{"id":"2f6eacca-f4c2-4fff-aa98-4a2ec8d849b8","arxiv_id":"1908.04550","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"New probabilistic representations yield unbiased Monte Carlo estimators for integration by parts and Bismut-Elworthy-Li formulas for one-dimensional killed diffusions.","lead":"This paper proves exact integration by parts formulas for diffusion processes killed at a boundary, using a reflection-principle Markov chain and a custom integration by parts calculus. These formulas lead to unbiased Monte Carlo estimators for barrier-type sensitivities, with finite variance achievable by importance sampling.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central representation (16) relies on an unproved, possibly false uniform C^{1,2} regularity bound for the killed semigroup in Section 10.1.","rationale":"Both the killed-law representation (Theorem 2) and the two IBP formulas (Theorems 5 and 7) are built on the semigroup regularity assumption in Section 10.1. The reader's weakest_assumption identifies exactly this imported C^{1,2} regularity. My stress-test agrees and sharpens it: for the paper's C^1_b assumptions, the stated uniform bound on second spatial derivatives up to t=0 is not a standard consequence and can fail for the Dirichlet heat kernel on a half-line. This is load-bearing because the Itô expansion and the induction step controlling remainders use the bound directly. Other parts of the argument, such as the Markov chain Malliavin duality (11), the boundary-merging computations, and the time-merging integrability estimates, appear internally coherent and are not where the central claim is most vulnerable. A concrete analytical check on the constant-coefficient case would settle whether the regularity claim holds; if it fails, the authors would need to either add weighted estimates or restrict the function class. Since the concern is testable and the rest of the proof is promising, the reader's CONDITIONAL verdict remains appropriate.","tokens_in":48030,"tokens_out":29657,"duration_ms":279851,"concrete_test":"For the model case σ≡1, b≡0, L=0, let f∈C^1_b([0,∞)) with f(0)=0 and f'(y)=1/log(1+y) (smoothed near 0). Compute P_t f(x)=∫_0^∞(g_t(x-y)-g_t(x+y))f(y)dy and evaluate sup_{0≤t≤T,x≥0}|∂_x^2 P_t f(x)|. If this is infinite, the uniform bound in Section 10.1 is false and the proof of Theorem 2 needs modification; if finite, the cited regularity result still needs to be verified for the actual half-line domain and C^1_b initial data assumed in Theorems 5 and 7.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The derivation of the central representation (16) in Section 10.1 applies Itô's formula to P_{T-s}f(\\bar Y_s). This requires the imported regularity result stated there: under (H), for smooth f with f(L)=0, P f ∈ C^{1,2}((0,T]×[L,∞)) and sup_{0≤t≤T}|∂_x^ℓ P_t f|∞ ≤ C for ℓ=1,2. The paper cites Garroni-Menaldi [15,16] for this but neither proves it nor verifies that the cited result covers the half-line domain [L,∞) with initial data f∈C^1_b (in the paper's nonstandard sense allowing linear growth). The claimed uniform bound up to t=0 is not a standard Schauder consequence for merely C^1_b data; for the Dirichlet heat kernel on a half-line, boundary layers can make ∂_x^2 P_t f unbounded of order t^{-1/2} as t↓0 (e.g., f with f(0)=0 and f' with a slow modulus of continuity). Since the one-step expansion (54), the estimates (56), and the induction (58) all rely on this sup bound, a failure invalidates Theorem 2 and hence Theorems 5 and 7. This is the most load-bearing uncertainty in the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper establishes probabilistic representations for two integration by parts formulas for a one-dimensional diffusion killed at a fixed level L: one of Bismut-Elworthy-Li type for the derivative of the killed semigroup with respect to the starting point, and one for the derivative with respect to the terminal point, i.e., a representation for E[f'(X_T)1_{τ≥T}] and ∂_x E[f(X_T)1_{τ≥T}]. The approach combines a Markov chain approximation of the killed process, based on the reflection principle, with a finite-dimensional Malliavin calculus built on the Gaussian increments of the chain and the jump times of an independent Poisson process. The main results, Theorem 5 and Theorem 7, express these derivatives as expectations of the test function evaluated on the chain multiplied by explicit random weights; the paper further shows how to apply importance sampling on