{"id":"fd937d98-9c20-4b2a-965f-5972b12d9254","arxiv_id":"2608.03916","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Acceleration Matching is a simulation-free, flow-matching-style algorithm for multi-marginal trajectory inference that regresses onto an explicit kinetic Brownian bridge acceleration field in phase space.","lead":"Acceleration Matching trains a neural network to predict a phase-space acceleration field that produces smooth random trajectories through unpaired snapshot data, without simulating trajectories during training. It offers a fast, spline-free alternative to existing trajectory inference methods for single-cell and other time-slice datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's Lipschitz hypothesis is not verified and fails for the paper's own construction: with Dirac two-marginal data the Markovianized drift is 3(1-x)/(1-t)^2 - 3v/(1-t), unbounded near t=1, so the exact-marginal guarantee is unproven for discrete empirical marginals.","rationale":"The reader's weakest assumption identifies the unproven Lipschitz condition on the Markovianized acceleration field as the key gap. My reading agrees and sharpens it: the condition is not merely unproven, it is violated by the paper's own construction for Dirac marginals, a case directly relevant to the empirical atomic measures treated in the benchmarks. Because Theorem 1 is the sole theoretical basis for the claimed exact matching of marginals, this is the most load-bearing concern. The issue is addressable: the authors could prove well-posedness under weaker conditions (e.g., integrable singular drifts) or add hypotheses on the marginals that ensure the conditional expectation is sufficiently regular; they could also state the theorem for the learned approximations and quantify the resulting marginal error. The numerical evidence is suggestive but mixed, and no code is released, so the practical status of the method remains somewhat open. For these reasons, the reader's CONDITIONAL verdict remains appropriate, and I do not recommend changing it. I agree with the reader's assessment rather than proposing a different load-bearing point.","tokens_in":21211,"tokens_out":17887,"duration_ms":161983,"concrete_test":"For d=1, set ρ0=δ_0, ρ1=δ_1, V0~N(0,σ_v^2 I). Derive a^M_t(x,v) from (7) and Lemma A.1 and verify that ∂_x a^M_t = -3/(1-t)^2, so the Lipschitz hypothesis of Theorem 1 fails. Then simulate the SDE dV_t = [3(1-X_t)/(1-t)^2 - 3V_t/(1-t)] dt + sqrt(ε) dB_t, dX_t=V_t dt, with X_0=0 and V_0~N(0,σ_v^2 I), using Euler–Maruyama with a fine grid and check whether the empirical law at t=1 concentrates at δ_1. If it does, the theorem's conclusion may still hold but the paper's proof does not apply; if it does not, the central marginal-matching claim is false in this elementary case.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central guarantee is Theorem 1, whose proof requires that the Markovianized drift a^M_t be Lipschitz in all variables so that the kinetic Fokker–Planck equation has a unique solution. The paper never checks this condition for the constructed field, and it is false already in the simplest nontrivial case. Take J=1, ρ0=δ_0, ρ1=δ_1, and V0~N(0,σ_v^2 I). Then PI is the kinetic Brownian bridge with X_0=0 and X_1=1 fixed. Using Lemma A.1's Gaussian conditioning, E[V_1|X_t=x,V_t=v] = 3(1-x)/(2(1-t)) - v/2, so the drift in (7) yields a^M_t(x,v) = 3(1-x)/(1-t)^2 - 3v/(1-t). Its x-derivative is -3/(1-t)^2, which diverges as t↑1; hence a^M is not Lipschitz uniformly in t, and the stated hypothesis of Theorem 1 fails for the exact interpolant the method proposes. For empirical atomic marginals the same singular behavior occurs near each atom, and no alternative well-posedness argument (e.g., singular SDE theory or uniqueness in a weighted space) is supplied. Consequently the introduction's claim that the method is 'guaranteed to satisfy the constraints (1)' is not established by the proof as written. The learned a_θ and q_φ replace the exact conditional expectation and initial law, so the implemented algorithm also inherits an unproven gap on and off the support of PI.