{"id":"b8d86f41-f4d7-462e-8432-88e7551ff73d","arxiv_id":"2505.23215","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"New closed-form drift, diffusion, and Gaussian jump generators for bridge processes enable simulation-free generative modeling of irregularly sampled time series with discontinuities, with learned generators that recover the target marginals when the model family is well specified.","lead":"Trajectory Generator Matching extends flow-matching style generative models to time series that jump and arrive at irregularly spaced times, by explicitly learning the drift, diffusion, and jump parameters of the underlying stochastic process. The method derives closed-form losses for Gaussian jump kernels and generates new realistic trajectories for applications like finance and patient monitoring.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Joint-distribution claim depends on unquantified per-step approximation error; restricted Gaussian bridge/jump family and finite memory can systematically miss real transitions, and Appendix C's Lipschitz bound is too weak to certify chaining.","rationale":"The main derivations (Proposition 2 KFE in both drift-diffusion and jump cases, Proposition 5 KL in d=1, Proposition 6) survive my re-derivation; the construction is non-circular and the experiments are honest but limited. The load-bearing gap is not a mathematical contradiction but an unquantified representational error: the training objective is exact only over the unrestricted generator family, and the paper does not show that the restricted family (linear-drift SDE with fixed eta, one Gaussian jump kernel, fixed memory m) can approximate the conditional transitions of the target process. Proposition 9/Remark 7 assumes the per-step error epsilon is small and that a Lipschitz constant K exists, but it does not estimate K and the resulting (1+20K^2)^n propagation is vacuous for long series unless K is very small. The d>1 slip in Proposition 5 is real but does not affect the d=1 experiments; it supports CONDITIONAL rather than ACCEPT. Because the reader's verdict already captures these issues, my read does not change the verdict.","tokens_in":18018,"tokens_out":13112,"duration_ms":135077,"concrete_test":"Run the method on a regime-switching AR process with bimodal innovation (e.g., X_{j+1}=0.9X_j+eps, eps ~ 0.5N(1,1)+0.5N(-1,1)) with memory m=10. For 100 fixed histories, estimate W2(P_true(·|history), P_generated(·|history)) from 10k Monte Carlo one-step samples; compare against the same W2 on a unimodal Gaussian-innovation process. If the per-step W2 for the bimodal process does not converge toward the unimodal baseline as training data increases, the restricted model family is misspecified and the joint-distribution claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Algorithm 2 samples from P_{X_{t0},...,X_{tn}} requires (i) each one-step learned generator to be close to P_{X_{t_{j+1}}|X_{t_j}=x_j,...} and (ii) per-step errors to be controlled under chaining. The paper only controls (ii) under assumptions: Remark 7/Appendix C suppose a per-step error epsilon and a Lipschitz constant K for both target and learned conditional laws, with bound (1+20K^2)^n; K is never estimated and the bound is vacuous for n=50-100 unless K is tiny. No control is given for (i). Proposition 4 gives exactness only for the unrestricted minimizer over all generators; the actual training restricts to a linear-drift SDE with fixed isotropic diffusion eta plus a single Gaussian jump kernel lambda N(mu, sigma^2 Id), and to a finite memory window m in Section 6.2. If the true transition is multi-modal, heavy-tailed, or has long memory, the fitted generator's conditional marginal is biased, and the unconditional MMD in Tables 1-2 cannot detect this because it averages over the time-series distribution. The two synthetic benchmarks (trend+noise, Black-Scholes) have unimodal or mildly non-Gaussian transitions, so they do not stress the restricted family. A secondary, concrete slip: Proposition 5's claimed minimizer sigma = trace(Sigma_J) should be sigma = sqrt(trace(Sigma_J)/d) for d>1, since solving dF/dsigma = 0 gives sigma^2 = trace(Sigma_J)/d; harmless for the d=1 experiments but wrong in the general statement.