{"id":"b68d7e7f-37aa-477d-a0d1-9e1bc93c5a69","arxiv_id":"2611.16995","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Entire sig-jump-diffusions are introduced, their expected signature is shown to satisfy linear ODEs, and this is applied to score-based generative modeling for time series and finite-dimensional data.","lead":"This paper develops a new class of stochastic processes called entire sig-jump-diffusions, whose expected signature can be computed by solving linear ordinary differential equations. The authors then use this framework to build score-based generative models for time series and finite-dimensional data, where learning the score reduces to a deterministic optimization problem.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The entire approximation of the generalized score function (Section 4.3) is not justified and the paper acknowledges that numerical validation is left to future work, making the central claim of a tractable learning procedure unsubstantiated.","rationale":"The reader's verdict is CONDITIONAL, with the weakest assumption being the entire approximation of the score function. I agree that this is the most load-bearing concern. However, I argue that the verdict should be REJECT rather than CONDITIONAL because the paper's central claim for generative modeling (Theorem 4.9) is presented as a reduction of score learning to a deterministic optimization problem, but this reduction relies entirely on an unverified and unjustified approximation. The paper does not provide any theoretical justification for the entire approximation (e.g., approximation bounds, density results, or conditions under which the score is entire). The paper explicitly states that numerical validation is left to future work, meaning there is no evidence that the proposed method works even in simple cases. A CONDITIONAL verdict would be appropriate if the paper had a plausible path to verification (e.g., a clear set of conditions under which the approximation holds, or a numerical experiment on a toy problem). The paper has neither. Therefore, the central claim is not just unverified but is based on an assumption that is likely false in general (score functions of diffusions are typically not entire functions of the signature components). The paper's own limitations (sentence [b51]) confirm this gap. The reader's weakest_assumption correctly identifies this, but the verdict should be stronger to reflect that the paper's main contribution is unsupported.","tokens_in":43295,"tokens_out":766,"duration_ms":8495,"concrete_test":"Implement the entire approximation for a simple known score function (e.g., the score of a Gaussian distribution under an Ornstein-Uhlenbeck forward process, as in Remark 4.10(ii)). Compute the expected signature-based loss (4.14) using the linear ODE (4.15) and optimize the coefficients u. Then, using the learned entire approximation g_u, simulate the reverse process (4.11) and compare the generated samples' distribution to the true target distribution via a standard metric (e.g., maximum mean discrepancy or Wasserstein distance). If the generated samples fail to match the target distribution within a reasonable tolerance, the entire approximation is insufficient for this simple case, casting doubt on the general claim.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim for generative modeling (Theorem 4.9) reduces score learning to a deterministic optimization problem by approximating the generalized score function with an entire function in time and signature components (Section 4.3). This entire approximation is a strong, unverified regularity assumption. The paper acknowledges that 'a numerical investigation of the generative models... is left to future work' (sentence [b51]) and that 'the practical validity of this assumption is untested' (reader's weakest_assumption). The entire approximation is used to express the denoising score matching objective as a linear functional of the expected signature, which is then computed via the linear ODE (Theorem 3.13). If the score function is not entire, the approximation error is uncontrolled, and the learned score may not correspond to the true reverse process. The paper provides no bounds on this approximation error, no discussion of the function space in which the approximation is made, and no convergence guarantees. This is not a minor technical gap; it is the core step that connects the theoretical machinery of expected signatures to the practical learning algorithm. Without it, the generative modeling framework remains a theoretical construction with no demonstrated viability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper introduces a class of jump-diffusions called entire sig-jump-diffusions, whose expected signature can be computed by solving linear ODEs in the extended tensor algebra. It then applies this framework to score-based generative modeling for time series and finite-dimensional data, proposing a method to learn a generalized score function via an entire approximation, which reduces the learning problem to a deterministic optimization problem. The paper also discusses polynomial jump-diffusions and an alternative inverse signature matching approach.","tokens_in":43711,"tokens_out":1190,"duration_ms":9102,"significance":"If the central claims hold, the paper provides a novel theoretical connection between expected signatures and generative modeling, potentially enabling tractable score learning for jump-diffusions. The theoretical development of entire sig-jump-diffusions and their expected signature computation is mathematically sound and extends prior work. However, the practical viability of the generative modeling framework is not demonstrated, and the key assumption that the generalized score function can be approximated by an entire function is unverified.","major_comments":[{"comment":"The entire approximation of the generalized