{"id":"2b168d5d-3a3b-48f0-b096-0657db389af8","arxiv_id":"2607.22958","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Quickest detection on rough paths reduces to thresholding a linear functional of the path signature, with learned coefficients.","lead":"The paper designs change-point detectors for rough, non-Markovian signals by checking whether a learned linear combination of the signal's iterated-integral 'signature' has crossed a threshold. The same rule is shown to be optimal, in a suitable sense, and robust to adversarial perturbations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Prop 4.1's geometric detector is not a functional of the observed path: f in eq. (22) uses the unobserved pre-change path X^(1) on [θ,t], so Lemma 2.6 cannot be invoked and the statistical guarantees in §4 rest on an unsupported premise.","rationale":"The paper's central optimal-stopping claim (Prop 3.5) is a credible adaptation of the BHR signature framework. The proof via randomized stopping times and the density of signature functionals appears internally consistent; even the non-adaptedness of the loss Y is handled implicitly by conditioning on the observed path and using the independence of the randomization variable Z. The geometry-based detector in Prop 4.1, however, has a genuine gap: the function f in eq. (22) depends on the latent pre-change path X^(1) after the change-point, which is not recoverable from the observed path. Lemma 2.6 requires a continuous functional of the observed path alone, so it cannot be applied to f. This is exactly the reader's identified weakest assumption, and it is load-bearing for the paper's advertised claim that the same structural form arises from intrinsic geometry, as well as for the statistical guarantees in Section 4. It does not undermine the numerical method, which learns the signature coefficient directly from the loss, nor does it invalidate Proposition 3.5. Thus the conditional verdict is appropriate; no adjustment is needed beyond what the reader already recommended.","tokens_in":34883,"tokens_out":22891,"duration_ms":230998,"concrete_test":"Set d=1, use the Brownian example (Example 3.8), fix 0<θ<t<T, and condition on a particular observed path X_{[0,t]}. Draw two latent continuations X^{(1),a}_{[θ,t]} and X^{(1),b}_{[θ,t]} independently from μ1. Because X^{(1)} and X^{(2)} are independent, both draws are compatible with the same observed path; compute f^{(a)}=d_{p-var;[0,t]}(X,X^{(1),a}) and f^{(b)} likewise. The conditional distribution of the CC/d_p-var distance between independent Brownian rough paths is non-degenerate, so P(f^{(a)}≠f^{(b)})>0. This contradicts the existence of a fixed function of the observed path equal a.s. to f, refuting the premise of Prop 4.1. Alternatively, estimate the conditional variance of f given X_{[0,t]} by Monte Carlo; if it is positive, the claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 defines f(\\hat Z)=f(Z,t)=d_{p-var;[0,t]}(Z,X^{(1)}). In the model of §3, X^{(1)} is latent; after θ the observed path equals X^{(2)} and carries no information about X^{(1)} on [θ,t] because X^{(1)} and X^{(2)} are independent. Thus for fixed t>θ, f(X,t) is not σ(\\hat X_s:0≤s≤t)-measurable; it has a non-degenerate conditional distribution given the observed path. Lemma 2.6 is a density theorem for elements of T=C(Λ_T,R), i.e. deterministic continuous functionals of the observed path. The proof of Prop 4.1 says 'Consider f∈T...' and applies Lemma 2.6 to obtain an l^{(1)} independent of the sample. This is where the argument breaks: f is not in T. Consequently the uniform approximation (26), the key inequalities (24)-(25), and Propositions 4.2-4.4 are not established. The numerical experiments optimize l directly for the loss and do not rely on Prop 4.1; Prop 3.5 is also independent. But the advertised claim that the same structural form 'arises from the intrinsic geometry of the observed path' is unsupported by the supplied proof.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a rough-path signature framework for Bayesian quickest detection. The observed signal is modeled as the concatenation of two independent geometric p-rough paths at an unknown change-point θ. The authors show (Prop. 3.5) that the optimal stopping rule for the natural loss processes Y^1 and Y^2 coincides with the first hitting time of a half-plane by a linear functional of the time-augmented signature, extending the signature optimal-stopping theory of [3]. A 'geometric detector' based on the p-variation distance to the pre-change path is introduced (§4), and statistical guarantees on delay and false-alarm probability are derived (Props. 