{"id":"efd1e44b-c6c6-44c2-836a-3c11ccb228fe","arxiv_id":"2607.19173","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"NKEs learn drift, diffusion, and jump operators of an SDE by fitting neural networks to the short-time evolution of Gaussian-mixture approximations to the Kolmogorov Forward equation.","lead":"This paper proposes Neural Kolmogorov Equations (NKEs), a way to train neural stochastic differential equations by learning the evolution of probability densities rather than individual noisy trajectories. NKEs handle coupled Gaussian and jump (Lévy) noise, and training can be parallelized over time steps, giving better accuracy and speed than several baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Jump/diffusion decomposition at a fixed observation scale is not identifiable from Eq. 31's MAD threshold: small jumps are equivalent to extra diffusion (Prop. 4), and unreported κ_sens can flag ~32% of pure-Gaussian increments as jumps, so the learned Lévy triplet is not unique.","rationale":"I read the paper as a serious attempt to move Neural SDE training from path simulation to density evolution. The KFE reformulation and the explicit projected advection/diffusion operators are well derived, and the limitations section honestly acknowledges fixed sampling and GMM projection issues. The empirical results are plausible, though code and hyperparameters are missing.\n\nThe central claim that would have to be true is that the Lévy triplet can be recovered from trajectory data without parametric assumptions. The load-bearing step is the jump thresholding in Eq. 31: it converts the continuous Lévy measure h into a small-jump part (absorbed into advection-diffusion) and a large-jump part (learned as a finite-time semigroup). Two independent problems make this step insecure.\n\n(1) Non-identifiability: Appendix C.2, Prop. 4 shows exactly that small jumps are equivalent, to second order in jump size, to an effective drift/diffusion. Thus for any fixed sampling interval, many (g,h) pairs yield the same or nearly the same increment distribution. The method cannot distinguish 'true diffusion plus small jumps' from 'larger diffusion alone.' The paper's acknowledgment that undetected jumps are 'modelled as effective diffusion' is not enough: it means the output of training is one element of an equivalence class, so the recovered stochastic model is not the true driver. This directly undercuts the abstract's promise of learning general Lévy-type stochastic forcing.\n\n(2) Threshold sensitivity: for Gaussian increments, the Eq. 31 threshold is κ_sens σ. At κ_sens=1, roughly 32% of increments from a pure diffusion would be flagged as jumps; at κ_sens=3, only about 0.3%, but then jumps of size ~1σ are missed. No κ_sens value is reported, no sensitivity analysis is provided, and the two-stage training cannot correct for misclassification. The experimental tables therefore depend on an unreported hyperparameter whose optimal value is exactly the separation between jump and diffusion scales the method is supposed to learn.\n\nThis is the same concern the reader identified, so agreement is 'agree'. The verdict should remain CONDITIONAL because the issue is empirical/identifiability-based and a controlled experiment could determine whether the method is useful in the identifiable regime. If the test shows pure-diffusion data are misclassified or recovered parameters shift with κ_sens, the claims must be narrowed to semigroup learning for large, well-separated jumps; if not, the conditional accept can proceed.","tokens_in":17967,"tokens_out":9193,"duration_ms":107803,"concrete_test":"On synthetic data with a known Lévy triplet, run two conditions: (a) pure diffusion (no jumps) and (b) jump-diffusion with jump std equal to the diffusion increment std at the observation step. Sweep κ_sens over {1,2,3,5}, using the paper's pipeline and M=1024 trajectories. Report the fraction of increments flagged by Eq. 31 and the recovered (f, gg^T, H^s_Θ). If condition (a) yields nonzero flagged increments or the recovered pair varies materially across κ_sens, then the Eq. 31 decomposition is not identifiable and the claims must be narrowed to 'large jumps at the chosen threshold.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires recovering the Lévy triplet (f, gg^T, h) from discretely observed trajectories. The only mechanism separating jumps from advection–diffusion is the MAD threshold in Eq. 31. This fails on two connected grounds.