{"id":"09f9c5d6-5a9b-4827-ba31-34de0f52a4cb","arxiv_id":"2412.18730","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Flow matching ODE trajectories are shown to become attracted to the data mean, then to clusters, and then to converge at the terminal time under weak geometric assumptions, including submanifold-supported data.","lead":"This paper proves that flow matching ODE sampling trajectories converge at the terminal time for data lying on low-dimensional submanifolds, and describes how they first move toward the data mean, then toward local clusters, and finally toward individual data points. It offers a theoretical foundation for stable sampling, one-step distillation, and understanding memorization in generative models.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5.3's O(σ_t^{ζ/2}) rate is not schedule-independent: with α_t=t, β_t=(1−t)^3 and p=δ_1, the exact error is Θ(1−t), while σ_t^{ζ/2}=Θ((1−t)^{3ζ/2}), so the bound fails for ζ>2/3.","rationale":"The reader's weakest assumption was Assumption 5.2(2) and the overbroad claim about infinite discrete measures. That is a real but secondary issue: it affects the expository remark, not the finite-discrete or compact-manifold cases that drive the main applications. The schedule-dependence of the rate is a stronger, load-bearing problem because it is a direct mathematical counterexample to the stated Theorem 5.3 under the paper's own assumptions. It does not destroy the per-trajectory convergence result, which is the core qualitative contribution, but it does invalidate the quantitative rate as written and the claim that the rate is schedule-independent. The fix is straightforward: either restrict the scheduling functions (e.g., require 1−α_t=O(σ_t), which holds for the common linear and variance-preserving schedules) or restate the terminal error as O(σ_t^{ζ/2}+1−α_t). Because the central convergence result appears salvageable and the paper's geometric analysis is otherwise carefully developed, the reader's conditional verdict remains appropriate; no change to the verdict is needed, though the required revision should include this correction.","tokens_in":62270,"tokens_out":15924,"duration_ms":140457,"concrete_test":"Compute the closed-form FM ODE for p=δ_1 with α_t=t, β_t=(1−t)^3: the solution is x_t=t+x_0(1−t)^3, so ∥Ψ_1(x_0)−Ψ_t(x_0)∥=|1−t−x_0(1−t)^3|. Taking x_0=0 gives error 1−t, while σ_t=(1−t)^3/t. For ζ=0.8, the ratio (1−t)/σ_t^{0.4} tends to infinity as t→1, contradicting Theorem 5.3. If confirmed, amend the theorem by adding a schedule assumption such as 1−α_t=O(σ_t) or restating the rate as O(σ_t^{ζ/2}+1−α_t).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The rate claim in Theorem 5.3 is false for the class of scheduling functions allowed in Section 2. The proof transfers the bound ∥z_λ−z_∞∥=O(e^{−ζλ/2}) in the σ-reparametrized coordinate to ∥Ψ_t(x)−Ψ_1(x)∥ 'up to the change of variable λ→t', but x_t=α_t z_{λ(t)}, so the factor α_t introduces an additional error ∥1−α_t∥·∥z_∞∥. For p=δ_1, α_t=t, β_t=(1−t)^3, the exact FM ODE is Ψ_t(x_0)=t+(1−t)^3x_0, giving Ψ_1(x_0)−Ψ_t(x_0)=(1−t)(1−(1−t)^2x_0), which is Θ(1−t). Meanwhile σ_t=β_t/α_t=(1−t)^3/t, so σ_t^{ζ/2}=Θ((1−t)^{3ζ/2}). For any fixed ζ>2/3, the claimed O(σ_t^{ζ/2}) bound fails because (1−t)/σ_t^{ζ/2}→∞. The schedule satisfies the paper's smoothness and monotonicity assumptions, and p=δ_1 satisfies Assumption 5.2 with k=0, so Theorem 5.3 as stated is internally inconsistent. The existence part of the theorem survives, but the quantitative rate needs a schedule condition or a corrected statement such as O(σ_t^{ζ/2}+1−α_t).","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes flow-matching (FM) ODE trajectories through the denoiser, introducing a noise-to-signal ratio parametrization and a pair of meta-theorems for attracting and absorbing dynamics. It characterizes three stages of ODE evolution: an initial approach to the data mean, an intermediate attraction to local clusters, and a terminal phase in which