4.1–4.4), along with a repeated-experiment aggregation rule and a distributionally robust minimax extension (§5). Numerical experiments compare the learned signature rules with CUSUM/Shiryaev in Brownian models and with Page–Hinkley in fractional Brownian models, and demonstrate robustness to adversarial total-variation perturbations.","tokens_in":35372,"tokens_out":17690,"duration_ms":172539,"significance":"If Prop. 3.5 is correct, the paper makes a useful contribution by showing that signature half-space hitting times are universal for a class of non-Markovian change-point problems; the proof in Appendix B is a substantial part of the paper. The numerical evidence is promising and the robust formulation is sensible. However, the geometric-detector argument in §4—one of the advertised main contributions—is not valid as written because the detector is not a functional of the observed path. The statistical guarantees in §4 therefore do not follow from the supplied proof. The paper has enough independent value in Prop. 3.5 and the experiments to warrant a major revision, but the theoretical claims in §4 must be either corrected or substantially weakened.","major_comments":[{"comment":"The function f(\\hat Z)=f(Z,t)=d_{p-var;[0,t]}(Z,X^{(1)}) is not a deterministic continuous functional of the observed path. In the model of §3, X^{(1)} is latent and independent of X^{(2)}; for t>θ the observed segment X_{[θ,t]} equals X^{(2)}_{[θ,t]} and contains no information about X^{(1)} on [θ,t]. Hence f(X,t) is not σ(\\hat X_s:0≤s≤t)-measurable. Lemma 2.6 applies only to elements of T=C(Λ_T,R), i.e., functions that map each observed path segment to a real number. The proof's assertion 'Consider f∈T' is therefore unjustified; the uniform approximation (26), inequalities (24)–(25), and Propositions 4.2–4.4 are not established. The inequality E[f(X,t)]≥g((t-θ)^+) in (23) concerns a true but unobserved distance and does not make f an observable path functional.","section":"§4.1, Eq. (22) and proof of Prop. 4.1"},{"comment":"The paper repeatedly claims that the same structural form 'arises independently from the intrinsic geometry of the observed path' (Abstract, §4, §7). Since Prop. 4.1 is the only support for this claim and it fails, the claim is unsupported. The numerical experiments in §6 train l directly via zeroth-order optimization and do not rely on Prop. 4.1; Prop. 3.5 is also independent. The authors should either (a) redefine the geometric detector so that it is a function of the observed path alone (e.g., distance to a known nominal pre-change path if one is available), or (b) present §4 as heuristic and remove or rework the statistical guarantees that depend on Prop. 4.1.","section":"Contribution (ii) and §7"}],"minor_comments":[{"comment":"The abstract says the signature rules 'outperform' classical methods. In the Brownian experiments (Table 1), Shiryaev has lower E[Y^1] and shorter delay than the signature rule; the outperformance is really against Page–Hinkley in the fractional Brownian setting. Please qualify the claim.","section":"Abstract and Table 1"},{"comment":"The text says Assumption 5.1 holds for the Wasserstein ambiguity sets because of weak compactness and upper semicontinuity, but no proof is supplied. Either prove the statement for Example 5.3 or present it as a condition to be verified.","section":"§5.2, Example 5.3"},{"comment":"The coefficient l* is defined as an arginf over T((R^{1+d})*). Existence of a minimizer is not established; Prop. 3.5 states equality of infima, not attainment. Please rephrase using an ε-optimal coefficient or prove existence.","section":"Corollary 4.3"},{"comment":"Restricting the adversary to perturbations with ||w||_{TV}=C_TV is described as 'without loss of generality'. For a general non-concave objective the supremum over the TV ball need not be attained on the boundary; please justify this reduction or soften the claim.","section":"§6.7"}],"recommendation":"major_revision","confidential_remarks":"The main theoretical novelty is Prop. 3.5, which appears credible and is independent of the flawed geometric-detector argument. The geometric section (§4) as currently written is not defensible: the detector depends on the latent pre-change path. A major revision should either replace the detector with one genuinely computable from observations or demote §4 to a heuristic and remove the statistical guarantees that rely on it. The numerical work is interesting but secondary to the theoretical claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: the concatenation model in eq. (3) and the adaptation of Bayer–Hager–Riedel–Schoenmakers' signature optimal stopping are genuine contributions; the geometric-detector guarantees in §4 are not justified as written.