{"id":"aa8f376c-4a2b-4ac0-9321-963b9844f27e","arxiv_id":"2411.15930","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Euler-Maruyama path sensitivity approximations for SDEs converge strongly with order O(h^{1/2}) under bounded derivative assumptions, matching the rate for the underlying process.","lead":"This paper proves that Euler-Maruyama approximations of the derivative of a stochastic process with respect to a parameter converge strongly at the same half-order rate as the process itself. The result fills a gap in stochastic numerical analysis that matters for computing option price sensitivities with multilevel Monte Carlo methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: Theorem 2 is sound under the stated assumptions; the bounded-derivative condition is restrictive but explicit.","rationale":"The reader's verdict identifies the bounded-derivative assumption as the weakest point, and I agree that it is the most restrictive condition. However, this is an explicit assumption of the theorem, not a hidden or circular one, and it is needed for the cited standard EM results. My scrutiny of the proof found no load-bearing flaw: the non-Lipschitz product terms are the core difficulty, and the proof controls them using the bounded derivatives plus finite moments of the sensitivity processes. The only issue I noticed is notational: some inequalities compare with |hat S_u - S_u| though the coefficients are evaluated at the rounded-down time; the corrected floor-time bounds still yield O(h^{p/2}) and do not change the conclusion. The higher-order and vector extensions are only sketched, but the main theorem for first-order scalar sensitivities is fully demonstrated. Therefore I see no reason to change the reader's ACCEPT verdict.","tokens_in":8285,"tokens_out":22058,"duration_ms":197559,"concrete_test":"Analytically re-derive the bound for the first pair in Theorem 2 with the rounded-down coefficient: verify E|dot a(hat S_floor u) - dot a(S_u)|^p <= L_a^p E|hat S_floor u - S_u|^p = O(h^{p/2}) by splitting into E|hat S_floor u - S_floor u|^p and E|S_floor u - S_u|^p. If this check fails, the published proof has a gap; otherwise it confirms the theorem.","verdict_should_be":"UNCHANGED","load_bearing_attack":"No significant objection identified. I retraced the proof of Theorem 2 and found the strategy valid: the error is decomposed into twelve terms, each of which is O(h^{p/2}). Terms involving differences of coefficients are controlled by the standard strong error E|hat S - S|^p = O(h^{p/2}); terms involving rounded-time increments of S or dot S are controlled by the classical increment bounds; products of the form (a'(hat S)-a'(S)) hat dot S are handled by Cauchy-Schwarz using finite moments of hat dot S. The final Gronwall step closes the argument. The boundedness of a', b' and second derivatives is exactly what converts the otherwise non-Lipschitz products into error-times-finite-moment terms. One minor notational issue: several displayed inequalities write |hat S_u - S_u| where the coefficient actually depends on |hat S_floor u - S_u|; replacing it with the floor value still gives O(h^{p/2}) by the triangle inequality, so the theorem remains correct. The higher-order and vector extensions are sketched rather than fully proved, but they do not affect the central first-order theorem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This note proves that the Euler-Maruyama discretisation of the pathwise sensitivity SDE (the derivative of the SDE solution with respect to a parameter) has strong L^p error of order 1/2 on finite time intervals, under the assumption that the drift and diffusion coefficients have bounded first and second derivatives. The main result is Theorem 2, which gives E[sup_{0<t<T} |\\hat{\\dot S}_t - \\dot S_t|^p] <= c_p^{(3)} h^{p/2} for any p>=2. The proof follows the classical EM strong convergence proof, decomposing the error into twelve drift and diffusion terms and bounding each with Jensen, Hölder, BDG, and Grönwall inequalities. The paper also sketches extensions to vector SDEs/parameters and higher-order sensitivities.","tokens_in":8432,"tokens_out":8006,"duration_ms":64689,"significance":"The result is significant because the combined process (S, \\dot S) does not satisfy the standard global Lipschitz condition (the product b'(S) \\dot S is not Lipschitz in the joint variable), so the usual EM convergence theorem cannot be applied directly. The note fills this gap with a self-contained proof