{"id":"73ded02d-9170-4ad3-a708-77fc083f14e8","arxiv_id":"2411.19132","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Conformal prediction is used to construct prediction regions for error trajectories, which then tighten joint chance constraints in a tractable optimal control problem for linear stochastic systems.","lead":"The paper uses conformal prediction to design feedback controllers that satisfy joint chance constraints on state and input for linear systems with unknown noise distributions. It provides a data-driven way to tighten constraints without assuming Gaussian or bounded noise, with marginal coverage guarantees that hold regardless of sample size.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 is proved only for marginal coverage over calibration data; the realized PR from Lemma 3 is not certified to satisfy the chance constraints in (3) for the deployed controller.","rationale":"The reader's weakest assumption—that calibrated PRs may not satisfy condition (6)—is real, but my sharpened concern is more fundamental: even when the PRs fit, Lemma 3 does not certify the realized PR's coverage. The paper itself acknowledges the marginal nature of CP in Remark 1, but Theorem 1 and the abstract's 'feasible for the original chance-constrained problem' do not carry the needed qualifier. A PAC-style correction such as the one in Remark 1 could repair the statement by adding a confidence level β and enlarging \\hatθ, but the paper does not adopt it. I therefore recommend CONDITIONAL acceptance: the algorithms and LMI derivations appear sound, but the headline guarantee must be restated as marginal over the calibration data (or complemented with a PAC confidence parameter), and Theorem 1 should clearly separate the joint probability space from the conditional one. The proposed simulation test would empirically confirm the gap.","tokens_in":12788,"tokens_out":18968,"duration_ms":186888,"concrete_test":"Use the paper's double-integrator example and the training set that yields Kdir. Draw B=1000 independent calibration datasets of size k1+1; for each, compute Ce and CKe by Lemma 3, solve (7), and estimate conditional coverage of (3c) and (3d) with M=10^6 fresh disturbance sequences. Record the fraction of the 1000 realized controllers whose estimated coverage is below 0.95 and compare it with the average coverage across runs. If a substantial fraction of realized controllers fall below 0.95 while the average is near 0.95, the guarantee is marginal and Theorem 1 does not certify the deployed controller.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Definition 1 defines a PR as a fixed set with Pr{ξ ∈ E} ≥ 1−θ. Lemma 3, however, certifies only the marginal statement Pr_{test,cal}{R_test ≤ C_e; R_test^u ≤ C_Ke} ≥ 1−θ, where the probability averages over the calibration data used to compute C_e and C_Ke. Conditioning on the realized calibration set, the coverage of B(C_e) and B(C_Ke) is a random variable with no certified lower bound; it can be substantially below 1−θ. Theorem 1's proof treats E as a fixed event with Pr{E} ≥ 1−θ and concludes that (u,x) is feasible for (3). This is valid only in the joint probability space over calibration and future disturbances, not for the realized controller. The chance constraints in (3c)–(3d) concern the deployed system, so the central claim overreaches unless the marginal interpretation is stated as the guarantee. The direct method's separate failure mode, that calibration radii may violate condition (6), is the same underlying issue: the realized PR is not certified either to have true coverage 1−θ or to fit inside the constraint ellipsoids.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers a finite-horizon linear stochastic optimal control problem with joint chance constraints on state and input and an unknown disturbance distribution. It decomposes the state into a nominal part and an error part, introduces prediction regions for error trajectories via split conformal prediction, and tightens the ellipsoidal constraints using the Pontryagin set difference. Two construction methods are proposed: a direct method that optimizes a feedback gain over training trajectories and then calibrates ball radii, and an indirect method that forms a disturbance ellipsoid by conformal prediction and synthesizes a robustly invariant error ellipsoid via the S-procedure. The main result states that the solution of the tightened deterministic problem is feasible for the original chance-constrained problem with probability at least 1−θ, with the guarantee understood as marginal over the calibration data.","tokens_in":13029,"tokens_out":11852,"duration_ms":111220,"significance":"The contribution is timely and relevant: it is distribution-free, handles joint chance constraints through nonconformity scores defined over whole trajectories, and provides a clean S-procedure derivation for the indirect robust-invariance step. The paper also releases code and includes a numerical comparison with scenario-based MPC. The central probabilistic argument is standard split-conformal prediction and is, apart from the interpretation issue below, formally sound. However, the exact status of the guarantee—marginal over calibration rather than conditional for the deployed controller—and the absence of an end-to-end feasibility certificate must be clarified before the claims as stated are acceptable.","major_comments":[{"comment":"The proof of Theorem 1 uses Pr{E} ≥ 1−θ, where E is the event