{"id":"b2b35819-a3b1-4139-b3dd-fab8107e85ec","arxiv_id":"2605.25567","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Ambient denoising score matching on manifold-supported data is biased by an O(σ^2) curvature term (1/2 W_H − Ric^#)∇ log q, which vanishes exactly on S^2.","lead":"The paper shows that when data live on a curved surface inside a high-dimensional space, standard denoising score matching learns a score function distorted by a curvature-dependent bias, and it derives an explicit formula for that bias. A practical takeaway is that on a 2-sphere the distortion vanishes, which explains why plain methods already work well on spherical data.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5.2's O(σ²) coefficient is missing a third-derivative term; the proof incorrectly asserts ∇∆mσ(0)=0 despite cubic terms in log Mσ.","rationale":"The reader's weakest assumption concerns the isotropy/exact-support setting, but my concern is an internal gap in the proof of the central expansion. The proof of Theorem 5.2 in Section F.3 applies Lemma F.7 to aσ = e^{f+mσ} and needs ∇∆mσ(0) to compute the σ² coefficient. It asserts this quantity vanishes because the leading s-dependent term of log Mσ is quadratic, but the O(||s||³) remainder contributes a generally nonzero constant after two derivatives. The explicit counterexample (graph y=x²+x³, uniform q) yields a direct 1D integral showing an O(σ²) bias, contradicting the claimed formula. This is load-bearing because the headline result is the explicit coefficient (16); if the coefficient is wrong for generic asymmetric embeddings, the central claim fails, though the S^2 and flat special cases may survive due to symmetry. The proposed test is cheap, exact at the level of 1D quadrature, and would settle the issue definitively.","tokens_in":23038,"tokens_out":30380,"duration_ms":337135,"concrete_test":"Compute the Rao-Blackwellized target for a compact C^5 arc of M={(x,x²+x³)} near 0, q uniform, at z=0, by 1D quadrature of rσ(0)=σ^{-2} ∫ s e^{-s²/(2σ²)} (1+2s²+6s³)(1-2s²-2s³) ds / ∫ e^{-s²/(2σ²)} (1+2s²+6s³)(1-2s²-2s³) ds for σ=10^{-2}, 10^{-3}, 10^{-4}; fit the σ² coefficient. If it is ≈12 (independent of q' and ∇log q), Theorem 5.2 is refuted; if it is 0, the concern is refuted.","verdict_should_be":"REJECT","load_bearing_attack":"In Section F.3, Proposition F.6 states log Mσ(s)=cσ - ½⟨Ric s,s⟩ + O(||s||³)+..., and then asserts ∇mσ(0)=0, ∇²mσ(0)=-Ric, and ∇∆mσ(0)=0. The last assertion is false for typical embeddings: the O(||s||³) remainder contains a cubic term C(s) (from the cubic part of the graph-coordinate normal component h(s) and from the tube Jacobian), and ∇∆C(0) is generally a nonzero constant vector. By Lemma F.7, ½∇∆mσ(0) contributes directly to the σ² coefficient of rσ(z). Example: M = graph of y=x²+x³ in R² at z=0, with uniform q. Jgr(s)=√(1+(2s+3s²)²)=1+2s²+6s³+..., Fσ(s)=1-2h(s)=1-2s²-2s³ (since W_0=2), so log Mσ(s)=4s³+...; hence ∇∆m(0)=24 and the exact conditional expectation gives rσ(0)=12σ²+O(σ⁴), while Theorem 5.2 predicts 0 (uniform q ⇒ ∇log q=0, b_q=0, g_ext=0). Thus Eq. (16) is incomplete; a third-derivative extrinsic term must be added. The S^d and T^2 cases happen to be symmetric/even in the graph coordinate, so their numerical checks do not detect this.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies ambient Gaussian denoising score matching (DSM) when the latent distribution is supported on a compact smooth submanifold M ⊂ R^D. It defines the Rao-Blackwellized tangent target rσ(z) = E[Tσ | π(X) = z], proves that among estimators depending on the observation only through the nearest-point projection π(X), rσ is the unique L2-optimal predictor, and shows that the raw tangent target Tσ has conditional variance d/σ^2 + O(1) while rσ has bounded variance. The paper proves the leading-order identification rσ(z) = ∇_M log q(z) + O(σ^2) and claims an explicit second-order expansion with an intrinsic Tweedie term and an extrinsic curvature term g_ext = (1/2 W_H − Ric^#) ∇_M log q. It specializes to spheres, observes that the extrinsic term vanishes on S^2, and provides a finite-sample local-averaging rate.","tokens_in":23385,"tokens_out":11165,"duration_ms":105532,"significance":"If the second-order expansion were correct, the paper would give a sharp and useful account of what ambient DSM learns on manifold-supported data, with a computable bias and a theoretical explanation for the benign behavior on S^2. The variance-collapse theorem (Thm 4.2) and the leading-order identification (Eq. 10) are solid and of clear