{"id":"534bd1ed-9ae2-476a-a1d5-3e629db44702","arxiv_id":"2507.09787","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A computable fixed-point drift estimator for multiplicative fractional-noise SDEs is proved to be well-defined, asymptotically normal with a confidence interval, and to achieve a 1/N mean-squared-error rate for every Hurst parameter H in (1/3,1).","lead":"Differential equations with rough random noise, including the difficult regime where the noise is rougher than Brownian motion, now admit a computable estimator for the drift parameter with proven accuracy. The estimator combines a pathwise-correction trick with a fixed-point step, and comes with a confidence interval and a 1/N risk bound valid for all Hurst exponents above 1/3.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proof of Prop 3.9 assumes I_N ≥ 0 to make the rough-regime map Θ̃_N R_+-valued, but I_N can be negative under Assumption 3.1 (e.g. fractional OU with x0=0), so well-posedness for H∈(1/3,1/2] is not established as written.","rationale":"The reader's weakest assumption identified the sign conditions of Assumption 3.1 and the uncheckable small-horizon condition (12). My stress-test found a more internal and more specific problem: even under Assumption 3.1, the proof of Proposition 3.9 uses the unstated and false assertion I_N ≥ 0 to show that Θ̃_N maps R_+ into itself. The fractional Ornstein-Uhlenbeck example with x0 = 0 satisfies Assumption 3.1 yet gives I_N ≤ 0. This is not a disagreement with the reader's general assessment — both point to the fragility of the rough-regime well-posedness — but the precise failing step is different. Because the flaw is in a proof of a central proposition rather than a demonstrated counterexample to the theorem, the appropriate verdict remains CONDITIONAL: the paper needs a corrected or additional argument (for example, modifying the estimator or enlarging the good event to control the sign of I_N) before the H ≤ 1/2 claims can be accepted. The other concerns raised by the reader — the omitted proof of Proposition 3.12, the lack of H ≤ 1/2 simulations, and the non-verifiable condition (12) — remain valid but are secondary to this proof gap.","tokens_in":28882,"tokens_out":21546,"duration_ms":257086,"concrete_test":"Specialize to N = 1, H = 0.4, b(x) = −x, σ = 1, x0 = 0, θ0 = 1, T = 1. Analytically, I_N = −X_T² / (2∫_0^T X_s² ds) < 0 whenever X_T ≠ 0, directly falsifying the proof's premise I_N ≥ 0. Then evaluate Θ̃_N(0) from Section 3.2: if I_N is negative but close to 0, the leading term −H∫_0^T t^{2H−1} dt dominates and Θ̃_N(0) < 0, so Θ̃_N is not R_+-valued. This would confirm that the Picard fixed-point argument for Proposition 3.9 collapses in an allowed model.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In the proof of Proposition 3.9 (Section 6.8), the self-map property of Θ̃_N is justified by 'since φ ⩽ 0 and I_N ⩾ 0' together with Λ_t^i(θ) ⩽ 0 for θ ∈ R_+. The assertion I_N ≥ 0 is never proved and is false under Assumption 3.1. For the allowed model b(x) = −x, σ = 1, x0 = 0, a primitive of b is B(x) = −x²/2, so I_N = (1/(N T D_N)) Σ_i (−(X_T^(i))²/2) ≤ 0, with strict inequality unless all copies end at 0. When I_N < 0, the argument r + I_N of Λ_t^i can be negative, the factor exp((r+I_N)∫_s^t ψ) in the integrand can exceed 1, and the inequality Λ_t^i ≤ 0 from (22) is no longer available; consequently Θ̃_N(r) can be negative and Θ̃_N is not shown to map R_+ into itself. Since the fixed point R_N is obtained by Picard's theorem from exactly this self-map property, the well-posedness of the fixed-point estimator in the rough regime H ∈ (1/3,1/2] — a headline contribution of the paper — is not proved as written. This is a concrete internal gap, independent of the separate small-horizon condition (12).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies estimation of the drift parameter θ0 in the SDE dX_t = θ0 b(X_t) dt + σ(X_t) dB_t driven by a multiplicative fractional Brownian motion with Hurst parameter H ∈ (1/3,1), using N independent copies observed on a fixed time interval [0,T]. The starting point is an infeasible least-squares-type estimator involving a Skorokhod integral, which the paper replaces by a computable fixed-point estimator. The transition relies on a new expression for the Malliavin derivative of X that does not depend explicitly on B (Proposition 2.2) and, for H ∈ (1/3,1/2], on a new reduction of a 2D Young integral to an explicit weighted integral (Propositions 2.3 and 2.5). The estimator is defined as θ_N = I_N + R_N, where R_N is the unique fixed point of a random map Θ_N (for H > 