trajectories converge to the data support. The main claims include well-posedness of the FM ODE on [0,1) under a finite second moment, terminal convergence for data supported on submanifolds, convergence-rate bounds, flow-map equivariance under similarity transformations, and a memorization analysis for discrete empirical measures. Proofs are deferred to a detailed appendix, and experiments on synthetic clusters, CIFAR-10, and FFHQ illustrate the predicted stage behavior.","tokens_in":62561,"tokens_out":6920,"duration_ms":69905,"significance":"If the stated results hold, this would be a substantial contribution: it provides the first per-trajectory terminal convergence analysis for FM ODEs with submanifold-supported data, a clean schedule-independent parametrization via σ, explicit rates for manifold and discrete cases, and falsifiable predictions about memorization and sampling-stage allocation. The paper also ships detailed appendix proofs, explicit tractable examples (e.g., Gaussian subspaces), and experiments that use the exact empirical optimal denoiser rather than only fitted neural denoisers. However, the central convergence-rate statement in Theorem 5.3 is false as stated for a legal choice of scheduling functions and a one-point data distribution. The framework is promising, and the existence part of the theorem may survive, but the rate claim and the proof that transfers rates from the σ-coordinate back to the t-coordinate need substantive correction.","major_comments":[{"comment":"The claimed rate ∥Ψ1(x)−Ψ_t(x)∥=O(σ_t^{ζ/2}) for every fixed 0<ζ<1 is internally inconsistent with the paper's own setup. Take p=δ_1, α_t=t, β_t=(1−t)^3. These scheduling functions are smooth, strictly monotonic, and satisfy α_0=β_1=0, α_1=β_0=1, so they are admissible in Section 2; p=δ_1 satisfies Assumption 5.2 with k=0. The exact FM ODE solution is x_t=t+(1−t)^3 x_0, so Ψ_1(x_0)−Ψ_t(x_0)=(1−t)(1−(1−t)^2x_0), which is Θ(1−t) for x_0≠0. But σ_t=(1−t)^3/t, so σ_t^{ζ/2}=Θ((1−t)^{3ζ/2}). For any fixed ζ>2/3, the claimed bound fails because (1−t)/σ_t^{ζ/2}→∞. The proof error is visible in Claim 8 of Section I.1: the O( e^{-ζλ/2}) bound is established for the σ-scaled variable z_λ=x_{t(λ)}/α_{t(λ)}, and then transferred to Ψ_t(x) 'up to the change of variable λ→t', but x_t=α_t z_{λ(t)}, so an additional error of order (1−α_t)∥z_∞∥ appears. The corrected statement should include this term, e.g., ∥Ψ_1(x)−Ψ_t(x)∥=O(σ_t^{ζ/2}+1−α_t), or impose a schedule condition such as 1−α_t=O(σ_t^{ζ/2}). The existence part of Theorem 5.3 may still be correct, but the rate statement as written is false.","section":"Section 5.1, Theorem 5.3"},{"comment":"The proof asserts that the continuity-equation argument of Gao et al. [2024, Theorem 3.1] remains valid verbatim when p is only assumed to have a finite 2-moment and may be singular (e.g., supported on a submanifold). This is a load-bearing step: it is exactly what lets the paper extend well-posedness beyond absolutely continuous data distributions. As written, the assertion is not proved; the standard derivation of ∂_t p_t + ∇·(u_t p_t)=0 typically requires approximation by absolutely continuous measures and a limiting argument to control the flux. Please either provide the full approximation/regularization argument or state the continuity equation as a separately proved lemma with the required assumptions.","section":"Appendix H.1, proof of Theorem 4.1"},{"comment":"The statement 'Any discrete distribution satisfies the assumptions with k=0' is false for infinite discrete measures with accumulation points. For example, on R take p=∑_{n≥1}2^{-n}δ_{1/n}; then 0∈supp(p), but p(B_r(0)) decays faster than any power of r as r→0, so the local mass lower bound p(B_r(x))≥C_R r^k cannot hold for any k. The subsequent proof of Theorem 5.3 uses this lower bound through Theorem D.4 to obtain the uniform denoiser rate, so the overstatement matters for the stated generality. The theorem is still applicable to finite discrete measures, which cover empirical