{"id":"6ba875ed-bbdd-48d9-a85b-b3e24d4ad5ac","arxiv_id":"2607.14965","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The scenario-based optimal value in uncertain optimal control converges at N^{-1/2} to a Gaussian limit, or to the lower envelope of a Gaussian process when multiple optima exist, enabling asymptotic confidence intervals.","lead":"This paper proves limit laws and confidence intervals for the value of sample-based optimal control under parameter uncertainty, and tests them on fed-batch reactor case studies. The key insight is a stability bound linking final outcomes to the time-integrated input profile.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 2 (uniform reachable-set bound) is the load-bearing condition for Proposition A.1/Theorem 3.1 and is never verified for the two case-study models; if it fails, the FCLT and confidence intervals are unjustified.","rationale":"The reader's weakest assumption is exactly Assumption 2, and I agree this is the most load-bearing condition for the central theorem. The proof of Proposition A.1 and the FCLT in Theorem B.2 both rely on a single compact reachable set X that contains every trajectory for every feasible control and every uncertainty realization. Without it, the terminal-loss family {F(x_u(t_f,·)):u∈U} need not be uniformly bounded or uniformly Lipschitz in the primitive metric, so the C(K) empirical-process convergence fails and Theorem 3.1 has no basis. The assumption is not a harmless regularity condition: it is a global a priori bound that can be violated by simple control-affine ODEs with finite-time blow-up. I do not see a logical gap in the proof of Theorem 3.1 under Assumptions 1–4; the theorem is internally consistent. The weakness is external: the paper does not verify Assumption 2 for the two numerical case studies, even though those case studies are used to illustrate the plug-in and subsampling intervals. A small amount of additional analysis (or a computational reachability certificate for the discretized control class) would close the gap. Since the central mathematical claim appears correct, I would not reject; I would condition acceptance on adding that verification or explicitly scoping the numerical claims. This is a partial adjustment of the reader's ACCEPT: the concern is identical to the reader's weakest assumption, but I give it enough weight to require a revision rather than leaving it as a caveat.","tokens_in":16000,"tokens_out":20437,"duration_ms":220045,"concrete_test":"For each of §5.1 and §5.2, use the stated control bounds and uncertainty support Ξ to compute a validated uniform bound M with sup_{u∈U_q, ξ∈Ξ, t∈[0,t_f]} ||x_u(t,ξ)||_2 ≤ M (e.g., interval/validated integration over the q=50 discretized control class, or an explicit Lyapunov/barrier bound). If such M cannot be established, or if trajectories are found leaving every proposed compact X, the case studies fall outside Theorem 3.1 and the interval guarantees are not justified; if M is established, substitute X={||x||_2≤M} into Assumption 2 and re-run the coverage study to confirm the asymptotic claims.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Proposition A.1 (eq. A.2) requires one finite constant C_{U,t_f} such that ||x_u(t_f,ξ)-x_v(t_f,ξ)|| ≤ C_{U,t_f} d(u,v) for all u,v∈U and ξ∈Ξ. This is obtained from Assumption 2 (a common compact reachable set X) plus Assumption 3. Local Lipschitz/continuous differentiability alone cannot yield such a uniform constant: for \\dot x = u x^2 on [0,1] with U={u∈L2: ||u||_2≤1} and x(0)=1, the control u≡1 is feasible and the trajectory blows up at t=1, so no compact X contains all trajectories. When this happens, the boundedness and entropy estimates in Theorem B.2 (and hence the limit theorem 3.1) collapse. The paper never proves Assumption 2 for the fed-batch models in §5; it only lists the control bounds. The specific kinetics may well satisfy it (volume is 1+∫u, x4 stays below 20 when ξ8<0, etc.), but no argument is given that every feasible u and ξ keeps x1,x2,x3 in a fixed compact set. Thus the numerical illustrations of Theorems 4.1–4.2 rest on an unverified global reachability certificate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops asymptotic statistical inference for sample average approximation (SAA) in finite-horizon open-loop optimal control under parametric uncertainty. The main theoretical result is Theorem 3.1: under assumptions of weak compactness of the control set, a uniform compact reachable set, smoothness of the dynamics, and Lipschitz terminal cost, the scaled SAA optimal-value error converges in distribution to the infimum of a centered Gaussian process over the population argmin set; when the argmin is unique, the limit is normal. The paper derives two confidence-interval procedures: a plug-in normal interval under uniqueness and a subsampling interval that does not require uniqueness. The proofs are based on a Volterra-operator stability