1/2) or Θ̃_N (for H ≤ 1/2). Under sign conditions (Assumption 3.1) and a small-horizon condition (12), the paper claims well-posedness of the fixed point on an event Δ_N whose complement has probability O(1/N), an asymptotic confidence interval (Propositions 3.6 and 3.10), and a non-asymptotic O(1/N) risk bound for doubly truncated versions (Propositions 3.8 and 3.12). A numerical study is reported for H = 0.7 and H = 0.9.","tokens_in":29199,"tokens_out":10557,"duration_ms":104411,"significance":"If the paper's claims are correct, it would be a substantial contribution: it provides a computable fixed-point drift estimator with parametric-rate guarantees for multiplicative fBm-driven SDEs over the full range H ∈ (1/3,1), extending earlier work that was restricted to additive noise and/or H > 1/2. The Malliavin-derivative reformulation of Proposition 2.2 and the 2D Young-integral representation of Proposition 2.3 are potentially useful tools beyond this specific estimation problem. The paper is transparent about many of its assumptions, gives worked proof outlines for most of the main results, and provides a numerical illustration. However, one of the central proof steps in the rough regime is not justified as written, and one main result is stated without proof, so the significance can only be assessed after those gaps are repaired.","major_comments":[{"comment":"The self-map property Θ̃_N(R_+) ⊂ R_+ is asserted using 'φ ⩽ 0 and I_N ⩾ 0' together with inequality (22). The claim I_N ⩾ 0 is never proved and is not true under Assumption 3.1. For example, take the allowed model b(x) = −x, σ = 1, x0 = 0; an antiderivative of b is B(x) = −x²/2, so I_N = (1/(NTD_N)) Σ_i (−(X_T^i)²/2) ≤ 0, and in fact I_N < 0 almost surely. When r + I_N < 0, the argument of the exponential in Λ_t^i(r + I_N) changes sign, inequality (22) is no longer available, and Θ̃_N(r) can take negative values; the displayed conclusion 'Θ̃_N(r) ≥ 0' is therefore unjustified. Since existence and uniqueness of R_N are obtained via Picard's fixed-point theorem from exactly this self-map property, the well-posedness of the fixed-point estimator for H ∈ (1/3,1/2] is not established as written. This is a load-bearing gap in the paper's headline claim.","section":"Section 6.8, proof of Proposition 3.9"},{"comment":"The proof of Proposition 3.12 is explicitly omitted, with the justification that it follows the same lines as Proposition 3.8. This is not a purely cosmetic omission: the rough-regime risk bound requires the well-posedness and contraction properties of Θ̃_N that are affected by the gap in Proposition 3.9, and the analogue of the transfer inequality (20) for Θ̃_N is not written out. A main theorem, the non-asymptotic risk bound for H ∈ (1/3,1/2], is thus left without support. The proof should be supplied, or the statement should be explicitly conditional on a repaired well-posedness result.","section":"Section 6, after Proposition 3.12"},{"comment":"The small-horizon condition (12) involves ||b||_f and E(b(X_T)) − b(x0), both of which depend on the law of X and hence on θ0 itself; it is not checkable from the observed data. This condition is load-bearing: it is used to prove P(Δ_N^c) ≤ c/N in Proposition 3.5, which in turn drives the asymptotic confidence intervals and the risk bounds. The numerical study in Section 4 never verifies condition (12), and it only covers H = 0.7 and H = 0.9. The paper should state clearly that the advertised guarantees are conditional on a non-verifiable model condition, and should discuss whether the small-horizon condition can be replaced by a data-driven or more explicit criterion.","section":"Section 3.1, condition (12)"}],"minor_comments":[{"comment":"The numerical section reports results only for H = 0.7 and H = 0.9, so it does not illustrate the rough regime H ∈ (1/3,1/2] that the paper presents as its main technical challenge; a simulation with, say, H ≈ 0.4 would be informative.","section":"Section 4.2"},{"comment":"The notation 'with b′ = b' is confusing: the drift b and its antiderivative should be denoted by different symbols throughout, for example B(x) = ∫_0^x b(y) dy.","section":"Section 3.1, Equation (8)"},{"comment":"The displayed definitions of Y_N and R_i contain typographical errors: repeated differentials such as 'dudvdudv' and ambiguous limits of the form ∫_0^v ∫_0^u make the intended integration variables unclear. These formulas should be rewritten with distinct variables for the two pairs of integration arguments.","section":"Proposition 3.6 and Remark 3.7"},{"comment":"The symbol n is overloaded: it denotes both the number of Riemann-sum discretization points and the number of Picard iterations in the composition (Θ_{N,n} ∘ ⋯ ∘ Θ_{N,n}). This makes the implementation description harder to follow than necessary.