\n\nFirst, identifiability. For any ε, the Lévy measure decomposes as h^{<ε}+h^{≥ε}. By Proposition 4, the small-jump part acts to O(ε^3) as an effective advection–diffusion with drift β=∫_{|y|<ε} y h(dy) and covariance γ=∫ yy^T h(dy). Consequently, at the observation scale s, many distinct pairs (g,h^{<ε}) produce indistinguishable short-time semigroups: the data can only identify an equivalence class, not a unique noise model. The learned 'jump semigroup' is therefore one arbitrary representative, and the abstract's 'learn general Lévy-type stochastic forcing directly' is unsupported.\n\nSecond, the detector itself is unreliable at the scales the paper considers. For a Gaussian diffusion increment with std σ, MAD ≈ 0.6745σ, so the threshold in Eq. 31 is κ_sens σ. With κ_sens=1, roughly 32% of pure-diffusion increments are flagged as jumps; with larger κ_sens, true jumps of size comparable to the diffusion scale are missed. The paper does not report κ_sens or any sensitivity analysis, and the two-stage training has no mechanism to correct misclassification. False positives bias the diffusion estimate by removing tails; false negatives collapse real jumps into effective diffusion. The reported benchmark numbers are therefore conditional on an unreported, scale-dependent, and non-identifiable threshold choice.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Neural Kolmogorov Equations (NKEs), a method for learning stochastic dynamics from discretely observed trajectories by fitting the Kolmogorov forward generator rather than simulating SDE paths. The generator is projected onto a Gaussian-mixture manifold, yielding ODEs for the evolution of mixture means and covariances under advection and diffusion, plus a finite-time jump semigroup modeled by a mixture-density network. Training uses a Strang-splitting objective, with a MAD-based threshold detector to separate jump increments from advection–diffusion increments. Experiments on Lorenz, Black–Scholes, double-well jump-diffusion, and molecular-dynamics coarse-graining benchmarks compare NKEs with Euler–Maruyama, Trajectory Flow Matching, and SDE-GAN baselines on mean, covariance, and Wasserstein errors.","tokens_in":18420,"tokens_out":3690,"duration_ms":41048,"significance":"The core idea is attractive and potentially useful: formulating neural-SDE learning as moment evolution on probability densities is a principled way to avoid autoregressive trajectory simulation and to handle coupled (non-diagonal) Gaussian noise. The basic advection and diffusion moment identities (Eqs. 13–15 and 17–19) are correct, and Appendix C gives plausible, even if sketched, error bounds for short-time Gaussianity, small-jump diffusion approximation, and jump freezing. If the central claims were fully supported, the work would be a meaningful step beyond simulation-free Gaussian-noise SDE training. However, the paper's headline claim of learning general Lévy-type forcing directly is currently not supported: the jump/diffusion separation rests on an unreported threshold sensitivity, and the small-jump/diffusion equivalence (Prop. 4) implies an identifiability limitation that is not addressed. The experiments are also not consistently faster than the parallel EM baseline.","major_comments":[{"comment":"The jump detector threshold depends on κsens, but κsens is never reported and no sensitivity analysis is given. This is load-bearing because the entire two-stage training procedure first separates jump increments from advection–diffusion increments using this detector. For pure-Gaussian diffusion increments, a threshold at κsens=1 flags roughly 32% of increments as jumps; larger κsens will miss true jumps comparable to the diffusion scale. The paper must report the chosen κsens, justify it, and show that the learned diffusion and jump models are stable over a range of κsens. Without this, the benchmark numbers in Tables 1–4 are conditional on an undisclosed hyperparameter.","section":"§5.2.2, Eq. (31)"},{"comment":"Proposition 4 shows that small jumps act, to O(ε^3), like an effective advection–diffusion with drift β=∫_{|y|<ε} y h(dy) and covariance γ=∫ yy^T h(dy). At any fixed observation scale, many different pairs (g, h^{<ε}) produce indistinguishable short-time semigroups, so the learned Lévy triplet is not identifiable from discretely sampled trajectories. This directly undermines the abstract's claim that NKEs 'learn general Lévy-type stochastic forcing directly.' The manuscript should either limit the claim to the identifiable equivalence class, or present