thm 1.5 (Cauchy–Schwarz), with equality iff the vectors are dependent; Thm 1.8 (parallelogram law).
- Thm 2.2 (Hilbert projection theorem). In a Hilbert space the distance to a nonempty closed convex set is attained at a unique point. For a closed subspace U, v − v_U ⊥ U, the decomposition is unique, and (U^⊥)^⊥ = U. A subspace is dense iff its orthogonal complement is {0}.
- §3, finite dimensions. With the dual basis, φ ↦ φ′ = Σ conj(φ_i) e_i is conjugate-linear and isometric, and φ(v) = ⟨φ′, v⟩ for any orthonormal basis.
- Thm 3.5 (Riesz). For continuous linear φ on a Hilbert space V there is a unique u with φ(v) = ⟨u, v⟩ for all v, and ‖u‖ = ‖φ‖. Proof: take a unit w ⊥ ker φ and u = conj φ(w) w.
- Thms 4.5–4.8. An orthogonal family is summable iff the squared norms are (an “infinite Pythagorean theorem”); Bessel’s inequality; the closed span of an orthonormal set is {Σ α_k e_k : Σ |α_k|² < ∞}, and v = Σ ⟨e_k, v⟩ e_k.
- Thm 4.9 (Riesz again). For bounded φ and any orthonormal basis {e_k}, u = Σ_k conj φ(e_k) e_k satisfies φ(v) = ⟨u, v⟩, and ‖φ‖ = ‖u‖ = (Σ_k |φ(e_k)|²)
- Thm 5.1 (Radon–Nikodym). If μ, ν are finite positive measures on the same σ-algebra and ν(A) = 0 whenever μ(A) = 0, there is a measurable h with ν(A) = ∫_A h dμ. Proof: φ(f) = ∫ f dν is bounded on L²(σ), σ = ν + μ, so ∫ f dν = ∫ f g dσ for some g; hence ∫ f(1 − g) dν = ∫ f g dμ (eq. 5.2); indicators show 0 ≤ g < 1 almost everywhere; truncations f_k of 1/(1 − g) and monotone convergence give h = g/(1 − g).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Every continuous linear functional on a Hilbert space is the inner product with a unique vector of equal norm | strong | Thm 3.5, full proof |
| C2 | The representer is Σ_k conj φ(e_k) e_k in any orthonormal basis, and ‖φ‖² = Σ_k |φ(e_k)|² | strong | Thm 4.9, full proof |
| C3 | The Radon–Nikodym theorem for finite positive measures follows from Riesz applied on L²(ν + μ) | strong | Thm 5.1, full proof |
| C4 | The extension to σ-finite, signed and complex measures is routine | weak | asserted in one sentence, not proved |
Concepts
- Unordered sum. L is the sum of a family if every ε has a finite Ω with every finite Ω′ ⊇ Ω within ε of L (Def. 4.4).
- Orthonormal basis. A maximal orthonormal set; its span is dense, not all of the space (the ℓ² example with a_i = 2
- Standard isomorphism. The conjugate-linear isometry V′ → V of §3.
Connections
The Wikipedia article (LIT-230) gives the same theorem, the same orthonormal-basis formula and a different convention. The Radon–Nikodym derivative proved here is the object MINE’s Donsker–Varadhan dual is tight at (NOTE-828, LIT-256); this paper’s proof is von Neumann’s.
Bearing on the record
- No THEORY, QUESTION or CLAIM cites this paper. Readings name it as the record’s secondary account of Riesz (NOTE-289, NOTE-239, NOTE-828).
- The seed’s summary needs one correction. LIT-243 and the skim say the representer is “the sum of φ(e_k)e_k”. The paper writes Σ_k conj φ(e_k) e_k (p. 12; the bar is lost in the text layer). The two agree on real spaces only.
- Cite it for what NOTE-289 wants. It proves uniqueness and the isometry for arbitrary Hilbert spaces, and gives the basis formula. It says nothing about kernels, RKHS or factorisation, as the skim noted.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Student expository work. Correct as read, but cite a textbook (Axler, Rudin, Stein–Shakarchi) for anything leaned on.
- Typographical slips:
- Thm 4.8 writes “span(e_Ψ) = Σ α_k e_k” for the closure of the span.
- In the projection-theorem proof one term lacks its factor 2 (“2‖p − q_m‖² + ‖p − q_n‖²”).
- In the proof of Thm 4.9, “⟨u, e_k⟩ = φ(e_k)” is right under its convention only because u carries the conjugated coefficients.
- The Radon–Nikodym proof is for finite measures; the general case is only asserted.
Open questions
- None the paper raises. Whether the record needs the σ-finite Radon–Nikodym theorem in full would decide whether a textbook source is wanted.