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is potentially significant, but the rough-regime well-posedness issue identified in the report is central and must be fixed before the headline claim can be accepted. The omitted proof of Proposition 3.12 is also a completeness problem. The manuscript is otherwise coherent and the H > 1/2 part appears sound; the authors should be given the opportunity to repair the H ≤ 1/2 argument or to restrict the claims accordingly. The numerical section should also be extended to the rough regime if the theoretical gap is resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a real step forward for H>1/2, but the rough regime H∈(1/3,1/2]—one of the two headline contributions—has a concrete gap in the well-posedness proof. The claim is that Θ̃_N maps R_+ into itself because φ≤0 and I_N≥0. The second assertion is false. For the allowed model b(x)=−x, σ=1, x0=0, the primitive of b used in defining I_N is −x²/2, so I_N = (1/(N T D_N))Σ(−(X_T^i)²/2)≤0, strictly negative almost surely. When I_N<0, the argument r+I_N in Λ_t^i can be negative, and the bound in (22) is no longer available. Thus the self-map property is not established, and fixed-point existence and uniqueness for H≤1/2 is not proved as written. This is a genuine internal gap, not a matter of taste.\n\nThat said, the paper does real work. Proposition 2.2, reformulating the Malliavin derivative of the solution without explicit dependence on the driving fBm, is a clean and useful trick. Proposition 2.3/2.5, handling the extra correction terms in the rough regime via 2D Young integrals, appears new and of independent interest. For H>1/2 the estimator is well-defined, the error-event bound, confidence interval, and O(1/N) risk bound are all proved with consistent rates, and the fixed-point architecture (fake estimator, contraction event, transfer inequality) is handled carefully. The numerical study, while modest, supports the H=0.7 and 0.9 behavior.\n\nThe secondary weaknesses are real but less serious. Proposition 3.12's proof is omitted; the authors say it mirrors 3.8, which is acceptable in principle, but they should at least sketch the parts that differ. The numerics never go below H=1/2, precisely where the new difficulty lives. Condition (12) is a small-horizon bound involving ||b||_f, an unknown population quantity, and the simulations do not verify it. The sign conditions φ≤0, ψ≤0, θ0>0 are restrictive, but the examples show a non-empty model class.\n\nBottom line: the H>1/2 contribution deserves publication, and the technical tools will be used by others. The H≤1/2 claims are not yet supported. This paper deserves a serious referee, but the referee should be asked for a major revision. If the I_N issue is fixed—perhaps by moving I_N into the good event or redefining the primitive—and the omitted details are supplied, the paper will be a strong contribution.","headline":"Solid H>1/2 theory and genuinely new Malliavin/Young tools, but a sign issue in the proof of Proposition 3.9 leaves the rough-regime well-posedness unproved.","tokens_in":29764,"tokens_out":4524,"would_cite":false,"duration_ms":48715,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M05","60G22","60H07","62F12"],"pacs":[],"model":"deepseek-v4-flash","headline":"For H∈(1/3,1), the paper constructs a computable fixed-point estimator of the drift θ0 of a multiplicative fractional-noise SDE from N independent copies, with an asymptotic confidence interval and mean-squared error of order 1/N.","keywords":["drift parameter estimation","fractional Brownian motion","multiplicative noise","fixed-point estimator","Skorokhod integral","Malliavin derivative","Young integral","rough paths"],"falsifier":"A concrete falsifier: with H=0.7, σ≡1, b(x)=-x, choose T so large that the small-horizon condition (12) is violated and run the estimator on N=$10^{4}$ simulated paths; if the mean-squared error still decays like 1/N and the coverage stays at the nominal level, the paper's condition (12) is not necessary as stated, whereas if the error fails to decay, the condition is doing real work.","tokens_in":28555,"feed_emoji":"🎯","tokens_out":7352,"duration_ms":71013,"temperature":0.7,"pith_summary":"This paper seeks to estimate the drift parameter θ0 of an SDE driven by multiplicative fractional Brownian noise from N independent copies observed over a fixed finite time horizon. The natural least-squares object is a 'fake