that the calibrated random set B(C_e) contains e(t) for all t; this probability is over both the calibration data and the future disturbance. The theorem statement, however, concludes that (u(0:N−1), x(1:N)) is a feasible solution to (3), whose probability is over disturbances only. For a realized calibration set, Pr{e(t) ∈ B(C_e) ∀t | D_cal} has no certified lower bound and can be below 1−θ. The paper should either state Theorem 1 explicitly as a marginal guarantee over calibration and disturbances, consistent with Remark 1 and the abstract, or use a PAC-style calibration (e.g., the bound in Remark 1) if the intended claim is about the deployed controller. As written, the central claim overreaches the proof.","section":"Section III-A, Theorem 1 and Lemma 3"},{"comment":"The design problem (11) only enforces training-quantile constraints η_e < η_max_e and η_u < η_max_u. The actual tightening radii used in (7) are the calibration quantiles C_e and C_Ke from Lemma 3. There is no certificate that these calibrated radii satisfy C_e < min_t 1/√λ_max(P_t) and C_Ke < 1/√λ_max(Q), which Lemma 2 and Theorem 1 require. The numerical example checks this after calibration, but the method can fail to produce any feasible instance of (7). This limitation should be stated explicitly, and the paper should discuss possible remedies such as increasing the calibration set, choosing a more conservative calibration level, or adding a feasibility verification step before solving (7).","section":"Section III-B, direct method (Eq. (11))"}],"minor_comments":[{"comment":"The name \"Vovk\" is typeset as \"V ovk\" in the Introduction and in References [12], [13], [22]; please correct the spacing.","section":"Throughout"},{"comment":"The symbol C_u is used in the statement of Lemma 2 while condition (6) uses C_Ke; unify the notation.","section":"Lemma 2"},{"comment":"The finite-sample correction \\theta = (1 + 1/(k−k1−1))(1−θ) is introduced with a citation but without derivation; a one-sentence justification would improve self-containedness.","section":"Section III-B, Eq. (11)"},{"comment":"The caption's \"E0.95(w(t))\" should be typeset as E^{0:N−1}_{0.95}(w(t)), and the comparison with scenario optimization should state explicitly that condition (6) may still require a sufficiently large calibration set.","section":"Fig. 1 and Section IV"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of L-CSS and the technical core is sound once the marginal-guarantee statement is made precise. The stress-test concern about Theorem 1 lands: the proof computes probabilities over calibration and future disturbance jointly, so the theorem statement must be qualified. I do not see a circularity problem; the conformal guarantee is an external statistical result. The main revision risk is that readers may mistake the marginal guarantee for a per-realization one, so the revision should address this explicitly and state the feasibility-gap limitation of the direct method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The useful contribution is the indirect method: conformal prediction on disturbance sequences gives an ellipsoidal prediction region, and the S-procedure turns that into a feedback gain plus an invariant ellipsoid for the error, with a convex LMI for fixed multipliers. That's clean and honestly distribution-free. The direct method is more of a bridge: it reformulates the problem as a quantile-based training step plus calibration, but the training step is nonconvex and solved by a genetic algorithm with no feasibility certificate. So treat the direct method as heuristic with a CP-certified calibration layer; the real intellectual content is the separation of training and calibration.\n\nThe soft spot that matters is the probabilistic interpretation. Lemma 3 certifies that the probability of the test trajectory staying in the realized prediction region, averaged over the calibration data, is at least 1−θ. That is a marginal guarantee. Theorem 1 treats the prediction region as a fixed event with Pr{E} ≥ 1−θ and concludes feasibility for (3). That composition is only correct in the joint probability space over calibration and future disturbances. When you deploy the controller after seeing a specific calibration set, the conditional coverage is a random variable that can be substantially below 1−θ. The authors mention this in Remark 1 and point to the PAC option, but they don't use it in the main theorem. If you read the abstract literally—'reliable probabilistic guarantees based on marginal coverage'—it's accurate. But the chance constraints in (3) are usually interpreted as a property of the realized policy. The paper should either state Theorem 1 as a marginal guarantee or add the PAC tightening to get a high-confidence conditional bound. This is not a mathematical error, but it is a precision problem that could mislead a reader.\n\nThe numerical study is a single double integrator, which is fine for an L-CSS paper but doesn't stress-test the marginal-versus-conditional gap. The comparison with scenario optimization is fair on sample complexity. The S-procedure derivation is standard and correct, and the citation pattern is appropriate.