distributions, but the assumption should be corrected to 'finite discrete distribution' or to an explicit condition on the local mass at accumulation points.","section":"Assumption 5.2"}],"minor_comments":[{"comment":"The phrase 'the acute angel condition' should read 'the acute angle condition'.","section":"Section 3.2"},{"comment":"The statement says 'For p satisfying Theorem 5.2'; the referenced object is an assumption, not a theorem, so it should say 'Assumption 5.2'.","section":"Section 5.1, Theorem 5.3"},{"comment":"The opening line 'With the same assumptions as in Theorem 4.3' refers to Proposition 4.3, not a theorem; the cross-reference should be corrected.","section":"Section 4.3, Proposition 4.4"},{"comment":"The discussion of the CIFAR-10 value σ0(V_i^ϵ)≈0.17 'corresponding to the final quarter of EDM's sampling steps' is schedule-dependent; the specific discretization and noise schedule used for this correspondence should be stated.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper is worth a serious referee, but the headline rate in Theorem 5.3 is wrong as stated. The stress-test example is correct: with α_t=t, β_t=(1-t)^3 and p=δ_1, the exact FM ODE gives Ψ_1(x0)-Ψ_t(x0)=Θ(1-t), while σ_t^{ζ/2}=Θ((1-t)^{3ζ/2}). For any ζ>2/3 the claimed O(σ_t^{ζ/2}) fails. This schedule satisfies the paper's smoothness and monotonicity assumptions, and the measure satisfies Assumption 5.2. The existence part survives; the rate needs a schedule-dependent correction, e.g. O(σ_t^{ζ/2}+1-α_t).\n\nWhat is genuinely new: the attracting/absorbing meta-theorems, the three-stage characterization of trajectories (mean, local clusters, terminal), and the terminal convergence proof for submanifold-supported data, which prior work could not handle. The σ-parametrization is a clean unifying device, and the memorization discussion with explicit σ0 constants is a nice practical byproduct.\n\nSoft spots in proportion: Theorem 4.1 relies on an unproven assertion that Gao et al.'s continuity-equation argument extends from absolutely continuous p to arbitrary finite-2-moment p. That is a real gap but likely fillable. The claim that any discrete distribution satisfies Assumption 5.2 with k=0 is only true for finite discrete measures; infinite discrete measures with accumulating atoms fail. Since the applications are empirical (finite) measures, this is a minor overstatement. No code or data are shipped, but the experiments are illustrative rather than load-bearing.\n\nBottom line: the central existence and convergence framework is solid enough to justify referee time. The rate bug is localized and fixable. This paper is for theorists working on diffusion/flow-matching ODEs; practitioners may skim the qualitative stage story. I would not cite the quantitative rate as stated, but I would cite the existence result after the fix.","headline":"Solid terminal-convergence framework, but Theorem 5.3's rate is false for a valid schedule; deserves review after a fix.","tokens_in":63166,"tokens_out":4368,"would_cite":false,"duration_ms":35625,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Flow-matching ODE trajectories converge per-sample at terminal time—not just in distribution—even when data lies on a low-dimensional submanifold, with the denoiser as the guiding mechanism.","keywords":["flow matching","per-trajectory convergence","denoiser","attracting and absorbing dynamics","data geometry","submanifold support","memorization","equivariance of flow maps"],"falsifier":"Two concrete checks settle the boundary of the claim. (1) Take an infinite discrete measure on a compact set with atoms accumulating at one point and weights decaying exponentially in their distance to that point: it has finite 2-moment and positive-reach support but violates Assumption 5.2(2) for every $k$, so integrating the exact FM ODE for it and observing whether trajectories still converge to a single atom, or instead wind near the accumulation point, tests whether the assumption is load-bearing or merely