{"id":"d3fb65e5-188f-4122-a5ae-3f82ced667a8","arxiv_id":"1909.00099","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"An explicit adaptive Milstein method with path-bounded time stepping is shown to converge strongly with order one for SDEs with one-sided Lipschitz drift and non-commutative noise.","lead":"This paper constructs an explicit adaptive Milstein method for stochastic differential equations with one-sided Lipschitz drift and proves strong L2 convergence of order one, including for non-commutative noise. The method shrinks the time step near large solution values and only rarely needs a fallback scheme, so it is both practical and analyzable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof of Theorem 4.1 contains a false equality in the bound for I_{1,2}; the term is not zero, though it appears to be O(h^2) and repairable.","rationale":"The reader's weakest-assumption analysis correctly flags that the tamed Milstein backstop is not proven to satisfy the mean-square consistency bound (3.5) in the non-commutative case, and that the numerical rate comparison uses hmax versus hmean. These are genuine concerns about the presentation and the strength of the experimental evidence, but they are peripheral to the analytical theorem. My reading of the proof of Theorem 4.1 found a more direct issue: the term I_{1,2} in the drift expansion is asserted to have zero conditional expectation. Since the error process E(r) contains the same Brownian increments as the factor Delta W_i, the conditional expectation is not generally zero, and a simple scalar example with f(x)=x, g(x)=x confirms a nonzero O(h^2) contribution. This is a real defect in the written proof. However, the defect is repairable: bounding |I_{1,2}| by (1/2)E||E(r)||^2 plus an O(h^2) term fits the existing Gronwall scheme and does not change the order-one conclusion. Thus the central claim is plausible and the proof can likely be fixed, but it should not be presented as fully correct without this repair. The verdict CONDITIONAL remains appropriate; I therefore leave the reader's verdict unchanged, while disagreeing with the reader's choice of weakest assumption.","tokens_in":21345,"tokens_out":11293,"duration_ms":113551,"concrete_test":"Re-derive the estimate for I_{1,2} in Section 6 without invoking independence from the Brownian increment. As a check, compute E[(X(r)-Ybar(r)) Z | F_tn] for the scalar SDE dX = X dt + X dW with X(0)=x0, where Z = int_{tn}^r dW and Ybar is the Milstein approximation. If the leading term is c h^2 with c != 0, the equality I_{1,2}=0 is false. Then verify that replacing the false equality by |I_{1,2}| <= (1/2)E[||E(r)||^2] + K h^2 still yields the Gronwall bound E[||X(T)-Ybar(T)||^2] <= C h_max^2, so the final claim survives with a repaired proof.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the proof of Theorem 4.1, after equation (6.6), the term I_{1,2} is declared to be zero 'by (3.6)'. This is not justified: E(r) = X(r) - Ybar(r) depends on the Brownian increments on [tn, r], so it is not F_tn-measurable, and E[langle E(r), c_i Delta W_i rangle | F_tn] need not vanish. A concrete scalar example with f(x)=x, g(x)=x, Yn=x0, and Z = int_{tn}^r dW gives Ybar(r) = x0(1 + Delta t + Z + (Z^2 - Delta t)/2), X(r) = x0 exp(Delta t/2 + Z), and E[(X(r)-Ybar(r))Z] = x0 Delta t(e^{Delta t}-1), which is nonzero. This shows the printed equality is false. The error is not order-destroying: for small Delta t the term is O(Delta t^2), so a standard Cauchy-Schwarz bound such as |I_{1,2}| <= (1/2)E[||E(r)||^2] + K h^2 would preserve the Gronwall argument and the final order-one convergence. But the proof as written contains an incorrect step, and this is more central to the paper's main claim than the numerical backstop and comparison-step issues identified by the reader.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces an explicit adaptive time-stepping Milstein method for Itô SDEs with one-sided Lipschitz drift and globally Lipschitz diffusion, without the usual commutativity assumption on the noise. The stepsize is chosen by a path-bounded strategy that shrinks steps when the numerical solution nears a sphere and invokes a backstop method if the step would fall below a minimum size. The main result (Theorem 4.1) states that, for any backstop satisfying the conditional mean-square consistency bound (3.5), the method is strongly L2 convergent of order one in the maximum step size; Theorem 4.2 asserts that the probability of using the backstop can be made arbitrarily small. The proofs rely on new moment bounds for Lévy areas (Lemma 2.3), proved in the appendix via the characteristic function. Numerical experiments on scalar and two-dimensional systems with multiplicative, additive, and non-commutative noise illustrate the convergence and efficiency claims.","tokens_in":21733,"tokens_out":15546,"duration_ms":136415,"significance":"If the proof is repaired, the paper makes a solid contribution: it provides an explicit order-one strong scheme for non-globally Lipschitz SDEs that does not require the commutativity condition, and it does so within a general adaptive framework