\n\nWhat is new and good: modeling the observed signal as a concatenation of two rough paths at a random change-point is natural and Lemma 3.7's observation that the p-variation distance to the pre-change path only records post-change divergence is clean. Proposition 3.5 is a faithful extension of the BHR framework, and the appendix's randomized-stopping treatment handles the discontinuous payoff properly. The numerics are also interesting: in the Brownian case the learned signature rule roughly matches CUSUM and Shiryaev, and in the fBm case it beats Page–Hinkley at comparable false alarm levels. The adversarial training experiment is a reasonable first pass.\n\nThe soft spot is real and load-bearing. In Prop 4.1, f(Z,t)=d_{p-var;[0,t]}(Z,X^(1)) depends on the latent X^(1) on [θ,t], and X^(1) is independent of the observed path after θ. So f is not a σ(X̂_s: s≤t)-measurable function of the observation, and Lemma 2.6 — a density result for continuous functionals of the observed path — cannot be applied to it. The stress-test note is right. Equations (24)–(25) and Propositions 4.2–4.4 are not established, and the advertised claim that the same structural form \"arises from the intrinsic geometry of the observed path\" is unsupported. Corollary 4.3 also quietly conflates the geometric l^(1) with the optimal l* and inherits the problem.\n\nThis does not kill the paper. The numerical experiments optimize l directly for the loss and never rely on Prop 4.1, so the empirical findings stand on their own. The central optimal-stopping result in Prop 3.5 is independent, and the concatenation model is worth keeping. But the §4 claims need real repair: either replace f with an observable proxy (e.g., distance to a reference path or to a windowed baseline) and redo the bounds, or clearly separate the existence-of-a-good-signature-functional question from the adapted-stopping-time question. As written, the guarantees in §4 are not just missing a detail; the object being approximated is not even of the right type.\n\nWho this is for: people working at the interface of sequential analysis and rough-path methods, and anyone who wants a concrete non-Markovian change-point model with a learnable stopping rule. It deserves peer review, not a desk reject, because the framework is timely and the flaw is localized and fixable. I would send it to a referee with a specific request to scrutinize Prop 4.1 and the measurability of f, and to require the authors to state clearly which results survive a corrected argument.","headline":"Concatenation model and BHR adaptation are solid, but §4's geometric detector has a load-bearing measurability gap; the numerics still stand but the statistical claims need rework.","tokens_in":35734,"tokens_out":2551,"would_cite":true,"duration_ms":28387,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62L15","60G40","62C20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The optimal rule for quickest detection on rough-path space is the first time a linear functional of the path signature crosses a threshold.","keywords":["quickest detection","change-point detection","rough paths","signatures","optimal stopping","non-Markovian","robust detection","fractional Brownian motion"],"falsifier":"Simulate the Brownian disorder model twice with the same post-change path X^{(2)} and same observed path up to θ, but different latent pre-change paths X^{(1)} that agree on [0,θ] and differ on (θ,T]. If the geometric detector f differs between the two runs while the observed path is identical, no signature functional of the observed path alone can approximate f uniformly, disproving Proposition 4.1's approximation claim.","tokens_in":34808,"feed_emoji":"🚨","tokens_out":6622,"duration_ms":62024,"temperature":0.7,"pith_summary":"The paper aims to show that quickest detection of a change in the law of an irregular, non-Markovian signal can be solved optimally by a single linear functional of the path's rough-path signature: the rule is to alarm the first time |⟨l, signature⟩| reaches one. If true, this reduces a path-space optimal stopping problem to learning one coefficient vector, with no parametric model and no Markovian sufficiency required. The paper also constructs a geometrically motivated detector from the p-variation distance to the pre-change path, proves finite-sample bounds on delay and false alarm probability, extends to repeated experiments and distributionally robust settings, and demonstrates numerically that the learned rules match or beat classical baselines in Brownian and fractional-Brownian settings.","feed_headline":"One signature functional solves quickest detection","feed_subtitle":"Learns a single linear coefficient from data; works for non-Markovian signals where classical parametric methods fail.","key_machinery":"The signature of a rough path — the sequence of iterated integrals encoding the entire path history — serves as a universal, model-free feature map. The paper's load-bearing mechanism is the half-space hitting time τ_l for a linear functional l of the signature, together with a density theorem (Lemma 2.6) stating that any continuous stopping policy