that uses only standard tools and explicitly stated bounded-derivative assumptions. The main theorem gives a rigorous foundation for the use of pathwise sensitivities in multilevel Monte Carlo and Greeks computations. The paper is honest about what is proved completely (the scalar first-order case) and what is only sketched (vector and higher-order extensions); the sketched extensions do not affect the central claim. The proof contains no fitted parameters or circular arguments.","major_comments":[],"minor_comments":[{"comment":"The proof of Lemma 3 is given as a single sentence ('The proof follows the same approach used with Theorem 1'); since this lemma is used in Theorem 2 to control moments of \\hat{\\dot S}, please expand the proof to show the Gronwall argument for the piecewise-constant coefficient process explicitly.","section":"Section 3, Lemma 3"},{"comment":"The displayed inequalities for the terms involving (\\hat a'_u - a'_u) \\hat{\\dot S}_u and similar write E|\\hat S_u - S_u|^p where the mean value theorem actually gives |\\hat S_{\\underline u} - S_u|; the two are interchangeable up to constants by the triangle inequality, but the notation should be made precise.","section":"Section 3, proof of Theorem 2"},{"comment":"The assumptions list bounded derivatives but do not state the finite-moment assumptions on S0 and \\dot S0; please add explicit conditions such as E|S0|^p < ∞ and E|\\dot S0|^p < ∞ (for the relevant p) to make the Gronwall arguments fully rigorous.","section":"Section 2, opening paragraph"},{"comment":"There are a few typos ('a n autonomous', 'pro ves') that should be corrected.","section":"Abstract and Introduction"},{"comment":"The extension to vector SDEs and higher-order sensitivities is only sketched; if the journal allows, a remark stating that the details are omitted for brevity would be helpful, or alternatively provide a brief indication of the inductive step.","section":"Section 4, Extensions"}],"recommendation":"minor_revision","confidential_remarks":"This is a well-written, compact note that fills a real gap. The central theorem's proof is sound. The main improvements needed are cosmetic and presentational. I recommend minor revision. I do not see any need for rejection. The sketched extensions are acceptable for a note."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a clean, short proof note that closes a real gap. It proves that Euler-Maruyama approximations of path sensitivities (the Greeks) converge strongly with order 1/2, matching the path approximation itself. The obstacle is real—the sensitivity SDE has coefficients that are not globally Lipschitz because b' times Sdot appears—and the proof handles it with the bounded-derivative assumption. That is exactly the sort of result you want before using pathwise sensitivities inside multilevel Monte Carlo.\n\nWhat's new: prior literature proved strong convergence of the Euler-Maruyama path but not for the pathwise derivative with respect to a parameter. Giles shows that if the first and second derivatives of drift and diffusion are uniformly bounded, then the sensitivity approximation converges in L^p with the usual half-order rate. The proof is a retracing of the Kloeden-Platen argument, using Jensen, Hölder, BDG, and Grönwall. I retraced the twelve-term decomposition in Theorem 2; the strategy works. The stress-test note flagged one notational slip where the text writes |S_u - S_u| but the coefficient depends on the floor value; that is harmless because the floor-value difference is still O(h^{p/2}) by the triangle inequality.\n\nThe soft spots are mostly about scope. The bounded-derivative assumption excludes many common financial models (e.g., CIR, Heston) where coefficients have unbounded derivatives near zero. That is stated openly, so it's a limitation, not a flaw. The vector and higher-order extensions are sketched, not fully proved, which is fine for a note whose main target is the scalar first-order theorem. There are no numerical experiments and no machine-checked formalization, but the proof is standard enough that this does not undermine confidence.