{"id":"19eb4dfc-d4ea-4e0d-ba71-fd407006f43a","arxiv_id":"2607.06652","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"An interarrival embedding makes signature methods work on TPPs, yielding SIGTPP trained on whole trajectories plus three pathwise distributional evaluation metrics.","lead":"The paper lifts discrete event sequences into continuous paths via an interarrival embedding so rough-path signatures can model them, then trains SIGTPP with a global path-level loss. It also supplies three mathematically justified distributional metrics for evaluating generative temporal point process models.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged determinacy gap; the central embedding and empirical claims hold under the paper's stated scope.","rationale":"The strongest claim rests on three pillars: (i) Theorem 6 (Lipschitz + injective embedding with constructive inverse), (ii) the resulting global Sig-W1 training objective for SIGTPP, and (iii) multi-metric empirical superiority (best average rank, ≥19% relative improvement). Pillars (i) and (iii) are carefully supported by the proofs in Appendix B and the bootstrap tables; the only soft link is the invocation of expected-signature determinacy after the pushforward, which the paper itself flags as unverified. Because that gap is already the reader's weakest_assumption and does not invalidate the embedding stability, the metric constructions under dN, or the reported rankings, no further adjustment of the CONDITIONAL verdict is warranted. The concrete radius check above is the natural next verification step already implied by Appendix E.","tokens_in":46907,"tokens_out":536,"duration_ms":7323,"concrete_test":"On the four synthetic processes (PS, IP, H1, H3), estimate the truncated expected signatures of Φ(N) up to degree M=8 (or the first dead degree from Table S4) and check whether the factorial decay of the tensor norms is consistent with infinite radius of convergence (e.g., plot log||E[S_k]|| + log(k!) and verify non-positive slope). If the series clearly diverges on any process, the separation argument for Sig-W1 on that law is unsupported; if it decays, the concern is empirically mitigated for the paper's main experiments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption (unverified infinite-radius moment condition of Theorem 3 for the pushforward laws Φ♯P after the interarrival embedding) is the genuine soft spot for the claim that Sig-W1 separates distinct laws on N. Section 3.4 and Appendix E already state this limitation explicitly; the paper only shows that injectivity of Φ plus the continuous-path determinacy theorem would imply separation for large enough M, without checking the radius-of-convergence condition on any dataset. This does not break the Lipschitz/injectivity results of Theorem 6, the well-definedness of E and W1 under dN, or the empirical ranking claims, which rest on the reported metrics rather than on full determinacy. No stronger internal inconsistency or experimental flaw is load-bearing for the strongest claim as written.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper develops a pathwise framework for generative modelling and evaluation of temporal point processes. It introduces the interarrival embedding Φ, a Lipschitz injective lift from càdlàg counting paths to continuous bounded-variation paths (Theorem 6), so that the signature transform and expected-signature tools apply. Building on Φ, SIGTPP is trained by matching truncated expected signatures of complete embedded trajectories (Sig-W1 loss, Eq. 4) rather than per-event conditional losses. The same counting-path metric d_N is used to justify three distributional discrepancies (energy, W1, Sig-W1) for evaluation. Empirically, SIGTPP is compared to VAE, DDPM, WGAN, GAMMA and DETER on four synthetic and five real datasets under eight metrics, reporting best average ranks, competitive pairwise wins, and relative-score gains of at least 19% against every baseline.","tokens_in":47171,"tokens_out":1081,"duration_ms":10339,"significance":"If the claims hold, the work supplies a usable bridge from rough-path signatures to discrete event sequences and a principled alternative to local likelihood or adversarial objectives for generative TPPs. The embedding proofs (Appendix B), the metric d_N with its OT interpretation, and the explicit energy/W1/Sig-W1 evaluation suite are concrete contributions that the community can reuse even if SIGTPP is not adopted as a default generator. Strengths include detailed stability proofs, bootstrap standard errors, multi-metric evaluation, truncation ablations, and released code. The main theoretical soft spot is the unverified infinite-radius moment condition needed for full expected-signature determinacy of the pushforwards; the paper already flags this in Section 3.4 and Appendix E, so the empirical ranking claims do not rest on it.","major_comments":[{"comment":"Section 3.4 and Theorem 3: the argument that Sig-W1 separates distinct laws on N relies on the pushforward laws Φ♯P satisfying the infinite-radius moment condition of the expected-signature determinacy theorem. The paper only shows injectivity of Φ plus continuous-path determinacy would imply separation for large enough M, and Appendix E correctly notes that such conditions are hard to verify even for continuous processes and are not checked here. This does not