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a generative model for time series based on generator matching. For each interval between consecutive observation times, the authors construct explicit Markov generators whose marginals interpolate between the observed data points: a drift-diffusion generator with a Gaussian bridge law N(m_t, τ_t) and a pure-jump generator with a related jump kernel. Training losses are derived from oracle regressions to these analytically known generators, and the jump kernel is parameterized by a scaled Gaussian so that the KL divergence in the loss has a closed form. The learned conditional generators are chained autoregressively to sample new trajectories, including on irregular time grids. Experiments on a synthetic trend-plus-noise process and a simulated Black-Scholes process compare the SDE, jump, and Markov-superposition variants against trajectory flow matching, reporting MMD on held-out test trajectories.","tokens_in":18332,"tokens_out":7438,"duration_ms":76387,"significance":"The paper contains several genuinely useful theoretical building blocks: Proposition 2 gives explicit, singularity-free generator formulas for Gaussian bridge marginals that can be verified directly against the Kolmogorov forward equation; Proposition 5 provides a closed-form KL divergence for Gaussian jump kernels, avoiding expensive binning; and the empirical comparison includes a held-out MMD metric and an ablation of the Markov superposition parameter. The main weakness is that the central claim of approximation of the full joint distribution rests on unverified assumptions (a Lipschitz constant K and a per-step error ε), and the validation is limited to two unimodal or mildly non-Gaussian benchmarks that do not stress the restricted model family. The theoretical result is a conditional statement rather than an unconditional guarantee, and the introduction's phrasing overstates what is proved.","major_comments":[{"comment":"The joint-distribution guarantee for Algorithm 2 relies on Proposition 9, which requires a per-step approximation error ε and a Wasserstein-2 Lipschitz constant K for both the target conditional laws and the learned kernels. No numerical estimate of K is given, and the bound (1+20K^2)^n W2^2(P_X0,P_Y0) + 4ε Σ_{i=0}^{n-1}(1+20K^2)^i is vacuous for the reported horizons n=50–100 unless K is extremely small. Moreover, condition (i) of Proposition 9 is assumed, not established, for the restricted model class. The introduction's statement that including memory 'proves' that the learned time series approximates the true joint distribution is therefore stronger than what the theorem actually shows; the claim should be qualified or supported by numerical estimates of K and ε.","section":"Section 5, Remark 7, Appendix C"},{"comment":"The training losses (4) and (5) characterize the exact conditional generator only when the minimizer is unrestricted. In the implemented method, the drift is restricted to a linear function, the diffusion is fixed isotropic η, the jump kernel is a single Gaussian λN(μ,σ^2 I_d), and conditioning is limited to a finite memory window m. For processes with multi-modal, heavy-tailed, or long-range-dependent transitions, the per-step conditional marginal will be systematically biased, and no error bound is provided for such misspecification. The synthetic benchmarks (trend-plus-noise and Black-Scholes) have unimodal or mildly non-Gaussian transitions, so the reported MMD values do not demonstrate robustness to this restriction. The paper should either include an experiment with a process that violates the Gaussian/linear assumptions or state this limitation more prominently in the abstract and conclusions.","section":"Section 4 and Section 6.2"}],"minor_comments":[{"comment":"The stated minimizer of F_{t,x} is (λ_t(x), μ_J, trace Σ_J), but the stationarity condition ∂σ F = 0 gives σ^2 = trace(Σ_J)/d, so the correct minimizer is σ = sqrt(trace(Σ_J)/d). In d=1 this equals σ_J, not σ_J^2 as suggested by the printed formula; the statement should be corrected.","section":"Proposition 5"},{"comment":"The definition of q_t contains a typographical error: 'q_t(y,x) =:= λ_t(x) J_t(y)' should be 'q_t(y,x) = λ_t(x) J_t(y)'.","section":"Proposition 2(ii)"},{"comment":"The replacement formula for I_k on the interval t ∈ [t_0, (t_0+t_1)/2) includes a factor '2 (k−1)!!' that appears incorrect for the unnormalized Gaussian integrals used elsewhere in the proof; please verify whether the correct expression should be (k−1)!! − I_k for even k and −I_k for odd k (possibly with normalization by sqrt(2π)).","section":"Appendix B, proof of Proposition 6"},{"comment":"The column headers in Tables 1 and 2 are ambiguous: the text refers to 'different numbers of subsampling times' and 'subsampling rates' interchangeably, while figure captions use percentages such as '50% subsampling