{"id":"c4f7505c-94ce-405a-ab23-1f88157afa65","arxiv_id":"2606.06179","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Only the gradient component of score errors affects marginal distributions in diffusion models, so L2 error can be arbitrarily large with perfect match; this yields an impossibility result, a gradient-only KL bound, and a Sobolev estimator that correlates better with quality.","lead":"The paper shows that diffusion models can have arbitrarily large L2 score matching errors yet perfectly match the target distribution, since only the gradient component of the error affects marginal Fokker-Planck dynamics while the solenoidal part is invisible. A smart generalist might read it to see why standard training losses in generative models do not directly guarantee sample quality and to learn about a new estimator focused on the observable error component.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Helmholtz-Hodge may fail to isolate components on R^d without decay conditions at infinity","rationale":"The reader's weakest_assumption correctly isolates the decomposition step as load-bearing. The concern is technical rather than philosophical: the geometric separation is asserted but its validity on the natural domain of diffusion models is not self-evident from the abstract and requires explicit verification in the proofs.","tokens_in":1770,"tokens_out":365,"duration_ms":34849,"concrete_test":"Extract the precise function space and boundary conditions stated for the Helmholtz-Hodge decomposition in the proof of the impossibility result; if none are given beyond L², construct a compactly supported solenoidal perturbation on R^d, extend it by a slow-decay tail, and numerically integrate the resulting Fokker-Planck operator to check whether the marginal density remains unchanged.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on the assertion that the Helmholtz-Hodge decomposition of the score error e = s_θ − ∇log p cleanly separates a gradient part (which alone enters the marginal Fokker-Planck) from a solenoidal part (structurally invisible). On unbounded domains the decomposition e = ∇ϕ + ∇×A is unique only under additional integrability or decay assumptions on e (typically |e(x)| = o(1/|x|) or e ∈ L^{2} with suitable weighted spaces). The abstract invokes this decomposition without stating such conditions; if they are absent, boundary terms at infinity can allow the solenoidal component to contribute to div(p·e) and therefore to the marginal evolution. This would invalidate both the impossibility result (arbitrarily large L² error with identical marginals) and the tightened KL bound that depends only on the gradient component.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper claims that the L² score-matching objective is not the right intrinsic measure of marginal quality for learned diffusion models: via a Helmholtz-Hodge decomposition of the score error into gradient and solenoidal components, only the gradient part enters the marginal Fokker-Planck dynamics while the solenoidal part is invisible to them. This geometry is used to prove (i) an impossibility result that no monotone function of the L² error can uniformly lower-bound any divergence between learned and target marginals, (ii) a tightened KL upper bound that depends only on the observable gradient component (improving the standard Girsanov bound), and (iii) a tractable estimator of the gradient component via a dual Sobolev identity that empirically correlates better with sample quality than the full L² error.","tokens_in":1951,"tokens_out":582,"duration_ms":36926,"significance":"If the decomposition is valid under the paper's conditions, the work supplies a geometric explanation for the frequent mismatch between L² score error and generative performance, supplies a strictly tighter marginal-space bound, and introduces a practical estimator whose empirical superiority is a concrete strength. The dual-Sobolev construction for the estimator is a technically clean contribution.","major_comments":[{"comment":"The Helmholtz-Hodge decomposition of the score error e = s_θ − ∇log p is applied on R^d (abstract and the section stating the three results) without any stated decay or integrability conditions at infinity. On unbounded domains the decomposition e = ∇ϕ + ∇×A is unique and the boundary terms vanish in ∫ div(p e) only under assumptions such as |e(x)| = o(1/|x|) or e belonging to a suitable weighted L² space; absent these, the solenoidal component can contribute to the marginal evolution and the impossibility result plus the tightened KL bound both fail.","section":"Section stating the decomposition and the three results"},{"comment":"The impossibility result (that no monotone function of the L² error uniformly lower-bounds divergences) is load-bearing on the claim that the solenoidal component can be made arbitrarily large while leaving the marginal Fokker-Planck equation unchanged. The construction must explicitly verify that the chosen vector field satisfies the decay conditions needed for the decomposition to isolate a purely invisible solenoidal part; otherwise the counter-example is not