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Acceleration Matching (AM), a method for trajectory inference from unpaired marginal snapshots at discrete times. The method lifts the problem to phase space, constructs an interpolation process from kinetic Brownian bridges conditioned on positional knots drawn from a coupling of the marginals, and derives an explicit conditional acceleration field. The central theoretical result, Theorem 1, states that the Markovianized acceleration field obtained by conditional expectation of the bridge acceleration yields a diffusion with the same marginals as the interpolant, provided the field is Lipschitz. The training objective regresses a neural network onto this field using only positional data, avoiding trajectory simulation during training. Numerical experiments on low- and high-dimensional benchmarks compare AM with flow-matching baselines and show competitive or superior performance on several metrics.","tokens_in":21535,"tokens_out":7711,"duration_ms":66712,"significance":"If the theoretical guarantee were fully established, the paper would make a valuable contribution: an explicit, simulation-free, preprocessing-light objective for multi-marginal trajectory inference with smooth paths. The closed-form bridge acceleration in Lemma 2 and the held-out marginal evaluation are genuine strengths, and the experimental design avoids circularity by testing on out-of-sample time points. However, the main theorem's Lipschitz hypothesis is not verified for the constructed field and fails for discrete empirical marginals, which are exactly the setting of all experiments. The central guarantee is therefore currently unproven, and the practical algorithm replaces the exact field and initial law with learned approximations, so the paper's stated guarantees do not apply to the implemented method without further argument.","major_comments":[{"comment":"The Lipschitz hypothesis on the Markovianized acceleration field a^M_t is not verified and fails for the paper's own construction in a two-marginal example. For J=1, ρ0=δ0, ρ1=δ1, and V0~N(0,σ_v^2 I), using the Gaussian conditioning of Lemma A.1 gives a^M_t(x,v)=3(1-x)/(1-t)^2 - 3v/(1-t). Its derivative with respect to x is -3/(1-t)^2, which diverges as t↑1, so a^M is not Lipschitz uniformly in t and the uniqueness hypothesis for the kinetic Fokker-Planck equation is not satisfied. Since all numerical experiments use empirical (atomic) marginals, this singular behavior is present in the exact interpolant for the data regime the paper targets. The proof of Theorem 1 therefore does not establish the claimed equality Law(X^M_t,V^M_t)=Law(X^I_t,V^I_t) or the marginal constraints (1) without an additional well-posedness argument for singular drifts.","section":"Section 3.1, Theorem 1"},{"comment":"Even under the assumptions of Theorem 1, the implemented pipeline replaces the exact conditional expectation a^M_t with a learned network a_θ and the exact initial law P^I_0 with a learned Gaussian sampler q_ϕ. Consequently the statements in the Introduction and Section 3.1 that AM 'is guaranteed to satisfy the constraints (1)' do not apply to the trained model; they hold at most for the idealized regression target in the limit of perfect approximation. The paper should either qualify these claims explicitly or provide an approximation-error analysis showing how the learned quantities inherit the marginal-matching property.","section":"Section 3.2, Algorithms 1 and 2"},{"comment":"The proof applies the tower property to E[a^j_t(X^I_t,V^I_t;X^I_{t_{j+1}},V^I_{t_{j+1}})|(X^I_t,V^I_t)] without establishing that the singular terms in (7), which grow like (t_{j+1}-t)^{-2} near the right endpoint, are integrable under the conditional law. This integrability is not automatic and needs a separate argument, especially when the marginals are atomic. This gap is separate from the Lipschitz issue but is likewise load-bearing for the derivation of the Markovianized field.","section":"Appendix A.3, proof of Theorem 1"}],"minor_comments":[{"comment":"The phrase 'is guaranteed to satisfy the constraints (1)' overstates the result given the learned approximations described in Section 3.2; a more cautious wording is needed.","section":"Section 1, paragraph 2"},{"comment":"On the GoM dataset, the held-out W2 of AM (0.163±0.010) is larger than the reported 3MSBM value (0.135), so the summary sentence 'competitive with or superior to existing algorithms' should be nuanced to reflect this comparison.","section":"Section 4.1, Table 1"},{"comment":"The sentence 'the pair (µ^a_t, a_t) satisfies the kinetic Fokker-Planck equation' is imprecise; it is the measure-valued path t↦µ^a_t that satisfies