{"id":"e039bfa1-4fc0-429b-aaa9-4f176e6ce0bd","arxiv_id":"2505.22646","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A signature-based moment-matching estimator is proved consistent for linear signature SDEs, with existence and uniqueness of solutions established for bounded driving rough paths.","lead":"This paper develops a method for estimating the unknown parameters of path-dependent stochastic differential equations using path signatures, and proves the estimator is consistent under regularity conditions. The method generalizes an existing moment-matching approach to a broader class of non-Markovian models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The consistency theorem proves only existence of a good root; the estimator as used in the experiments is selected with the true parameter, so ESMM is not validated as a standalone consistent estimator.","rationale":"The paper contains substantial mathematical contributions: the rough-path existence and uniqueness results for signature SDEs, the polynomial expression for Picard iterations in Theorem 5.4, and a consistency statement with explicit rates in Theorem 5.11 are nontrivial and largely carefully argued. The reader's verdict of CONDITIONAL is appropriate because the core theory is likely repairable. My concern is distinct from but adjacent to the reader's weakest assumption. The reader emphasizes that the invertible-Jacobian condition is not derived; I agree, and Section 7 shows non-identifiability can make it fail. However, the more decisive issue for the paper's central claim is that Theorem 5.11 does not actually define an estimator: it asserts that good roots exist, but the method section does not specify how to choose among multiple roots of (5.8). The empirical section resolves this by using the true parameter in Eq. (6.4), a selection rule that is unavailable in practice. Thus the numerical evidence supports the existence of an accurate root when aided by the answer key, not the consistency of ESMM as a standalone procedure. This does not change the verdict from CONDITIONAL, but it sharpens the condition: the paper should either specify and analyse a computable root-selection rule, or clearly restrict the claim to 'there exists a consistent root-finding oracle.'","tokens_in":36926,"tokens_out":2990,"duration_ms":38714,"concrete_test":"Re-run Experiments 6.2–6.4 with the GitHub code, replacing Eq. (6.4) by a rule that does not use θ0, e.g., choose the real solution to (6.3) with minimal residual norm (or minimal parameter norm). Also compute the Jacobian JP(θ0) numerically for the reported models and word sets to verify invertibility. If no θ0-free rule recovers parameters to the reported accuracy, or if JP(θ0) is singular, the ESMM as a standalone consistent estimator is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing gap is that Theorem 5.11 establishes only existence of a root of (5.8) near θ0 under an invertibility condition; it does not define the ESMM estimate. The polynomial system can have multiple real solutions: Remark 6.5 already exhibits a sign-flipped diffusion root, and Section 7 exhibits distinct parameters with identical laws. The only numerical selection rule, Eq. (6.4), chooses the real solution closest to the true θ0 in L1, which is unavailable in any real estimation problem. Consequently the experiments in Section 6 cannot be read as validating ESMM: they validate the existence of a good root when the answer key is used. Moreover, the invertibility assumption at θ0 is not checked or derived for the models in Experiments 6.2–6.4, and Proposition 5.12 only gives differentiability almost everywhere, so the consistency theorem is conditional on a nondegeneracy hypothesis that may fail precisely in the non-identifiable cases the paper documents.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a theory of path-dependent signature SDEs, proves existence and uniqueness of solutions on short time intervals via rough path methods, and introduces an Expected Signature Matching Method (ESMM) for estimating drift and diffusion parameters. The main theoretical result, Theorem 5.11, states that if the restricted expected signature map P is differentiable at the true parameter with invertible Jacobian, then for sufficiently many samples N and Picard iterations r, the polynomial system (5.8) has a solution arbitrarily close to the true parameter. The paper also includes numerical experiments, a non-identifiability construction in Section 7, and explicit rates in the consistency theorem.","tokens_in":37174,"tokens_out":11046,"duration_ms":136395,"significance":"If the central consistency claim were fully established, the paper would be a valuable contribution to statistical inference for path-dependent SDEs. The paper has clear strengths: it gives an explicit polynomial representation of Picard iterations (Theorem 5.4), provides uniform noise bounds for solutions, states an explicit convergence rate in (5.11), and ships reproducible code. The non-identifiability analysis in Section 7 is a useful caveat. However, the consistency theorem as stated proves only existence of a good root, not that the ESMM estimator as used in practice converges, and the proof of the key uniform bounds contains a gap. These issues are load-bearing for the paper's headline claim.","major_comments":[{"comment":"The proof that ω(s,t)=C d_p(X|_{Δ_{s,t}},0)^p is a control is not valid as written. The argument shows, for each fixed level i, that V_i(s,u)+V_i(u,t) ≤ V_i(s,t) ≤ d_p(X|_{Δ_{s,t}},0)^p, where V_i is the sum over a partition of the i-th level increments. But d_p^p is the maximum over i of V_i, and the sum of two such maxima need not be bounded by the maximum of the individual sums. Since the superadditivity of ω is used to apply the Universal Limit Theorem and to obtain the uniform bounds in (B.6)-(B.7), this gap propagates to Lemma 5.7 and Theorem 5.11. The authors should either prove the claimed superadditivity, or replace ω by a true control built from level-wise variations and adjust the constants.","section":"Proposition 4.1 and Appendix B"},{"comment":"The consistency theorem establishes only that the polynomial system (5.8) has a solution in B_ε(θ0); it does not define a selection rule for the ESMM estimate. The experiments, however, define the estimate as the real solution closest to the true θ0 in (6.4), which is unavailable in any real estimation problem. Moreover, Remark 6.5 and Section 7 show that multiple real solutions can exist, including sign-flipped diffusion roots and distinct parameters with identical laws. Consequently, the numerical results in Section 6 do not validate ESMM as a standalone consistent estimator; they validate the existence of a good root when the answer key is used. The authors should either provide a selection rule that does not depend on θ0 and prove convergence for that rule, or explicitly reframe the contribution as existence of consistent roots rather than a consistent estimator.","section":"Theorem 5.11 and Section 6, Eq. (6.4)"},{"comment":"The experiments solve the polynomial system (6.3), which omits the indicator χ_{E_ξ} and replaces E[X^J χ_{E_ξ}] by E[X^J] in the coefficients. This differs from the estimating equations (5.8) and (5.7) for which consistency is proven. Remark 6.1 justifies this by the smallness of T, but no rate or verification is provided for the values T=0.1 and T=0.2 used in the experiments. As a result, the experiments are not a direct numerical check of Theorem 5.11, and the discrepancy should be addressed or the claims about empirical validation should be softened.","section":"Section 6, Eq. (6.3) and Remark 6.1"}],"minor_comments":[{"comment":"The caption of Figure 3 refers to word sets W1 and W2, but the sets defined in (6.6) are W3 and W4; this should be corrected.","section":"Figure 3 caption"},{"comment":"In the displayed system for Experiment 6.4, the equation for dY_t^{(3)} is missing the dt factor in the drift term; it should read θ_5(Y^{(2,1)}_t - Y^{(1,2)}_t)dt + θ_6 ◦ dW^{(3)}_t.","section":"Experiment 6.4"},{"comment":"Remark 5.1 asserts an error in [35, Equation 3.19] and proposes a correction, but no detailed derivation is given. Since this is a strong claim about a published Annals of Statistics paper, the authors should either provide the full counterexample or derivation, or soften the wording.","section":"Remark 5.1"},{"comment":"In the statement of Theorem 5.11, 'N 0, r0' should read 'N_0, r_0' consistently with the rest of the paper.","section":"Theorem 5.11 statement"},{"comment":"Proposition 5.12 proves only local Lipschitzness and hence differentiability almost everywhere; the experiments do not check the stronger hypothesis of Theorem 5.11 that the Jacobian of P is invertible at the true θ0. A sentence acknowledging this gap for the simulated models would be helpful.","section":"Proposition 5.12"}],"recommendation":"major_revision","confidential_remarks":"The root-selection problem is the most serious issue: the consistency theorem does not match the estimator actually implemented in the experiments. In my view this is fixable by reframing the paper's contribution as proving existence of consistent roots and presenting the experiments as an exploration of the method's behavior, rather than as validation of a fully specified consistent estimator. The control-function gap in Proposition 4.1 also needs a rigorous repair, but it appears local and repairable. I would not recommend rejection, since the existence theory and polynomial structure are novel and potentially useful, but the current claims overreach what is proved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the existence and uniqueness theorems for signature SDEs (Theorems 3.9 and 4.1) are genuinely new and fill a gap in Cuchiero et al. Second, the consistency theorem (5.11) is real but only shows that a good root exists; the numerical evaluation then picks that root using the true parameter. So the paper is a solid theoretical contribution wrapped around an estimation method that, as presented, is not validated as standalone.\n\nWhat is good: The extension of the Expected Signature Matching Method from polynomial path-independent SDEs to linear signature SDEs is natural, and the polynomial Picard iteration expressions in Theorem 5.4 are a nice computational feature. The proof of consistency under weaker assumptions than Papavasiliou-Ladroue—only invertibility at θ0 rather than uniform control on all Jacobians—is a genuine improvement. The authors also deserve credit for Section 7, where they openly show distinct parameters can give identical laws; that is an honest treatment of identifiability.