technical. (2) For data uniform on a circle (an $m = 1$ manifold), a high-precision integration measuring the exponent of $\\|\\Psi_1(x) - \\Psi_t(x)\\|$ versus $\\sigma_t$ distinguishes the theorem's $O(\\sqrt{\\sigma_t})$ bound from the paper's conjectured $O(\\sigma_t)$, pinning down whether the manifold rate is tight.","tokens_in":61983,"feed_emoji":"🎯","tokens_out":16921,"duration_ms":139192,"temperature":0.7,"pith_summary":"Flow matching (FM) is a generative framework that samples by integrating a learned vector field, but what individual trajectories do at the end of the process—whether they actually land on the data—has been open, especially when the data sits on a low-dimensional submanifold. This paper proves that under mild assumptions (finite second moment, positive reach, and a local mass lower bound on the support), the flow map limit $\\Psi_1(x)$ exists for almost every starting point, is measurable, pushes the Gaussian prior exactly onto the data distribution, and is approached at rate $O(\\sigma_t^{\\zeta/2})$ for any fixed $0 < \\zeta < 1$. The same attracting-and-absorbing mechanism yields a three-stage description of sampling: trajectories first drift toward the data mean, then are drawn into local clusters, and finally are absorbed onto the data support. Per-trajectory convergence is what one-step distillation and consistency-style models implicitly require, and the terminal analysis also yields quantitative statements about memorization and about how flows transform under rotations, scalings, and translations. The paper's central object is the denoiser—the conditional mean of the data given a noisy observation—which turns out to be the only data-dependent part of the dynamics.","feed_headline":"Flow-matching ODEs converge sample-by-sample, even on manifolds","feed_subtitle":"Per-trajectory terminal convergence is proven, with rates that also explain memorization.","key_machinery":"The engine is the denoiser $m_\\sigma(x) = E[X \\mid X_\\sigma = x]$—the conditional mean of the data $X$ given a noisy observation $X_\\sigma = X + \\sigma Z$—which is the only data-dependent ingredient of the FM vector field. Reparameterizing the ODE by the noise-to-signal ratio $\\sigma = \\beta_t/\\alpha_t$ puts the dynamics in the form $dx_\\sigma/d\\sigma = -(1/\\sigma)(m_\\sigma(x_\\sigma) - x_\\sigma)$, so the trajectory always points toward the current denoiser value. Two meta-theorems convert properties of the denoiser into trajectory control: an attracting theorem and an absorbing theorem, both driven by the acute-angle condition $\\langle m_\\sigma(x) - x, \\mathrm{proj}_\\Omega(x) - x\\rangle > 0$ against a closed set $\\Omega$. For the terminal stage the key input is a posterior-concentration estimate giving $\\|m_\\sigma(z) - \\mathrm{proj}_\\Omega(z)\\| = O(\\sigma^\\zeta)$ uniformly near the support (with rate $\\sigma\\sqrt{m}$ on an $m$-dimensional manifold and exponential decay for finite discrete data), which feeds an absorbing argument that keeps trajectories inside bounded tubes around the support and away from the medial axis, where the denoiser's limit is discontinuous. These rates make the $1/\\sigma$ drift integrable, which is precisely what turns distributional convergence into per-trajectory convergence.","core_discovery":"The paper's central claim is Theorem 5.3: for any data distribution with finite second moment whose support has positive reach and whose local mass does not thin out faster than a power law (Assumption 5.2), the flow map of the FM ODE has a limit $\\Psi_1(x) := \\lim_{t\\to 1} \\Psi_t(x)$ for Lebesgue-almost every $x$. The limit is measurable, its pushforward of the Gaussian prior is exactly the data distribution $p$, and the per-trajectory error satisfies $\\|\\Psi_1(x) - \\Psi_t(x)\\| = O(\\sigma_t^{\\zeta/2})$ for any fixed $0 < \\zeta < 1$, where $\\sigma_t = \\beta_t/\\alpha_t$ is the noise-to-signal ratio. The rate sharpens to $O(\\sqrt{\\sigma_t})$ when the data is supported on a smooth submanifold and