{"id":"ace14147-f3f7-4124-b962-90e5d8715698","arxiv_id":"2512.19647","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Milstein schemes for hyperbolic SPDEs with contractive (non-analytic) C0-semigroups converge at rate α∈(1/2,1], up to √log(T/h) for rational schemes and with no log factor for the exponential scheme, in pathwise-uniform L^p error.","lead":"This paper proves that Milstein time-stepping schemes for hyperbolic stochastic partial differential equations converge at order up to one — h for exponential schemes, h√log(T/h) for rational schemes — the first rigorous result of this kind, resolving an open problem posed in 2015. It applies to Schrödinger, Maxwell, and transport equations with multiplicative noise.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified; main proof is internally coherent, with only conventional verifications (Grönwall normalization, cited lemmas) and the acknowledged iterated-integral implementation gap.","rationale":"I reviewed the proof of Theorems 5.3 and 5.7 in detail, including the error split, the semigroup-difference lemmas, the stochastic Fubini step in (5.17), the Hölder estimates (5.18) and (5.24), and the discrete Grönwall closure. The exponent bookkeeping is consistent and I could not identify an internal inconsistency. The reader's weakest-assumption designation (A4) is reasonable as a restrictiveness concern, but it is an explicit assumption of the theorem, not a hidden flaw. The cited external results (Theorem 2.3, Proposition 2.4, Theorem 4.2) are standard in the area; the deferred pathwise stability remark is not used in the main argument. The implementation gap for non-commutative noise is real and acknowledged, but it does not invalidate the mathematical convergence claim. Therefore I do not change the conditional verdict, though I recommend the Grönwall normalization check and independent verification of the cited maximal inequalities.","tokens_in":49918,"tokens_out":30318,"duration_ms":283452,"concrete_test":"Recompute the discrete Grönwall step in Proposition 4.4 by substituting β = C√h into Lemma 2.11 on a two-step example, verifying that β²j ≤ C²T and that the resulting stability constant matches the displayed (1+C²T)^{1/2} exp((1+C²T)/2). If the h factor is actually missing, stability would grow like exp(C²/h) and the convergence proof would collapse; if it is present as argued, the proof stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I find no load-bearing mathematical flaw in the central claim. Within Assumptions A1–A4, the proof skeleton is coherent: the three-term error split (5.10), Lemmas 5.5/5.6, the Taylor-expansion estimates (5.16)–(5.25), and the discrete Grönwall closure all check out. The logarithmic factor enters only through Lemma 5.6 (rational semigroup approximation) and correctly disappears for the exponential scheme because E3 = 0. The most subtle point is the discrete Grönwall application in Proposition 4.4: the displayed inequality φ(j) ≤ c + C(h∑φ²)^{1/2} must be read with Lemma 2.11 parameter β = C√h, giving β²j ≤ C²T and the stated (1+C²T) factor. The text omits the explicit √h but the conclusion uses C²T, so the step is correct; a referee should still confirm it. The only substantive limitation is the one the paper itself flags in Section 7: exact iterated stochastic integrals are simulable only under a commutativity condition, so for non-commutative noise the implemented scheme is not the analyzed scheme. This narrows the practical scope but does not falsify the theorem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a temporal Milstein theory for semilinear hyperbolic SPDEs of the form dU + AU dt = F(U)dt + G(U)dW on a Hilbert space X, where -A generates a contractive, not necessarily analytic C0-semigroup. It introduces rational and exponential Milstein schemes, proves pathwise-uniform strong error bounds in L^p(Ω) of order min{α,1} up to a logarithmic factor for rational schemes and without the logarithmic factor for the exponential scheme, under a two-space Kato framework Y↪X with α-Hölder type regularity assumptions (Assumptions A1-A4). The central result, Theorem 5.3 and its exponential counterpart Theorem 5.7, is applied to stochastic Schrödinger, Maxwell, and transport equations, and numerical experiments for the stochastic Schrödinger equation confirm the predicted rates.","tokens_in":50075,"tokens_out":15982,"duration_ms":140064,"significance":"If the result holds, it resolves an open problem posed in [26, p. 324] and [31, Rem. 6.7] by giving the first rigorous convergence rates above 1/2 for Milstein schemes in the hyperbolic, non-analytic semigroup