rate'. Please clarify whether the entries 5, 10, 25, 50, 100 are percentages or absolute counts of retained time steps.","section":"Section 6.2, Tables 1 and 2"},{"comment":"The notation 'ALG1(x_i, L^{ξ_i,t_{i+1}}_t)' is not defined in the pseudocode; presumably it means one step of Algorithm 1 applied with the given generator, but this should be stated explicitly.","section":"Algorithm 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a numerical-analysis-oriented machine learning venue. The core generator identities appear correct and are a solid contribution, but the theoretical guarantee is weaker than the introduction suggests, and the experimental validation does not stress the model family. The issues are fixable: correct the minimizer in Proposition 5, soften or qualify the joint-distribution claim, and add a discussion (or numerical estimate) of the per-step approximation error and the Lipschitz constant. I would not reject the paper, but it needs revision before the claims can be accepted at face value."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a serious look. What is actually new: a rho-stabilized Brownian-bridge generator that removes the singular drift in Zhang et al.'s trajectory flow matching, a jump-kernel parameterization by scaled Gaussians with an analytic KL divergence (instead of binning as in Holderrieth et al.), and a clean conditional, irregular-grid autoregressive framing. I re-checked the key identities in d=1: the drift-diffusion KFE computation, the jump generator, and the KL formula in Proposition 5 are all correct. The Markov superposition trick also appears to work as advertised. The experiment section is thin but honest: two synthetic datasets, one baseline, MMD as the only metric. The improvement over TFM in the subsampled regime is real, and the superposition often helps. That said, there are three soft spots. First, Proposition 5 states the sigma-minimizer as trace(Sigma_J), which is only right in d=1. The correct value is sqrt(trace(Sigma_J)/d); harmless for the experiments but a clear slip. Second, the joint-distribution guarantee in Remark 7 / Appendix C is too weak to carry the paper's central claim. The chaining bound grows like (1+20K^2)^n, K is never estimated, and for n=50-100 it is vacuous unless K is tiny. No per-step approximation error is quantified. The authors should either prove a tighter bound or soften the claim to \"approximates the joint distribution under unverified but plausible regularity.\" Third, the restricted model family (single Gaussian jump kernel, fixed diffusion, finite memory window) can systematically miss multi-modal or heavy-tailed transitions, and the two benchmarks do not stress it. The Black-Scholes process is essentially log-normal; the synthetic trend+noise is unimodal. So the empirical evidence does not establish robustness for the kind of financial or clinical series the paper motivates. Overall, the core math is sound and the closed-form KL is a practical contribution. I recommend sending to peer review, with a request to fix the d>1 slip and to either add a realistic error bound or temper the joint-distribution claim. This paper is for researchers working on simulation-free generative modeling of continuous-time or irregularly sampled processes. I'd bring it to reading group and cite the stabilized bridge and KL formula.","headline":"A solid generator-matching paper with a genuinely useful closed-form KL for Gaussian jump kernels, but the joint-distribution guarantee rests on an unquantified Lipschitz assumption and a model family the experiments never stress.","tokens_in":18934,"tokens_out":1598,"would_cite":true,"duration_ms":19698,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J25","60J60","60J75","60H10","62M10","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper constructs explicit drift-diffusion and jump generators with the same Gaussian marginals, enabling simulation-free training of conditional time-series models that handle jumps and irregular sampling.","keywords":["generator matching","time series generation","stochastic differential equations","jump processes","irregular sampling","Kullback-Leibler divergence","Markov processes","trajectory flow matching"],"falsifier":"Generate a dataset from a transition kernel that is a mixture of two well-separated Gaussians, or a jump process with two distinct jump sizes, train the method with the single-Gaussian jump ansatz, and compare the conditional densities sampled by Algorithm 2 with the true ones under Wasserstein-2; if the distance does not shrink