value: the d/σ^2 risk floor is an exact, parameter-free statement, and the flat-case reduction (Prop 5.1) is clean. The paper also makes falsifiable coefficient predictions (Fig. 2), which is a strength. However, the central second-order formula (Thm 5.2) is incomplete: a cubic term in the geometric weight contributes to the σ^2 coefficient through ∇Δ m(0), as a simple graph example shows. Thus the paper's main quantitative claim is not reliable as stated, although the leading-order and variance results appear unaffected.","major_comments":[{"comment":"The proof asserts ∇Δmσ(0)=0 on the grounds that the leading s-dependent term in log Mσ is quadratic. This is false: the O(||s||^3) remainder generally contains a cubic term C(s), and ∇ΔC(0) is a nonzero constant that contributes to rσ(z) at order σ^2 through Lemma F.7. Concretely, take a compact extension of the graph y = x^2 + x^3 at z=0 with uniform q. Then Jgr(s)=1+2s^2+6s^3+... and Fσ(s)=1−2s^2−2s^3 (since W=2), so log Mσ(s)=4s^3+...; hence ∇Δm(0)=24 and the exact posterior mean gives rσ(0)=12σ^2+O(σ^4), whereas Eq. (16) predicts 0. Thus Theorem 5.2 is missing a third-derivative term.","section":"Appendix F.3, Prop F.6 and Eq. (64)"},{"comment":"The stated second-order expansion is not correct in general; the missing term involves third derivatives of the embedding and is not captured by W_H or Ric^#. The numerical checks in Figure 2 are on S^d and T^2, which are even/odd symmetric at the evaluation point so the cubic term averages to zero; they therefore do not detect the error. The manuscript should give a corrected expansion with a rigorous remainder and test it on a non-symmetric manifold (e.g., an ellipsoid or the graph example).","section":"Theorem 5.2 / Corollary 5.4"},{"comment":"The proof strategy expands Jgr and Fσ only to quadratic order and then, in Prop F.6, concludes that the remainder 'contains no term linear in s at order σ^2.' That condition does not imply ∇Δmσ(0)=0. A valid second-order derivation must expand the geometric weight to cubic order and include the σ^2/2 ∇(Δm)(0) term in the Gaussian-moment expansion (Lemma F.7). The algebra needs to be redone; the stated operator (1/2 W_H − Ric^#) is at most part of the correct coefficient.","section":"Section F.3, Lemmas F.4-F.6"}],"minor_comments":[{"comment":"There are several reference inconsistencies: the proof of Prop F.6 refers to 'Theorem F.4 and Theorem F.5' where these are Lemmas F.4 and F.5, and in the application of Lemma F.7 it is called 'Theorem F.7'. These should be corrected.","section":"Appendix F.3"},{"comment":"The text states the variance is d/σ^2 + O(1), but Figure 1 displays the second moment E||Tσ||^2. Since E||Tσ||^2 = Var(Tσ) + ||E[Tσ]||^2 and the mean is O(1), the plot is consistent; the caption should clarify which quantity is shown.","section":"Section 4.3 and Figure 1"},{"comment":"The claim for the product of two circles S^1(R1)×S^1(R2) in R^4 that the operator is ½ diag(1/R1^2, 1/R2^2) is stated without derivation. A short verification would improve readability and trust in the frame formula.","section":"Remark F.3 / Eq. (59)"}],"recommendation":"major_revision","confidential_remarks":"The referee report identifies a concrete counterexample to Theorem 5.2, so the central second-order expansion is not correct as stated. The variance-collapse theorem and leading-order identification appear sound and publishable, but the σ^2 coefficient must be corrected by including the cubic contribution to ∇Δ m(0) or by proving it vanishes under the stated assumptions. The authors should also update the numerical verification to include a manifold without the even symmetry of S^d or T^2."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Divit has a genuinely good idea here: the Rao-Blackwellized target is the right object for ambient DSM on manifold data, the d/sigma^2 variance floor and collapse theorem are clean, and the leading-order identification r_sigma = grad_M log q + O(sigma^2) is solid. The flat-case reduction is also exact and useful. But the paper's main theorem, the explicit sigma^2 coefficient, is not correct as stated. In the graph-coordinate proof (Appendix F.3, Proposition F.6), log M_sigma(s) is expanded as const - 1/2 <Ric s,s> + O(||s||^3), and the paper then asserts grad laplacian m_sigma(0) = 0. That is false for typical embeddings: the O(||s||^3) remainder has a cubic part, and grad laplacian of that cubic part is generally a nonzero constant that feeds directly into the sigma^2 term via the Gaussian-moment expansion.