with a probabilistic guarantee on backstop usage. The self-contained Lévy-area moment estimates with explicit constants, the detailed one-step error analysis, and the demonstration that the backstop is rarely triggered are strengths. The central convergence result depends on a hypothesis on the backstop that is not verified for the non-commutative numerical example, and the proof contains a false equality in the local error analysis; both issues are local and repairable, so the core claim remains plausible.","major_comments":[{"comment":"In the estimate for I_{1.2} (the second line after Eq. (6.9)), the equality to zero is not justified by (3.6). The error E(r) = X(r) - Ybar(r) depends on the Brownian increments on [tn, r], so it is not F_tn-measurable, and E[<E(r), V>|F_tn] with V = sum_i Df(Yn)g_i(Yn)∫_{tn}^r dW_i does not generally vanish. A concrete scalar example with f(x)=g(x)=x and Yn=x0 gives E[(X(r)-Ybar(r))∫_{tn}^r dW] = x0 (r-tn)(e^{r-tn}-1), which is nonzero. This is load-bearing because it removes a term from the local error bound (6.15); as written, the proof does not establish (6.15). The term appears to be O((r-tn)^2) and can be bounded via Cauchy-Schwarz and an 'a b' inequality, preserving the Gronwall argument and the final order-one convergence, but the proof must be corrected.","section":"Section 6, proof of Theorem 4.1, estimate for I_{1.2}"},{"comment":"The theorem requires the backstop map φ to satisfy the conditional mean-square bound (3.5). In the numerical experiments the backstop is the tamed Milstein method (2.15) from [27], which is analysed in [27] only under the commutativity condition (2.16). For the non-commutative test in Section 5.2, (2.16) fails and (3.5) is not verified, so the experiment in Figure 2(e,f) is not covered by Theorem 4.1. The authors should either prove (3.5) for a tamed/truncated Milstein scheme without commutativity, or use a backstop for which (3.5) is known to hold, or explicitly state that the non-commutative experiment is illustrative rather than a consequence of the theorem.","section":"Definition 3.2 / Remark 3.1 / Section 5.2"}],"minor_comments":[{"comment":"The word 'asssume' should be 'assume'.","section":"Assumption 3.1"},{"comment":"In the display of (3.4), the integrals in the backstop term are written with limits ∫_s^{tn}, which appears to be reversed; they should presumably read ∫_{tn}^{s}.","section":"Definition 3.2, Eq. (3.4)"},{"comment":"The proof invokes (4.1) for the error X(tn)-Yn at intermediate mesh points, but (4.1) is stated only for the terminal time T. The argument would be complete if the authors state that the same bound holds at all mesh points, or justify the extension.","section":"Theorem 4.2, proof"},{"comment":"In the caption of Figure 2(e), 'Aaptive' should be 'Adaptive'.","section":"Figure 2 caption"}],"recommendation":"major_revision","confidential_remarks":"The false equality in Section 6 is the most serious concern; it is repairable, so I recommend major revision rather than rejection. The backstop-hypothesis gap for the non-commutative experiment also needs to be addressed before publication. The paper's core idea is sound and the Lévy-area estimates are a useful contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely new adaptive Milstein method for non-globally Lipschitz SDEs without commutativity, and the main theorem is almost right—but there is a false equality in the proof of Theorem 4.1 that needs repairing, not just polishing.\n\nWhat it does well: the path-bounded step-size strategy is a real extension of the authors' earlier adaptive Euler framework. The proof is structured honestly: explicit constants, separate treatment of drift and diffusion errors, and a self-contained Lévy area moment lemma via the characteristic function. The claim that the backstop is used with arbitrarily small probability (Theorem 4.2) follows from Markov and the convergence bound. That is a useful package.\n\nThe soft spots, in order of severity. First, I1.2 in Section 6: the paper sets E[<E(r), Σ Df(Yn)gi(Yn)∫dWi>|F_tn]=0 \"by (3.6)\". That is not justified. E(r)=X(r)-Ybar(r) depends on the Brownian increments on [tn,r], so it is not F_tn-measurable, and the conditional expectation of the product does not vanish. The stress-test example is correct: for f(x)=g(x)=x, one gets nonzero x0 Δt(e^{Δt}-1), which is O(Δt^2). So the printed equality is false. The good news is the error is not order-destroying; it can be bounded by (1/2)E‖E(r)‖^2 + C Δt^2 and absorbed into the existing K2 h^2 terms. The proof is repairable, but a referee should insist the step be rewritten.\n\nSecond, the backstop gap the reader flagged: Theorem 4.1 assumes the backstop φ satisfies (3.5), and Remark 3.1 says in practice they use tamed Milstein (2.15), which is only proved for commutative noise. In the non-commutative experiment (Section 5.2), the implemented backstop is therefore not covered by existing theory, and (3.5) is never checked. This is a real gap, though it may be minor in practice because the backstop is rarely triggered. It should be stated as an assumption plus numerical evidence, not left implicit.