the equation given the field a_t.","section":"Section 2.2, Equation (KFP)"},{"comment":"The displayed loss uses a single sampled time s_j per interval per minibatch; the paper does not state that this gives an unbiased estimator of the population loss (9), which would help clarify the stochastic optimization.","section":"Algorithm 1, line 9"},{"comment":"The number of paired samples (x^(i)_0, v^(i)_0) used to train the initial-velocity sampler q_ϕ is not reported; Table 10 gives hyperparameters but not the dataset size for this step.","section":"Appendix B.1"},{"comment":"The Tong et al. (2024) reference ends with 'Expert Certification.' which appears to be a BibTeX artifact rather than part of the title.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The central theoretical gap is substantial: Theorem 1's Lipschitz hypothesis fails for the atomic marginals used in all experiments, so the paper's headline guarantee is unproven in the intended regime. The methodological idea is promising, and the empirical results are encouraging, but the authors need to either prove well-posedness for the singular drift (e.g., in a weighted space), modify the construction to produce a genuinely Lipschitz field, or substantially weaken the claimed guarantees. With such a revision the paper could be suitable for publication; as written, the gap affects the core claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"AM is a genuinely new approach: it lifts multi-marginal trajectory inference to phase space, defines an explicit conditional acceleration field for kinetic Brownian bridges (Lemma 2, Equation 7), and Markovianizes it so the training loss is a simple regression on positional data, with no trajectory simulation. That is a real extension of flow matching to second-order dynamics, and the paper works out the Gaussian conditioning details carefully. The numerical experiments are also fairly compared against MMFM, OT-CFM, OT-MFM, and 3MSBM; AM is competitive or better than the MMFM variants on the low-dimensional benchmarks, though it trails 3MSBM on GoM/LV and OT-MFM on some single-cell settings. The writing is clear, and the claims about being simulation-free are accurate.\n\nThe soft spot is Theorem 1. The theorem's proof assumes the Markovianized acceleration field a^M_t is Lipschitz in all variables, so the kinetic Fokker–Planck equation has a unique solution. The paper never checks this condition, and the stress-test example shows it is false already in the simplest case: two Dirac marginals at 0 and 1. The conditional drift becomes 3(1-x)/(1-t)^2 - 3v/(1-t), whose x-derivative diverges as t->1. So the exact-marginal guarantee, which the introduction advertises as \"guaranteed to satisfy the constraints (1),\" is not established for atomic or empirical marginals. No alternative well-posedness argument (weighted spaces, singular SDE theory) is provided. This is a real gap, but I think it is patchable: the method may still work numerically, and the theorem could likely be recovered under weaker well-posedness assumptions or with slightly smoothed marginals. The implemented algorithm also replaces the exact initial law with a learned Gaussian sampler and the exact field with a learned network, so the practical guarantee is approximate; that is common in flow matching, but it should be stated more prominently. Finally, no code is released, which limits reproducibility.\n\nAll that said, this is a solid, honest paper. The math is standard and the references are appropriate; the citation pattern shows no inflation. The central idea is worth pursuing, and the stress-test counterexample should be addressed head-on. This paper deserves a serious referee: send it to peer review, but expect a major revision that either proves the Lipschitz condition under reasonable assumptions on the marginals or replaces it with a valid well-posedness argument for the singular drift.","headline":"A clever simulation-free phase-space trajectory inference method with an explicit kinetic bridge drift, but the central theorem's Lipschitz hypothesis is unverified and fails for the atomic marginals it targets, leaving the exact-marginal guarantee unproven.","tokens_in":22081,"tokens_out":6144,"would_cite":true,"duration_ms":50546,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes Acceleration Matching, a simulation-free regression algorithm that learns smooth phase-space dynamics