invalidate Lipschitz/injectivity of Φ or the well-definedness of E and W1 under d_N, but it does leave the theoretical status of Sig-W1 as a separating metric incomplete. A short discussion of what can be said without the radius condition (e.g., that Sig-W1 is always a pseudometric, and when it is positive in practice) would make the claim precise.","section":null},{"comment":"Section 4.1 and Tables 1–2: model selection uses a rank aggregate over validation diagnostics while checkpoints are chosen by validation L_log(τ). Because several reported metrics (including L_log(τ) and L_λ) enter both selection and evaluation, and because CRPS consistently favours the conditional baselines, it would strengthen the central empirical claim to report a sensitivity check under an alternative selection criterion (e.g., validation W1 or Sig-W1 only) or to hold out one metric family from selection. The current protocol is transparent but leaves open whether the average-rank advantage is partly selection-driven.","section":null}],"minor_comments":[{"comment":"Definition 4 / Figure 1: the figure caption and surrounding text are clear, but a one-line statement that the signature is applied to the time-augmented path t ↦ (t, Φ(η)_t) would help readers who skip the paragraph after Eq. (2).","section":null},{"comment":"Table 3: the ablation is only on TX and SO; a sentence on whether M=3 was also preferred on the synthetic suite would make the truncation choice more uniform.","section":null},{"comment":"Appendix E: the linear-interpolation lift is listed as a limitation; a brief pointer to why step or other schemes were not used (beyond the standard signature literature) would be useful for follow-up work.","section":null},{"comment":"Notation: N is used both for the space of counting paths and for a random counting path; a consistent distinction (e.g., script N vs. N) would reduce occasional ambiguity in Section 3.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The determinacy gap is already disclosed by the authors and is not load-bearing for the embedding theorems or the empirical ranking claims as written. I would not treat it as grounds for major revision or reject. The paper is a solid methods contribution for a machine-learning venue; the main editorial question is whether the journal wants a methods-plus-theory package of this length."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful core is the interarrival embedding Φ (Def. 4) plus Theorem 6: Lipschitz into L1 continuous paths, constructive inverse via the backshift level-set recursion, and Hölder stability on the δ-sieve. That is the piece that was missing; without it, expected-signature tools do not apply to counting paths. They then train SIGTPP with a global Sig-W1 loss on whole trajectories rather than per-event NLL, and they put three distributional discrepancies (E, W1, Sig-W1) on a proper footing under the path metric d_N. Code and data are public.\n\nWhat works: Appendix B is careful (affine L1 bounds, time-warp inequality, discrete formula for d_N, OT interpretation). Experiments are multi-metric with bootstrap SEs, synthetic regimes plus EasyTPP-style real sets, and an ablation on truncation degree. Relative scores and average ranks favour SIGTPP on the pathwise criteria; the paper is honest that CRPS still prefers the conditional baselines and that pointwise MAE rewards the deterministic regressor. That honesty is useful for the subfield.\n\nSoft spots, in proportion: the separation claim for Sig-W1 still leans on the infinite-radius moment condition of the continuous-path determinacy theorem after the pushforward Φ♯P. They state this limitation in §3.4 and Appendix E and do not check it on data. It does not break Lipschitz/injectivity, the well-definedness of E and W1, or the reported rankings. Other limits are ordinary: exponential cost in signature degree, linear interpolation is a modelling choice, and the generator is a standard LSTM-style AR decoder. Baselines are the natural generative peers of similar size; broader comparison would be nice but is not required for the claim as written.\n\nThis is for people who already care about neural TPPs or signature methods for event data. The math and the multi-metric evaluation are solid enough that a serious referee should see it. I would engage, cite the embedding and the metrics, and keep the determinacy caveat in mind.","headline":"Solid pathwise fix for signatures on TPPs: new embedding with real proofs, first signature generative model, and three justified metrics; determinacy gap is real but already flagged and not load-bearing for the empirical claim.","tokens_in":47722,"tokens_out":548,"would_cite":true,"duration_ms":8730,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"An interarrival embedding turns event sequences into continuous paths so signature methods can generate and score whole TPP trajectories.","keywords":["temporal point processes","rough path signatures","interarrival embedding","generative models","signature Wasserstein distance","counting paths","distributional discrepancy"],"falsifier":"Find a pair of distinct TPP laws whose interarrival embeddings have identical expected truncated signatures at every finite degree (or whose empirical Sig-W1 distance collapses to zero while energy or W1 distances remain large), which would show that the signature loss fails to separate