{"id":"0815c653-5989-4d41-b5a2-dd01f220d7a2","arxiv_id":"2509.02971","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"For generative flows on multiscale scientific fields, the noise spectrum should be at least as rough as the data's, and a scale-adaptive schedule can tame the terminal-time stiffness of rougher noise.","lead":"This paper designs the noise and time schedule for flow-based generative models on scientific data with sharp multiscale Fourier spectra. It shows that matching the noise roughness to the data improves stability and accuracy, and proposes a schedule that handles rougher-than-data noise.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Schedule theorem is Gaussian-only; the NS success rests on a hand-set μ⋆ with no check that the ½|log μ⋆| Lipschitz bound applies to the learned non-Gaussian drift.","rationale":"I read this as a theory-plus-experiment paper. The Gaussian results (Prop 3.1, Prop 5.1) are internally consistent: the computations in Appendix B check out, and the matched-spectrum experiments on Gaussian and Allen-Cahn data support the roughness heuristic. The concern is not that the math is wrong; it is that the paper's most striking empirical result—accurate Navier-Stokes spectra in 10 RK4 steps with a designed schedule—is outside the proven regime. Prop 3.3 requires compact support in the Cameron-Martin space and only gives t ≤ 1−δ, while the schedule's quantitative bound is specifically Gaussian. The paper is honest about the compact-support issue in Sec. 4.2.3, but it does not address the transfer from the Gaussian schedule to the nonlinear NS drift. That gap is addressable by a sensitivity/robustness check. The reader's weakest assumption identified the compact-support hypothesis and the hand-fitted μ⋆; I partly agree, but I locate the load-bearing issue one step further along: even if the compact-support assumption were satisfied for white noise, the quantitative schedule guarantee would still be unproven for the non-Gaussian target. The empirical code is available, so the proposed check is feasible. No change to the conditional verdict; the paper should be accepted with the condition that the NS experiment be shown robust to μ⋆ or the claim be softened.","tokens_in":18697,"tokens_out":10719,"duration_ms":115194,"concrete_test":"Rerun the 128×128 Navier-Stokes experiment with schedule (5.1) for μ⋆ ∈ {10^{-4}, 10^{-5}, 10^{-6}} and with the linear schedule, all at 10 RK4 steps and identical UNet training. If the designed schedule is not uniformly more accurate than linear in enstrophy at k = 2^6, or if the best μ⋆ is far from the empirically read 10^{-5}, then the schedule's benefit does not depend on the proposed scale-adaptive principle. To make contact with Prop 5.1, also measure the Lipschitz constant of the trained drift by finite differences along high-frequency perturbations at several t; if it exceeds ½|log μ⋆| by a large factor, the Gaussian bound is not the operative mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central practical claim—that scale-adaptive schedules keep generation efficient when the noise is rougher than the data—rests on Proposition 5.1, which bounds the drift Jacobian by ½|log μ⋆| for Gaussian targets with C0 and C1 mutually diagonalizable and μ⋆ the smallest eigenvalue of C1 C0^{-1}. In the Navier-Stokes experiment (Sec. 4.2.3 and 5.2), the target is non-Gaussian, the drift is a trained UNet rather than the linear Gaussian map, and μ⋆ = 10^{-5} is read from the data-to-noise enstrophy ratio at the highest resolved mode. Proposition 3.3 supplies only existence of a Lipschitz constant for compactly supported targets on t ∈ [0,1−δ]; it does not quantify that constant or cover t near 1. Thus no stated theorem gives the quantitative efficiency gain claimed for the non-Gaussian case. The observed 10-step RK4 improvement could be specific to this μ⋆, this network, and this noise normalization. Section 6 lists automated spectral estimation as future work, implicitly conceding μ⋆ is hand-tuned. This unvalidated transfer is the weakest link in the paper's central argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies noise distribution and interpolation schedule design within the stochastic interpolants / flow matching framework, targeting scientific data with multiscale Fourier spectra. For Gaussian targets, it shows that if the noise covariance C0 is smoother than the target covariance C1 in the sense that C1 C0^{-1} is unbounded, the drift field of the generative ODE has unbounded operator