score function (Section 4.3) is the central step that connects the theoretical machinery to a practical learning algorithm. The paper acknowledges that numerical investigation is left to future work (sentence [b51]) and provides no bounds on the approximation error, no discussion of the function space in which the approximation is made, and no convergence guarantees. Without this, the claim that score learning reduces to a deterministic optimization problem is unsubstantiated. This is a load-bearing issue for the generative modeling application.","section":"Section 4.3, Theorem 4.9","authors_acknowledge":"Finally, we stress that the goal of the paper is to establish the framework and the analytic learning procedure, while a numerical investigation of the generative models, on time series as well as on finite-dimensional data, is left to future work.","authors_acknowledge_p":0.62},{"comment":"The derivation of the denoising score matching objective (4.14) relies on the conditional generalized score identity (4.12). However, the existence and uniqueness of such a conditional generalized score for the lifted truncated signature process Z^≤N is not established. The paper assumes that (4.12) holds for all admissible test functions f, but this is a nontrivial condition that requires justification, especially given that the law P^N_t is supported on a submanifold (group-like elements) of T^N(R^d).","section":"Section 4.3, equation (4.14)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Theorem 4.9 states that the optimization problem (4.13) is equivalent to (4.17) under the assumption that all conditions of Corollary 3.15 are satisfied. However, the conditions of Corollary 3.15 include that the coefficient sequence ṽ(u) belongs to D(L) and that the corresponding integrability conditions hold. These conditions are not verified for the specific ṽ(u) arising from the entire approximation of the score. The paper does not provide sufficient conditions on the coefficients u to ensure these hold.","section":"Section 4.3, Theorem 4.9","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Theorem 4.11 constructs an entire sig-diffusion with a unique invariant measure. However, this construction is for a pure diffusion (δ=0) and does not include jumps. The paper's title and main framework include jumps, but the invariant measure result is limited to the diffusion case. The applicability to jump-diffusions in the generative modeling context is therefore restricted.","section":"Section 4.4, Theorem 4.11","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The paper is well-written and mathematically rigorous. The notation is consistent and the structure is clear.","section":"General","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The assumption that F(E) < ∞ (finite measure space) is used to simplify the proof of Lemma 3.18. The paper mentions this can be relaxed, but it would be helpful to discuss how.","section":"Section 3.3.2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The definition of the generalized score function S_t is given in terms of the operator K and the marginal density p_t. It would be helpful to clarify the domain of S_t and the precise integrability conditions required for (4.2) to hold.","section":"Section 4.1, Definition 4.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The inverse signature matching approach relies on the assumption that the expected signature characterizes the law of the process (Chevyrev and Oberhauser, 2022). This is a strong assumption that is not discussed in the context of the specific polynomial diffusions used.","section":"Section 4.5","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The numerical validation in Example 3.28 is for a specific polynomial jump-diffusion and shows agreement with Monte Carlo. This is a nice illustration of the expected signature computation, but it does not validate the generative modeling framework.","section":"Example 3.28","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper uses the time-extended signature to incorporate time dependence. It would be helpful to explicitly state how the time component is handled in the ODE (4.15).","section":"Section 4.3","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Christa,\n\nThis paper does two things. First, it extends the class of signature SDEs (Cuchiero et al. 2026b) to processes with jumps, calling them entire sig-jump-diffusions. The main theoretical result is that the expected signature of such processes can be computed by solving a linear ODE in the extended tensor algebra (Theorem 3.13). The extension to jumps is nontrivial because the series convergence is harder, and the authors give conditions under which it holds. They also show that polynomial jump-diffusions embed into this framework and that their expected signature reduces to a finite-dimensional matrix exponential (Proposition 3.24, Corollary 3.25). This part is clean and appears mathematically sound. The numerical example in Section 3.3.3 (Example 3.28) is a nice validation of the expected signature formula for a specific Ornstein-Uhlenbeck process with jumps, and the code is provided.\n\nSecond, the paper sketches an application to score-based generative modeling for time series and finite-dimensional data. The idea is to use the expected signature formula to reduce learning a generalized score function to a deterministic optimization problem (Theorem 4.9). This is where the soft spots are. The core step is approximating the generalized score function by an entire function in time and signature components (Section 4.3). This is a strong regularity assumption that is not justified. The paper acknowledges that numerical investigation is left to future work (sentence [b51]), so the practical validity of the entire approximation is untested. The stress-test note is right to flag this as a load-bearing gap: without it, the generative modeling framework remains a theoretical construction with no demonstrated viability. The paper also acknowledges that recovering time series from generated signatures is subject to approximation issues not addressed here (sentence [b44]), and that the reverse process is not the true pathwise time reversal but only marginal-preserving (Remark 4.5). These are honest limitations, but they add up.