estimate (Proposition A.1) and an empirical-process FCLT (Theorem B.2). Numerical case studies on fed-batch reactor and ethanol fermentation problems illustrate the confidence intervals, and the authors report honest finite-sample undercoverage of the plug-in interval.","tokens_in":1434,"tokens_out":1653,"duration_ms":147729,"significance":"If the results hold, this is a useful and nontrivial contribution. The key analytic idea is Proposition A.1: under control-affine dynamics and a compact reachable set, terminal states depend Lipschitz-continuously on the cumulative input profile, allowing the control set to be equipped with a compact metric without imposing norm compactness on the controls. This cleanly yields the FCLT and the limiting distribution of the optimal-value error. The distinction between the Gaussian plug-in regime and the non-Gaussian nonunique-argmin regime is practically relevant for scenario-based dynamic optimization. The paper ships reproducible code and archived data, and the numerical study is transparent about the fact that the plug-in interval undercovers at the tested sample sizes. The main gap is formal verification of the reachable-set assumption for the numerical models; the argmin step in the proof of Theorem 3.1 is also deferred to a citation and should be spelled out.","major_comments":[{"comment":"Assumption 2 is stated but never verified for the two fed-batch models in §5. The uniformity constant C_{U,t_f} in Prop. A.1 is obtained from a common compact reachable set X containing x_u(t,ξ) for all u∈U, ξ∈Ξ, and t∈[0,t_f]; this same boundedness drives the entropy estimate and Theorem B.2. The case studies only list control bounds and the parametric uncertainty model; no invariant-set or boundedness argument is given for the state components. Since a Lipschitz/continuity assumption alone is not enough to guarantee such a uniform bound (the growth-rate terms are polynomial/exponential in the states), the numerical illustrations of Theorems 4.1–4.2 are not formally covered unless Assumption 2 is checked. Please add a short verification (e.g., monotonicity/invariance boxes using the sign of ξ_8, the feed concentration, and the terminal-volume bound) or explicitly downgrade the examples","section":"§5.1–5.2, Assumption 2"},{"comment":"The transition from Theorem B.2 to (3.1) is the sentence that the arguments from Shapiro et al. (2021), Theorem 5.7, apply verbatim. This is the load-bearing argmin step. Please state the theorem and verify its hypotheses in the present infinite-dimensional setting: compactness of K=(U,d) has been established, but the existence/measurability of SAA minimizers and the treatment of the nonunique argmin set U* should be made explicit. A short direct proof using the continuity of the infimum map and an argmin expansion would make the paper self-contained. As written, the central limit result inherits all of its difficulty from this citation.","section":"Appendix B, proof of Theorem 3.1"}],"minor_comments":[{"comment":"The word 'convex' in Assumption 2 is not used anywhere in the proofs (compactness suffices). If convexity is intended for later use, say so; otherwise remove it.","section":"§2, Assumption 2"},{"comment":"The terminal-state constraint x_4(t_f) ≤ 200 should be incorporated explicitly into the definition of the feasible set U for Example 5.2 (e.g., as an integral bound on u), since the theory's U is described in §2 by control bounds alone.","section":"§5.2"},{"comment":"The space C(0,t_f;R^s) should be written as C([0,t_f];R^s), since all estimates take suprema over the closed interval, including t_f.","section":"Notation"},{"comment":"The reference value J*_{N_ref} is itself a random proxy for J*, so the reported coverage is coverage of J*_{N_ref} rather than of J*; a sentence noting this additional noise source would be useful.","section":"§5.1, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically sound and likely publishable after a focused revision. The most important point is the verification of Assumption 2 for the two case studies; without it, the numerical claims are not formally supported. The argmin step in Theorem 3.1 should also be made explicit for the infinite-dimensional setting. Neither issue appears to threaten the core analytic result, and the numerical honesty about finite-sample undercoverage is a strength. I would be willing to accept after these revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real contribution here is Proposition A.1 — the Lipschitz stability of terminal states in the cumulative-input metric d(u,v) = max_t |∫_0^t (u-v)|. That single estimate turns an infinite-dimensional control problem into a compact empirical-process problem, and it lets the authors bring the classical SAA machinery (FCLT, argmin limit, subsampling) to a problem class where it hadn't been applied before. The main theorem — N^{1/2}(Ĵ_N^* − J^*) ⇒ inf_{u∈U*} Z(u), with Gaussian reduction under uniqueness — is exactly the right result, and the two confidence interval constructions (plug-in and subsampling) are sensible and clearly motivated. The numerics on the two fed-batch examples are transparent: they show finite-sample undercoverage, report the discretization, and the code is archived. That is reproducible work.