the jump times to obtain finite variance and provides numerical tests.","tokens_in":48329,"tokens_out":12013,"duration_ms":127419,"significance":"The methodology is original and potentially valuable: using the Poisson jump times to merge boundary terms is an elegant way to regularize Dirac-type singularities without the usual localisation arguments, and the resulting representations are directly usable for unbiased Monte Carlo simulation. The boundary merging lemmas in the appendix are nontrivial and are supported by explicit Gaussian convolution computations. If the technical gaps identified below are fixed, the paper would be a meaningful contribution to Monte Carlo methods for killed diffusions and to Malliavin calculus on discrete-time approximations. The numerical experiments give concrete evidence for the practical behaviour of the estimators, although the reported variances remain large for higher oscillation of the coefficients.","major_comments":[{"comment":"The proof of Theorem 2 rests on the imported regularity assertion that under (H), for smooth f with f(L)=0, the killed semigroup satisfies P f ∈ C^{1,2}((0,T]×[L,∞)) and sup_{0≤t≤T}|∂_x^ℓ P_t f|∞ ≤ C for ℓ=1,2 (Section 10.1, before Eq. (54)). The manuscript does not verify that the cited results [15,16] apply to the half-line domain with initial data in the paper's class C^1_b, which allows unbounded functions with bounded derivative. For such data the second-derivative bound is not a standard consequence: for a compactly supported C^1 function that behaves like (y-L)^{3/2} near L, ∂_x^2 P_t f diverges like t^{-1/4} as t↓0. Since the one-step expansion (54), the estimates (56) and the induction (58) all depend on this bound, the derivation of the representation (16) for the full class in Theorem 2 is incomplete. The authors should either prove a version of the bound under their stated assumptions or provide an explicit approximation argument that proves (16) for arbitrary f in that class without assuming the uniform second-derivative bound.","section":"10.1"},{"comment":"Lemma 7, the forward transfer-of-derivatives formula, is the key ingredient in the proof of the Bismut-Elworthy-Li formula (Theorem 7), but its proof is omitted with only the remark that it is similar to Lemma 4. Given that the proof of Lemma 4 itself relies on the nontrivial identities (61) and Corollary 3, the omission leaves Theorem 7 without a verifiable derivation. Please include a complete proof of Lemma 7 or point to a fully specified appendix where every step is carried out.","section":"6 (Lemma 7)"},{"comment":"The central practical claim of Section 7 is that Beta-distributed jump times achieve finite moments of all orders for the estimators from Theorems 2, 5 and 7. For the Markov-chain representation (45) a calculation is sketched, but for the IBP weights the text states that the proof 'follows similar lines of reasonings ... and is thus omitted'. This is a load-bearing claim because the weights in (48)-(50) are more singular than the weights in (46). The authors should provide the moment estimate for the IBP estimators, including the verification of the time-degeneracy estimates (51)-(52) for the reweighted boundary-merging weights.","section":"7 (importance sampling)"},{"comment":"The final Step 6 of the proof of Theorem 5 only sketches the absolute convergence and the L^p-integrability (p<2) of the right-hand side, referring to 'a similar argument' as in Section 10. The argument for the terms involving G1 and G2 and the application of Lemma 13 require a careful justification of the uniform integrability over the infinite sum over n. Moreover, Lemma 13, which is used crucially in the jump-reduction procedure, is stated without proof. Please provide the proof of Lemma 13 and a complete derivation of the moment bounds for the tree weights.","section":"5.3 (Theorem 5, Step 6)"}],"minor_comments":[{"comment":"The reference in the Introduction to 'Anderson and Kohatsu-Higa [3]' should be 'Andersson and Kohatsu-Higa'; the same typo appears in the bibliography, where reference [4] is dated '20 18'.","section":"1"},{"comment":"The symbol E= is used without a formal definition in Section 10.1 and elsewhere; it should be defined explicitly in the notation section, as it may be confused with equality in law.","section":"Notation"},{"comment":"The sentence 'Under the condition ... P_t f(L)=f(L)=0 together with P_0 f(x)=f(x)' overloads the symbol P (semigroup vs probability) and would be clearer if