matching prescribed marginals at every observation time.","keywords":["trajectory inference","kinetic Brownian bridge","phase-space dynamics","simulation-free training","multi-marginal constraints","flow matching","stochastic differential equations","single-cell RNA sequencing"],"falsifier":"Simulate the ideal Markovianized SDE for a single interval with known endpoints and compare its terminal marginal to the kinetic Brownian bridge terminal law as the Euler step shrinks; disagreement would disprove the marginal-matching mechanism. Separately, on a dataset where the distribution of initial velocity given position is clearly non-Gaussian, the learned Gaussian sampler should fail to satisfy held-out marginals if the practical pipeline is the actual claim.","tokens_in":20964,"feed_emoji":"🧬","tokens_out":9732,"duration_ms":77561,"temperature":0.7,"pith_summary":"Acceleration Matching (AM) is a proposed answer to a concrete problem: from unpaired position snapshots at several time points, produce smooth random trajectories whose position distribution at each snapshot time equals the observed one. The paper's central claim is that this can be done by picking the acceleration field of a second-order stochastic differential equation to be the conditional expectation of the explicit kinetic Brownian bridge acceleration. With that field, the non-Markovian interpolant and a Markov diffusion driven by the field have identical marginals at every time (Theorem 1), so a squared-residual regression against the explicit bridge acceleration is a valid, simulation-free training objective. The method needs only positional data, avoids trajectory simulation during training, and requires no smoothing preprocessing. The paper reports numerical evidence on ocean-current, predator-prey, and single-cell benchmarks indicating that AM matches or beats existing flow-matching and phase-space methods on held-out marginals.","feed_headline":"No simulation: Acceleration Matching learns smooth trajectories","feed_subtitle":"A regression onto explicit kinetic Brownian bridge drifts satisfies every marginal constraint.","key_machinery":"The load-bearing object is the kinetic Brownian bridge: kinetic (underdamped) Brownian motion $dX_t = V_t\\,dt$, $dV_t = \\sqrt{\\varepsilon}\\,dB_t$ conditioned to start and end at prescribed phase-space points. Doob's $h$-transform makes its drift explicit, equation (7), with inverse-square and inverse-time terms in the remaining time $t_{j+1}-t$. The paper's central mechanism is the Markovianization step: averaging this bridge drift over the future endpoint conditionally on the current position and velocity yields an acceleration field that depends only on the present state, and Theorem 1 shows the resulting diffusion and the original non-Markovian interpolant share the same marginals for all $t$. This is what converts trajectory inference into regression onto a known field.","core_discovery":"The paper establishes, as a mathematical statement, that the multi-marginal trajectory interpolation problem can be solved by a single Markovian second-order SDE. Starting from a kinetic Brownian motion conditioned to pass through sampled phase-space knots, the bridge drift on each interval is explicit: $a^j_t(x,v;x_{j+1},v_{j+1}) = \\frac{6(x_{j+1}-x)}{(t_{j+1}-t)^2} - \\frac{2(v_{j+1}+2v)}{t_{j+1}-t}$. Replacing this future-dependent drift by its conditional expectation given the current phase-space point produces a Markov process that, under a Lipschitz assumption, has the same law at every time as the pinned interpolant; in particular the position marginals satisfy the prescribed constraints. Therefore the quadratic regression loss on the explicit field has the correct Markovianized field as its unique minimizer, and the paper presents this as the first-order principle behind the training algorithm.","pith_inferences":["An extension not pursued in the paper is to use a richer conditional initial-velocity sampler than the diagonal Gaussian; the theorem's exactness hinges on reproducing $\\mathrm{Law}(V_0\\mid X_0)$, so any failure of the Gaussian model should show up as a systematic marginal error.","The explicit bridge drift blows up like $(t_{j+1}-t)^{-2}$ near the right endpoint, so discretization error there is the most plausible numerical bottleneck; truncating or reparameterizing the interval near the end is a testable modification.","The same conditional-expectation