with more training data, the restricted generator family is the reason. A second check: estimate the Lipschitz constant $K$ in Remark 7 from the data and evaluate $(1+20K^2)^n$ for the trajectory length $n$; if the bound exceeds the observed approximation error by many orders, the stated joint guarantee is too loose to explain the empirical success.","tokens_in":17729,"feed_emoji":"📈","tokens_out":9394,"duration_ms":93542,"temperature":0.7,"pith_summary":"The paper develops a generative model for time-continuous stochastic processes from irregularly sampled observations, targeting settings like high-frequency finance and clinical vital signs. Its central move is to construct explicit Markov-process bridges between consecutive observations: a linear-drift SDE and a pure-jump process that both have the same Gaussian marginals at every intermediate time. This makes conditional time-series generation into a simulation-free regression problem, and parameterizing the jump kernel by a scaled Gaussian yields a closed-form KL loss. A memory window plus an autoregressive sampling scheme then approximates the joint distribution of the observed process rather than only its one-step marginals. The approach works with jumps and irregular grids, which previous trajectory flow matching handled poorly.","feed_headline":"Explicit generators make discontinuous, irregular time series trainable","feed_subtitle":"Both drift and jump generators are closed-form, so training needs no simulation and handles irregular sampling.","key_machinery":"The machinery is the pair of generators in Proposition 2, together with the KL identity for Gaussian jump kernels. The drift-diffusion generator $L^{\\mathrm{Diff}}_t f(x)=\\nabla f(x)^T u_t(x)+\\frac{\\eta^2}{2}\\Delta f(x)$ and the jump generator $L^{\\mathrm{Jump}}_t f(x)=\\int(f(y)-f(x))q_t(y,x)\\,dy$ each solve the Kolmogorov forward equation for the same Gaussian density $p_t$, which is what makes a Markov process with exactly $N(m_t,\\tau_t)$ marginals available in closed form. The jump rate uses the positive and negative parts of the same function $\\xi_t$, so the two generators are two sides of one identity; Proposition 5 turns the KL between $q_t(\\cdot,x)$ and a learned $\\lambda N(\\mu,\\sigma^2 I)$ into a formula involving only the mean and covariance of the truncated-Gaussian measure $J_t$, and Proposition 6 supplies those moments in closed form. This is what removes binning and simulation from training.","core_discovery":"For fixed endpoints x0 and x1, the paper proves that the Gaussian interpolation family $N(m_t, \\tau_t)$ with $m_t=(1-t)x_0+tx_1$ and $\\tau_t=\\eta^2 t(1-t)+\\rho^2$ is generated exactly by two different Markov processes. The drift-diffusion generator has drift $u_t = x_1 - x_0 - (x - m_t)\\eta^2 t/\\tau_t$ and diffusion $\\eta$; the jump generator has rate $q_t(y,x)=\\lambda_t(x)J_t(y)$, where $\\lambda_t(x)=\\max(0,-\\xi_t(x))$ and $J_t$ is the renormalized positive part of $\\xi_t$ times the Gaussian density. Both satisfy the Kolmogorov forward equation, so both produce the same marginals; setting $\\rho>0$ replaces the singular Brownian bridge of the earlier approach with a stabilized one. Because the losses in Proposition 4 are Bregman divergences, training the drift and jump networks against these explicit generators yields conditional generators, and Proposition 5 gives the KL between the target jump rate and a learned Gaussian jump kernel in closed form. Algorithm 2 glues these per-interval processes using a memory window, which by Remark 7 approximates the full joint distribution under a Lipschitz condition.","pith_inferences":["Not claimed by the paper: the closed-form KL identity only needs the first two moments of $J_t$, so replacing the single Gaussian jump kernel by a Gaussian mixture or a normalizing flow should keep the loss structure intact.","Not claimed by the paper: marking events with an extra coordinate, such as trade sizes or alarm types, should let the same generator pair model marked point processes; this is a natural extension of the jump generator.","Not claimed by the paper: part of the reported gain over the earlier method may come from predicting velocity rather than denoising; a controlled comparison holding the bridge fixed would separate the two choices.","Not claimed by the paper: because the joint error bound grows as $(1+20K^2)^n$, the practical guarantee on long trajectories depends on $K$ being small; fitting $K$ from data would show whether the bound is informative."],"forward_implications":["Because both generators are