{"id":"724adefb-fa5a-4719-9ac2-08c2c9029076","arxiv_id":"2505.23023","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A kernel density estimator using the intrinsic dimension d achieves MSE rate O(n^{-2m/(d+2m)}) on d-rectifiable sets with an m-order tangent approximation, and O(n^{-4/(d+4)}) on sets that are smooth almost everywhere.","lead":"This paper proves that a modified kernel density estimator, which uses the intrinsic dimension of the data, converges to the true density on d-rectifiable sets including algebraic varieties and semi-algebraic sets. The result extends earlier manifold-only guarantees and shows that ambient dimension does not slow the rate for such sparse or structured data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1.1's rate requires Assumption A.2, which is not an automatic consequence of d-rectifiability and is omitted from the statement; as written, the theorem is false.","rationale":"The reader's weakest assumption is exactly the load-bearing concern: Assumption A.2 is unstated in Theorem 1.1 but is required for the claimed rate. I agree with this identification. The paper's proof under A.2 uses a standard bias-variance decomposition and is essentially correct, so the appropriate response is a conditional acceptance with revision, not rejection. I also examined Lemma 4.1's Jacobian expansion and found a concrete error in the power of h, but the lemma's O(h^2) conclusion appears salvageable because the Jacobian correction contributes at order O(h^2) after integration and the evenness of the radial kernel kills the first-order tangential term. The D=50 numerical experiment is suspect because random sparse supports make the kernel have almost no same-support neighbors at small n, but this affects the illustration rather than the central theorem. On balance, no independent fatal flaw beyond the overstated theorem statement was found, so the reader's CONDITIONAL verdict should stand unchanged.","tokens_in":9587,"tokens_out":25946,"duration_ms":246999,"concrete_test":"Construct Ω as the graph of f(x) = ∫_0^x (1/log(2/t)) dt on a bounded interval, smoothed if necessary, which is a C^1 rectifiable curve whose derivative has local modulus of continuity ~1/log(1/h) on a positive-measure set. For a smooth radial kernel K, compute the bias integral in Lemma 3.1 at such points and verify that the error is O(1/log^2(1/h)), not O(h^m) for any m > 0. This directly falsifies Theorem 1.1 as stated and confirms that Assumption A.2 must be an explicit hypothesis.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central rate claim, Theorem 1.1, is stated for every d-rectifiable set Ω, with m described only as 'a measure of how well Ω is approximated by its approximate tangent spaces.' Yet Section 3 proves the theorem only under Assumption A.2, which postulates that the integral approximation error in the blow-up is O(h^m) for some fixed m > 0. Rectifiability alone (Definition 2.1 and Theorem 2.3) guarantees only the limit as h → 0, not any polynomial rate. This is not a minor omission: there exist d-rectifiable sets for which no m > 0 satisfies A.2 on a positive H^d-measure set. For example, take a C^1 curve that is the graph of an antiderivative of a continuous function whose local modulus of continuity is ~1/|log(1/h)| on a fat Cantor set; the blow-up error for a smooth radial kernel is then O(1/log^2(1/h)), which is not O(h^m) for any m > 0. Thus Theorem 1.1 as stated is false. The correct theorem is conditional on A.2, and the abstract and theorem statement must be revised to say this. A secondary issue is that the Jacobian computation in Lemma 4.1 has an incorrect power: the lower block of Dψ_h should be O(h), so the metric correction is O(h^2), not O(h^4); the lemma's O(h^2) conclusion still appears to survive, but the proof needs correction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the kernel density estimator \\hat p_n(x) = (1/(n h_n^d)) \\sum_i K(||X_i-x||/h_n) for probability measures supported on a d-rectifiable set \\Omega \\subset R^D. Under a quantitative tangent-approximation assumption (A.2) that postulates an O(h^m) error in the blow-up integrals of K and K^2, Section 3 derives the bias-variance bound MSE[\\hat p_n(x)] = O(h^{2m}) + O(1/(n h^d)), leading to the rate O(n^{-2m/(d+2m)}) for h_n \\asymp n^{-1/(d+2m)}. Section 4 proves that on sets that are smooth H^d-almost everywhere the exponent m=2 holds, yielding the classical rate O(n^{-4/(d+4)}) and covering algebraic varieties and semi-algebraic sets. Section 5 reports a numerical experiment for d-sparse vectors in ambient dimensions 5, 10, and 50. The main result is sound only as a conditional theorem; the unconditional statement of Theorem 1.1 is not justified.","tokens_in":9882,"tokens_out":19082,"duration_ms":205649,"significance":"The conditional rate result is a clean and useful generalization of manifold