\n\nWho is this paper for? Readers working on signature methods for stochastic processes will find the extension to jumps useful. The generative modeling part is more speculative and will need substantial empirical work before it can be taken as a practical method. The paper deserves a serious referee because the theoretical core (expected signature of entire sig-jump-diffusions) is a genuine contribution and the mathematics is carefully laid out. The referee should press on the entire approximation assumption and ask for either justification or a relaxation. The paper would be stronger if it included even a simple numerical experiment demonstrating the generative modeling pipeline on a toy problem.\n\nMy recommendation: accept for peer review, but flag that the generative modeling claims are conditional on unverified assumptions and that the paper would benefit from numerical validation before publication.\n\nBest,\n[Your name]","headline":"The paper extends signature SDEs to jump processes and shows that the expected signature solves a linear ODE; the generative modeling application is a plausible sketch but rests on an unverified entire approximation and has no numerical validation.","tokens_in":44166,"tokens_out":682,"duration_ms":6460},"pith_extraction":{"msc":["60L10","60J76","68T05"],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"Expected signature of entire sig-jump-diffusions solves linear ODEs, enabling score-based generative modeling","keywords":["signature","expected signature","jump-diffusion","entire sig-jump-diffusion","score-based generative model","time series generation","polynomial process","inverse signature matching"],"falsifier":"Compute the expected signature of a simple polynomial jump-diffusion (e.g., an Ornstein-Uhlenbeck process with jumps) using the matrix exponential formula (Corollary 3.25) and compare it to a Monte Carlo estimate. If the two disagree beyond Monte Carlo error for a range of parameters, the theoretical framework is contradicted. (The paper provides such a comparison in Example 3.28 for a specific process, showing agreement.)","tokens_in":44810,"feed_emoji":"📈","tokens_out":1715,"duration_ms":156859,"temperature":0.7,"pith_summary":"This paper introduces a class of jump-diffusions called entire sig-jump-diffusions, whose characteristics depend linearly on potentially infinitely many components of their own signature. The main theoretical result is that the expected signature of such a process can be computed by solving a linear ordinary differential equation in the extended tensor algebra (Theorem 3.13). This tractability is then exploited to develop score-based generative models for time series and finite-dimensional data. The learning of a generalized score function is reduced to a deterministic optimization problem that involves solving linear ODEs, bypassing neural network training. The paper also shows that polynomial jump-diffusions embed into this framework and proposes an alternative inverse signature matching method for finite-dimensional data generation.","feed_headline":"Expected signature of jump-diffusions solved by linear ODE","feed_subtitle":"A new class of entire sig-jump-diffusions makes score-based generative modeling for time series tractable without neural networks.","key_machinery":"The central object is the signature of a càdlàg semimartingale, an element of the extended tensor algebra whose components are iterated integrals. The key identity is the linear ODE (3.5) for the expected signature: $\\frac{d}{dt} \\langle c(t), x \\rangle = \\langle L(c(t)), x \\rangle$, where $L$ is an operator defined in terms of the sig-coefficients (Definition 3.6, equation (3.3)). This operator $L$ encodes the drift, diffusion, and jump characteristics of the entire sig-jump-diffusion, and its linearity in the tensor algebra is what makes the expected signature tractable.","core_discovery":"The paper establishes that entire sig-jump-diffusions, defined as semimartingales whose drift, diffusion, and jump characteristics are linear functionals of their own signature (Definition 3.3), have expected signatures governed by a linear ODE in the extended tensor algebra (Theorem 3.13). This generalizes previous work on continuous signature SDEs to the jump-diffusion setting. The result is then applied to score-based generative modeling: by approximating the generalized score function (Definition 4.1) with an entire function in time and signature components, the score-matching objective reduces to solving a linear ODE and a deterministic optimization problem (Theorem 4.9). The paper further proves that polynomial jump-diffusions are a subclass of entire sig-jump-diffusions, so their expected signature reduces to a matrix exponential (Proposition 3.24, Corollary 3.25). For finite-dimensional data, an alternative inverse signature matching method is proposed that learns the score by matching expected signatures of forward and reverse processes (Section 4.5).","pith_inferences":["If the entire approximation of the generalized score is valid in practice, the method could dramatically reduce the computational cost of score-based generative modeling for time series, as the learning phase becomes a deterministic optimization problem rather than a stochastic one.","The inverse signature matching method (Section 4.5) suggests a new paradigm for generative modeling where one matches expected signatures of forward and reverse processes, which could be extended beyond polynomial diffusions to any class where