written as 'the semigroup P_t satisfies the Dirichlet condition P_t f(L)=0'.","section":"10.1"},{"comment":"The description of the symbol sets S_{n+1}, \\bar S^k_{n+1} and \\hat S^k_{n+1} is terse; a small table with the meaning of each symbol (0, e, c, I, B*e, Bfe) would improve readability.","section":"5.1"}],"recommendation":"major_revision","confidential_remarks":"This is a technically ambitious paper with interesting ideas and a substantial amount of correct-looking computation. The main concern is not the overall strategy but the incomplete support for several load-bearing claims: the imported semigroup regularity in Section 10.1, the omitted proof of Lemma 7, the sketched finite-variance analysis in Section 7, and the unproved time-merging Lemma 13. I believe these gaps are fillable within the manuscript's scope, so I recommend major revision rather than rejection. The authors should be encouraged to provide complete proofs, possibly in a supplementary appendix, and to state clearly the function-space assumptions under which the semigroup regularity is used."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper has real substance, and the boundary-merging mechanism is the most interesting part. But the proof of Theorem 2, and with it Theorems 5 and 7, leans on a regularity claim in Section 10.1 that I do not think is true as stated. The paper asserts that for smooth f with f(L)=0, the killed semigroup satisfies sup_{0≤t≤T}|∂_x^ℓ P_t f|_∞ ≤ C for ℓ=1,2. For the Dirichlet heat kernel on the half-line, take f(x)=x^2. The solution has a boundary layer that makes ∂_x^2 P_t f(0) of order t^{-1/2}, so the uniform bound fails while f(0)=0. This is not an exotic counterexample; it is the standard one-dimensional Brownian motion, which satisfies assumption (H). The one-step Itô expansion (54) and the estimates (56) feed directly off that bound, and without it the induction (58) does not close. So the proof as written has a load-bearing gap.\n\nWhat the paper does well: the boundary merging lemmas are a real piece of work. Using the Poisson jump time to convolve the Gaussian density against the boundary Dirac term is clever, and the explicit computations in the appendix look correct. The Markov chain representation via the reflection principle is a solid extension of the Bally–Kohatsu-Higa framework, and the unbiased Monte Carlo angle is honestly presented, including the variance problem and the importance-sampling fix. The numerical tests are limited but candid about deterioration as the coefficients oscillate more.\n\nMinor issues: Lemma 7 is dismissed with “proof is similar”; given the length of the paper, a sketch would help. No code is supplied, which is a small reproducibility hit for a simulation paper.\n\nMy overall take: the main construction is likely salvageable, and the IBP tree algebra is coherent, but the analytic foundation needs repair. I would not desk-reject this; it deserves serious refereeing, with the specific instruction to check the semigroup regularity and to fix the proof, possibly with time-weighted norms or by smoothing the test function before applying Itô's formula.","headline":"The construction is genuinely new and the boundary merging is clever, but the proof of the central representation relies on a semigroup regularity estimate that appears false as stated.","tokens_in":48839,"tokens_out":4686,"would_cite":false,"duration_ms":48455,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that derivatives of killed-diffusion expectations can be written as exact expectations over a reflected Markov chain sampled at Poisson jump times, yielding unbiased Monte Carlo estimators.","keywords":["integration by parts","killed processes","Malliavin calculus","Markov chain approximation","Monte Carlo simulation","boundary merging","Bismut-Elworthy-Li formula","importance sampling"],"falsifier":"Use killed Brownian motion with constant drift, where $E[f(X_T)1_{\\tau\\ge T}]$ and its derivatives have closed-form expressions, and Monte Carlo evaluate the right-hand side of Theorem 7 with increasing sample sizes; any systematic mismatch between the estimator's average and the closed-form derivative would refute the identity.","tokens_in":47854,"feed_emoji":"🎲","tokens_out":9299,"duration_ms":89611,"temperature":0.7,"pith_summary":"For a one-dimensional diffusion killed when it hits a level $L$, the paper proves exact probabilistic representations for two integration-by-parts formulas: one for $E[f'(X_T)1_{\\tau\\ge