# Review: one-parameter theory that fits every proper subset of n observations but not all n ## 1. Precise restatement For every integer n ≥ 2 there is a setup with the following parts: - n observations y_1 = … = y_n = 0. - A one-parameter model f_i(θ), with θ in a compact interval Θ. - Each f_i is continuous in θ. The setup has three properties: - **(a) Proper subsets fit exactly.** For every proper subset S ⊊ {1,…,n}, some θ_S satisfies f_i(θ_S) = 0 for all i ∈ S. - **(b) The full set cannot fit exactly.** No θ satisfies f_i(θ) = 0 for all i. - **(c) The misfit is robust.** min_θ max_i |f_i(θ)| = δ_n > 0. So at any tolerance below δ_n the full dataset rejects the model, while every proper subset still fits. Hidden or implicit assumptions: - **The model depends on n.** Both the prediction functions and the parameter range change with n. The quantifier is "for each n there exists a theory", not "one theory for all n". - **Complexity grows with n.** The prediction functions are polynomials of degree 2(n−1). - **"Fit" means sup-norm residual.** It is either exact equality or residual within a tolerance. Any norm on ℝⁿ gives the same qualitative result. - **The data are idealised zeros.** The choice of zero data is inessential, since it is just a shift. - **Rival theory A plays no logical role.** It only sets the philosophical context. ## 2. Checking the argument The parameter is θ ∈ [1,n], and the model is f_i(θ) = ∏_{j≠i} (θ−j)². **Proper subsets fit.** A proper subset S omits some index k. Set θ = k. - For each i ∈ S we have i ≠ k, so the product defining f_i includes j = k. - That factor is (k−k)² = 0, so f_i(k) = 0. - An empty S is trivial. ✓ **The full set does not fit.** - If θ is not an integer, every factor (θ−j)² is positive, so every f_i(θ) > 0. - If θ = k is an integer, then f_k(k) = ∏_{j≠k} (k−j)² = ((k−1)!(n−k)!)² > 0. ✓ **The margin is positive.** g(θ) = max_i f_i(θ) is continuous, as a maximum of finitely many polynomials. It is positive at every point of the compact interval [1,n]. So g attains a positive minimum δ_n. ✓ The compactness assumption is not even needed. On all of ℝ, g(θ) → ∞ as |θ| → ∞, so a positive minimum still exists. **Independent re-derivation.** Let Z_i be the zero set of f_i, so Z_i = {1,…,n} \ {i}. - The sets Z_i have empty total intersection. - Every n−1 of them share a common point. - Positivity of the margin follows from closedness and compactness. Equivalently, the sets {θ : f_i(θ) ≤ τ} also have empty intersection for small τ > 0, by the finite-intersection property. **Numerical check, n = 3.** - The predictions are f_1 = (θ−2)²(θ−3)², f_2 = (θ−1)²(θ−3)², and f_3 = (θ−1)²(θ−2)². - At θ = 1: f_2 = f_3 = 0 and f_1 = 4. - At θ = 2: f_2 = 1. At θ = 3: f_3 = 4. - At θ = 1.5: (f_1, f_2, f_3) = (0.5625, 0.5625, 0.0625). - The minimum of g on [1,3] is roughly 0.5. This is positive, so the claim is confirmed. ## 3. Attempts to break it **Could a single fixed theory work for nested, growing datasets?** No. - Suppose one fixed model had the property for the first n settings and also for the first n+1. - The (n+1)-case requires the first n settings to be jointly satisfiable. - The n-case says they are not. Contradiction. - So the result genuinely needs the theory to be constructed anew for each n. This is consistent with the text's remark that the unbounded resource is algebraic complexity, but the one-line claim does not say it. **Is the margin non-degenerate?** Yes. - δ_n is positive and in fact grows with n in absolute terms. - Rescaling f_i would change the size of δ_n, not its positivity. **Is the parameter genuinely continuous and only one?** Yes. It is a single real parameter in an interval. **Does the philosophical conclusion follow?** Yes, in a limited sense. - Parameter count alone does not bound the size of a minimal inconsistent subset of the data. - The broader conclusion, that redundancy arguments need structural assumptions, follows as a modest negative point. No counterexample to the stated claim survives. ## 4. Novelty This is a variant of well-known phenomena. - **Failure of Helly-type theorems for nonconvex sets.** In one dimension, convex sets have Helly number 2. The zero sets here are nonconvex finite sets whose Helly number is n. - **Helly-type bounds for algebraic sets.** For polynomials of bounded degree, the Helly number is bounded by the dimension of the space of polynomials of that degree. A degree-2(n−1) univariate polynomial space has dimension 2n−1 ≥ n, so the construction fits within known bounds. - **Parameter count does not bound capacity.** The classic example is that the one-parameter family sin(θx) has infinite VC dimension. The specific Lagrange-style construction is a neat illustration, not a new result. ## 5. Fairness of the summary The sentence is literally true under the natural reading "for every n, some one-parameter theory…". However: - It omits that the theory must be tailored to n. - It omits that the theory's degree, or complexity, grows without bound. - A fixed theory cannot achieve this for nested datasets. The supporting text discloses these points, so the summary is acceptable, with that caveat. ```json {"verdict": "correct with caveats", "novelty": "special case or variant of known results", "fair_summary": "yes", "confidence": 5, "one_line": "The construction is valid (every proper subset fits at θ=k, the full set has a strictly positive minimax misfit), but the theory must be rebuilt with growing polynomial degree for each n, and the result is a Helly-number/VC-style illustration rather than a new finding."} ```