setting. The proof is substantial: it introduces two auxiliary processes, a three-term error split, stochastic Fubini, maximal inequalities for stochastic convolutions, a logarithmic square-function estimate, and a discrete Grönwall closure. The manuscript is unusually careful about constants, Gâteaux vs. Fréchet differentiability, and the distinction between pointwise and pathwise-uniform errors. It also provides publicly available code (DOI 10.5281/ZENODO.18229440) and openly discusses limitations, including the exact-simulation problem for iterated integrals in the non-commutative case and the linear-growth-on-Y restriction that excludes some Nemytskii nonlinearities. These caveats narrow the practical scope but do not contradict the internal mathematical claims.","major_comments":[],"minor_comments":[{"comment":"The assumption states g ∈ L∞(0,T;L^p(Ω;L_2^{(2)}(H,X))), but g is the linear noise value taking values in L_2(H,X), and the proof uses |||g|||∞,p,Y, i.e. the L_2(H,Y)-norm. The bilinear space L_2^{(2)} appears to be a typo; it should be L_2(H,X) (or L_2(H,Y) under the Y-invariance already assumed in A2(c)).","section":"Lemma 5.2"},{"comment":"The paper openly states that for non-commutative noise the exact iterated integrals cannot currently be simulated, so the implemented scheme in that case is not the analyzed scheme. Since this is a significant practical caveat, it should be stated more prominently in the introduction or abstract, not only in the numerical section.","section":"Section 7, last paragraph"},{"comment":"The application of the discrete Grönwall lemma is compressed. After φ(j) ≤ c + C(h∑φ(i)^2)^{1/2}, the identification β = C√h gives β²j ≤ C²T, which is what yields the (1+C²T)^{1/2} exp((1+C²T)/2) factor. Stating this identification explicitly would improve readability.","section":"Proposition 4.4"},{"comment":"The introduction states Y↪D(A^α) continuously, while Assumption A1 uses the real interpolation space D_A(α,∞) for α<1. The relationship should be clarified, since D(A^α) embeds into D_A(α,∞) but the two conditions are not identical in the statement of the theorem.","section":"Theorem 1.1"},{"comment":"The numerical rates are obtained from 100 samples and a reference solution at h=2^{-14}. Reporting confidence intervals or standard errors would strengthen the comparison against the theoretical rates, especially for the small step sizes where the reference error may bias the observed rates.","section":"Section 7.1, Figure 1 and Table 2"}],"recommendation":"minor_revision","confidential_remarks":"The central theorem is internally coherent; I found no load-bearing mathematical error. The main limitation—exact simulation of iterated stochastic integrals for non-commutative noise—is acknowledged by the authors but is worth emphasizing to the editor as a scope restriction. The paper depends heavily on the framework of [31], but the extension to Milstein schemes is nontrivial and the added machinery (Taylor expansion of F and G, the auxiliary processes, the E3 analysis) is worked out in detail. I recommend minor revision to address the presentation issues above."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read this as a genuine advance. For a decade the Milstein scheme for SPDEs was a parabolic story — Jentzen–Röckner and successors relied on analytic semigroup regularization — and the hyperbolic case was explicitly left open in [26, p. 324]. This paper closes that gap for contractive C0-semigroups: rational Milstein schemes converge at rate min{α,1} up to a log factor in the pathwise-uniform L^p norm, and exponential Milstein schemes at the same rate without the log. The extension from exponential to rational schemes via the logarithmic square function estimate is the right technical step, and the pathwise-uniform upgrade from pointwise mean-square bounds is a genuine improvement, not a repackaging.\n\nThe proofs are detailed enough that I could check the exponent bookkeeping: the three-term error split in Theorem 5.3, the auxiliary processes in Definition 5.4, Lemmas 5.5–5.6, and the discrete Grönwall closure are all coherent. The ζ^{2α−1} integral producing 1/(2α) in (5.18), the h^{3/2} and h^{α+1/2} terms from Lemma 5.2, and the disappearance of the log factor in the exponential case because E3=0 — all consistent. I did not find a load-bearing mathematical flaw, and the stress-test's note about the discrete Grönwall application matches my reading: the √h is implicit but the resulting (1+C²T) factor is correct.