KDE: it separates the geometry of the domain (the tangent approximation exponent m) from the usual bias-variance tradeoff, and Theorem 1.2 gives a short proof of the classical rate for almost-everywhere-smooth sets without reach or exponential-map assumptions. The paper also correctly identifies algebraic varieties and semi-algebraic sets as natural examples. However, the central theorem as stated overclaims: for general rectifiable sets no positive m need exist, so the rate is not a theorem about all rectifiable sets. The fix is local (restate Theorem 1.1 and the abstract conditionally on A.2), and the remaining issues are proofread-level, so the contribution is viable after a major revision.","major_comments":[{"comment":"The theorem is stated for every d-rectifiable set, but the proof relies on Assumption A.2, which postulates a fixed exponent m>0 with O(h^m) in the blow-up integral (6). Rectifiability (Theorem 2.3) guarantees only the limit h to 0, not any polynomial rate. There exist d-rectifiable sets (for example, C^1 curves whose tangent direction has logarithmic modulus of continuity on a fat Cantor set) for which the error in (6) is not O(h^m) for any m>0, so the asserted rate in Theorem 1.1 is false as stated. The formal theorem and the abstract must be restated conditionally: if A.2 holds for some m>0, then the rate holds; without A.2, only consistency is claimed (as the text itself notes in Section 3).","section":"Section 1, Theorem 1.1; Section 3, Assumption A.2"},{"comment":"The Jacobian expansion in the proof is incorrect. Since phi(0)=0 and D phi(0)=0, the Taylor expansion gives D phi(hz)=h D^2 phi(0)(z)+O(h^2), so the lower block of D psi_h is O(h), not O(h^2). Consequently D psi_h(z)^T D psi_h(z)=Id+h^2(D^2 phi(0)(z))^T(D^2 phi(0)(z))+O(h^3), and J_h(z)=1+O(h^2), not 1+O(h^4). The displayed equalities in the proof involving h^4 and (D phi(0)(z))^T(D phi(0)(z)) are false. The lemma's final O(h^2) conclusion still survives because the true O(h^2) Jacobian correction is compatible with the stated error, but the proof must be corrected.","section":"Section 4, Lemma 4.1"}],"minor_comments":[{"comment":"The displayed decomposition is misprinted; it should read MSE[\\hat p_n(x)] = E[(\\hat p_n(x)-p(x))^2] = (E[\\hat p_n(x)]-p(x))^2 + E[(\\hat p_n(x)-E[\\hat p_n(x)])^2].","section":"Section 3, bias-variance decomposition"},{"comment":"Several integrals over Omega use dP(y) where p(y) dH^d(y) is meant (for example 'p(y)dP(y)' in Lemma 3.1 and Lemma 3.2); since P has density p with respect to H^d, the dP should be dH^d in those lines.","section":"Lemmas 3.1 and 3.2"},{"comment":"The phrase 'with probability 1' following the deterministic MSE bound is not meaningful; the intended statement is presumably 'for H^d-almost every x'.","section":"Theorems 1.1 and 1.2"},{"comment":"The displayed truncated Gaussian kernel is not the kernel used in the paper's estimator: it contains a factor 1/h inside K, and its normalizing constant is the one-dimensional Gaussian CDF, so the integral of K over R^d is not 1 for d=3. Please give the correctly normalized d-dimensional kernel actually used in the code.","section":"Section 5, kernel formula"},{"comment":"With d=3 and n at most 10,000, the expected number of samples in a given coordinate 3-plane is n / binom(50,3) approximately 0.51, so the reported small MSE requires an explanation of the test-point selection and how the estimator is evaluated when no sample lies near the test point.","section":"Section 5, D=50 experiment"},{"comment":"There are several typos that should be corrected, including 'Suppse', 'denstiy', 'Coordiante', 'outisde', 'interecting', and 'to to'.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper's main theorem as stated is false, but the conditional theorem and the smooth case are sound, and the issues are repairable without new ideas. I recommend major revision rather than rejection. The author should be asked to make Theorem 1.1 explicitly conditional on Assumption A.2 and to correct the proof of Lemma 4.1 and the numerical section."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Main take: the paper has a useful idea—KDE on rectifiable sets with a rate controlled by the blow-up approximation exponent—but the headline theorem as stated is false because it omits the quantitative assumption A.2. The proof works once A.2 is assumed, but rectifiability alone does not give any polynomial rate.\n\nThe genuinely new piece is the conditional rate n^{-2m/(d+2m)} for general d-rectifiable sets under a quantitative tangent approximation condition, and the observation that algebraic varieties and semi-algebraic sets fall out as the smooth-a.e. case with m=2. The paper is honest that the smooth case is covered with minor modifications to existing manifold KDE arguments. The bias-variance computation is standard and correct under A.2.