T}]$ and one, of Bismut-Elworthy-Li type, for $\\partial_x E[f(X_T)1_{\\tau\\ge T}]$. The trick is to replace the diffusion by a reflected Markov chain sampled at the jump times of an independent Poisson process, so that the infinite-dimensional integration-by-parts problem becomes finite-dimensional. Boundary terms that appear when derivatives are transferred backward in time are smoothed by averaging over the intermediate Poisson jump time, a step the paper calls boundary merging. Because the final formulas are exact expectations, they give unbiased Monte Carlo estimators; choosing a Beta distribution for the jump times makes every moment finite.","feed_headline":"Unbiased Monte Carlo for killed-diffusion derivatives","feed_subtitle":"A Poisson-time reflected chain plus a boundary-merging step makes integration-by-parts formulas exact.","key_machinery":"The central object is the reflected one-step Markov chain $\\bar X_{i+1}=\\rho_{i+1}\\bar X_i+(1-\\rho_{i+1})(2L-\\bar X_i)+\\sigma(\\bar X_i)(W_{\\zeta_{i+1}}-W_{\\zeta_i})$, where $\\rho_{i+1}$ is a Bernoulli(1/2) coin flip and the $\\zeta_i$ are Poisson jump times; the reflection principle makes its law match a killed Brownian step. On this chain the paper builds a finite-dimensional Malliavin calculus: a derivative $D_{i+1}$ in the Gaussian increment $Z_{i+1}$ and an adjoint integral operator $I_{i+1}$ satisfying the duality $E[D_{i+1}f(\\bar X_{i+1})H]=E[f(\\bar X_{i+1})I_{i+1}(H)]$. The load-bearing mechanism is the boundary-merging lemma: when a transferred derivative leaves a weight multiplied by $\\delta_L(\\bar X_i)$, one conditions on the intermediate jump time $\\zeta_i$, uses explicit Gaussian time-convolution identities, and replaces the two transitions by a single transition of the merged boundary chain $\\bar X^B$. This removes the time singularity and keeps the final estimator in $L^p$ for $p<2$, or for every $p\\ge1$ under Beta importance sampling.","core_discovery":"Under Assumption (H) (smooth, bounded, uniformly elliptic coefficients), for test functions $f\\in C_b^1(\\mathbb{R})$ with $f(L)=0$, the paper proves Theorem 5: $T E[f'(X_T)1_{\\tau\\ge T}]$ equals an expectation over the reflected chain $\\bar X$ and its boundary-merged variant $\\bar X^B$, with explicit weights built from the chain increments and Poisson jump times. Theorem 7 gives the parallel Bismut-Elworthy-Li formula for $T\\partial_x E[f(X_T)1_{\\tau\\ge T}]$, obtained by transferring derivatives forward in time; there the boundary terms vanish because $f(L)=0$, so no merging is needed. Corollaries 1 and 2 convert these representations into formulas for the derivatives of the killed transition density with respect to the terminal point and the starting point. The whole construction is designed so that an unbiased Monte Carlo simulation follows directly, with no discretization bias.","pith_inferences":["The boundary-merging recipe is likely a general template: whenever a local Malliavin weight collides with a Dirac boundary term, averaging over the intermediate jump time convolves two Gaussian densities and weakens the singularity; the same template may apply to local times, occupation times, or running maxima of one-dimensional diffusions.","The paper's restriction to one dimension is driven by the reflection principle and explicit Gaussian kernels; in multidimensional settings where a reflection principle or explicit transition density for the approximating chain exists, the perturbation-plus-merging argument may extend.","The numerical tests show variance growing sharply as the sinusoidal modulation of the diffusion coefficient increases, which suggests the method is most effective for nearly constant coefficients and would need higher-order variance reduction for strongly state-dependent noise."],"forward_implications":["Theorem 5 gives an exact, unbiased Monte Carlo estimator for $T E[f'(X_T)1_{\\tau\\ge T}]$ without any time discretization bias.","Theorem 7 gives the corresponding Bismut-Elworthy-Li estimator for the derivative with respect to the starting point, with no boundary-merging step needed.","Corollaries 1 and 2 provide probabilistic representations of the derivatives of the killed transition density with respect to the terminal and initial points.","Choosing Beta-distributed jump times makes the estimator's moments finite for every $p\\ge 1$, whereas exponential jump times guarantee only $p<2$.","The condition $f(L)=0$ can be removed by replacing $f$ with $f-f(L)$, so the