a concrete identifiability analysis and a test showing that varying jump sizes around the diffusion scale does not materially change the recovered advection–diffusion operator.","section":"Appendix C.2, Prop. 4 and Eq. (20)"},{"comment":"The paper explicitly acknowledges that the projection onto Gaussian mixtures is non-unique and often ill-posed for K>1, and that this 'leads to discontinuities and poor convergence when learning complex jump patterns.' Yet the jump-learning step uses a Gaussian-mixture MDN with K_J components, i.e., K>1 projection. The manuscript does not report K_J, the component-pruning thresholds, or the covariance regularization mentioned in Appendix B.3.1. As a result, the jump semigroup in the double-well experiment (Table 3) is not a well-defined object. The authors need to specify the projection regularization and demonstrate that the learned jump law is stable under changes in K_J and initialization.","section":"Appendix B.3 and §7"},{"comment":"The paper claims 'improved training efficiency,' but the reported training times do not support this against the Euler–Maruyama baseline: NKE is slower than EM in Table 1 (42.4 vs 33.9), Table 3 (1545 vs 1152), and Table 4 (323.1 vs 206.3); only in Table 2 is it faster (4.1 vs 35.7). If the claim is meant only relative to SDE-GAN, it should be stated as such. The efficiency argument should be reframed as parallel-in-time capability rather than wall-clock speed, or the comparison should be made on equal footing with the same hardware and implementation maturity.","section":"Tables 1–4, training time columns"}],"minor_comments":[{"comment":"The abstract's 'learn general Lévy-type stochastic forcing directly' is inconsistent with the limitations stated in §7 and with the identifiability issue raised in Appendix C.2. The wording should be softened or accompanied by the required identifiability conditions.","section":"Abstract and §7"},{"comment":"Typo: 'provie' should be 'provide' in the first paragraph of Appendix A.","section":"Appendix A"},{"comment":"Typo: 'Trajecotry Flow Matching' should be 'Trajectory Flow Matching'.","section":"Appendix E"},{"comment":"The jump loss is written as a negative log-likelihood of target points under the convolved mixture. It would help to state explicitly that this is a conditional maximum-likelihood objective and to describe how the MDN's output covariances are constrained to be positive definite.","section":"§5.4, Eq. (42)"},{"comment":"The notation H^s for the finite-time jump distribution is introduced, but it is not immediately clear that H^s is a probability measure on jump sizes for the interval [t,t+s] and that the no-jump mass is included. A one-sentence clarification would improve readability.","section":"§4.4, Eq. (21)"}],"recommendation":"major_revision","confidential_remarks":"This is a promising framework with correct moment machinery, but the central general-noise claim is currently under-supported. The most important fixes are reporting κsens and a sensitivity study, addressing the small-jump/diffusion identifiability limitation explicitly, and stabilizing/justifying the K>1 projection used for jump learning. These are substantial but addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real idea — learning the projected Kolmogorov forward operator on a Gaussian-mixture basis instead of simulating neural SDE paths — and the authors work out the moment identities and splitting scheme carefully. But the paper oversells the generality of the noise learning, and the jump/diffusion split rests on an unreported threshold sensitivity. I’d send it to review; I’d want revisions before I’d cite the strong claims.\n\nWhat’s genuinely new: treating the KFE itself as the learning object, parameterizing advection, diffusion, and jump operators with neural networks on a Gaussian-mixture projection. That framing buys parallel-in-time training and a unified treatment of coupled Gaussian and jump noise. The derivations in Sections 4.2–4.4 are correct as far as I can tell, and Appendix C gives plausible error bounds for the Gaussian closure, small-jump diffusion approximation, and jump freezing. The empirical story is also consistent: NKEs beat EM, TFM, and SDE-GAN on the benchmarks shown, sometimes by a lot.