\n\nThe real soft spots are the ones the paper itself admits. First, several key tools are cited rather than proved: the log-square estimate (Prop. 2.4), the maximal inequality (Thm. 2.3), and well-posedness (Thm. 4.2) come from the literature, including the second author's own [31]. A referee should verify those citations carry exactly the stated constants and hypotheses; I did not spot a mismatch, but the reliance is real. Second, Remark 4.5(iii) asserts a pathwise uniform stability result with only a sketch of a proof. That is a gap worth closing, though not one that threatens the main theorem. Third, the exact iterated stochastic integrals are simulable only under the commutativity condition used in Section 7; for non-commutative noise the implemented scheme is not the analyzed one. The paper is honest about this, but it narrows the practical scope. Fourth, the numerics use only 100 Monte Carlo samples and an acknowledged reference-solution error floor; the observed rates are plausible but not decisive.\n\nOne structural limitation I'd flag: the linear-growth-on-Y assumption genuinely excludes standard Nemytskii nonlinearities for second-order equations, forcing Y=H^1 and capping the rate at 1/2. The authors say this openly and offer a nonlocal nonlinearity as a workaround. That is a fair statement of scope, not a hidden flaw.\n\nOverall: this deserves a serious referee and, assuming the cited lemmas hold up, publication. I would cite it and would bring it to a reading group on higher-order numerical methods for SPDEs.","headline":"This paper delivers the first rigorous rate-1 (up to log) pathwise-uniform error analysis for Milstein schemes in a hyperbolic SPDE setting, resolving a real open problem; the proof skeleton checks out, and the main caveats are the ones the authors themselves flag.","tokens_in":50758,"tokens_out":1050,"would_cite":true,"duration_ms":12882,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H15","65C30","65M12","65M15","35R60"],"pacs":[],"model":"deepseek-v4-flash","headline":"Milstein schemes reach first-order convergence for hyperbolic SPDEs.","keywords":["Milstein scheme","hyperbolic SPDE","contractive semigroup","pathwise uniform convergence","strong convergence rate","stochastic Schrödinger equation","iterated stochastic integrals","rational time discretization"],"falsifier":"Simulate the linear stochastic Schrödinger setting of Section 7.1 with reference step size 2^{-16} and step sizes down to 2^{-12}; if the measured exponential-Milstein slope is below 0.9, the claimed order-one rate fails to show up numerically; if the rational-Milstein slope matches h√log(T/h) rather than h, the logarithmic factor is real.","tokens_in":49620,"feed_emoji":"📈","tokens_out":6991,"duration_ms":64916,"temperature":0.7,"pith_summary":"This paper proves that Milstein-type time discretizations, not just Euler-type methods, converge at up to first order for hyperbolic stochastic partial differential equations, where the leading operator generates a contractive but non-analytic semigroup. Previous rigorous results in this setting were stuck at rate 1/2 for schemes using only Wiener increments; the paper shows the extra iterated-integral correction in the Milstein scheme overcomes that barrier, with a pathwise-uniform L^p error of order h^α, up to a logarithmic factor for rational variants and with no logarithmic factor for the exponential variant. The proof covers stochastic Schrödinger, Maxwell, and transport equations, and numerical experiments for Schrödinger equations report rates close to 1. If correct, this resolves a decade-old open question about higher-order schemes for hyperbolic SPDEs and extends the Milstein theory beyond the parabolic case.","feed_headline":"Milstein schemes reach first-order convergence for hyperbolic SPDEs","feed_subtitle":"First proof that the Milstein correction beats the 1/2-order barrier set by Wiener-increment schemes for contractive semigroups.","key_machinery":"The central object is the Milstein correction (G′G)(u_i)Δ²W_{i+1} — the iterated stochastic integral of the noise coefficient's derivative along the noise — added to each Euler step. The argument is carried by a two-space framework: the scheme is stable on a smoother subspace Y, while differentiability is used in a mixed way (directional differentiability on X, full differentiability from Y to X, and (2α−1)-order continuity of derivatives), which allows genuinely nonlinear Nemytskii-type operators to be treated. The rational-scheme analysis hinges on a logarithmic square-function estimate for discrete stochastic convolutions, which produces the √log(T/h) factor, and a discrete comparison lem","core_discovery":"Under a set of regularity assumptions — a smoother