formulas extend to general smooth test functions."],"supporting_citations":[{"why":"Supplies the $C^{1,2}$ regularity of the killed semigroup up to the boundary, used in the Ito expansion in Section 10.1.","marker":"[15]"},{"why":"Companion analytic regularity result for the parabolic integro-differential problem, also imported for the semigroup bounds.","marker":"[16]"},{"why":"Provides the probabilistic parametrix/perturbation construction for marginal laws that the Markov chain representation adapts to killed processes.","marker":"[7]"},{"why":"Earlier analytic representation for the killed diffusion's marginal law; the chain representation here is the more workable version.","marker":"[14]"},{"why":"Originates the importance-sampling scheme on jump times used in Section 7 to obtain finite variance.","marker":"[3]"},{"why":"Gives the generalized integration-by-parts framework for Gaussian transition densities used to define the Dirac boundary terms.","marker":"[22]"},{"why":"The reflection principle that underlies the construction of the reflected Markov chain in Lemma 1.","marker":"[24]"},{"why":"The continuous-time Bismut-Elworthy-Li integration-by-parts formula for killed diffusions that this paper recovers in unbiased simulation form.","marker":"[25]"}],"fun_headline_variants":["Exact IBP for killed diffusions via reflected chains","Unbiased Monte Carlo for killed diffusion derivatives","Reflected chain yields exact integration by parts","Poisson-time reflected chain: unbiased IBP for killed processes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument needs the survival expectation $P_t f(x)=E[f(X_t)1_{\\tau>t}]$ to be twice differentiable in space and once in time, with bounded derivatives up to the killing level; if this regularity fails, the one-step Ito expansion that connects the diffusion to the reflected chain breaks down.","fun_headline_variants_meta":{"raw":{"variants":["Exact IBP for killed diffusions via reflected chains","Unbiased Monte Carlo for killed diffusion derivatives","Reflected chain yields exact integration by parts","Poisson-time reflected chain: unbiased IBP for killed processes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000431,"raw_usage":{"total_tokens":2138,"prompt_tokens":821,"completion_tokens":1317,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":1255}},"tokens_in":437,"tokens_out":1317,"duration_ms":11348,"temperature":1.0,"reasoning_tokens":1255,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:39:18.659550+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use killed Brownian motion with constant drift, where $E[f(X_T)1_{\\tau\\ge T}]$ and its derivatives have closed-form expressions, and Monte Carlo evaluate the right-hand side of Theorem 7 with increasing sample sizes; any systematic mismatch between the estimator's average and the closed-form derivative would refute the identity.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the $C^{1,2}$ regularity of the killed semigroup up to the boundary, used in the Ito expansion in Section 10.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Companion analytic regularity result for the parabolic integro-differential problem, also imported for the semigroup bounds."},{"cited_title":"Bally and A","cited_arxiv_id":null,"evidence_quote":"Provides the probabilistic parametrix/perturbation construction for marginal laws that the Markov chain representation adapts to killed processes."},{"cited_title":"On the first hitting times of one dimensional elliptic diffusions","cited_arxiv_id":"1609.09327","evidence_quote":"Earlier analytic representation for the killed diffusion's marginal law; the chain representation here is the more workable version."},{"cited_title":"Andersson and A","cited_arxiv_id":null,"evidence_quote":"Originates the importance-sampling scheme on jump times used in Section 7 to obtain finite variance."},{"cited_title":"Ikeda and S","cited_arxiv_id":null,"evidence_quote":"Gives the generalized integration-by-parts framework for Gaussian transition densities used to define the Dirac boundary terms."},{"cited_title":"Karatzas and S","cited_arxiv_id":null,"evidence_quote":"The reflection principle that underlies the construction of the reflected Markov chain in Lemma 1."},{"cited_title":"Malliavin and A","cited_arxiv_id":null,"evidence_quote":"The continuous-time Bismut-Elworthy-Li integration-by-parts formula for killed diffusions that this paper recovers in unbiased simulation form."}],"review_version":1}