rigorous.","section":"Proof of the impossibility result"}],"minor_comments":[{"comment":"The empirical section reports that the gradient-component estimator correlates substantially better with sample quality, but does not state the precise datasets, number of runs, or statistical tests used; adding these details would strengthen reproducibility without affecting the central claims.","section":"Empirical evaluation of the estimator"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the careful reading and the precise identification of the technical conditions required for the Helmholtz-Hodge decomposition on unbounded domains. We address both major comments below and will strengthen the manuscript accordingly.","responses":[{"response":"We agree that explicit decay and integrability conditions are necessary for the uniqueness of the decomposition and the vanishing of boundary terms on R^d. In the revised manuscript we will add a new subsection (placed immediately before the three main results) that states the precise assumptions: the score error e belongs to a weighted L² space with weight (1+|x|)^{d+ε} for some ε>0, which implies |e(x)|=o(1/|x|) at infinity. Under these conditions the solenoidal component is divergence-free with respect to the weighted measure and does not enter the Fokker-Planck equation. We will also verify that all densities and score networks considered in the paper (smooth, sub-Gaussian tails) satisfy the required integrability, so the three results remain valid. These additions make the geometric argument fully rigorous.","revision_made":"yes","referee_comment":"The Helmholtz-Hodge decomposition of the score error e = s_θ − ∇log p is applied on R^d (abstract and the section stating the three results) without any stated decay or integrability conditions at infinity. On unbounded domains the decomposition e = ∇ϕ + ∇×A is unique and the boundary terms vanish in ∫ div(p e) only under assumptions such as |e(x)| = o(1/|x|) or e belonging to a suitable weighted L² space; absent these, the solenoidal component can contribute to the marginal evolution and the impossibility result plus the tightened KL bound both fail."},{"response":"We accept that the counter-example construction must be accompanied by an explicit verification of the decay conditions. In the revision we will replace the current sketch with a fully detailed construction: we take a compactly supported, divergence-free vector field A (hence e=∇×A) that is C^∞ and identically zero outside a large ball. Compact support immediately satisfies |e(x)|=o(1/|x|) and all weighted integrability requirements. We then show that the corresponding score perturbation leaves the marginal Fokker-Planck equation invariant while making the L² norm arbitrarily large. The revised proof will cite the new subsection on decay conditions to confirm that the decomposition isolates a purely solenoidal, invisible component.","revision_made":"yes","referee_comment":"The impossibility result (that no monotone function of the L² error uniformly lower-bounds divergences) is load-bearing on the claim that the solenoidal component can be made arbitrarily large while leaving the marginal Fokker-Planck equation unchanged. The construction must explicitly verify that the chosen vector field satisfies the decay conditions needed for the decomposition to isolate a purely invisible solenoidal part; otherwise the counter-example is not rigorous."}],"tokens_in":1553,"tokens_out":626,"duration_ms":26202,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main point is that standard L2 score matching does not control distributional quality in diffusion models. A model can have huge L2 error yet match the target distribution perfectly because the error vector field splits into a gradient component that enters the marginal dynamics and a solenoidal component that does not.\n\nThe paper applies a Helmholtz-Hodge decomposition to the score error and uses it to prove three things: no monotone function of the full L2 error can uniformly bound any divergence between learned and target distributions; the KL divergence is bounded by a term that depends only on the gradient component, which tightens the usual Girsanov bound; and a dual Sobolev identity gives a practical estimator for that gradient component which correlates better with sample quality than the raw L2 error.\n\nThese results are new. The geometric split and the impossibility claim do not appear in earlier diffusion analyses, and the estimator is a concrete addition.\n\nThe soft spot is the decomposition itself. On unbounded domains the Helmholtz-Hodge split is unique only under decay or integrability conditions at infinity. The abstract states the decomposition without mentioning those conditions. If boundary terms allow the solenoidal part to affect div(p e), both the impossibility result and the gradient-only KL bound lose their force. The paper would need to state and verify the required assumptions for the claims to hold.