\n\nThird, the numerical rate plots compare hmax for the adaptive scheme against hmean for fixed-step schemes. That is a defensible convention, but it slightly obscures the comparison; worth an explicit sentence.\n\nBottom line: the central idea is new and the convergence framework is mostly sound, but the false equality means the paper is not accept-ready. I would send it out—a good referee can get the proof fixed in one round.","headline":"A genuinely new adaptive Milstein scheme without commutativity, with a main convergence proof that contains one repairable false equality and a backstop gap; worth refereeing, not accept-ready.","tokens_in":22142,"tokens_out":2949,"would_cite":true,"duration_ms":42922,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65C30","60H35"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that an explicit adaptive Milstein scheme, using path-bounded step sizes and a backstop method, converges strongly with first order for SDEs with one-sided Lipschitz drift and no commutativity condition on the noise.","keywords":["stochastic differential equations","adaptive time-stepping","Milstein method","strong convergence","one-sided Lipschitz drift","Lévy area","non-commutative noise","backstop method"],"falsifier":"Estimate the conditional mean-square error in (3.5) for one step of the tamed Milstein backstop at $h_{\\min}$ in the non-commutative setup: if Monte Carlo averages show it scales worse than $h_{\\min}^3$, the backstop does not meet the theorem's hypothesis, and a direct strong-convergence plot for the full scheme in that regime should show order loss.","tokens_in":21104,"feed_emoji":"🎲","tokens_out":6919,"duration_ms":58894,"temperature":0.7,"pith_summary":"This paper proves a first-order strong-convergence result for an explicit adaptive Milstein method applied to SDEs whose drift is only one-sided Lipschitz and may grow polynomially, with no commutativity condition on the noise. The method shrinks its step size when the numerical path approaches a sphere of radius $R$, and falls back to a small-step backstop scheme if the step would become too small. The main theorem bounds the $L_2$ error at the terminal time by $C h_{\\max}$, so halving the maximum step halves the error. This matters because explicit methods on uniform meshes are known to diverge for such SDEs, and Milstein-type methods usually need commutative noise; the adaptive scheme overcomes both. The paper also shows the probability of ever invoking the backstop can be made arbitrarily small.","feed_headline":"Adaptive Milstein method reaches order-one convergence","feed_subtitle":"Step sizes shrink near large states and a rarely used backstop handles non-Lipschitz drift.","key_machinery":"The central object is a path-bounded time-stepping strategy (Definition 3.3): whenever the step size is interior, the numerical path satisfies $Q \\leq \\|Y_n\\| < R$, in practice achieved by $h_{n+1} = \\max(h_{\\min}, \\min(h_{\\max}, h_{\\max}/\\|Y_n\\|))$, so the step shrinks as the norm grows. Bounding the path by $R$ makes the non-globally Lipschitz drift bounded along the numerical path, turning the one-sided Lipschitz condition into a usable error estimate. The argument also relies on the mean-square consistency bound (3.5) for the backstop method used at $h_{\\min}$, and on Lemma 2.3, which bounds conditional moments of L\\'evy areas by powers of the step length; that lemma is what lets the proof handle non-commutative noise.","core_discovery":"The central claim is Theorem 4.1: for a solution $X$ of (1.1) and an adaptive Milstein solution $\\bar{Y}$ built with a path-bounded time-stepping strategy (Definition 3.3) and a backstop satisfying the mean-square consistency bound (3.5), the root-mean-square terminal error satisfies $(\\mathbb{E}[\\|X(T)-\\bar{Y}(T)\\|^2])^{1/2} \\leq C(R,\\rho,h_{\\max}) h_{\\max}$, with $C(R,\\rho,h_{\\max}) = C(R,\\rho) + O(h_{\\max})$. Thus, for fixed radius $R$ and step ratio $\\rho$, the scheme is strongly convergent of order one. A second result (Theorem 4.2) says the probability that a step falls below $h_{\\min}$ and triggers the backstop can be made smaller than any tolerance by taking $\\rho$ large enough. The proof does not impose the commutativity condition $Dg_i g_j = Dg_j g_i$; instead it uses conditional moment bounds on L\\'evy areas (Lemma 2.3) to control the non-commutative terms.","pith_inferences":["A natural extension is to prove the mean-square bound (3.5) for the tamed Milstein backstop in the non-commutative setting; that would close the gap between the theorem's hypotheses and the paper's two-dimensional numerical example.","The path-bounded strategy suggests a practical calibration rule: choose $\\rho$ large enough that the backstop is never triggered, then choose $h_{\\max}$ small enough that the $O(h_{\\max})$ term in the error constant stays negligible.","Because the error bound is exponential in $T$ and in constants depending on $R$, the method is likely best suited to moderate time horizons; the adaptive step rule could be combined with other base methods whenever a backstop satisfying (3.5) is