\n\nI think the result is correct and worth having in the literature. The main theorem is properly proved, and the gap it fills is real. It deserves a serious referee; I'd send it out. For a journal, minor revision would be enough—clarify the floor notation and maybe add a remark on what would be needed to relax the bounded-derivative condition. For someone working on MLMC Greeks, this is a useful citation; for others in numerical SDEs, it's a tidy, correct result that resolves a known open point.","headline":"Short proof note that closes a real gap: Euler-Maruyama path sensitivities converge strongly with order 1/2 under bounded derivative assumptions.","tokens_in":8978,"tokens_out":1761,"would_cite":true,"duration_ms":15799,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65C30","60H35","65C05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Differentiating the Euler-Maruyama approximation with respect to a parameter yields path sensitivity estimates whose strong error is O(h^{1/2}), matching the path approximation itself.","keywords":["Euler-Maruyama","pathwise sensitivities","strong convergence","stochastic differential equations","multilevel Monte Carlo","Greeks","infinitesimal perturbation analysis","Burkholder-Davis-Gundy inequality"],"falsifier":"Run the sensitivity estimator on a scalar SDE with bounded, smooth coefficients, for example $a(\\theta,S)=\\sin(S+\\theta)$, $b(\\theta,S)=\\cos(S+\\theta)$, computing $\\dot S$ exactly from the same Brownian increments (or a very fine reference simulation), and estimate $E[\\sup_{0<t<T}|\\hat{\\dot S}-\\dot S|^2]$ for $h=2^{-4},...,2^{-8}$. If the error does not decay proportionally to $h$, the theorem's rate would be contradicted. Conversely, replacing the diffusion by an unbounded-derivative coefficient, such as $b(\\theta,S)=1+S^2$, and observing a slower rate would confirm that the bounded-derivative assumption is load-bearing.","tokens_in":8032,"feed_emoji":"📉","tokens_out":11063,"duration_ms":100855,"temperature":0.7,"pith_summary":"This note proves a convergence result that stochastic numerical analysis had left open: if you estimate the derivative of an SDE path with respect to a model parameter by differentiating the Euler-Maruyama scheme, the estimate converges strongly with order $1/2$ on a finite time interval. That is the same strong order as the Euler-Maruyama approximation of the path itself, despite the fact that the differentiated SDE violates the global Lipschitz condition normally required for the standard proof. The theorem covers scalar and vector parameters and paths, and by induction covers second- and higher-order sensitivities whenever the needed derivatives of drift and diffusion exist and are bounded. Because multilevel Monte Carlo error analysis relies on strong orders, the result supplies the missing ingredient for a rigorous treatment of pathwise Greeks in that setting.","feed_headline":"Sensitivity estimators converge at strong order one half","feed_subtitle":"The missing O(h^{1/2}) strong-convergence bound for pathwise Greeks in SDE simulation is now proved.","key_machinery":"The central object is the pathwise sensitivity SDE, $d\\dot S_t = (\\dot a_t + a'_t \\dot S_t)\\,dt + (\\dot b_t + b'_t \\dot S_t)\\,dW_t$, obtained by differentiating the original SDE in the parameter $\\theta$. Its Euler-Maruyama discretisation is exactly the derivative of the Euler-Maruyama discretisation of the original SDE, which is what makes the estimator cheap to compute. The proof machinery is the continuous-time interpolation of both schemes followed by a term-by-term decomposition of the error $E_t = \\hat{\\dot S}_t - \\dot S_t$ into twelve integrals, each bounded by Jensen, H\\\"older, the Burkholder-Davis-Gundy inequality, the known $O(h^{p/2})$ path error from [10], and Gr\\\"onwall's inequality; boundedness of first and second derivatives supplies the constants $L_a, L_b$ that keep every product controlled.","core_discovery":"Theorem 2 is the central result: under the assumption that the first and second derivatives of the drift $a$ and diffusion $b$ with respect to $S$ and $\\theta$ exist and are uniformly bounded, for any $p \\ge 2$ there is a constant $c_p^{(3)}$ such that $E[\\sup_{0<t<T}|\\hat{\\dot S}_t - \\dot S_t|^p] \\le c_p^{(3)} h^{p/2}$, where $\\hat{\\dot S}$ is the continuous-time interpolation of the Euler-Maruyama approximation to the sensitivity SDE and $\\dot S$ is the exact pathwise sensitivity. In words, the sensitivity estimator has $L^p$ strong order $1/2$, matching the classical Euler-Maruyama path bound. The proof re-traces the standard Euler-Maruyama analysis and controls the extra terms through the boundedness of $a'$, $b'$ and the cross-differences; the same argument extends to vector SDEs, vector parameters, and $k$-th order sensitivities, with Lemma 4 bounding products of differences at each induction