\n\nThis work is for people who analyze score-based models theoretically. It gives a clear reason why training loss and generation quality can diverge and supplies a better diagnostic. It deserves peer review because the argument is sharp and directly challenges a quantity used in most theoretical bounds, even if the functional-analytic details need checking.","headline":"L2 score error can be arbitrarily large while marginals match exactly because only the gradient part of the error drives Fokker-Planck dynamics.","tokens_in":2450,"tokens_out":403,"would_cite":false,"duration_ms":31720,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A diffusion model can perfectly match the target distribution while incurring arbitrarily large L² score error because only the gradient part of the error affects the marginal dynamics.","keywords":["diffusion models","score matching","Helmholtz-Hodge decomposition","Fokker-Planck equation","generative modeling","distributional divergence","score error geometry"],"falsifier":"Train a score network whose error vector field has a large solenoidal component yet produces samples whose empirical distribution converges to the target; if the samples fail to match whenever the full L² error is large, the claim is falsified.","tokens_in":2685,"feed_emoji":"","tokens_out":655,"duration_ms":27466,"temperature":0.7,"pith_summary":"The paper shows that minimizing the L² score matching error is not guaranteed to produce good marginal distributions in diffusion models. A model can match the target distribution exactly even when the full L² error grows without bound. The reason is a geometric split of the score error vector field: its gradient component drives the Fokker-Planck evolution of the probability density, while the solenoidal component has no effect on it. This decomposition explains why standard bounds that use the full L² error are loose and why a new estimator focused on the gradient component tracks sample quality more closely.","feed_headline":"Diffusion models match targets with unbounded L2 score error","feed_subtitle":"Only the gradient component of the error affects marginal dynamics; the solenoidal part is invisible.","key_machinery":"Helmholtz-Hodge decomposition of the score error vector field, which isolates the gradient component that governs marginal Fokker-Planck dynamics from the solenoidal component that does not.","core_discovery":"Applying the Helmholtz-Hodge decomposition to the difference between the learned and true score functions isolates a gradient component that alone enters the marginal Fokker-Planck equation; the orthogonal solenoidal component is invisible to the evolution of the probability measure. Consequently no monotone function of the L² score error can serve as a uniform lower bound on any divergence between the learned and target distributions, while an upper bound on the KL divergence can be stated using only the gradient component.","pith_inferences":["Training objectives could be redesigned to penalize only the gradient component rather than the full L² norm.","The same decomposition may apply to other generative models that rely on estimated vector fields or scores.","Monitoring the size of the solenoidal component during training could diagnose unnecessary model capacity without harming distributional accuracy."],"forward_implications":["No monotone function of the L² score error can uniformly lower-bound any divergence between learned and target distributions.","The KL divergence between learned and target distributions admits an upper bound that depends only on the gradient component of the score error.","A dual Sobolev identity yields a tractable estimator of the gradient component that correlates more strongly with sample quality than the full L² error."],"fun_headline_variants":["Diffusion models match targets despite arbitrary L2 score errors","Only gradients in score errors affect marginal dynamics","L2 score errors cannot bound distributional divergences","Solenoidal components render score errors invisible"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The Helmholtz-Hodge decomposition of the score error correctly identifies a gradient component that alone controls the marginal probability evolution.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion models match targets despite arbitrary L2 score errors","Only gradients in score errors affect marginal dynamics","L2 score errors cannot bound distributional divergences","Solenoidal components render score errors invisible"]},"model":"grok-4.3","cost_usd":0.007471,"raw_usage":{"total_tokens":3451,"prompt_tokens":710,"num_sources_used":0,"completion_tokens":55,"cost_in_usd_ticks":74712000,"prompt_tokens_details":{"text_tokens":710,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2686,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":710,"tokens_out":55,"duration_ms":33124,"temperature":1.0,"reasoning_tokens":2686,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-30T10:57:29.744452+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Train a score network whose error vector field has a large solenoidal component yet produces samples whose empirical distribution converges to the target; if the samples fail to match whenever the full L² error is large, the claim is falsified.","supporting_citations":[],"review_version":2}