\n\nOverall, this is a solid, useful paper. I'd send it to peer review with a request to clarify the probabilistic interpretation and to soften the direct method's claims. It deserves a serious referee.","headline":"A clean distribution-free method for joint chance constraints in linear control, but the main guarantee is marginal over calibration data, not a certified bound for the deployed calibration set.","tokens_in":668,"tokens_out":823,"would_cite":true,"duration_ms":47648,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93E20","93C05","62G15"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that conformally calibrated prediction regions can replace distribution knowledge in linear stochastic optimal control, converting a chance-constrained problem into a deterministic one with guaranteed coverage.","keywords":["conformal prediction","chance-constrained optimal control","stochastic systems","distribution-free uncertainty quantification","prediction regions","linear systems","joint chance constraints","tube-based control"],"falsifier":"Exhibit a linear system and i.i.d. calibration data for which the quantile radii of Lemma 3 satisfy the nonemptiness condition of Lemma 2, but the resulting closed-loop trajectory violates the state or input chance constraint with empirical frequency above $\\theta$ over $10^4$ test disturbances; if such a case exists, Theorem 1's guarantee is false.","tokens_in":12625,"feed_emoji":"🎛️","tokens_out":5808,"duration_ms":58789,"temperature":0.7,"pith_summary":"The paper addresses optimal control of linear stochastic systems whose noise distribution is unknown, with joint chance constraints on states and inputs. It proves that if one can construct prediction regions for the closed-loop error trajectories with marginal coverage at least $1-\\theta$, then solving a tightened deterministic problem yields a controller whose constraint violations occur with probability at most $\\theta$. Two constructive methods are given: a direct one that trains a feedback gain on disturbance samples and calibrates error-region radii, and an indirect one that first bounds disturbances by an ellipsoid from conformal prediction and then synthesizes an invariant ellipsoid through the S-procedure. The probabilistic guarantee is distribution-free and independent of the number of calibration samples.","feed_headline":"Chance constraints hold by conformal prediction, no noise model needed","feed_subtitle":"Two data-driven methods turn unknown-disturbance control into a deterministic problem, keeping violations under the chosen risk.","key_machinery":"The central machinery is a two-step train-calibrate split together with nonconformity scores defined over error trajectories. In the direct method, the scores are $R_e = \\max_t \\|e(t)\\|$ and $R_u = \\max_t \\|Ke(t)\\|$; a training set selects the feedback gain $K$ by quantile constraints, and a separate calibration set produces the radii $C_e$ and $C_{Ke}$ as empirical quantiles, yielding ball-shaped prediction regions $B(C_e)$ and $B(C_{Ke})$ with marginal coverage. In the indirect method, a minimum-volume ellipsoid enclosing training disturbances is combined with a conformal quantile to obtain an ellipsoidal disturbance prediction region $\\mathcal{W}$, and the S-procedure converts the robust-invariance condition into an LMI/BMI that is solved for an ellipsoidal prediction region $\\mathcal{E}$ for the error state. In both methods, Lemma 2 ensures the tightened constraint sets are nonempty, which makes Theorem 1 applicable.","core_discovery":"On its own terms, the paper's discovery is that the intractable chance-constrained optimal control problem (3) with unknown noise distribution can be replaced by the deterministic problem (7), provided one has prediction regions $E^{1:N}_{1-\\theta}(e(t))$ and $E^{0:N-1}_{1-\\theta}(Ke(t))$ that cover the error process and the feedback term with marginal probability at least $1-\\theta$. The key reduction is Theorem 1: whenever the prediction regions fit strictly inside the constraint ellipsoids in the sense of Lemma 2, any optimal solution of (7) defines a policy $u(t)=Ke(t)+v(t)$ whose trajectory satisfies the joint state and input chance constraints with probability at least $1-\\theta$. The paper then shows two ways to build such regions from data without knowing the distribution, both ending in a conformal quantile step that gives the coverage.","pith_inferences":["Editorial inference: the coverage proof only needs the calibration error trajectories to be i.i.d., so the train-calibrate scheme could in principle be lifted to nonlinear simulation models, with a different, model-specific method for gain synthesis.","Editorial inference: because the guarantee is marginal, a fixed calibration set can yield a conditional violation probability above $\\theta$; applying a PAC-type adjustment would convert the claim into a probably-approximately-correct one at the price of a larger quantile.","Editorial inference: the direct method's training problem is nonconvex in $K$, and the conformal guarantee starts only after $K$ is fixed, so the paper does not certify that the numerically found gain is near-optimal, only that the final policy meets the chance constraints.","Editorial inference: infinity-norm nonconformity scores would produce box-shaped prediction regions that fit polyhedral constraints more tightly than balls; the paper mentions this possibility but does not develop it, and testing it on the double-integrator example would quantify the conservatism gap."],"forward_implications":["Any feasible solution of the tightened problem (7) is a feasible solution of the original chance-constrained problem (3), so the probabilistic constraints are certified with probability at least $1-\\theta$ without knowing the