\n\nSoft spots:\n1. Theorem 1.1 and the abstract overstate the hypotheses. Assumption A.2 is not a consequence of rectifiability. There are rectifiable sets (e.g., a C^1 curve whose tangent varies with logarithmic modulus) where the blow-up error is not O(h^m) for any m>0. So the theorem needs to be rewritten as conditional on A.2, with m as part of the hypothesis. This is not cosmetic; as written, the theorem is false.\n2. Lemma 4.1 contains a Jacobian expansion error. In the parametrization ψ_h(z) = (z, φ(hz)/h), the lower block of Dψ_h is Dφ(hz), which is O(h) (since Dφ(0)=0), not O(h^2). So the metric correction is O(h^2) and the Jacobian is 1+O(h^2), not 1+O(h^4). The O(h^2) conclusion still holds, but the proof needs correction.\n3. Minor: Lemmas 3.1 and 3.2 write dP(y) where dH^d(y) is intended (or the density should not appear twice). Also the D=50 numerical experiment is underspecified; it would be good to clarify how test points are chosen and why the MSE curves are so well-behaved.\n\nWho this is for: statisticians working on low-dimensional structure in high dimensions, especially sparse vectors and low-rank matrices. The paper is short, mostly readable, and the main idea is worth discussing. It deserves peer review, but a serious referee should require the theorem statement and abstract to be corrected, the Jacobian proof fixed, and the experiment clarified.\n\nRecommendation: send to referee, but expect substantial revision.","headline":"Conditional on a strong unstated assumption, the KDE rate argument is correct; the headline theorem overclaims and the Jacobian proof has an error, but the rectifiable-set perspective is worth refereeing.","tokens_in":10415,"tokens_out":4511,"would_cite":false,"duration_ms":43065,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62G07","28A75"],"pacs":[],"model":"deepseek-v4-flash","headline":"A modified kernel estimator converges to the true density on rectifiable sets, with the classical rate when the set is smooth almost everywhere.","keywords":["kernel density estimation","rectifiable sets","approximate tangent spaces","intrinsic dimension","algebraic varieties","semi-algebraic sets","convergence rate","sparse data"],"falsifier":"In the Section 5 experiment, the theory predicts that log MSE plotted against log $n$ has slope $-4/(d+4) = -4/7$ for $d=3$; a slope measurably different from this would contradict Theorem 1.2's rate. For Theorem 1.1's premise, one can test Assumption A.2 directly on a rectifiable curve whose tangent angle oscillates non-polynomially in scale, for instance like $1/|\\log h|$, and check whether the bias in Lemma 3.1 is truly $O(h^m)$ for some fixed $m$.","tokens_in":9361,"feed_emoji":"📈","tokens_out":10667,"duration_ms":115813,"temperature":0.7,"pith_summary":"This paper extends kernel density estimation from smooth manifolds to $d$-rectifiable sets, the measure-theoretic generalization of manifolds that includes sparse vectors, low-rank matrices, and positive semidefinite matrices. The estimator replaces the ambient dimension $D$ with the intrinsic dimension $d$ in the bandwidth normalization. The paper proves that its mean squared error is $O(n^{-2m/(d+2m)})$ whenever the set's blow-ups approach the approximate tangent space at rate $O(h^m)$. For sets that are smooth almost everywhere, including algebraic varieties and semi-algebraic sets, this gives the classical rate $O(n^{-4/(d+4)})$, independent of the ambient dimension. A numerical experiment on $d$-sparse vectors illustrates the predicted convergence.","feed_headline":"Rectifiable sets get manifold-speed density estimates","feed_subtitle":"A dimension-aware kernel estimator hits the classical rate on sparse, low-rank, and semidefinite data.","key_machinery":"The load-bearing object is the approximate tangent space $T_x\\Omega$: a $d$-plane in $\\mathbb{R}^D$ such that integrals over the blown-up set $(\\Omega - x)/h$ converge to integrals over $T_x\\Omega$ as $h\\to 0$. Rectifiability guarantees such spaces exist for $H^d$-almost every $x$ (Theorem 2.3), and Assumption A.2 quantifies the convergence as $O(h^m)$ for the kernel $K$ and $K^2$. An isometry from $T_x\\Omega$ to $\\mathbb{R}^d$ turns the tangent integral into $\\int_{\\mathbb{R}^d} K(\\|u\\|)\\,du = 1$, which closes the bias computation; Lemma 4.1 uses a second-order Taylor expansion in normal coordinates to identify $m=2$ on locally smooth sets.","core_discovery":"The central claim is Theorem 1.1: for a $d$-rectifiable set $\\Omega \\subset \\mathbb{R}^D$ carrying a probability measure with density $p$ with respect to $H^d$, the estimator $\\hat{p}_n(x) = \\frac{1}{h_n^d} \\sum_{i=1}^n K(\\|X_i - x\\|/h_n)$ satisfies $\\mathrm{MSE}[\\hat{p}_n(x)] = O(n^{-2m/(d+2m)})$ with probability one, provided the approximate tangent spaces of $\\Omega$ approximate the blow-ups at polynomial rate $O(h^m)$ (Assumption A.2). Theorem 1.2 specializes