{"id":"f83eca8d-eab8-4a61-907e-c295d8c4bbd2","arxiv_id":"2607.18192","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The small-noise limit of path-dependent risk-sensitive exit-time control is a deterministic control problem with Cameron–Martin cost.","lead":"This paper proves that risk-sensitive exit-time control problems for stochastic differential equations with coefficients that depend on the whole past path converge, as the noise vanishes, to a deterministic control problem. The result gives a rigorous bridge between a randomly perturbed system with memory and a simpler optimization problem, useful for safety analysis of engineered systems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Main theorem depends on Lemma 3.2, which defers to unpublished comparison theorem [9] for convex expectations on path space; unverified extension to relaxed-control framework is the key correctness risk.","rationale":"The reader's weakest assumption correctly identifies the pivotal unverified step. The central claim Theorem 2.4 is a convergence result whose proof is built on Theorem 3.1, a variational formula for log-transformed control problems with path-dependent coefficients. Theorem 3.1's proof reduces the problem to Lipschitz payoffs via Lemma 3.2, which invokes an unpublished comparison theorem for convex expectations [9]. Since the exit-time payoffs in the main proof are discontinuous (upper/lower semicontinuous) and not Lipschitz, this reduction is essential. The paper does not contain a self-contained proof or even a verification that E* and E' satisfy the convex-expectation axioms necessary for [9, Theorem 2.12] in the relaxed-control framework with unbounded controls and infinite-dimensional path space. This is a genuine correctness risk. Other omissions (supersolution property in Lemma 3.6(b), martingale-problem verification in Proposition 4.2, and the use of Theorem 3.1 for ε_n≠1) are either standard or easily repaired, and do not threaten the central argument as directly. Thus I agree with the reader's CONDITIONAL verdict; no verdict change is needed.","tokens_in":24853,"tokens_out":16180,"duration_ms":134839,"concrete_test":"Independently verify whether E'(φ)=sup_{P∈R(t,ω)} E^P[φ - 1/2∫∥z∥^2] satisfies the axioms of [9, Definition 2.1] on the path space—especially any continuity-from-below or tightness assumption—and confirm that [9, Theorem 2.12] applies to the pair (E*, E'). If an axiom fails, construct a bounded upper semianalytic φ (e.g., g(τ_D)) for which (3.1) is violated; this would falsify Lemma 3.2 and hence Theorem 2.4. If the axioms are verified, the reduction is sound and the main proof is complete modulo the minor technical omissions noted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1, Lemma 3.2 is the pivotal reduction: it claims the variational representation (3.1) for all bounded upper semianalytic φ follows once it holds for bounded Lipschitz φ with finite horizon. The proof reads: 'It follows precisely as in the proofs of [9, Theorems 2.21, 5.2] that E* and E' are convex expectations on the path space as defined in [9, Definition 2.1]. Equipped with this observation, the claimed equivalence follows directly from the comparison [9, Theorem 2.12].' Here [9] is an unpublished preprint (to appear in Math. Oper. Res., arXiv:2503.10572). The present paper does not verify the axioms of [9, Definition 2.1] for E'—the relaxed-control value with unbounded control z and quadratic cost—nor the hypotheses of the comparison theorem on the infinite-dimensional path space C(R_+;R^d). If [9, Theorem 2.12] does not apply, Theorem 3.1 is only proved for Lipschitz payoffs. Yet the proof of Theorem 2.4 in Section 5 applies Theorem 3.1 to φ = g(τ_{D_δ}^{t_n}) in the upper bound (5.1) and φ = g(τ_D^t) in the lower bound (5.2); these exit-time payoffs are not Lipschitz. Hence both the upper and lower semicontinuity arguments in Section 5 would lack a variational foundation, breaking the central convergence result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the small-noise asymptotics of a risk-sensitive exit-time control problem for SDEs with path-dependent coefficients. The main result, Theorem 2.4, asserts that for initial conditions in a set T, the log-transformed value ε log sup E[exp(g(τ_D)/ε)] converges to a deterministic path-dependent control value V^0, defined through Cameron–Martin controls h and relaxed controls m. The proof proceeds by first establishing a variational representation (Theorem 3.1) using path-dependent PDE techniques and convex expectations on path space, then proving a relaxed-control convergence result (Corollary 4.3), and finally combining these with semicontinuity arguments in Section 5. The paper also contains a worked one-dimensional example with memory.","tokens_in":25275,"tokens_out":23887,"duration_ms":236779,"significance":"If correct, the paper would provide a path-dependent extension of the Boué–Dupuis variational approach and of classical small-noise limits for risk-sensitive escape control, with an explicit computable example. The combination of PPDE comparison