counting-path distributions.","tokens_in":47844,"feed_emoji":"⏱","tokens_out":890,"duration_ms":9064,"temperature":0.7,"pith_summary":"Temporal point processes produce sequences of event times, but their natural sample paths are discontinuous step functions. Signature methods from rough path theory give universal features and distributional characterisations for continuous paths of bounded variation, so those tools have been largely out of reach for event data. This paper builds a stable, injective lift—the interarrival embedding—that maps each counting path to a continuous piecewise-linear path that records successive waiting times, then applies the signature on the time-augmented curve. With that lift in hand, the authors train SIGTPP, a generative model that minimises a global signature-Wasserstein loss between complete generated and observed trajectories rather than a sum of per-event losses. The same path metric also yields three mathematically justified distributional discrepancies (energy, Wasserstein-1, and signature-Wasserstein-1) for evaluating generative TPPs. On synthetic and real-world benchmarks, SIGTPP obtains the best average rank across eight complementary metrics and improves relative scores against every baseline by at least 19 percent on average.","feed_headline":"Signatures generate whole event sequences, not one jump at a time","feed_subtitle":"A continuous lift of jump paths lets a path-level loss train and score temporal point processes","key_machinery":"The interarrival embedding Φ: it interpolates each counting path so that the value at successive event times equals the successive interarrival durations, producing a continuous piecewise-linear path of bounded variation on which the (time-augmented) signature is well-defined and injective.","core_discovery":"The interarrival embedding is a Lipschitz, injective map from the space of unit-jump counting paths into continuous paths of bounded variation, with a constructive inverse that is Hölder continuous under a mild separation of interarrival times. This lift makes the expected-signature characterisation available for TPP laws and supports SIGTPP, the first signature-based generative model for temporal point processes trained with a single path-level Sig-W1 loss on complete trajectories.","pith_inferences":["The same embedding could serve as a drop-in feature map for supervised or forecasting models on event sequences, not only for generative training.","Because the lift is constructive and invertible, one could decode signature-space interventions back into event times, enabling controllable generation of sequences with prescribed higher-order statistics.","Marked or high-dimensional TPPs may require a carefully chosen multi-dimensional analogue of the interarrival lift if signature dimension is not to explode."],"forward_implications":["Signature methods can be applied to discrete event sequences without treating them as continuous time series or forcing a parametric intensity.","Generative TPP training can target a single global discrepancy between complete trajectories instead of a sum of local conditional losses.","Energy, W1, and Sig-W1 distances on counting paths become rigorously justified evaluation metrics for generative TPPs.","Pointwise errors such as MAE and MSE are shown to favour deterministic regressors and should not be primary metrics for generative quality."],"fun_headline_variants":["Interarrival lift turns jump paths continuous for full-sequence signatures","sigTPP: path-level Sig-W1 loss generates complete TPP trajectories","Expected signatures now characterise discrete event sequence laws","Lipschitz embedding of counting paths unlocks generative signature TPPs","Three new discrepancies score whole TPP distributions via signatures"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"That the laws of the embedded continuous paths satisfy the infinite-radius moment condition needed for the expected signature to uniquely determine the distribution, so that matching truncated signatures separates distinct event-sequence laws.","fun_headline_variants_meta":{"raw":{"variants":["Interarrival lift turns jump paths continuous for full-sequence signatures","sigTPP: path-level Sig-W1 loss generates complete TPP trajectories","Expected signatures now characterise discrete event sequence laws","Lipschitz embedding of counting paths unlocks generative signature TPPs","Three new discrepancies score whole TPP distributions via signatures"]},"model":"grok-4.5","effort":"low","cost_usd":0.004882,"raw_usage":{"total_tokens":1407,"prompt_tokens":794,"num_sources_used":0,"completion_tokens":88,"cost_in_usd_ticks":48820000,"prompt_tokens_details":{"text_tokens":794,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":525,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":794,"tokens_out":88,"duration_ms":6109,"temperature":1.0,"reasoning_tokens":525,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T00:09:56.490829+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Find a pair of distinct TPP laws whose interarrival embeddings have identical expected truncated signatures at every finite degree (or whose empirical Sig-W1 distance collapses to zero while energy or W1 distances remain large), which would show that the signature loss fails to separate counting-path distributions.","supporting_citations":[],"review_version":1}