can be uniformly approximated by such linear signature functionals on a compact set of probability arbitrarily close to one. Chen's identity lets concatenation at the change-point be represented as tensor product of signatures, and the homogeneous p-variation distance is used both to measure post-change divergence and to define the geometric detec","core_discovery":"The central claim, Proposition 3.5, is that the infimum over all (F^X_t)-stopping times of E[Y_{τ∧T}] equals the infimum over linear functionals l of E[Y_{τ_l∧T}], where τ_l is the first hitting time of the half-space |⟨l, X^{<∞}_{0,t}⟩| ≥ 1 and X is the concatenation of the pre- and post-change rough paths at the unknown change-point θ. This holds for both Bayesian loss objectives considered (false alarm plus delay penalty, and symmetric early/late penalties). The paper argues the same structural form emerges independently from the intrinsic geometry of the observed path, namely from the p-variation distance to the pre-change path, and derives statistical guarantees: with probability at lea","pith_inferences":["If Proposition 3.5 holds for all rough-path laws, it suggests the signature hitting time is a universal architecture for change-point detection, with the truncation level N as the only tuning parameter — a testable claim beyond the paper's specific examples.","The geometric-detector argument would be falsifiable in practice: check whether a single signature functional can uniformly approximate the p-variation distance when the latent pre-change path is only observed up to θ; the paper's Lemma 2.6 justification does not obviously cover dependence on the unobserved post-θ continuation.","The robust formulation hints that adversarial training of the signature coefficient could serve as a general-purpose robustification layer for any sequential decision rule, not just quickest detection.","The Brownian example suggests that the learned coefficient l should approximately recover the classical likelihood-ratio statistic; verifying that link could connect the rough-path approach to classical theory in a precise way."],"forward_implications":["Quickest detection becomes a finite-dimensional learning problem: estimate one linear functional from data, then monitor its absolute value against a threshold.","The method applies to non-Markovian, non-semimartingale signals such as fractional Brownian motion, where classical sufficient statistics like likelihood ratios are unavailable.","Explicit probabilistic bounds tie detection delay to the inverse of the separation function g and the approximation tolerance, giving a clear trade-off between speed and accuracy.","With R independent paths sharing a change point, both the expected delay and false-alarm probability shrink toward their lower bounds exponentially in R.","Under the least-favorable-pair assumption, the distributionally robust problem is solved by the same signature half-space rule, extending the approach to adversarial perturbations."],"fun_headline_variants":["Rough path signatures detect change in one linear hit","One linear functional spots distributional change","Quickest change detection via rough path signatures","Robust change detection with rough path geometry","Hitting a half-space: the optimal change detector"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the geometric detector f—the p-variation distance from the observed path to the unobserved pre-change path X^{(1)}—can be uniformly approximated by a single linear signature functional on a set of probability close to one, even though f depends on the post-change continuation of X^{(1)} that the observed path does not reveal.","fun_headline_variants_meta":{"raw":{"variants":["Rough path signatures detect change in one linear hit","One linear functional spots distributional change","Quickest change detection via rough path signatures","Robust change detection with rough path geometry","Hitting a half-space: the optimal change detector"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000535,"raw_usage":{"total_tokens":2397,"prompt_tokens":722,"completion_tokens":1675,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":1616}},"tokens_in":466,"tokens_out":1675,"duration_ms":11721,"temperature":1.0,"reasoning_tokens":1616,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T04:02:10.294220+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the Brownian disorder model twice with the same post-change path X^{(2)} and same observed path up to θ, but different latent pre-change paths X^{(1)} that agree on [0,θ] and differ on (θ,T]. If the geometric detector f differs between the two runs while the observed path is identical, no signature functional of the observed path alone can approximate f uniformly, disproving Proposition 4.1's approximation claim.","supporting_citations":[],"review_version":1}