known in closed form for every interval, training never has to simulate the SDE or jump process; sampling is deferred to Algorithm 2.","Taking a convex combination of the learned drift-diffusion and jump generators produces another valid generator, so one model can interpolate between purely continuous and purely jumpy dynamics.","The closed-form KL loss for Gaussian jump kernels removes the need for binning the jump rate, making the loss cheap and differentiable.","With a fixed memory window, the autoregressive sampler approximates the full joint distribution of the observations, not only the one-step marginals, up to an error governed by the Wasserstein-Lipschitz constant of the conditional laws.","The stabilized bridge with $\\rho>0$ removes the $t=1$ singularity of the previous method, so training and inference do not require an ad hoc noise floor."],"supporting_citations":[{"why":"Supplies the generator-matching framework, the Bregman loss identities, and the Markov superposition principle that the paper adapts to conditional time-series generation.","marker":"[14]"},{"why":"Is the trajectory flow matching baseline whose singular Brownian-bridge drift is replaced by the stabilized generators.","marker":"[32]"},{"why":"Provides the generator decomposition into drift, diffusion, and jump terms that underlies Proposition 2.","marker":"[28]"},{"why":"Used in Appendix C to bound the Wasserstein effect of the added Gaussian noise when proving joint-distribution approximation.","marker":"[2]"},{"why":"Defines the energy-distance MMD metric used to compare generated and ground-truth trajectories.","marker":"[26]"}],"fun_headline_variants":["Closed-form generators train irregular time series without simulation","Jump and drift generators match irregular time series exactly","Time series generation without simulation via explicit generators","Explicit generators handle jump discontinuities in time series"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes each one-step conditional distribution can be matched by a linear-drift SDE with fixed diffusion plus a single Gaussian jump kernel using a fixed memory window, and for the joint-distribution guarantee it also assumes the true conditional laws are Lipschitz with a known constant; real transitions that are multimodal, heavy-tailed, or long-range dependent fall outside this and come with no error bound.","fun_headline_variants_meta":{"raw":{"variants":["Closed-form generators train irregular time series without simulation","Jump and drift generators match irregular time series exactly","Time series generation without simulation via explicit generators","Explicit generators handle jump discontinuities in time series"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00024,"raw_usage":{"total_tokens":1504,"prompt_tokens":917,"completion_tokens":587,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":528}},"tokens_in":533,"tokens_out":587,"duration_ms":5824,"temperature":1.0,"reasoning_tokens":528,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:53:40.931812+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a dataset from a transition kernel that is a mixture of two well-separated Gaussians, or a jump process with two distinct jump sizes, train the method with the single-Gaussian jump ansatz, and compare the conditional densities sampled by Algorithm 2 with the true ones under Wasserstein-2; if the distance does not shrink with more training data, the restricted generator family is the reason. A second check: estimate the Lipschitz constant $K$ in Remark 7 from the data and evaluate $(1+20K^2)^n$ for the trajectory length $n$; if the bound exceeds the observed approximation error by many orders, the stated joint guarantee is too loose to explain the empirical success.","supporting_citations":[{"cited_title":"Holderrieth, M","cited_arxiv_id":null,"evidence_quote":"Supplies the generator-matching framework, the Bregman loss identities, and the Markov superposition principle that the paper adapts to conditional time-series generation."},{"cited_title":"Zhang, Y","cited_arxiv_id":null,"evidence_quote":"Is the trajectory flow matching baseline whose singular Brownian-bridge drift is replaced by the stabilized generators."},{"cited_title":"von Waldenfels","cited_arxiv_id":null,"evidence_quote":"Provides the generator decomposition into drift, diffusion, and jump terms that underlies Proposition 2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the energy-distance MMD metric used to compare generated and ground-truth trajectories."}],"review_version":1}