to $O(\\sigma_t)$ for finite discrete measures, and the denoiser $m_\\sigma(x) = E[X \\mid X_\\sigma = x]$ is shown to converge to the projection onto the support for every point outside the medial axis. These results upgrade the previously known distributional convergence of FM sampling to genuine per-trajectory convergence, covering the submanifold case that earlier well-posedness results excluded and providing the terminal map that one-step distillation needs.","pith_inferences":["The local mass lower bound (Assumption 5.2(2)) is the true crux: the paper's remark that 'any discrete distribution' satisfies it with $k = 0$ holds for finite discrete measures but not for infinite ones with accumulating atoms, so terminal convergence for such 'sticky' supports is neither proved nor disproved here and is a natural stress test of the theorem's boundary.","Because the rates are stated for the exact, unlearned denoiser, a trained network whose near-terminal error shrinks slower than $O(\\sigma^\\zeta)$ falls outside the absorbing argument; this suggests a training diagnostic—monitoring $\\|m_\\sigma^\\theta(x_\\sigma) - x_\\sigma\\|$ through the final $O(\\sqrt{\\sigma_t})$ window—as a predictor of whether sampling will land on or off the data support.","The three-stage picture implies a scheduling principle the paper only gestures at: estimate $\\sigma_{\\mathrm{init}}$ and $\\sigma_{\\mathrm{cluster}}$ from dataset quantities (diameter, cluster separation, cluster weights) and allocate integration steps in proportion to denoiser change; because the proven constants are worst-case, the stage boundaries are better treated as tunable hyperparameters th","If the conjectured $O(\\sigma_t)$ manifold rate is correct, the terminal window shrinks further, strengthening the case for early stopping and coarse final schedules; a closed-form computation on the $m$-sphere with uniform density would likely settle the conjecture before a general proof."],"forward_implications":["Per-trajectory terminal convergence supplies the missing object for one-step distillation: the limit map $\\Psi_1$ that consistency-style models train against is proved to exist, be measurable, and exactly transport the prior to the data.","The $O(\\sigma_t^{\\zeta/2})$, $O(\\sqrt{\\sigma_t})$, and $O(\\sigma_t)$ rates mean trajectories barely move in the terminal stage, so sampling steps spent there are nearly wasted; the paper's practical corollary is to shift integration effort toward the intermediate stage where the denoiser changes most.","For finite discrete (empirical) data, every training point $x_i$ attracts trajectories inside its shrunk Voronoi cell $V_i^\\epsilon$, with per-point attraction strength $\\sigma_0(V_i^\\epsilon)$ that grows with the point's weight and isolation—which explains why duplicated or isolated training samples are the ones a trained model memorizes.","An asymptotically optimal denoiser—even one severely corrupted at intermediate noise levels—still drives every trajectory to a training point, so avoiding memorization requires the near-terminal denoiser itself to generalize, that is, to approximate projection onto the true data manifold rather than onto training points.","The flow map $\\Psi_1$ commutes with rotations, scalings, and translations of the ambient space, so a model trained on a transformed dataset is exactly the transformed flow of the original model; this makes augmentation-based reasoning about flows safe and permits training in a low-dimensional affine subspace and then lifting the result."],"supporting_citations":[{"why":"Defines the flow-matching framework and the conditional vector field whose expected minimizer is the learned vector field; the object whose ODE dynamics the paper studies.","marker":"[Lipman et al., 2022]"},{"why":"Provides the rectified-flow construction and the theorem that the conditional vector field generates the probability path, which the paper reparameterizes through the noise-to-signal ratio.","marker":"[Liu et al., 