expected signatures are tractable.","The paper's reliance on the entire approximation (Section 4.3) is a strong assumption that may limit practical applicability; numerical experiments are needed to assess whether this approximation holds for realistic data distributions.","The connection between entire sig-jump-diffusions and the universal approximation property of signatures implies that any sufficiently regular jump-diffusion can be approximated by an entire sig-jump-diffusion, potentially making the framework widely applicable if the convergence conditions can be verified."],"forward_implications":["Time series generation can be performed by perturbing the empirical distribution of truncated signatures with an entire sig-jump-diffusion, then running a reverse process with a learned generalized score, all without training a neural network for the score.","The expected signature of any polynomial jump-diffusion can be computed via a matrix exponential, enabling fast and accurate moment computations for such processes.","The inverse signature matching method provides a way to learn the reverse drift of a polynomial diffusion by matching expected signatures, reducing the problem to a deterministic optimization that avoids score matching entirely.","The existence of invariant measures for lifted entire sig-diffusions (Theorem 4.11) guarantees that the forward perturbation process forgets the initial data distribution, a prerequisite for score-based generative modeling.","The framework unifies and extends prior work on signature SDEs, holomorphic jump-diffusions, and polynomial processes under a single formalism for expected signature computation."],"supporting_citations":[{"why":"Provides the continuous (no-jump) version of signature SDEs and the duality method for expected signature computation that this paper extends to the jump-diffusion setting.","marker":"Cuchiero et al. (2026b)"},{"why":"Introduces holomorphic jump-diffusions, which are shown to embed into the entire sig-jump-diffusion framework; the ODE method for expected signatures is related.","marker":"Cuchiero et al. (2026a)"},{"why":"Defines polynomial jump-diffusions, which are shown to be a subclass of entire sig-jump-diffusions and whose expected signature reduces to a matrix exponential.","marker":"Cuchiero et al. (2012); Filipović and Larsson (2020)"},{"why":"Establishes that the expected signature characterizes the law of a stochastic process under suitable conditions, justifying the use of expected signature matching in generative modeling.","marker":"Chevyrev and Oberhauser (2022)"},{"why":"Provides the theoretical foundation for invariant measures of hypoelliptic diffusions on Lie groups, used in Theorem 4.11 to construct invariant measures for lifted entire sig-diffusions.","marker":"Baudoin et al. (2008)"},{"why":"Establishes the score-based generative modeling framework via SDEs that this paper extends to jump-diffusions and signature space.","marker":"Song et al. (2021)"},{"why":"Introduces the family of marginal-preserving reverse dynamics for jump-diffusions with isotropic α-stable noise, which this paper generalizes to arbitrary jump-diffusions via the generalized score.","marker":"Popov et al. (2025)"}],"fun_headline_variants":[],"cache_read_input_tokens":0,"weakest_assumption_plain":"The generalized score function can be approximated arbitrarily well by an entire function in time and signature components (Section 4.3). This entire approximation is used to reduce score learning to a deterministic optimization problem; if it fails, the learning procedure may not converge to the true score.","fun_headline_variants_meta":{"raw":{"variants":[]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":7.6e-05,"raw_usage":{"cost":7.57302e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":7.57302e-05,"upstream_inference_prompt_cost":7.53662e-05,"upstream_inference_completions_cost":3.64e-07},"total_tokens":1035,"prompt_tokens":1033,"completion_tokens":2,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":0}},"tokens_in":1033,"tokens_out":2,"cost_source":"openrouter_usage","duration_ms":3025,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":0,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-13T07:35:37.276800+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Compute the expected signature of a simple polynomial jump-diffusion (e.g., an Ornstein-Uhlenbeck process with jumps) using the matrix exponential formula (Corollary 3.25) and compare it to a Monte Carlo estimate. If the two disagree beyond Monte Carlo error for a range of parameters, the theoretical framework is contradicted. (The paper provides such a comparison in Example 3.28 for a specific process, showing agreement.)","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines polynomial jump-diffusions, which are shown to be a subclass of entire sig-jump-diffusions and whose expected signature reduces to a matrix exponential."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that the expected signature characterizes the law of a stochastic process under suitable conditions, justifying the use of expected signature matching in generative modeling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the theoretical foundation for invariant measures of hypoelliptic diffusions on Lie groups, used in Theorem 4.11 to construct invariant measures for lifted entire sig-diffusions."},{"cited_title":"P., Kumar, A., Ermon, S., and Poole, B","cited_arxiv_id":null,"evidence_quote":"Establishes the score-based generative modeling framework via SDEs that this paper extends to jump-diffusions and signature space."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the family of marginal-preserving reverse dynamics for jump-diffusions with isotropic α-stable noise, which this paper generalizes to arbitrary jump-diffusions via the generalized score."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}