estimator' because it contains an unobservable Skorokhod integral, so the authors replace that integral with a corrected pathwise one and define a computable fixed-point estimator whose fixed point is the unique solution of a random contraction map. The paper proves that for every Hurst parameter H in (1/3,1), including the rough regime H in (1/3,1/2], the fixed point exists on a high-probability event, yields an asymptotic confidence interval with guaranteed coverage, and has mean-squared error of order 1/N. If correct, this is the first parametric-rate drift estimation procedure for multiplicative fractional noise across the full range of H, extending earlier work that was limited to additive noise, to H>1/2, or to a single trajectory over a long time horizon.","feed_headline":"New estimator handles rough fractional-noise drift at a 1/√N rate","feed_subtitle":"Malliavin-calculus identities turn an unobservable Skorokhod estimator into a computable fixed point with confidence intervals.","key_machinery":"The machinery rests on three pieces. First, Proposition 2.2 rewrites the Malliavin derivative of the solution as D_s X_t = σ(X_t) exp(θ0 ∫_s^t (b' - σ'b/σ)(X_u) du) 1_{[0,t)}(s), an expression with no explicit dependence on the driving fractional Brownian motion. Second, for H∈(1/3,1/2], Proposition 2.3 reduces the two-dimensional Young integral of a regular function x against the fBm covariance to α_H ∫∫ x(s,t)|t-s|^{2H-2} ds dt with α_H = H(2H-1), which turns the rough-regime correction into a computable kernel integral. Third, using these identities, the gap between the fake estimator and a pathwise estimator becomes a fixed-point equation for the maps Θ_N (when H>1/2) and Θ̃_N (when H≤1/2), both built from the coefficient functions φ = σ(σb' + σ'b) and ψ = (σb' - σ'b)/σ, the empirical denominator D_N, and the pathwise increment I_N; under the paper's sign conditions these maps are contractions on R_+, so their fixed point R_N exists and is unique and θ_N = I_N + R_N is the estimator.","core_discovery":"The paper establishes that the gap between the unobservable Skorokhod-based estimator and a computable pathwise one can be closed by a fixed-point equation, for all H∈(1/3,1). The key identity is a Malliavin derivative formula D_s X_t = σ(X_t) exp(θ0 ∫_s^t ψ(X_u) du) that is free of explicit dependence on the driving fractional Brownian motion, together with a new 2D Young-integral reduction (Proposition 2.3) that converts the rough-regime correction into a Riemann-integrable form. Under sign conditions on the coefficients and a small-horizon condition, the resulting maps Θ_N and Θ̃_N are contractions on R_+, so the fixed point exists and is unique; the paper then proves P(Δ_N^c) ≤ c/N, an asymptotic confidence interval with coverage at least 1-2λα, and a non-asymptotic bound E(|$θ_N^{{c,d}}$ - θ0|²) ≤ c/N.","pith_inferences":["The small-horizon condition (12) involves ||b||²_f, an expectation under the unknown law of X, so a data-driven check or plug-in estimate of this quantity would be needed to turn the procedure into an off-the-shelf method; the paper does not develop such a check.","The B-free Malliavin derivative formula of Proposition 2.2 is likely to simplify other inference problems for fBm-driven SDEs, such as nonparametric drift estimation or estimation from discretely sampled paths, beyond the fixed-point estimator studied here.","The 2D Young integral reduction of Proposition 2.3 may be useful in other statistical contexts for rough noise, for instance in constructing moment estimators or in weak convergence arguments, because it reduces a double integral against the fBm covariance to an explicit kernel integral.","The extension to d-dimensional fBm is only sketched in Section 5 via a finite-difference Jacobian estimator; testing that estimator numerically would be a natural next step, but it is not part of the paper's claims."],"forward_implications":["For H>1/2 the estimator θ_N = I_N + R_N is fully computable from the data, since I_N is an explicit function of the increments of b(X) and R_N is obtained by iterating the contraction Θ_N.","For H∈(1/3,1/2] the same computable fixed-point scheme works, at the price of the bounded-drift assumption (2.4) and a more conservative variance bound Y_N.","The asymptotic confidence interval in Proposition 3.6 (resp. 3.10) covers θ0 with probability at least 1-2λα, which can be made close to 1-α by choosing λ close to 1 and α small.","The doubly truncated estimator θ_N^{c,d} has mean-squared error at most c/N, so the estimator converges at the parametric rate.","The bad event on which the fixed point is