\n\nWhere the paper is soft: The proof of Proposition 4.1 contains a questionable step. It asserts that dp(X|Δs,u,0)^p + dp(X|Δu,t,0)^p ≤ dp(X|Δs,t,0)^p, but dp is a maximum over levels, and the usual p-variation metric does not obviously satisfy that superadditivity. This is a real gap in the uniform bounds that Lemma 5.7 and Theorem 5.11 rely on. More importantly, Theorem 5.11 proves existence of a solution to (5.8) near θ0, but the ESMM estimator is not defined as a specific root. In the experiments, Eq. (6.4) picks the real solution closest to the true θ0—information no user has. So the experiments demonstrate that good roots exist, not that ESMM recovers parameters without the answer key. The invertibility condition on P at θ0 is also not checked for the models in Experiments 6.2–6.4, and the paper's own Section 7 shows cases where it may fail.\n\nNet: The theoretical core is worth refereeing. The gap in Proposition 4.1 needs fixing or a modified argument. The empirical section should be reframed—either propose a concrete selection rule, perhaps via homotopy continuation or additional moment conditions, or explicitly present the experiments as evidence of feasibility under oracle selection. This is a revise-and-resubmit, not a desk reject. I would bring it to a reading group and would accept a referee assignment.","headline":"The existence/uniqueness theory for signature SDEs is new and the consistency proof improves on prior work, but the estimation method is undercut by a proof gap and by experiments that pick the root using the true parameter.","tokens_in":37685,"tokens_out":2193,"would_cite":true,"duration_ms":25859,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60L20","60L90","62M99","62M09"],"pacs":[],"model":"deepseek-v4-flash","headline":"Path-signature SDEs get a provably consistent estimator.","keywords":["path signatures","rough paths","path-dependent stochastic differential equations","signature SDEs","Expected Signature Matching Method","consistent estimator","parameter estimation","identifiability"],"falsifier":"Construct a linear signature SDE and a word set for which the restricted expected signature map $P$ has a singular Jacobian at the true parameter or is not differentiable there, and show that the polynomial system (5.8) then either has no solution in small neighborhoods of $\\theta_0$ or that two different parameters produce the same solution law on $E_\\xi$; the paper's own Section 7 construction already gives candidate parameters with identical restricted laws, so testing whether ESMM can separate them would settle the practical reach of the theorem.","tokens_in":36730,"feed_emoji":"📈","tokens_out":4390,"duration_ms":51126,"temperature":0.7,"pith_summary":"This paper develops and proves the consistency of a parameter-estimation method for a broad class of path-dependent stochastic differential equations, the linear signature SDEs, whose drift and diffusion are linear functionals of the path signature. The method, Expected Signature Matching, matches the empirical expected signature of observed trajectories to the expected signature of Picard iterations of the model, expressed as explicit polynomials in the unknown parameters. The main theorem states that under a nondegeneracy condition, with enough samples and enough Picard iterations the estimated parameters land arbitrarily close to the true ones, almost surely, with an explicit exponential convergence rate in the number of Picard iterations. If correct, this extends moment-matching parameter estimation from Markovian polynomial SDEs to a flexible class of path-dependent models.","feed_headline":"Path-signature SDEs get a provably consistent estimator","feed_subtitle":"New method matches expected path signatures to recover drift and diffusion in path-dependent models.","key_machinery":"The central objects are signature SDEs, differential equations whose drift and diffusion are linear functionals of the truncated path signature, and the Expected Signature Matching Method (ESMM). The argument is carried by three pieces: the lifted RDE formulation that turns a path-dependent equation into a path-independent one on the tensor algebra, so the Universal Limit Theorem gives existence, uniqueness, and Picard convergence; the polynomial representation of Picard iterations in the parameter $\\theta$, with degree at most $Q(r,|I|)=O(2^r)$; and the Miranda fixed-point theorem, which converts sign conditions on the boundary of a small cube into existence of a nearby solution of the polynomial system.","core_discovery":"The central discovery is that for linear signature SDEs, the expected signature of the r-th Picard iteration is a polynomial in the parameter $\\theta$ whose coefficients are determined by the expected signature of the driving signal, and that this polynomial approximation converges uniformly to the restricted expected signature of the true solution. The paper then proves that if the limiting map $P(\\theta)$ is differentiable at the true parameter with an invertible Jacobian, the polynomial system formed by matching these polynomials to empirical expected signatures admits a solution in any $\\varepsilon$-ball around the true parameter, almost surely, for sufficiently large sample size $N$ and Picard depth $r$. Along the way it establishes existence and uniqueness of solutions to signature SDEs on short time intervals through the rough-path Universal Limit Theorem, which legitimizes the Picard expansion on which the estimator relies.","pith_inferences":["The choice of words $I_1,\\ldots,I_d$ is a practical tuning parameter: the invertibility of the Jacobian and hence the success of ESMM depends on it, so a natural extension is to study data-driven word selection.","The non-identifiability examples in Section 7 suggest that ESMM actually identifies an equivalence class of parameters rather than a unique one; characterizing this class or adding