norm near t=0 (Prop. 3.1); conversely, matched-spectrum noise keeps the drift bounded and, with the proposed scalar schedule (5.1), gives a drift Jacobian bounded by (1/2)|log μ⋆| (Prop. 5.1). For general targets, a Lipschitz bound is stated under a compact-support assumption in the Cameron–Martin space of the noise (Prop. 3.3). Numerical experiments on Gaussian random fields, stochastic Allen–Cahn invariant measures, and stochastic Navier–Stokes invariant measures illustrate that spectrum-matched noise improves efficiency in the tractable cases, while for the Navier–Stokes case a scale-adaptive schedule with white noise improves the enstrophy spectrum with only 10 RK4 steps.","tokens_in":19025,"tokens_out":8317,"duration_ms":97269,"significance":"If the results hold, the paper makes a useful contribution by connecting noise roughness, drift regularity, and integration cost in flow-based generative models, and by giving an explicit, resolution-robust scalar schedule for Gaussian targets whose Lipschitz constant grows only logarithmically with the spectral range. The Gaussian derivations in Props. 3.1 and 5.1 are clean, and the release of code supports reproducibility. The function-space perspective and the demonstration that standard white noise can be inferior to spectrum-matched noise for smooth Gaussian/Allen–Cahn targets are valuable. However, the central practical claim for complex non-Gaussian targets such as Navier–Stokes is not backed by a quantitative theorem: Prop. 5.1 is Gaussian-only, while Prop. 3.3 only gives existence of a Lipschitz constant under assumptions that likely fail for the Navier–Stokes invariant measure. The Navier–Stokes experiment also uses a hand-fitted μ⋆ read from the data spectrum. These gaps limit the significance of the paper as a general theory, though they may be addressable by reframing and additional experiments.","major_comments":[{"comment":"The quantitative efficiency claim for the scale-adaptive schedule is proved only for Gaussian targets with C0 and C1 mutually diagonalizable (Prop. 5.1). In the Navier–Stokes experiment (Sec. 4.2.3 and 5.2), the target is non-Gaussian, the drift is a trained UNet rather than the linear Gaussian map, and μ⋆ = 10^{-5} is read off the data-to-noise enstrophy ratio at k=26. Prop. 3.3 supplies only existence of a Lipschitz constant on [0,1−δ] under compact support in the Cameron–Martin space V; it does not quantify the constant, and the compact-support assumption is not verified for the Navier–Stokes invariant measure. Indeed, the paper itself notes in Sec. 4.2.3 that the NS invariant measure is not in the Cameron–Martin space of the spectrum noise, and no analogous statement is made for white noise. Thus no stated theorem justifies applying the ½|log μ⋆| bound to the learned non-Gaussian dri","section":"§5.2, Prop. 5.1, Prop. 3.3"},{"comment":"The parameter μ⋆ = 10^{-5} is selected directly from the data spectrum (the enstrophy ratio at the highest resolved mode). This is a fitted value, not a prediction, and Section 6 lists automated spectral estimation as future work. The observed 10-step RK4 improvement may be specific to this μ⋆, this network, and this noise normalization. To make the scale-adaptive claim credible, the paper should report sensitivity of the results to μ⋆ (e.g., varying it over several orders of magnitude) and, ideally, propose a data-driven estimator for μ⋆. Without this, the Navier–Stokes experiment is a demonstration on a hand-tuned parameter rather than a validation of a design principle.","section":"§5.2, Fig. 6"}],"minor_comments":[{"comment":"In the experiment descriptions, the interpolant is written as It = αt z + βt z; the second factor should be βt x1 (or βt x0, depending on notation). This typo appears in both sections.","section":"§4.2.1, §4.2.2"},{"comment":"The displayed formula for Cov(˙I_t, I_t) Cov(I_t)^{-1} omits the C0 factors: it should be (˙α_t α_t C0 + ˙β_t β_t C1)(α_t^2 C0 + β_t^2 C1)^{-1}. The subsequent line is correct, but the intermediate expression is inconsistent.","section":"Appendix B, proof of Prop. 3.1"},{"comment":"The Lipschitz bound for the conditional expectation F is missing the factor β_t/α_t^2 that appears in (C.7). While this factor can be absorbed into the constant Cδ in (C.3), the displayed bound (C.10) is not correct as written and the proof should be adjusted.","section":"Appendix C, Eq. (C.7)–(C.10)"},{"comment":"The statement that 'with more RK4 steps, the result using spectrum noise does not improve' is not supported by a figure