not defined has probability O(1/N), so the truncation at Δ_N does not change the asymptotics."],"supporting_citations":[{"why":"Supplies the fixed-point estimation strategy for additive fractional noise with H∈(1/2,1) that this paper extends to multiplicative noise and to H≤1/2.","marker":"[50]"},{"why":"Theorem 3.1 gives the Skorokhod-to-pathwise integral relationship for H∈(1/3,1/2] used in Proposition 2.5.","marker":"[68]"},{"why":"Theorem 3.11.1 provides the variance formula for the Skorokhod integral used to construct Y_N in the H>1/2 confidence interval.","marker":"[10]"},{"why":"Supplies the Malliavin calculus background and the Skorokhod-pathwise relationship for H>1/2 used in Proposition 2.2 and Equation (3).","marker":"[58]"},{"why":"Provides the rough path framework for the well-posedness of the SDE and the controlled-path representation used in the proof of Proposition 2.5.","marker":"[33]"},{"why":"Supplies the definition and Riemann-sum convergence of 2D Young integrals used in Proposition 2.3.","marker":"[34]"},{"why":"Addresses the multiplicative fBm drift estimation only with one trajectory and T→∞, the baseline this paper improves on for the multi-copy regime.","marker":"[41]"},{"why":"Gives the density existence condition used to define the time-averaged density f and the norm ||·||_f.","marker":"[6]"}],"fun_headline_variants":["Fixed-point drift estimator for rough fractional-noise SDEs","Computable drift estimator for multiplicative rough noise","Rough-noise drift: fixed-point method with 1/√N rate","New fixed-point approach to drift in rough SDEs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the sign condition (Assumption 3.1: b', φ, ψ ≤ 0 and θ0 > 0) joined with the small-horizon condition (12); if those fail, fixed-point existence and all the statistical guarantees are unproven, and (12) depends on the unknown law of X, so it cannot be checked from the data alone.","fun_headline_variants_meta":{"raw":{"variants":["Fixed-point drift estimator for rough fractional-noise SDEs","Computable drift estimator for multiplicative rough noise","Rough-noise drift: fixed-point method with 1/√N rate","New fixed-point approach to drift in rough SDEs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000739,"raw_usage":{"total_tokens":3323,"prompt_tokens":990,"completion_tokens":2333,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":2265}},"tokens_in":606,"tokens_out":2333,"duration_ms":19819,"temperature":1.0,"reasoning_tokens":2265,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:49:31.708044+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete falsifier: with H=0.7, σ≡1, b(x)=-x, choose T so large that the small-horizon condition (12) is violated and run the estimator on N=$10^{4}$ simulated paths; if the mean-squared error still decays like 1/N and the coverage stays at the nominal level, the paper's condition (12) is not necessary as stated, whereas if the error fails to decay, the condition is doing real work.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the fixed-point estimation strategy for additive fractional noise with H∈(1/2,1) that this paper extends to multiplicative noise and to H≤1/2."},{"cited_title":"& Tindel, S","cited_arxiv_id":null,"evidence_quote":"Theorem 3.1 gives the Skorokhod-to-pathwise integral relationship for H∈(1/3,1/2] used in Proposition 2.5."},{"cited_title":"& Zhang, T","cited_arxiv_id":null,"evidence_quote":"Theorem 3.11.1 provides the variance formula for the Skorokhod integral used to construct Y_N in the H>1/2 confidence interval."},{"cited_title":"(2006).The Malliavin Calculus and Related Topics.Springer, Berlin-Heidelberg","cited_arxiv_id":null,"evidence_quote":"Supplies the Malliavin calculus background and the Skorokhod-pathwise relationship for H>1/2 used in Proposition 2.2 and Equation (3)."},{"cited_title":"& Hairer, M","cited_arxiv_id":null,"evidence_quote":"Provides the rough path framework for the well-posedness of the SDE and the controlled-path representation used in the proof of Proposition 2.5."},{"cited_title":"& Victoir, N","cited_arxiv_id":null,"evidence_quote":"Supplies the definition and Riemann-sum convergence of 2D Young integrals used in Proposition 2.3."},{"cited_title":"& Zhou, H","cited_arxiv_id":null,"evidence_quote":"Addresses the multiplicative fBm drift estimation only with one trajectory and T→∞, the baseline this paper improves on for the multi-copy regime."},{"cited_title":"& Tindel, S","cited_arxiv_id":null,"evidence_quote":"Gives the density existence condition used to define the time-averaged density f and the norm ||·||_f."}],"review_version":1}