principles, convex expectations, and relaxed-control convergence is original and the overall strategy is well motivated. The paper is clearly written and the semicontinuity manipulations in Section 5 are transparent. However, the main theorem rests on two load-bearing points that are not fully justified: the reduction to Lipschitz payoffs is deferred to an unpublished comparison result, and the deterministic limit V^0 defined in Section 2 does not obviously coincide with the relaxed-control limit actually obtained in Section 5. These issues affect the central claim and require attention.","major_comments":[{"comment":"After applying Corollary 4.3, the limsup is bounded by sup_{P∈R^0_M(t,ω)} E_P[g(τ^t_{Dδ}) - cost]. The next inequality '≤ V^{0,δ}_{t,ω}' is asserted without argument. Here V^{0,δ} is defined in Section 2 as a supremum over (h,m) with product structure m(ds,dλ)⊗δ_{h'(s)}(dz), whereas R^0 consists of relaxed controls with arbitrary joint measures M*(ds,dλ,dz). For a general M*, the marginal m on Λ and the averaged h'(s)=∫ z M*_s(dλ,dz) do not reproduce the drift ∫σ(s,X,λ) z M*; the equality would require, e.g., σ independent of λ or z independent of λ conditionally. The product controls are a strict subset of R^0, and the supremum over R^0 can be strictly larger. No convexity or density argument is provided to justify the reduction. This makes the upper bound (5.1) unproved and Theorem 2.4, as stated, unsupported.","section":"Section 3.1, Lemma 3.2"},{"comment":"The proof of Lemma 3.2 is one sentence: it asserts that E* and E' are convex expectations as in [9] and then invokes the comparison theorem [9, Theorem 2.12]. This is load-bearing because Theorem 3.1 is later applied in Section 5 to exit-time payoffs g(τ_D^{t_n}), which are not Lipschitz. The present paper does not verify that E', defined via relaxed control rules with unbounded control z and quadratic cost on the infinite-dimensional path space C(R_+;R^d), satisfies the axioms of [9, Definition 2.1] or the hypotheses of the comparison theorem. The dependence on an unpublished preprint makes this gap more serious. Please either include a self-contained verification of the convex-expectation axioms and the comparison, or provide a direct proof of the reduction for the specific structure of E'.","section":"Section 3.2, Lemma 3.6(b)"},{"comment":"Lemma 3.6(b) proves only the subsolution property and explicitly leaves the supersolution property and the d-Lipschitz continuity to the reader. These properties are necessary for Lemma 3.9, where v=e^{tilde v} is shown to solve the G-backward equation, and for the uniqueness argument. The omitted parts are not routine in the presence of an unbounded action variable z and the relaxed-control topology. Similarly, Proposition 4.2(a) omits the martingale-problem argument and the compactness proof that underpin Corollary 4.3. Since these results are essential for both the variational formula and the convergence argument, the full proofs or precise statements of how they follow from the cited references should be included.","section":"Section 4, Proposition 4.2(a)"}],"minor_comments":[{"comment":"The reduction to λ≡1 is justified heuristically ('Intuitively, ...', 'Approximating controls in A by piecewise constant controls, we find ...'). Since this is a claimed explicit characterization of the limit, the argument should either be made rigorous or explicitly labeled as heuristic.","section":"Discussion 2.5"},{"comment":"The proof that the stated geometric conditions imply (t,ω)∈T is informal; some estimates gloss over measurability and the treatment of the case τ_D=∞ is abbreviated. Please clarify these points.","section":"General"},{"comment":"The notation V^0 vs V^{0,δ} and the distinction between the deterministic control problem in Section 2 and the relaxed-control value in Sections 4–5 is confusing; a remark explaining the intended equivalence would help.","section":"Appendix A"},{"comment":"The definition of C^{1,2}_{pol} is delegated to the appendix, but the appendix defines derivatives only on the space of càdlàg paths. The passage between continuous and càdlàg paths should be stated more carefully.","section":"Section 3.1, proof of Lemma 3.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies heavily on the authors' unpublished preprint [9] for a load-bearing comparison theorem. I recommend that the editor ensure [9] is available and that its hypotheses indeed cover the relaxed-control setting here; otherwise the variational representation is only proved for Lipschitz payoffs. More importantly, the deterministic limit V^0 in Section 2 appears not to match the relaxed-control limit R^0 that the proof actually yields. This is not a presentation issue: the paper's main theorem may need to be restated with a different limiting control problem. A revision should either prove the missing reduction or correct the definition of V^0 and rework the example accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate generalization of the Boué–Dupuis variational formula and the small-noise exit-time limit to path-dependent coefficients. Theorem 3.1 is new, and Theorem 2.4 is the stated convergence. If both hold, they fill a real gap in the literature.