2023]"},{"why":"Introduces the denoiser as the conditional mean and the EDM design whose 18-step noise schedule and pretrained denoiser serve as the empirical reference in the CIFAR-10 experiments.","marker":"[Karras et al., 2022]"},{"why":"Established well-posedness of the FM ODE under restrictive assumptions that exclude submanifold-supported data; Theorem 4.1 extends exactly this existence and uniqueness result.","marker":"[Gao et al., 2024]"},{"why":"Supplies the geometric foundations—reach, medial axis, projection structure, and the angle lemmas—used by the attracting, absorbing, and projection-convergence arguments.","marker":"[Federer, 1959]"},{"why":"Continuity-equation theory (Section 8.1) is invoked to pass from the local Lipschitz vector field to existence, uniqueness, and continuity of the flow map in Theorem 4.1.","marker":"[Ambrosio et al., 2008]"},{"why":"Prior result that the denoiser converges to the projection near the data support; the paper strengthens this to almost-everywhere convergence with rates and uses it for trajectory control.","marker":"[Pidstrigach, 2022]"},{"why":"Prior analysis of FM sampling from discrete data and local cluster absorption that implicitly assumed trajectory convergence; the paper supplies the rigorous convergence proof it lacked.","marker":"[Gao and Li, 2024]"},{"why":"Consistency models motivate the terminal flow map $\\Psi_1$; the paper's existence result for $\\Psi_1$ is what a one-step distillation objective of this kind presupposes.","marker":"[Song et al., 2023]"}],"fun_headline_variants":["Flow-matching ODEs converge per-trajectory, even on manifolds","Per-trajectory convergence proven for flow-matching ODEs","Flow matching ODEs: terminal convergence with rates","Denoiser-guided ODE dynamics: per-trajectory limits","Per-trajectory limits for flow matching, submanifold-safe"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Assumption 5.2(2): near every point of the data support, the probability mass inside a ball of radius $r$ must be at least $C_R r^k$ for some $k \\ge 0$, and the terminal-convergence proof needs this uniform lower bound to turn posterior concentration into the denoiser rate that makes the $1/\\sigma$ drift integrable; the paper's remark that this holds for any discrete distribution is only correct for finite discrete measures, not for infinite ones with accumulating atoms.","fun_headline_variants_meta":{"raw":{"variants":["Flow-matching ODEs converge per-trajectory, even on manifolds","Per-trajectory convergence proven for flow-matching ODEs","Flow matching ODEs: terminal convergence with rates","Denoiser-guided ODE dynamics: per-trajectory limits","Per-trajectory limits for flow matching, submanifold-safe"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000566,"raw_usage":{"total_tokens":2738,"prompt_tokens":1060,"completion_tokens":1678,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":1586}},"tokens_in":676,"tokens_out":1678,"duration_ms":9602,"temperature":1.0,"reasoning_tokens":1586,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:32:06.546222+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Two concrete checks settle the boundary of the claim. (1) Take an infinite discrete measure on a compact set with atoms accumulating at one point and weights decaying exponentially in their distance to that point: it has finite 2-moment and positive-reach support but violates Assumption 5.2(2) for every $k$, so integrating the exact FM ODE for it and observing whether trajectories still converge to a single atom, or instead wind near the accumulation point, tests whether the assumption is load-bearing or merely technical. (2) For data uniform on a circle (an $m = 1$ manifold), a high-precision integration measuring the exponent of $\\|\\Psi_1(x) - \\Psi_t(x)\\|$ versus $\\sigma_t$ distinguishes the theorem's $O(\\sqrt{\\sigma_t})$ bound from the paper's conjectured $O(\\sigma_t)$, pinning down whether the manifold rate is tight.","supporting_citations":[],"review_version":1}