\n\nA concrete check: take M as the graph of y = x^2 + x^3 in R^2, z = 0, uniform q. Then J_gr(s) = 1 + 2s^2 + 6s^3 + ..., F_sigma(s) = 1 - 2s^2 - 2s^3, so M_sigma(s) = 1 + 4s^3 + ..., and the exact conditional expectation gives r_sigma(0) = 12 sigma^2 + O(sigma^4). Theorem 5.2 predicts 0 because the intrinsic score is zero. So Eq. (16) is missing a third-derivative extrinsic term. The S^d and T^2 cases are symmetric enough that the cubic term vanishes, which is why the numerical checks look good, but the general formula does not hold.\n\nThe rest of the paper should not be judged by this mistake. The variance-collapse theorem and the Rao-Blackwell canonicality calculation are independent and appear correct; the finite-sample rate is only sketched, and there is no code or error bars, but those are minor compared with the missing term. I would send this to peer review—the core ideas deserve referee time, and the error is fixable by adding the missing term and rechecking. But the current version is not acceptable. A reader who wants the S^2/T^2 coefficients should be directed to the symmetric cases, not to the general theorem.","headline":"Solid canonicality and variance-collapse core, but the headline sigma-squared formula is false as stated: it is missing a third-derivative extrinsic term.","tokens_in":23870,"tokens_out":9107,"would_cite":true,"duration_ms":90647,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62R30","53B25"],"pacs":[],"model":"deepseek-v4-flash","headline":"Ambient Gaussian denoising score matching on manifold-supported data, after conditioning on the nearest-point projection, estimates the true Riemannian score to first order in noise, with an explicit σ² curvature bias that vanishes on S².","keywords":["denoising score matching","manifold hypothesis","Rao-Blackwellization","Riemannian score","nearest-point projection","curvature bias","Weingarten operator","sphere S^2"],"falsifier":"Compute rσ by high-accuracy quadrature on a compact embedded manifold with known intrinsic score, such as the flat torus T² embedded in R³ with a wrapped-Gaussian density, at several small σ, and form the ratio (rσ − ∇_M log q − σ² b_q)/(σ² ∇_M log q). Theorem 5.2 predicts the constant ½ W_H − Ric^# (on T², +½ for the chosen embedding); any measurable σ-dependence at leading order, or any value inconsistent with the Weingarten and Ricci computation, would refute the expansion.","tokens_in":22909,"feed_emoji":"📐","tokens_out":6338,"duration_ms":69166,"temperature":0.7,"pith_summary":"Denoising score matching is well-posed on manifold-supported data only at positive noise, and the object it learns has been unclear. This paper identifies the canonical object: the conditional expectation of the tangent denoising target given the nearest-point projection π(X), the unique L²-optimal predictor among all estimators that see the data only through π(X). Conditioning removes a singular normal-fiber noise channel whose variance diverges as d/σ², replacing it with a target of bounded variance. The paper then computes the small-noise expansion of this target: it equals the intrinsic Riemannian score up to first order, with an explicit second-order bias combining an intrinsic smoothing term and an extrinsic curvature operator built from the Weingarten and Ricci tensors. On the 2-sphere the extrinsic term cancels exactly, giving a quantitative explanation for why ambient denoising score matching has worked well on spherical scientific data despite the manifold singularity.","feed_headline":"Rao-Blackwellization recovers the true score on manifolds","feed_subtitle":"Conditioning on the nearest-point projection kills the d/σ² noise; the curvature bias vanishes exactly on S².","key_machinery":"The load-bearing object is rσ(z) = E[ P_{T_z M}(Z−X)/σ² | π(X)=z ], the conditional expectation of the tangent denoising target given the nearest-point projection π(X) onto M. This Rao-Blackwellized target is the unique L²-optimal predictor among estimators depending on X through π(X), and it removes the raw target's irreducible d/σ² variance floor. The expansion is derived by changing to tubular coordinates around M and normal coordinates at z, obtaining a fiber posterior that is a centered Gaussian in tangent displacements times smooth geometric corrections, and applying a manifold Stein identity. The curvature operator ½ W_H − Ric^# — mean-curvature Weingarten minus Ricci endomorphism — c","core_discovery":"The paper's central discovery is Theorem 5.2: for a compact C^5 embedded manifold of positive reach, with strictly positive