\n\nWhat the paper does well: the strategy — variational representation, then control-rule convergence via Hausdorff metric — is the right template, and the authors are transparent about where they skip. The example (2.6) with memory is genuinely computable, and the verification of (0,0)∈T is a nice touch. The semicontinuity arguments in Section 5 are clear.\n\nThe soft spot, and it is a real one, is Lemma 3.2. The proof says E* and E' are convex expectations 'precisely as in the proofs of [9, Theorems 2.21, 5.2]' and then uses [9, Theorem 2.12]. That comparison is doing a lot of work: it is the only bridge from Lipschitz payoffs to the upper semianalytic exit-time payoffs used in Section 5. The paper does not verify that E' — a relaxed-control rule on C(R_+;R^d) with unbounded z and quadratic cost — satisfies the axioms of [9, Definition 2.1], nor that the comparison theorem applies on that infinite-dimensional path space. And [9] is an unpublished preprint. This is not a fatal flaw on its own; the claim may well be true. But the proof as written does not establish it. A referee should ask for the verification to be written out or for [9] to appear in a published form before Theorem 2.4 is taken as proved.\n\nOther omissions are minor by comparison. Lemma 3.6(b) skips the supersolution property and d-Lipschitz continuity; Proposition 4.2(a) abbreviates the martingale-problem argument. Both are standard, likely fillable. The reliance on [10], [11] is fine; they are published and the cited results are in-scope.\n\nVerdict: conditional accept, in substance. The central argument is coherent and the missing pieces appear to be technical rather than conceptual. Who should read it: people working on risk-sensitive exit-time control, large deviations for non-Markovian systems, or PPDE methods. It is not a desk reject. Send it to a referee, and make the main instruction this: get Lemma 3.2 fully verified, or gate the theorem on the published version of [9].","headline":"Real path-dependent extension of Boué–Dupuis with a genuine gap in the Lipschitz reduction; deserves refereeing, but the key lemma needs to be either fixed or deferred to a published [9].","tokens_in":25718,"tokens_out":2741,"would_cite":true,"duration_ms":25805,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93E03","93E20","35D40","60H10","60F10","49L25"],"pacs":[],"model":"deepseek-v4-flash","headline":"For risk-sensitive exit-time control problems with path-dependent coefficients, this paper proves that the log-transformed value converges to the value of a deterministic control problem as the noise intensity vanishes.","keywords":["risk-sensitive control","exit-time problems","small-noise asymptotics","path-dependent coefficients","variational representation","path-dependent PDE","convex expectations","relaxed controls"],"falsifier":"Exhibit a path-dependent SDE satisfying Assumption 2.1 and a bounded upper semianalytic payoff φ for which the claimed variational equality (3.1) fails, or find an initial condition (t,ω)∈T and sequences (t_n,ω_n), ε_n→0 for which V^{ε_n}_{t_n,ω_n} does not converge to V^0_{t,ω}. More narrowly, checking whether the comparison theorem from the authors' preprint holds for relaxed control rules with unbounded z would settle the load-bearing step.","tokens_in":24776,"feed_emoji":"🎯","tokens_out":7574,"duration_ms":63535,"temperature":0.7,"pith_summary":"This paper establishes what happens to a risk-sensitive exit-time control problem when the noise intensity in a controlled stochastic differential equation goes to zero. The coefficients may depend on the whole past of the path, not just the current state, and the controller maximizes an exponentially weighted exit time. The main theorem shows that, after taking logarithms and scaling by the noise level, the value converges to the value of a deterministic control problem with path-dependent coefficients and a quadratic penalty on control effort. To prove this, the paper derives a variational representation for log-transformed control problems that combines path-dependent PDE methods with convex expectations, then identifies the limit through convergence of relaxed control rules. A worked example with a memory-dependent drift gives the limiting value and optimal control explicitly.","feed_headline":"Path-dependent exit-time control: small-noise limit is deterministic","feed_subtitle":"Risk-sensitive exit value converges to a deterministic control problem even with path-dependent coefficients.","key_machinery":"The load-bearing object is the variational representation E^*_t(φ)=E'_t(φ) (Theorem 3.1) for the log-transform of a path-dependent stochastic control problem. It expresses the entropically transformed value as a supremal expectation over relaxed control rules carrying an extra drift control z and a quadratic penalty ∥z∥²/2. The proof identifies both sides as viscosity solutions of the same path-dependent Hamilton–Jacobi–Bellman equation—one with the maximization Hamiltonian