regularizers would extend the method's usefulness.","The consistency proof appears adaptable to the earlier path-independent polynomial SDE setting, potentially repairing the gap the paper identifies in that prior proof; re-deriving the theorem in that setting would be a concrete test extension.","For small time horizons the indicator term $\\chi_{E_\\xi}$ can be dropped, and the simulations confirm the method still works; probing longer horizons where the bound becomes active would test the scope of the theoretical restriction."],"forward_implications":["ESMM is consistent: for any $\\varepsilon>0$, almost surely there exist $N_0, r_0$ such that for all $N \\geq N_0$ and $r \\geq r_0$, the polynomial system has a solution within $\\varepsilon$ of the true parameter.","Existence and uniqueness of solutions of signature SDEs hold on short time intervals, with Picard iterations converging to the unique solution.","The method generalizes expected signature matching from path-independent polynomial SDEs to path-dependent signature SDEs, with smaller polynomial degrees than in the earlier setting.","The restricted expected signature map $P$ is locally Lipschitz, hence differentiable almost everywhere, so the differentiability assumption is automatically satisfied at Lebesgue-almost every parameter point.","The estimator converges exponentially in the Picard depth $r$, with an explicit threshold given in the paper."],"supporting_citations":[{"why":"Supplies the Expected Signature Matching template and the path-independent polynomial SDE class that this paper generalizes.","marker":"[35]"},{"why":"Introduces linear signature SDEs, the model class whose parameter estimation is the object of this paper.","marker":"[12]"},{"why":"Provides the Universal Limit Theorem and Picard-iteration convergence that underlie existence, uniqueness, and the polynomial representation.","marker":"[27]"},{"why":"Establishes the characteristic property of expected signatures used to justify matching expected signatures for parameter recovery.","marker":"[10]"},{"why":"States the generalized Miranda theorem that turns boundary sign conditions into existence of a nearby solution in the consistency proof.","marker":"[30]"},{"why":"Supplies the rough-path theory and the local Lipschitz dependence of solutions used to prove that $P$ is locally Lipschitz.","marker":"[15]"}],"fun_headline_variants":["Path signatures yield consistent SDE parameter estimates","New method matches signatures to infer SDE drift and diffusion","Provably accurate estimation for path-dependent SDEs","ESMM: consistent parameter recovery for signature SDEs","Signature SDEs: existence, uniqueness, and consistent fitting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the restricted expected signature map $P$ is differentiable at the true parameter with an invertible Jacobian, a condition the model itself does not guarantee because Section 7 exhibits distinct parameters giving the same distribution.","fun_headline_variants_meta":{"raw":{"variants":["Path signatures yield consistent SDE parameter estimates","New method matches signatures to infer SDE drift and diffusion","Provably accurate estimation for path-dependent SDEs","ESMM: consistent parameter recovery for signature SDEs","Signature SDEs: existence, uniqueness, and consistent fitting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000223,"raw_usage":{"total_tokens":1446,"prompt_tokens":921,"completion_tokens":525,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":447}},"tokens_in":537,"tokens_out":525,"duration_ms":5619,"temperature":1.0,"reasoning_tokens":447,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:03:30.328070+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a linear signature SDE and a word set for which the restricted expected signature map $P$ has a singular Jacobian at the true parameter or is not differentiable there, and show that the polynomial system (5.8) then either has no solution in small neighborhoods of $\\theta_0$ or that two different parameters produce the same solution law on $E_\\xi$; the paper's own Section 7 construction already gives candidate parameters with identical restricted laws, so testing whether ESMM can separate them would settle the practical reach of the theorem.","supporting_citations":[{"cited_title":"Parameter estimation for rough differential equations","cited_arxiv_id":null,"evidence_quote":"Supplies the Expected Signature Matching template and the path-independent polynomial SDE class that this paper generalizes."},{"cited_title":"Differential Equations Driven by Rough Paths","cited_arxiv_id":null,"evidence_quote":"Provides the Universal Limit Theorem and Picard-iteration convergence that underlie existence, uniqueness, and the polynomial representation."},{"cited_title":"Signature moments to characterize laws of stochastic processes","cited_arxiv_id":null,"evidence_quote":"Establishes the characteristic property of expected signatures used to justify matching expected signatures for parameter recovery."},{"cited_title":"A generalized theorem of Miranda and the theorem of Newton–Kantorovich","cited_arxiv_id":null,"evidence_quote":"States the generalized Miranda theorem that turns boundary sign conditions into existence of a nearby solution in the consistency proof."},{"cited_title":"Friz and Nicolas B","cited_arxiv_id":null,"evidence_quote":"Supplies the rough-path theory and the local Lipschitz dependence of solutions used to prove that $P$ is locally Lipschitz."}],"review_version":1}