latent density q and isotropic Gaussian corruption, the Rao-Blackwellized tangent target satisfies rσ(z) = ∇_M log q(z) + σ²( b_q(z) + g_ext_M(z) ) + o(σ²), where b_q is the intrinsic Tweedie correction ½∇_M(Δ_M log q + ‖∇_M log q‖²) and g_ext_M = (½ W_H − Ric^#_z) ∇_M log q, with W_H the Weingarten operator in the mean-curvature direction and Ric^# the Ricci endomorphism. The first term recovers the true intrinsic Riemannian score; the second-order correction is the finite-σ bias introduced specifically by ambient Gaussian corruption on curved supports. On spheres the","pith_inferences":["If the expansion holds at moderate σ, the same formula suggests a practical debiasing step for ambient score models on product manifolds and rotation groups such as SO(3), where the dimension-dependent coefficient 1 − d/2 would predict systematic shrinkage or amplification of the learned score.","The S² cancellation is a coincidence of the Einstein identity at dimension two, not a sign that ambient denoising score matching is generally unbiased; using only S² to validate manifold-aware score matching may understate the bias present on other manifolds.","A natural testable extension is anisotropic Gaussian corruption: replacing the isotropic covariance in the noise model should change both the variance floor to a weighted trace of the covariance and the curvature operator, and the same quadrature approach could verify the modified formula.","The finite-sample bandwidth rate in the appendix, roughly (σ⁻²N)^(-2/(d+2)), suggests Rao-Blackwellized preprocessing translates directly into improved sample complexity; direct empirical comparison on non-spherical product manifolds would quantify the gain."],"forward_implications":["Provides a population-level answer to what ambient denoising score matching learns on manifold-supported data: the intrinsic Riemannian score, up to an explicit order-σ² curvature bias.","Rao-Blackwellization is not a constant-factor improvement: it removes the singular d/σ² variance channel, shown to be an irreducible Bayes-risk floor for any estimator using only a fiber-collapsing summary.","The ambient-versus-intrinsic bias is computable from the Weingarten and Ricci tensors and the learned score, so practitioners can accept, subtract, or avoid it depending on target accuracy.","On S² the extrinsic bias is exactly zero, so ambient denoising score matching there matches intrinsic methods up to the intrinsic smoothing bias; on S¹, S³, and higher spheres the bias is nonzero and grows with dimension.","Flat cases reduce exactly to standard d-dimensional Gaussian denoising score matching, pinning down the baseline against which curvature effects must be measured."],"fun_headline_variants":["Rao-Blackwellized score matching fixes manifold noise bias","Nearest-point conditioning eliminates DSM divergence on manifolds","True intrinsic score recovered; S² curvature bias vanishes exactly","Ambient noise on manifolds: Rao-Blackwellized target recovers true score"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the latent distribution sits exactly on a compact C^5 manifold with positive reach and is corrupted by isotropic Gaussian noise; if the data only approximate the manifold or the noise is anisotropic, the d/σ² variance floor and the curvature coefficient ½ W_H − Ric^# need not hold.","fun_headline_variants_meta":{"raw":{"variants":["Rao-Blackwellized score matching fixes manifold noise bias","Nearest-point conditioning eliminates DSM divergence on manifolds","True intrinsic score recovered; S² curvature bias vanishes exactly","Ambient noise on manifolds: Rao-Blackwellized target recovers true score"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1186,"prompt_tokens":758,"completion_tokens":428,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":355}},"tokens_in":502,"tokens_out":428,"duration_ms":5287,"temperature":1.0,"reasoning_tokens":355,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T13:10:44.034112+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute rσ by high-accuracy quadrature on a compact embedded manifold with known intrinsic score, such as the flat torus T² embedded in R³ with a wrapped-Gaussian density, at several small σ, and form the ratio (rσ − ∇_M log q − σ² b_q)/(σ² ∇_M log q). Theorem 5.2 predicts the constant ½ W_H − Ric^# (on T², +½ for the chosen embedding); any measurable σ-dependence at leading order, or any value inconsistent with the Weingarten and Ricci computation, would refute the expansion.","supporting_citations":[],"review_version":2}