\n\nNow the soft spots, in rough order of importance. First, the identifiability issue raised in the stress test is real. Equation 31 decomposes jumps from diffusion using a MAD threshold at the observation scale, but for any cutoff small jumps are equivalent to extra drift and diffusion. The learned Lévy triplet is not unique, and κ_sens is unreported. The paper even half-admits this in Section 5.2.2, but the abstract's claim about learning 'general Lévy-type stochastic forcing directly' goes beyond what is supported. The fix is to reframe the contribution as learning a finite-time semigroup that reproduces density evolution, which is still valuable and doesn't require triplet identifiability.\n\nSecond, the K>1 Gaussian-mixture projection is ill-posed and non-unique. The authors know this — Section 7 and Appendix B say so — and they add regularization, but it remains a genuine weakness for state-dependent jump learning. Third, the practical reproducibility gap: no code, no κ_sens, no hyperparameter table, and the baselines are 'adapted' to Julia with details deferred. Given how much of the method's credibility rests on the experiments, that needs to be fixed.\n\nWho this is for: researchers working on neural SDEs, simulation-free training, or stochastic system identification. They will get a useful new tool and a clean set of approximation arguments, as long as they read the claims carefully.\n\nRecommendation: engage with it and send it to serious referees. The core construction is original, the math is mostly sound, and the limitations are at least partially acknowledged. But peer review should demand code, the missing sensitivity parameter, and a rewritten set of claims that matches what the method actually learns.","headline":"A solid, original training scheme for neural SDEs via projected Kolmogorov operators, weakened only by overclaimed noise identifiability and unreported threshold parameters.","tokens_in":18853,"tokens_out":2590,"would_cite":true,"duration_ms":32942,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H10","60J74","65C30","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Neural Kolmogorov Equations learn stochastic dynamics by modeling the evolution of probability densities rather than individual trajectories, enabling parallel-in-time training and recovery of general Lévy-type noise, including coupled Gaus","keywords":["Neural stochastic differential equations","Kolmogorov Forward Equation","Lévy processes","operator splitting","Gaussian mixtures","simulation-free training","jump processes","Lagrangian Galerkin methods"],"falsifier":"Simulate a compound-Poisson jump-diffusion where the jump scale equals the diffusion increment scale (e.g., jump standard deviation equals sqrt(dt) times the diffusion coefficient), train NKEs, and compare the learned drift, diffusion, and jump mixture to the ground truth. If the recovered parameters deviate systematically or the generated trajectory Wasserstein distance does not beat an Euler-Maruyama baseline, the jump-thresholding decomposition is the culprit.","tokens_in":17889,"feed_emoji":"🎲","tokens_out":3988,"duration_ms":35757,"temperature":0.7,"pith_summary":"The paper proposes Neural Kolmogorov Equations (NKEs), a method that learns stochastic differential equations by modeling how probability densities evolve under the Kolmogorov Forward equation. The key claim is that the generator of this equation—advection, diffusion, and jump terms—can be directly parameterized by neural networks, allowing the model to infer the structure of the driving noise without parametric assumptions. By projecting density evolution onto Gaussian mixtures and using operator splitting, NKEs avoid autoregressive trajectory simulation and become parallelizable in time. If correct, the method provides a unified treatment of Gaussian and jump noise with competitive accuracy on the paper's benchmarks, including systems with coupled noise and discontinuous jumps.","feed_headline":"Learn stochastic dynamics from probability flows, not paths","feed_subtitle":"The method parallelizes in time and recovers jump and coupled-noise terms without parametric assumptions.","key_machinery":"The central mechanism is a Lagrangian Galerkin projection onto K Gaussian mixture components, which reduces the infinite-dimensional Kolmogorov forward generator to an ODE on the parameters (weights, means, covariances). The generator splits as A + D + J: advection transports means, diffusion spreads covariances, and the jump semigroup convolves the mixture with a learned conditional Gaussian mixture. Strang splitting (e^{Ls} ≈ e^{As/2} e^{Js} e^{Ds} e^{As/2}) is used to train each operator separately, making the training loss parallelizable in time and second-order accurate in