disturbance distribution.","The guarantee holds with any number of calibration samples; unlike scenario optimization, no minimum scenario count that grows with horizon and confidence is required because the calibration step is separate from gain synthesis.","With the feedback gain fixed, the tightened problem is convex and, in the paper's example, solves in under 0.02 seconds, making the approach suitable for online re-planning.","Both the direct and indirect methods produce prediction regions that satisfy Lemma 2 when their radii stay below the smallest semi-axis of the constraint ellipsoids, so the two methods can be combined or used independently.","Constraint tightening based on maximum-over-time nonconformity scores preserves the joint nature of the chance constraints, so simultaneous satisfaction across the whole horizon is what is certified, not just per-step margins."],"supporting_citations":[{"why":"Supplies the exchangeability lemma (Lemma 1) that gives the marginal coverage guarantee for the empirical quantile used to certify the prediction regions.","marker":"[21]"},{"why":"Motivates the adjusted confidence level in (11) for split-conformal calibration, which the direct method uses when training the feedback gain.","marker":"[19]"},{"why":"Provides the state-error decomposition and tube-based constraint tightening that Theorem 1 extends to data-driven prediction regions.","marker":"[3]"},{"why":"Gives a scenario-based constraint-tightening formulation that this paper generalizes by replacing scenario bounds with conformal prediction regions.","marker":"[8]"},{"why":"Supplies the convex minimum-volume ellipsoid formulation (13) and the S-procedure used in the indirect method.","marker":"[23]"},{"why":"Defines the robust controlled invariance concept used to ensure the ellipsoidal prediction region remains invariant under disturbances in $\\mathcal{W}$.","marker":"[24]"},{"why":"Justifies the S-lemma step that converts the disturbance-containment condition (18) into the BMI (19) in the indirect method.","marker":"[25]"},{"why":"Serves as the scenario-optimization stochastic MPC baseline that the numerical comparison targets, where the required scenario count motivates the proposed approach.","marker":"[7]"}],"fun_headline_variants":["Conformal prediction makes control design distribution-free","Unknown noise? Conformal prediction guarantees chance constraints","Distribution-free optimal control via conformal prediction","Chance constraints enforced without knowing noise distribution","Conformal prediction: control under unknown noise, reliably safe"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the calibration data are i.i.d. draws from the same disturbance distribution that will act on the system, and that the computed prediction regions are small enough to leave the tightened constraint sets nonempty; conformal prediction guarantees the coverage but not the size, so there is no a priori guarantee that a given dataset yields a feasible design.","fun_headline_variants_meta":{"raw":{"variants":["Conformal prediction makes control design distribution-free","Unknown noise? Conformal prediction guarantees chance constraints","Distribution-free optimal control via conformal prediction","Chance constraints enforced without knowing noise distribution","Conformal prediction: control under unknown noise, reliably safe"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000261,"raw_usage":{"total_tokens":1542,"prompt_tokens":845,"completion_tokens":697,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":627}},"tokens_in":461,"tokens_out":697,"duration_ms":52768,"temperature":1.0,"reasoning_tokens":627,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:30:30.239612+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Exhibit a linear system and i.i.d. calibration data for which the quantile radii of Lemma 3 satisfy the nonemptiness condition of Lemma 2, but the resulting closed-loop trajectory violates the state or input chance constraint with empirical frequency above $\\theta$ over $10^4$ test disturbances; if such a case exists, Theorem 1's guarantee is false.","supporting_citations":[{"cited_title":"Conformal prediction under covariate shift,","cited_arxiv_id":null,"evidence_quote":"Supplies the exchangeability lemma (Lemma 1) that gives the marginal coverage guarantee for the empirical quantile used to certify the prediction regions."},{"cited_title":"Stochastic tubes in model predictive control with probabilistic const raints,","cited_arxiv_id":null,"evidence_quote":"Provides the state-error decomposition and tube-based constraint tightening that Theorem 1 extends to data-driven prediction regions."},{"cited_title":"Con straint- tightening and stability in stochastic model predictive co ntrol,","cited_arxiv_id":null,"evidence_quote":"Gives a scenario-based constraint-tightening formulation that this paper generalizes by replacing scenario bounds with conformal prediction regions."},{"cited_title":"A survey of the S-lemma,","cited_arxiv_id":null,"evidence_quote":"Justifies the S-lemma step that converts the disturbance-containment condition (18) into the BMI (19) in the indirect method."},{"cited_title":"A randomized ap proach to stochastic model predictive control,","cited_arxiv_id":null,"evidence_quote":"Serves as the scenario-optimization stochastic MPC baseline that the numerical comparison targets, where the required scenario count motivates the proposed approach."}],"review_version":1}