\n\nThe soft spot is the one the stress-test flags, and it is real. Assumption 2 — a compact reachable set X containing all trajectories for all feasible u and all ξ — is load-bearing for Proposition A.1 and therefore for Theorem 3.1. Without it, the boundedness and entropy estimates collapse. Local Lipschitz data are not enough; the blow-up example \\x = u x^2 with a feasible u ≡ 1 shows that. The paper never proves that the two fed-batch models satisfy Assumption 2. The specific kinetics and control bounds probably do keep states bounded, but it is not automatic, and the numerical illustrations of Theorems 4.1–4.2 depend on it. This is not a fatal flaw in the mathematics — the theorem is stated as conditional on the assumption — but it is a gap in the application, and a serious referee should ask for either a proof of the reachable-set bound for the case studies or a discussion of how a user could verify it.\n\nMinor points: the proof of Theorem 3.1 hands off to Shapiro et al. Theorem 5.7, which is fine but should spell out why the metric-space conditions are satisfied. The subsampling method is computationally heavy, and the paper wisely omits coverage simulations for it.\n\nWho is this for? People working in stochastic optimal control, SAA for infinite-dimensional decision spaces, and batch-process optimization. It deserves peer review. The theory is solid; the gap is in the verification of the examples. With a reachability argument added, I would be comfortable with acceptance. Worth your time to read.","headline":"A genuinely new stability estimate in the cumulative-input metric yields the SAA optimal-value CLT and confidence intervals for infinite-dimensional control-affine Mayer problems; the theory is sound, but the empirical claims rest on an unverified global reachable-set assumption.","tokens_in":16807,"tokens_out":1705,"would_cite":true,"duration_ms":17256,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49K45","60F17","62F12","90C15"],"pacs":[],"model":"deepseek-v4-flash","headline":"The error of a scenario-based optimal value follows a Gaussian-process infimum, and reduces to a normal distribution when the true optimum is unique.","keywords":["sample average approximation","optimal control under uncertainty","functional central limit theorem","confidence intervals","batch process optimization","parametric uncertainty","optimal value estimation","subsampling"],"falsifier":"Construct a control-affine system that satisfies the smoothness and box-control assumptions but has finite-time blow-up for some bounded control, so the compact reachable set in Assumption 2 does not exist; if the √N Gaussian-process limit still appears empirically, the theorem's stated hypotheses are not sharp. Conversely, on a problem engineered with two exactly optimal controls, check whether the plug-in interval's coverage drops below nominal while the subsampling interval's coverage stays near nominal.","tokens_in":15890,"feed_emoji":"📊","tokens_out":7600,"duration_ms":74764,"temperature":0.7,"pith_summary":"The paper studies the sample average approximation (SAA) method for finite-horizon open-loop optimal control problems with uncertain parameters. Its central claim is that the scaled optimal-value error N^{1/2}(J_N^* - J^*) converges in distribution to the infimum of a centered Gaussian process over the set of population-optimal controls. When the population problem has exactly one optimal control, that limit is normal, with variance equal to the variance of the terminal loss at the optimal control. The paper uses this limit to construct two asymptotically valid confidence intervals for the true optimal value: a cheap plug-in interval under uniqueness, and a subsampling interval that works even when multiple optimizers exist. Two fed-batch case studies show that the intervals quantify the sampling error, with the plug-in interval undercovering at small sample sizes and approaching nominal coverage as the scenario count grows.","feed_headline":"Scenario-based optimization errors obey a √N Gaussian-process limit","feed_subtitle":"More scenarios shrink the error in the reported optimum; two confidence intervals now tell you how much to trust it.","key_machinery":"The carrying object is the time-integrated control profile, U(u)(t) = ∫_0^t u(s) ds, and the associated distance d(u,v) = max_t |U(u)(t) - U(v)(t)|. The key stability estimate states that terminal states are Lipschitz continuous in this distance, uniformly over the uncertainty set. This estimate turns the sample objective into a stochastic process indexed by a compact metric space, enabling a functional