subspace Y embedded in the state space X, Lipschitz drift and noise on X, linear growth and full differentiability on Y, and a (2α−1)-order continuity condition on the derivatives — the rational Milstein scheme converges at rate min{α,1} in pathwise uniform L^p norm, with the error bounded by a constant times h^α√log(T/h), while the exponential Milstein scheme achieves the same rate without the logarithmic factor. This is the first rigorous error analysis showing that Milstein schemes can break the 1/2-order barrier for hyperbolic SPDEs, improving on results for exponential Euler and rational schemes that use only Wiener increments.","pith_inferences":["Editorial inference: The same two-space template should extend to locally Lipschitz nonlinearities via a stopping-time argument; the paper's own Section 7.3 simulation with a Nemytskii nonlinearity outside the assumptions already shows numerical rates near 1, suggesting the regularity conditions are sufficient rather than necessary.","Editorial inference: The logarithmic factor for rational schemes appears tied to the logarithmic square-function estimate; a sharper analysis of the discrete stochastic convolution might remove it for A-stable rational schemes as well, yielding a unified no-log theory.","Editorial inference: For stochastic wave equations, the authors note their approach does not improve on rate 3/2 obtained by a specialized scheme; combining the Milstein correction with the wave equation's extra smoothing is a natural next step to recover such higher rates within a unified framework.","Editorial inference: For non-commutative noise, exact simulation of the iterated integral is not currently known; combining the present error analysis with a derivative-free Milstein approximation would give a fully implementable first-order scheme for hyperbolic SPDEs."],"forward_implications":["For hyperbolic SPDEs with contractive non-analytic semigroups, Milstein schemes achieve convergence rate min{α,1} in pathwise uniform L^p sense, improving the previously best-known 1/2 rate.","The exponential Milstein scheme attains rate 1 without any logarithmic factor when Y = D(A), matching the order of the Itô–Taylor expansion in the SDE case.","Rational A-stable schemes such as implicit Euler and Crank–Nicolson Milstein variants inherit the same rates, with Crank–Nicolson reaching 2/3 for Maxwell's equations under natural regularity assumptions.","The error estimates are pathwise uniform (max over time inside the expectation), not just pointwise root-mean-square bounds, and hold in every L^p moment for p in [2,∞).","For linear stochastic Schrödinger equations with sufficiently smooth potential and noise, the exponential Milstein scheme converges at rate 1, and numerical experiments confirm rates close to 1."],"fun_headline_variants":["Milstein scheme cracks 1/2-order barrier for hyperbolic SPDEs","Milstein hits order one for hyperbolic SPDEs, beats Euler","First-order convergence proven for hyperbolic SPDE Milstein","Milstein beats Wiener increment schemes for hyperbolic SPDEs","Hyperbolic SPDEs: Milstein scheme achieves order-one rate"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The result depends on the derivatives of the drift and noise being smooth enough on a smaller subspace Y — continuity of order 2α−1 — and on the combined noise sensitivity G′G being Lipschitz on the state space; without that, the Milstein correction cannot be controlled and only the old 1/2 rate is known.","fun_headline_variants_meta":{"raw":{"variants":["Milstein scheme cracks 1/2-order barrier for hyperbolic SPDEs","Milstein hits order one for hyperbolic SPDEs, beats Euler","First-order convergence proven for hyperbolic SPDE Milstein","Milstein beats Wiener increment schemes for hyperbolic SPDEs","Hyperbolic SPDEs: Milstein scheme achieves order-one rate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000152,"raw_usage":{"total_tokens":1069,"prompt_tokens":798,"completion_tokens":271,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":183}},"tokens_in":542,"tokens_out":271,"duration_ms":3130,"temperature":1.0,"reasoning_tokens":183,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T14:38:46.301897+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the linear stochastic Schrödinger setting of Section 7.1 with reference step size 2^{-16} and step sizes down to 2^{-12}; if the measured exponential-Milstein slope is below 0.9, the claimed order-one rate fails to show up numerically; if the rational-Milstein slope matches h√log(T/h) rather than h, the logarithmic factor is real.","supporting_citations":[],"review_version":1}