or table. If this is an important negative result, it should be documented; otherwise it should be removed or qualified.","section":"§4.2.3"},{"comment":"The proposition assumes the eigenvalues of C1 C0^{-1} satisfy 1 ≥ μ1 ≥ ... ≥ μd. The paper should state what happens if the largest eigenvalue exceeds 1, since the schedule formula (5.1) and the monotonicity step in the proof rely on μ ≤ 1.","section":"Prop. 5.1"},{"comment":"The energy/enstrophy spectra are averaged over ensembles, but no error bars or ensemble sizes are reported. Given that some differences between methods appear modest at low wavenumbers, error bars would strengthen the empirical claims.","section":"Figures 2–6"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely acceptable after revision. The main risk is overclaiming in the abstract/conclusion relative to what is proved: the non-Gaussian schedule transfer is heuristic, and μ⋆ is hand-fitted. I would ask for either a non-Gaussian quantitative result (even for a restricted class, e.g., log-concave perturbations) or a clear reframing, plus sensitivity analysis for μ⋆ in the Navier–Stokes example. Also note that reference [10] appears to be closely related work by the same group on Lipschitz-guided schedule design; the differentiation between the present contribution and [10] should be made more explicit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper earns its place at the table. The formal Gaussian results are done properly, the scale-adaptive schedule is a genuinely useful addition, and the experiments show a real effect. But the gap between what is proved and what the Navier-Stokes experiment claims is larger than the paper lets on.\n\nWhat is new: the roughness condition for noise, made precise via C1 C0^{-1} boundedness (Prop 3.1), and the scalar schedule (5.1) with Jacobian bound 1/2 |log μ⋆| (Prop 5.1). The latter is simple and clever, and the proof is short. The Allen-Cahn and Gaussian experiments with spectrum-matched noise are convincing, and the code is released. Credit where due: the paper says plainly in Sec 4.2.3 that Proposition 3.3 does not apply to Navier-Stokes. That is honest.\n\nSoft spots. First, the general target theory (Prop 3.3) assumes compact support in the noise Cameron-Martin space and only gives existence of a Lipschitz constant on t in [0,1−δ]; it does not quantify the constant or cover the terminal-time stiffness that the rest of the paper is about. Second, the key Navier-Stokes result with rough white noise uses μ⋆ = 10^{-5}, read off the data spectrum. That is fitting the schedule to the data, not predicting it; the paper even lists automated spectral estimation as future work. So the central practical claim — 10-step RK4 with designed schedule works for non-Gaussian Navier-Stokes — is supported by one experiment with one hand-set number, no error bars, and no comparison to multiscale baselines. I don't think that kills the work; the principle is plausible and the Gaussian proof backs it up. But the Navier-Stokes section is a demonstration, not a validation of the theory.\n\nAlso note: the claim that matched-spectrum noise improves numerical efficiency over standard white noise for near-Gaussian targets is well supported. The broader abstract claim about complex non-Gaussian targets is only heuristically supported.\n\nBottom line: for someone working in generative models for scientific data, this is a useful read. I would cite the Gaussian schedule result. It deserves a serious referee, but the referee should push the authors to either extend Prop 5.1 to a quantitative non-Gaussian bound or soften the Navier-Stokes claims, and to add repeats and error bars.","headline":"Worth a real read: clean Gaussian analysis and a useful schedule, with the Navier-Stokes win resting on a hand-tuned parameter and a loose non-Gaussian bridge.","tokens_in":19467,"tokens_out":1801,"would_cite":true,"duration_ms":19942,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H15","62M40","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Flow-based generative models of multiscale scientific fields require noise whose Fourier spectrum decays no faster than the target spectrum; otherwise the drift becomes unbounded near t=0, and a scale-adaptive schedule restores numerical ef","keywords":["stochastic interpolants","multiscale Fourier spectra","noise roughness","Cameron–Martin space","Lipschitz drift","scale-adaptive schedules","function-space generative models","Navier–Stokes invariant measure"],"falsifier":"For a Matérn target