step.","pith_inferences":["The uniform boundedness of derivatives is likely stronger than needed; the proof only uses $L^p$ control of products, so polynomial-growth coefficients with finite moments may give the same order. This is an editorial inference, not a claim of the paper.","The same re-tracing argument should transfer to other one-step schemes whose path approximation has strong order $1/2$ or $1$, such as Milstein; a direct test would be to verify the sensitivity error slope numerically on a model with bounded derivatives.","A simple numerical check would settle the practical rate: for a model like $dS_t = \\sin(S_t+\\theta)\\,dt + \\cos(S_t+\\theta)\\,dW_t$, estimate $E[\\sup|\\hat{\\dot S}-\\dot S|^2]$ over many paths at $h=2^{-4},...,2^{-8}$ and compare the slope to $1/2$; the paper gives no numerical experiments."],"forward_implications":["With the strong order $1/2$ established for sensitivity estimators, the standard multilevel Monte Carlo variance and cost analysis applies to pathwise sensitivity estimates, not just to path estimates.","Second- and higher-order sensitivities converge at the same strong order, provided the corresponding derivatives of the drift and diffusion are bounded, so higher-order Greeks inherit the same convergence behaviour.","The argument covers multi-dimensional SDEs and multi-parameter sensitivities, so the result is not limited to scalar test problems.","The author conjectures that the same analysis extends to the Milstein scheme, leaving that extension for future work."],"supporting_citations":[{"why":"Supplies the standard Euler-Maruyama strong convergence theorem and its proof template, which the sensitivity proof re-traces.","marker":"[10]"},{"why":"Supplies the Burkholder-Davis-Gundy martingale inequality used to bound stochastic integrals throughout the proof.","marker":"[4]"},{"why":"Defines multilevel Monte Carlo estimation, whose error analysis is the stated reason the O(h^{1/2}) sensitivity bound matters.","marker":"[6]"}],"fun_headline_variants":["Sensitivity estimators hit strong order 1/2","Pathwise Greeks match path convergence rate","Strong order 1/2 for SDE sensitivities","Euler-Maruyama sensitivity error O(h^1/2) proven"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes the first and second derivatives of the drift and diffusion coefficients with respect to both the state and the parameter exist and are uniformly bounded for all parameter and state values; if any of those derivatives can grow without bound, the constants used to control every term in the error decomposition do not exist and the argument collapses.","fun_headline_variants_meta":{"raw":{"variants":["Sensitivity estimators hit strong order 1/2","Pathwise Greeks match path convergence rate","Strong order 1/2 for SDE sensitivities","Euler-Maruyama sensitivity error O(h^1/2) proven"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00014,"raw_usage":{"total_tokens":1112,"prompt_tokens":847,"completion_tokens":265,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":200}},"tokens_in":463,"tokens_out":265,"duration_ms":3239,"temperature":1.0,"reasoning_tokens":200,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:43:17.987740+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the sensitivity estimator on a scalar SDE with bounded, smooth coefficients, for example $a(\\theta,S)=\\sin(S+\\theta)$, $b(\\theta,S)=\\cos(S+\\theta)$, computing $\\dot S$ exactly from the same Brownian increments (or a very fine reference simulation), and estimate $E[\\sup_{0<t<T}|\\hat{\\dot S}-\\dot S|^2]$ for $h=2^{-4},...,2^{-8}$. If the error does not decay proportionally to $h$, the theorem's rate would be contradicted. Conversely, replacing the diffusion by an unbounded-derivative coefficient, such as $b(\\theta,S)=1+S^2$, and observing a slower rate would confirm that the bounded-derivative assumption is load-bearing.","supporting_citations":[{"cited_title":"Kloeden and E","cited_arxiv_id":null,"evidence_quote":"Supplies the standard Euler-Maruyama strong convergence theorem and its proof template, which the sensitivity proof re-traces."},{"cited_title":"Burkholder, B","cited_arxiv_id":null,"evidence_quote":"Supplies the Burkholder-Davis-Gundy martingale inequality used to bound stochastic integrals throughout the proof."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines multilevel Monte Carlo estimation, whose error analysis is the stated reason the O(h^{1/2}) sensitivity bound matters."}],"review_version":1}