central limit theorem for empirical processes. The same distance controls the convergence of the estimated variance in the plug-in interval and the consistency of the subsampling distribution.","core_discovery":"The principal result is that, under compactness and smoothness assumptions, the SAA optimal value converges at the standard Monte Carlo rate N^{-1/2}, with the exact limiting distribution equal to inf_{u in U*} Z(u), where Z is a centered Gaussian process on the control space with covariance Cov(Z(u), Z(v)) = Cov(F(x_u(t_f, ξ)), F(x_v(t_f, ξ))). If the population problem has a unique optimal control u*, the limit simplifies to N(0, Var(F(x_{u*}(t_f, ξ)))). The paper shows that this distributional statement directly justifies two confidence intervals for the population optimal value: one plug-in interval relying on the Gaussian limit under uniqueness, and one subsampling interval that does no","pith_inferences":["If the compact reachable-set assumption cannot be verified for a given model — for example, kinetics that admit finite-time blow-up — the interval constructions have no theoretical grounding; a practitioner would need a reachability certificate before trusting either interval.","Because the stability estimate is in terms of cumulative input rather than the control value itself, the sampling uncertainty should be insensitive to high-frequency control chatter as long as the cumulative delivery is close; this is consistent with the paper's singular-control examples.","A direct next experiment is to engineer a problem with a flat objective plateau (two or more exactly optimal controls) and compare the two intervals: the subsampling interval should hold its coverage, while the plug-in interval should fail even at large N.","The same Gaussian-process machinery likely extends to risk measures such as conditional value-at-risk if a Lipschitz terminal cost is retained, giving the error distribution for risk-averse versions of the problem."],"forward_implications":["The statistical error in an SAA optimal value is of order N^{-1/2}; quadrupling the number of scenarios halves the error, independent of the dimension of the uncertain parameter vector.","When the true problem has a unique optimal policy, the plug-in normal interval gives an asymptotically calibrated 1−β confidence interval around the SAA value at negligible extra computational cost.","When several optimal policies exist, the limiting distribution is the lower envelope of a Gaussian process, so normal intervals fail; the subsampling interval maintains asymptotic coverage without requiring uniqueness.","In the two fed-batch studies, the plug-in interval's finite-sample coverage is below nominal for N between 8 and 64 and improves as N grows, so small-N plug-in intervals should be treated as diagnostics rather than calibrated statements.","The theory covers open-loop policies with risk-neutral terminal costs; the paper identifies risk-averse objectives, state-path constraints, and feedback policies as open extensions."],"fun_headline_variants":["Scenario-based optimization error follows √N Gaussian limit","Sample-average optimal value converges at N^{-1/2} rate","Confidence intervals for scenario-based optimal value","Two confidence intervals for scenario-based optimization derived"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Everything rests on the assumption that a single compact set contains every state trajectory for every admissible control and every uncertain parameter (Assumption 2); without a verifiable bound like that, the Lipschitz stability estimate and both confidence interval guarantees have no foundation.","fun_headline_variants_meta":{"raw":{"variants":["Scenario-based optimization error follows √N Gaussian limit","Sample-average optimal value converges at N^{-1/2} rate","Confidence intervals for scenario-based optimal value","Two confidence intervals for scenario-based optimization derived"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000878,"raw_usage":{"total_tokens":3618,"prompt_tokens":715,"completion_tokens":2903,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":459,"completion_tokens_details":{"reasoning_tokens":2842}},"tokens_in":459,"tokens_out":2903,"duration_ms":21188,"temperature":1.0,"reasoning_tokens":2842,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T00:34:25.363705+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a control-affine system that satisfies the smoothness and box-control assumptions but has finite-time blow-up for some bounded control, so the compact reachable set in Assumption 2 does not exist; if the √N Gaussian-process limit still appears empirically, the theorem's stated hypotheses are not sharp. Conversely, on a problem engineered with two exactly optimal controls, check whether the plug-in interval's coverage drops below nominal while the subsampling interval's coverage stays near nominal.","supporting_citations":[],"review_version":1}