with regularity s1 and noise with lower roughness s0 < s1, Proposition 3.1 predicts ∥B(t)∥ ~ 1/t near t=0; measuring the minimum RK4 step size needed to keep the generated energy spectrum within a fixed tolerance at N=32, 64, and 128 should show a clear resolution-dependent blow-up if the claim is right. Conversely, if a Gaussian-noise generative model with s0 > s1 still produces accurate fine-scale spectra with a fixed small number of steps, the drift-blowup mechanism is not the controlling factor.","tokens_in":18621,"feed_emoji":"🌊","tokens_out":9347,"duration_ms":97561,"temperature":0.7,"pith_summary":"Flow-based generative models build a field sample by integrating a learned drift from noise into data. The paper's core point is that for scientific data with wide Fourier spectra, the noise must not be smoother than the target: if the noise covariance decays faster in Fourier space than the data covariance, the drift operator becomes unbounded near the initial time and the generative ODE is ill-conditioned as resolution grows. For Gaussian targets this is a precise operator condition; for general targets, compact support in the noise's Cameron–Martin space guarantees a bounded, Lipschitz drift. When the correct roughness is known, matching the noise spectrum to the target spectrum cuts the number of integration steps needed to reproduce fine-scale spectra. When the target is too non-Gaussian to be matched by Gaussian noise, a scalar scale-adaptive schedule keeps the drift's Lipschitz constant at half the log of the spectral ratio and yields accurate Navier–Stokes enstrophy spectra with only 10 RK4 steps.","feed_headline":"Noise must be at least as rough as the data in generative flows","feed_subtitle":"Matching noise roughness to data keeps flow drift bounded and captures fine scales cheaply.","key_machinery":"The central object is the stochastic interpolant and its conditional-expectation drift b_t(x)=E[İ_t | I_t=x], which defines a generative ODE whose law matches the data law at t=1. Gaussian analysis reduces it to the operator B(t)=(α̇α C0 + β̇β C1)(α^2 C0 + β^2 C1)^{-1}; the ratio C1 C0^{-1} controls whether the drift stays bounded near t=0, and the Cameron–Martin space of the noise plays the analogous role for general targets. The other load-bearing piece is the scalar scale-adaptive schedule (5.1), α_t = sqrt((μ*−μ*^t)/(μ*−1)), β_t = sqrt((μ*^t−1)/(μ*−1)), built from the smallest eigenvalue ratio μ* of C1 C0^{-1}; it makes the drift's Lipschitz constant 1/2 |log μ*| instead of growing polyn","core_discovery":"Working in the stochastic interpolant picture, the paper establishes a roughness-matching principle: for the interpolant I_t = α_t z + β_t x_1, the drift b_t(x)=E[İ_t | I_t=x] is well posed only if the noise z is at least as rough as the target x_1. For Gaussian measures this is exact (Proposition 3.1): b_t(x)=B(t)x with B(t)=(α̇α C0 + β̇β C1)(α^2 C0 + β^2 C1)^{-1}, and an unbounded C1 C0^{-1} forces ∥B(t)∥→∞ as t→0. For general targets, compact support of the data in the noise's Cameron–Martin space gives a bounded, Lipschitz drift (Proposition 3.3). The paper then shows two regimes: matched-spectrum noise reproduces fine scales with very few steps when the target's spectrum is known, while","pith_inferences":["The logarithmic dependence on μ* suggests the schedule's benefit should persist at higher resolution and in 3D, where the smallest resolved eigenvalue ratio shrinks only logarithmically; this is an extrapolation the paper does not run.","The heuristic that sets μ* from the ratio of the smallest target spectral value to the noise plateau could be automated, e.g. by estimating the spectrum and choosing μ* before training, rather than fixing it by hand.","The same Cameron–Martin reasoning may guide non-Gaussian noise families: choosing a noise whose Cameron–Martin space exactly contains the data support could combine the efficiency of spectrum-matched noise with the flexibility needed for non-Gaussian targets.","Integrated with per-scale or hierarchical modeling, the schedule could be applied separately to different wavenumber bands rather than globally, possibly improving accuracy on intermittent fields."],"forward_implications":["White noise is too smooth for smooth data: using it as the noise in a flow model on a Matérn-like field makes the drift unbounded as t→0, so errors at fine scales cannot be controlled as resolution increases.","When the data's fine-scale Fourier structure is known, noise with a matched spectrum reaches accurate spectra with