available."],"forward_implications":["For additive noise, the adaptive Milstein scheme coincides with an adaptive Euler-Maruyama scheme, and the same proof gives order-one strong convergence for that scheme.","The method applies without the commutativity condition (2.16), so the L\\'evy-area terms are controlled rather than assumed away.","The probability of using the backstop at any step can be driven below any fixed tolerance by increasing $\\rho$, so in typical runs the backstop is rarely activated.","The error constant grows as $\\rho \\to \\infty$, so a larger step ratio makes backstop use rarer but worsens the error constant; $h_{\\max}$ and $\\rho$ must be balanced.","The error bound depends on $R$ through constants that grow with $R$, so paths should be confined to the smallest radius compatible with the dynamics."],"supporting_citations":[{"why":"Supplies the tamed Milstein method used as the backstop and the main fixed-step comparison method; its commutativity restriction defines the gap this paper removes.","marker":"[27]"},{"why":"Introduces the adaptive time-stepping framework and the backstop idea that the path-bounded strategies here extend to the Milstein case.","marker":"[15]"},{"why":"Extends adaptive Euler methods to SDE systems with monotone coefficients, providing the admissible strategy class adapted in Definition 3.3.","marker":"[14]"},{"why":"Shows divergence of Euler-type methods on uniform meshes for non-globally Lipschitz coefficients, motivating the adaptive approach.","marker":"[11]"},{"why":"Defines the explicit Milstein discretisation on which the adaptive scheme is built.","marker":"[16]"},{"why":"Provides the characteristic-function based view of L\\'evy-area sampling used in the proof of Lemma 2.3.","marker":"[21]"},{"why":"Supplies the moment bounds for the exact solution (Lemma 2.2) used throughout the error analysis.","marker":"[8]"},{"why":"Gives the L\\'evy-area sampling procedure used in the two-dimensional non-commutative numerical experiment.","marker":"[7]"}],"fun_headline_variants":["Adaptive Milstein hits order-one strong convergence","Path-bounded steps give Milstein order-one rate","Milstein with adaptive steps tames non-Lipschitz SDEs","Step-size trick gives Milstein order-one"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the backstop method satisfying a mean-square accuracy condition, and in the non-commutative numerical test that condition is assumed rather than checked, because the tamed Milstein backstop was only proven for commutative noise.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive Milstein hits order-one strong convergence","Path-bounded steps give Milstein order-one rate","Milstein with adaptive steps tames non-Lipschitz SDEs","Step-size trick gives Milstein order-one"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000986,"raw_usage":{"total_tokens":4171,"prompt_tokens":926,"completion_tokens":3245,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":3192}},"tokens_in":542,"tokens_out":3245,"duration_ms":21833,"temperature":1.0,"reasoning_tokens":3192,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T06:01:35.797273+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Estimate the conditional mean-square error in (3.5) for one step of the tamed Milstein backstop at $h_{\\min}$ in the non-commutative setup: if Monte Carlo averages show it scales worse than $h_{\\min}^3$, the backstop does not meet the theorem's hypothesis, and a direct strong-convergence plot for the full scheme in that regime should show order loss.","supporting_citations":[{"cited_title":"W ang and S","cited_arxiv_id":null,"evidence_quote":"Supplies the tamed Milstein method used as the backstop and the main fixed-step comparison method; its commutativity restriction defines the gap this paper removes."},{"cited_title":"Kelly and G","cited_arxiv_id":null,"evidence_quote":"Introduces the adaptive time-stepping framework and the backstop idea that the path-bounded strategies here extend to the Milstein case."},{"cited_title":"Adaptive Euler methods for stochastic systems with non-globally Lipschitz coefficients","cited_arxiv_id":"1805.11137","evidence_quote":"Extends adaptive Euler methods to SDE systems with monotone coefficients, providing the admissible strategy class adapted in Definition 3.3."},{"cited_title":"Hutzenthaler, A","cited_arxiv_id":null,"evidence_quote":"Shows divergence of Euler-type methods on uniform meshes for non-globally Lipschitz coefficients, motivating the adaptive approach."},{"cited_title":"Kloeden and E","cited_arxiv_id":null,"evidence_quote":"Defines the explicit Milstein discretisation on which the adaptive scheme is built."},{"cited_title":"Malham and A","cited_arxiv_id":null,"evidence_quote":"Provides the characteristic-function based view of L\\'evy-area sampling used in the proof of Lemma 2.3."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the moment bounds for the exact solution (Lemma 2.2) used throughout the error analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the L\\'evy-area sampling procedure used in the two-dimensional non-commutative numerical experiment."}],"review_version":1}