construction should carry over to damped kinetic Brownian motion ($\\gamma>0$), since the appendix derives the analogous bridge drift; if so, the method gains a free parameter for tuning trajectory roughness while keeping the same simulation-free loss.","Because the loss is a simple regression, applying the same acceleration-matching principle to other reference processes, or to manifolds with explicit bridge drifts, would transfer the simulation-free property to those settings."],"forward_implications":["In the population limit, the regression loss is exact: its unique minimizer is the field from Theorem 1, so training never needs to simulate trajectories or enforce smoothness with splines.","The learned dynamics are second-order, so generated position paths have continuous velocities and are smooth by construction, unlike piecewise flow-matching stitches.","Multi-marginal constraints at observation times are built into the interpolant, and uneven time points are handled by $\\Delta_j$ weights in the loss.","Because velocities are sampled jointly given all position knots, the objective couples intervals rather than decomposing into independent segment-wise regressions.","If Theorem 1's Lipschitz hypothesis holds, the same field can be integrated at inference by any SDE solver, with initial velocities drawn from the learned conditional sampler."],"supporting_citations":[{"why":"Supplies the well-posedness conditions for the kinetic Fokker–Planck equation invoked by Theorem 1 to identify marginal laws.","marker":"Bogachev et al. (2022)"},{"why":"Supplies the commutation identity used in the proof of Lemma 2 to compute the bridge generator and the explicit drift (7).","marker":"Bakry et al. (2014)"},{"why":"The flow-matching framework that AM lifts to second-order dynamics with multiple marginal constraints.","marker":"Lipman et al. (2022); Albergo and Vanden-Eijnden (2022); Liu et al. (2022)"},{"why":"Defines the MMFM baselines and the constant/vanishing noise variants used in the numerical comparison.","marker":"Rohbeck et al. (2025)"},{"why":"The phase-space multi-marginal Schrödinger bridge matching method whose conditional acceleration field and simulation-based training are contrasted with AM.","marker":"Theodoropoulos et al. (2025)"},{"why":"Supplies the preprocessed Gulf of Mexico and Lotka–Volterra benchmark datasets used for the low-dimensional experiments.","marker":"Shen et al. (2024)"}],"fun_headline_variants":["Explicit bridge drifts let Acceleration Matching skip trajectory simulation","Learn interpolating trajectories by regressing onto conditional acceleration","Acceleration Matching: a single SDE regression for multi-marginal interpolation","Markovian acceleration field makes trajectory inference simulation-free"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes the averaged acceleration field is Lipschitz so the kinetic Fokker–Planck equation has a unique solution, and it assumes the exact initial phase-space law; the practical algorithm replaces both with learned approximations, and the paper does not show the averaged field is Lipschitz near the interval endpoints where the bridge drift diverges.","fun_headline_variants_meta":{"raw":{"variants":["Explicit bridge drifts let Acceleration Matching skip trajectory simulation","Learn interpolating trajectories by regressing onto conditional acceleration","Acceleration Matching: a single SDE regression for multi-marginal interpolation","Markovian acceleration field makes trajectory inference simulation-free"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1424,"prompt_tokens":891,"completion_tokens":533,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":464}},"tokens_in":507,"tokens_out":533,"duration_ms":5645,"temperature":1.0,"reasoning_tokens":464,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:44:55.174958+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the ideal Markovianized SDE for a single interval with known endpoints and compare its terminal marginal to the kinetic Brownian bridge terminal law as the Euler step shrinks; disagreement would disprove the marginal-matching mechanism. Separately, on a dataset where the distribution of initial velocity given position is clearly non-Gaussian, the learned Gaussian sampler should fail to satisfy held-out marginals if the practical pipeline is the actual claim.","supporting_citations":[],"review_version":2}