G and one with its Legendre-transformed counterpart G̃—and uses a comparison principle for convex expectations on path space to reduce the class of admissible payoffs from upper semianalytic to Lipschitz functions.","core_discovery":"On the paper's own terms, the discovery is Theorem 2.4: for every initial condition (t,ω) belonging to a set T of regular initial conditions, every sequence of initial conditions converging to (t,ω), and every noise level ε_n→0, the logarithmically scaled risk-sensitive exit-time value V^{ε_n}_{t_n,ω_n} converges to V^0_{t,ω}, the value of a deterministic path-dependent control problem. The deterministic limiting problem has two ingredients not visible at finite noise: an added drift control drawn from the Cameron–Martin space and a quadratic penalty for its energy. The regularity set T excludes boundary behaviour where the δ-blow-up value V^{0,δ} does not converge to V^0, and the paper give","pith_inferences":["If the comparison principle invoked in Lemma 3.2 extends as the authors expect, the same variational formula should hold for bounded upper semianalytic payoffs far beyond exit times, including quantile-like and constraint-type criteria in path-dependent control.","The condition (t,ω)∈T is the real substance of the theorem: the paper's boundary analysis suggests that, in Markovian models, T is essentially the set of starting points from which the boundary can be left with a finite-energy control; testing this characterization in fully path-dependent models is a natural next step.","The Hausdorff convergence of relaxed control rules in Corollary 4.3 is proved under quadratic moment bounds; one could expect analogous convergence under p-th moment bounds, which would extend the method to rewards with polynomial growth.","The memory example indicates that the limiting control problem sometimes reduces to a finite-dimensional Hilbert-space projection; similar reductions could be systematically derived for affine or linearized path-dependent dynamics."],"forward_implications":["Small-noise asymptotics for risk-sensitive exit problems are now available when drift and diffusion depend on the entire past, not just the current state.","The limiting deterministic control problem is explicit, so optimal limiting strategies can be computed: a path-dependent ODE with an added Cameron–Martin drift and a quadratic control cost.","The variational formula Theorem 3.1 applies to measurable functions of the path beyond the exit-time reward g(τ_D), making it a standalone tool for entropic transforms.","The memory example yields a fully computed limit and explicit optimal control, demonstrating that the abstract condition (t,ω)∈T holds in a concrete non-Markovian model.","Because convergence holds along sequences of initial conditions, the result has a stability property: small perturbations of initial data do not break the limit."],"fun_headline_variants":["Path-dependent exit-time risk limit is deterministic control with drift penalty","Small-noise limit of risk-sensitive exit-time: deterministic control","Exit-time control converges to deterministic path-dependent problem","Risk-sensitive exit-time limit adds drift cost, becomes deterministic","Path-dependent exit-time: log value limit is deterministic with quadratic cost"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof chain assumes that the comparison theorem for convex expectations on path space, proved by the authors for a Markovian setting, remains valid in the present path-dependent relaxed-control framework; Lemma 3.2 invokes it without a self-contained proof, and the reduction from general payoffs to Lipschitz payoffs—a step Theorem 3.1 cannot do without—collapses if that comparison fails.","fun_headline_variants_meta":{"raw":{"variants":["Path-dependent exit-time risk limit is deterministic control with drift penalty","Small-noise limit of risk-sensitive exit-time: deterministic control","Exit-time control converges to deterministic path-dependent problem","Risk-sensitive exit-time limit adds drift cost, becomes deterministic","Path-dependent exit-time: log value limit is deterministic with quadratic cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000655,"raw_usage":{"total_tokens":2792,"prompt_tokens":658,"completion_tokens":2134,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":402,"completion_tokens_details":{"reasoning_tokens":2066}},"tokens_in":402,"tokens_out":2134,"duration_ms":53025,"temperature":1.0,"reasoning_tokens":2066,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T15:43:28.411660+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Exhibit a path-dependent SDE satisfying Assumption 2.1 and a bounded upper semianalytic payoff φ for which the claimed variational equality (3.1) fails, or find an initial condition (t,ω)∈T and sequences (t_n,ω_n), ε_n→0 for which V^{ε_n}_{t_n,ω_n} does not converge to V^0_{t,ω}. More narrowly, checking whether the comparison theorem from the authors' preprint holds for relaxed control rules with unbounded z would settle the load-bearing step.","supporting_citations":[],"review_version":1}