as few as 5 RK4 steps, while white noise needs 20–80 steps and still degrades on finer grids.","The well-posedness criterion transfers to general targets: if the data lives compactly inside the noise's Cameron–Martin space, the drift is bounded and Lipschitz on [0,1−δ].","For rougher-than-data noise, the scale-adaptive schedule (5.1) caps the drift Lipschitz constant at 1/2 |log μ*|, turning a polynomially stiff integration problem into a logarithmically mild one.","Using white noise plus the designed schedule with μ* read from the spectrum gives accurate Navier–Stokes enstrophy spectra at 128×128 with only 10 RK4 steps."],"supporting_citations":[{"why":"Introduces stochastic interpolants; supplies the interpolant process and the drift-as-conditional-expectation formulation used throughout.","marker":"[2]"},{"why":"Gives the general stochastic interpolant framework, including the ODE and SDE formulations and the drift estimation loss.","marker":"[3]"},{"why":"Provides the Gaussian measure and Cameron–Martin space theory used in Proposition 3.1 and in defining 'rougher than' in Proposition 3.3.","marker":"[5]"},{"why":"Supplies the Cameron–Martin space definitions and properties the general-target argument relies on.","marker":"[20]"},{"why":"Provides the infinite-dimensional diffusion-model setting and the conditional-expectation formula used in the proof of Proposition 3.3.","marker":"[43]"},{"why":"Establishes that schedule (5.1) minimizes the averaged squared drift gradient, supporting the schedule's efficiency claim.","marker":"[10]"},{"why":"Proves ergodicity of the 2D Navier–Stokes equation with degenerate stochastic forcing, making the target invariant measure well defined.","marker":"[21]"}],"fun_headline_variants":["Roughness matching: noise must be as rough as data","Matched noise roughness stabilizes flow-based generation","Scale-adaptive flows: noise roughness key to fine scales","Rough noise for rough data: a principle for generative flows","Noise roughness principle cuts cost in multiscale generation"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The general Lipschitz-drift guarantee assumes the target distribution is compactly supported inside the noise's Cameron–Martin space; when that assumption fails, as it does for the Navier–Stokes invariant measure under matched-spectrum noise, the paper falls back on rougher noise plus a schedule whose scale parameter μ* is fitted from the observed data spectrum rather than derived.","fun_headline_variants_meta":{"raw":{"variants":["Roughness matching: noise must be as rough as data","Matched noise roughness stabilizes flow-based generation","Scale-adaptive flows: noise roughness key to fine scales","Rough noise for rough data: a principle for generative flows","Noise roughness principle cuts cost in multiscale generation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000146,"raw_usage":{"total_tokens":1053,"prompt_tokens":810,"completion_tokens":243,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":162}},"tokens_in":554,"tokens_out":243,"duration_ms":3117,"temperature":1.0,"reasoning_tokens":162,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:13:32.918509+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a Matérn target with regularity s1 and noise with lower roughness s0 < s1, Proposition 3.1 predicts ∥B(t)∥ ~ 1/t near t=0; measuring the minimum RK4 step size needed to keep the generated energy spectrum within a fixed tolerance at N=32, 64, and 128 should show a clear resolution-dependent blow-up if the claim is right. Conversely, if a Gaussian-noise generative model with s0 > s1 still produces accurate fine-scale spectra with a fixed small number of steps, the drift-blowup mechanism is not the controlling factor.","supporting_citations":[{"cited_title":"Building normalizing flows with stochastic interpolants","cited_arxiv_id":null,"evidence_quote":"Introduces stochastic interpolants; supplies the interpolant process and the drift-as-conditional-expectation formulation used throughout."},{"cited_title":"Gaussian Measures","cited_arxiv_id":null,"evidence_quote":"Provides the Gaussian measure and Cameron–Martin space theory used in Proposition 3.1 and in defining 'rougher than' in Proposition 3.3."},{"cited_title":"Ergodicity of the 2D Navier-Stokes equations with de- generate stochastic forcing","cited_arxiv_id":null,"evidence_quote":"Proves ergodicity of the 2D Navier–Stokes equation with degenerate stochastic forcing, making the target invariant measure well defined."}],"review_version":1}