to sets that are locally smooth manifolds almost everywhere, where the tangent approximation error is $O(h^2)$ by a Taylor expansion, yielding the classical rate $O(n^{-4/(d+4)})$; Whitney stratification shows every $d$-dimensional algebraic variety and semi-algebraic set falls into this class. The proof is the standard bias-variance split: the tangency assumption bounds the bias by $O(h^m)$, variance is $O(1/(n h^d))$, and $h_n \\propto n^{-1/(d+2m)}$ balances the two.","pith_inferences":["Because the paper gives no bound on $m$ for a non-smooth rectifiable set, the practical content of Theorem 1.1 depends on a future geometric analysis of tangency rates; for many fractal-like rectifiable sets the rate may fail to be polynomial.","The theorem only guarantees convergence at $H^d$-almost every point, so behaviour at singular points such as the origin in the sparse-vector variety is open; a pointwise analysis there could require boundary-aware kernels.","The bias term is $O(h^m)$ and variance is $O(1/(n h^d))$, so the optimal bandwidth could in principle be selected by estimating $d$ and $m$ from data, turning this into an adaptive procedure."],"forward_implications":["On algebraic varieties and semi-algebraic sets, including sparse vectors, low-rank matrices, and positive semidefinite matrices, density estimates converge at $O(n^{-4/(d+4)})$ with no dependence on the ambient dimension $D$.","The bandwidth choice $h_n \\propto n^{-1/(d+4)}$, familiar from manifold density estimation, remains the right scaling for these non-manifold spaces.","When the set is merely rectifiable without a tangency rate, the estimator is still consistent but no convergence speed is available.","The method does not require local parametrizations or exponential maps, so it applies on spaces where manifold chart arguments break down."],"supporting_citations":[{"why":"Establishes the classical univariate kernel density estimator and its bias-variance analysis, which the paper generalizes.","marker":"[Par62]"},{"why":"Extends kernel density estimation to multivariate data in $\\mathbb{R}^D$, providing the baseline whose rate slows with ambient dimension.","marker":"[Cac66]"},{"why":"Supplies the rectifiability characterization via approximate tangent spaces (Theorem 2.3), the foundation of Assumption A.2.","marker":"[Sim83]"},{"why":"Introduced the intrinsic-dimension kernel density estimator on known Riemannian submanifolds, the setting this paper extends.","marker":"[Pel05]"},{"why":"Generalized submanifold density estimation to unknown manifolds; the paper builds on its kernel-localization analysis.","marker":"[OG09]"},{"why":"Studied kernel density estimation on unknown submanifolds with reach conditions and used tangent-space parametrizations, the technique Lemma 4.1 adapts.","marker":"[BH20]"},{"why":"Provides the Whitney stratification used in Lemma 2.4 to show algebraic and semi-algebraic sets are rectifiable.","marker":"[Whi57]"}],"fun_headline_variants":["Intrinsic-dimension kernel density hits classical rates","Classical density rates recovered on rectifiable sets","Algebraic varieties get manifold-speed density estimation","Density estimator adapts to data's intrinsic shape","Sharp density rates for sparse and low-dimensional data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire rate rests on the unverified exponent $m$ in Assumption A.2: the set's scaled-up neighborhood must approach its tangent plane with error $O(h^m)$, but ordinary rectifiability guarantees only that the error goes to zero, not that it does so polynomially, and the paper gives no way to compute $m$ for a general non-smooth set.","fun_headline_variants_meta":{"raw":{"variants":["Intrinsic-dimension kernel density hits classical rates","Classical density rates recovered on rectifiable sets","Algebraic varieties get manifold-speed density estimation","Density estimator adapts to data's intrinsic shape","Sharp density rates for sparse and low-dimensional data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000694,"raw_usage":{"total_tokens":3118,"prompt_tokens":902,"completion_tokens":2216,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":2145}},"tokens_in":518,"tokens_out":2216,"duration_ms":20107,"temperature":1.0,"reasoning_tokens":2145,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:57:28.523466+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the Section 5 experiment, the theory predicts that log MSE plotted against log $n$ has slope $-4/(d+4) = -4/7$ for $d=3$; a slope measurably different from this would contradict Theorem 1.2's rate. For Theorem 1.1's premise, one can test Assumption A.2 directly on a rectifiable curve whose tangent angle oscillates non-polynomially in scale, for instance like $1/|\\log h|$, and check whether the bias in Lemma 3.1 is truly $O(h^m)$ for some fixed $m$.","supporting_citations":[],"review_version":1}