the absence of jumps.","core_discovery":"The core discovery is that the Kolmogorov Forward generator, decomposed into advection, diffusion, and jump components, can be learned as a neural operator on the finite-dimensional space of Gaussian mixtures, turning SDE identification into a deterministic regression problem on density moments. Advection is recovered from the mean dynamics of localized Gaussian clusters, diffusion from their covariance growth, and the jump semigroup from thresholded jump increments via conditional Gaussian mixture deconvolution. This operator-level formulation avoids the need for autoregressive simulation and naturally supports non-diagonal diffusion tensors and general jump laws.","pith_inferences":["Editorial inference: If the density-operator view is correct, the method could be extended to non-uniform sampling by learning the jump generator directly in a basis that supports the generator, although the paper notes this is out of reach for its Gaussian mixture representation.","Editorial inference: The jump-thresholding assumption implies a stress test: performance should degrade gracefully as jump sizes shrink toward the diffusion scale; identifying that boundary would define the method's practical application envelope.","Editorial inference: The same Lagrangian projection idea could be applied to other parametric families (e.g., Student-t or skewed mixtures) to handle fat tails or state-dependent jump distributions with less ill-posed projection.","Editorial inference: Because the projection for K>1 Gaussian mixtures is non-unique, state-dependent jump learning may be the first place to fail in higher dimensions; regularizing the mixture parameters across the state space would be a natural extension."],"forward_implications":["NKEs can learn coupled, non-diagonal diffusion tensors without maximum-likelihood estimation, which is not possible with standard EM-based neural SDE training.","Training is parallelizable across time steps because each loss term depends only on local clusters and finite differences, eliminating the need for autoregressive rollout.","The same framework handles Gaussian and jump noise in one generative model, learning the jump distribution itself rather than assuming a fixed parametric family.","On the paper's benchmarks, NKEs achieve lower mean, covariance, and Wasserstein errors than Euler-Maruyama matching, trajectory flow matching, and SDE-GAN baselines, while keeping training times comparable."],"fun_headline_variants":["Learn SDEs via density flows, not path simulations","Neural density flows: parallel SDE learning with jumps and noise","Kolmogorov forward: neural SDEs made deterministic and parallel","From paths to probabilities: neural SDEs handle jumps and coupled noise","Stop simulating paths: learn density evolution for general noise"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method's separation of jumps from diffusion relies on jumps being visibly larger than ordinary random fluctuations at the observation time step; if a jump is about the same size as a typical diffusion increment, the split cannot be identified and both the learned jump semigroup and the advection–diffusion operator become biased.","fun_headline_variants_meta":{"raw":{"variants":["Learn SDEs via density flows, not path simulations","Neural density flows: parallel SDE learning with jumps and noise","Kolmogorov forward: neural SDEs made deterministic and parallel","From paths to probabilities: neural SDEs handle jumps and coupled noise","Stop simulating paths: learn density evolution for general noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1186,"prompt_tokens":672,"completion_tokens":514,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":416,"completion_tokens_details":{"reasoning_tokens":426}},"tokens_in":416,"tokens_out":514,"duration_ms":5704,"temperature":1.0,"reasoning_tokens":426,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T13:12:52.529853+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a compound-Poisson jump-diffusion where the jump scale equals the diffusion increment scale (e.g., jump standard deviation equals sqrt(dt) times the diffusion coefficient), train NKEs, and compare the learned drift, diffusion, and jump mixture to the ground truth. If the recovered parameters deviate systematically or the generated trajectory Wasserstein distance does not beat an Euler-Maruyama baseline, the jump-thresholding decomposition is the culprit.","supporting_citations":[],"review_version":1}