{"id":"9c940aea-df03-49aa-9559-42c1f7fafc4e","arxiv_id":"2501.18039","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"OGD-BZC is a projected online gradient descent controller with a buffer zone that keeps convex safety constraints satisfied at all times under adversarial disturbances and achieves O~(sqrt(T)) regret against the best safe linear policy.","lead":"This paper gives an online controller for linear systems that must obey convex state and input limits at every step, even when disturbances and costs are adversarial. It proves the controller always stays safe and has regret that grows like the square root of the time horizon, an improvement over earlier convex-safe schemes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's buffer constants are inconsistent with the appendix bounds, so the stated parameter choices do not provably yield the claimed safety guarantee.","rationale":"The reader's verdict is CONDITIONAL with the weakest assumption listed as Assumption 2. I agree that Assumption 2 is important, but it is an explicitly acknowledged feasibility assumption. A more immediate and concrete gap is the inconsistency between Theorem 1's stated ϵ1, ϵ2, ϵ3 and the bounds actually derived in the appendix. This is precisely the reason the safety theorem 'as stated is not fully proven,' as the reader's own rationale notes, though it was not placed in the weakest_assumption field. Since correcting constants likely restores the theorem, the appropriate verdict remains CONDITIONAL rather than ACCEPT or REJECT. The central asymptotic claim O~(√T) appears robust to these constant corrections; the issue is a proof-discipline defect in the statement of the main safety result.","tokens_in":25137,"tokens_out":19367,"duration_ms":221106,"concrete_test":"Re-derive Theorem 1 using the appendix's actual bounds from Lemma 5 (ϵ1 = c1√mn H(1−γ)^H) and Lemma 6 (ϵ2 as derived, with √(m^2n^3)H^2η). Then, for a concrete instance (e.g., n=2, m=1, γ=0.5, κ=2, w=0.3) compute H and η from Corollary 1 and check whether ϵ1+ϵ2 ≤ ϵ* − ϵ1 − ϵ3 holds with the appendix formulas. If the margin is negative while the theorem's formulas give a positive margin, the stated parameter selection is unsafe.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The safety theorem's parameter conditions do not match the supporting lemmas. Theorem 1 states ϵ1(H) := c1 H(1−γ)^H, but Lemma 5 proves ∥x_t−\\tilde{x}_t∥∞ ≤ c1√mn H(1−γ)^H, an extra √mn factor. Similarly, Theorem 1 defines ϵ2(η,H) := c2√(mn^3)ηH^2, while Lemma 6's derivation yields a bound with different m,n exponents (apparently √(m^2n^3) or worse). Because Corollary 1 chooses H, η, and ϵ based on these stated formulas, an implementer following the paper may select parameters that fail to satisfy the actual buffer inequality ϵ1+ϵ2 ≤ ϵ ≤ ϵ*−ϵ1−ϵ3. The safety proof's key containment (X_ϵ)_{−ϵ1−ϵ2} ⊆ X then does not close, so deterministic safety under adversarial disturbances is not established as stated. This is load-bearing because safety is the paper's central contribution; the error is likely correctable by absorbing the missing powers into constants, but the theorem as written is not proven.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers online nonstochastic control of a known linear time-invariant system with convex state and input constraints that must hold at every time step under bounded adversarial disturbances. It proposes OGD-BZC, a projected online gradient descent algorithm over a set of disturbance-action controllers, and claims deterministic safety plus O~(√T) regret with respect to the best safe linear policy in hindsight. The algorithm uses a safe policy set Ω_ϵ defined through set shrinkage/expansion, an approximate convex cost, and a buffer zone whose size is set by Theorem 1. The regret analysis splits the regret into a performance gap, bounded by standard OCO arguments, and a policy-set gap, bounded using new convex-geometric lemmas. Numerical experiments on a two-dimensional double-integrator-like system illustrate safety and regret behavior.","tokens_in":25357,"tokens_out":11284,"duration_ms":121299,"significance":"If correct, the paper would be a meaningful step beyond prior work that handles affine safety constraints: it would provide a deterministic anytime safety guarantee for general convex constraints under adversarial disturbances while retaining a sublinear regret rate. The strategy of using shrinkage/expansion of convex sets to define the safe policy set is a genuinely new analytical ingredient relative to [3], and the derivation is essentially self-contained, with explicit constants and no fitted parameters. The paper also makes concrete, falsifiable predictions: deterministic constraint satisfaction for all disturbance sequences and an O~(√T) regret bound for the stated parameter choices. These strengths make the results worth pursuing, but the mismatches described below currently prevent the stated theorems from being accepted as proven.","major_comments":[{"comment":"The definition ϵ1(H) := c1 H(1−γ)^H in Theorem 1 is not the quantity proven in Lemma 5, which establishes ∥xt−˜xt∥∞ ≤ c1√(mn) H(1−γ)^H with the constant c1 defined inside the lemma. Since the safety proof closes the containment only through the condition ϵ1(H)+ϵ2(η,H) ≤ ϵ ≤ ϵ*−ϵ1(H)−ϵ3(H), using the smaller Theorem-1 value of ϵ1 can select a buffer ϵ that does not actually compensate for the approximation error, so Theorem 1 as stated is not proven. The mismatch is local and likely fixable by absorbing √(mn) into the constant or into the stated ϵ1, but Corollary 1's choices of H, η, and ϵ must then be re-derived.","section":"§V-A, Theorem 1; Appendix B, Lemma 5"},{"comment":"Theorem 1 states ϵ2(η,H) := c2√(mn^3)ηH^2, but Lemma 6's derivation yields ∥hx(M_{t−H+1:t})−˚hx(M_t)∥_∞ w ≤ O(√(mn^3))H^2η for the state block and ∥hu(M_{t−H:t})−˚hu(M_t)∥_∞ w ≤ O(mn)H^2η for the input block, and the lemma sets ϵ2 with a √(m^2 n^3) factor, i.e., m n^{3/2}. The stated ϵ2 is therefore smaller by a factor of √m than the guaranteed bound, so the containment (hx−˚hx)·B̄(w) ⊆ B̄(ϵ2) used in the proof of Theorem 1 is not justified with the theorem's parameters. This also propagates to Corollary 1: with the corrected ϵ2, the term ϵT√(m^2n^2)H^3/ϵ* in Theorem 2 becomes O~(m^2 n^{3/2}√T) rather than O~(m^{3/2}n^{3/2}√T), so the stated exponent in Corollary 1 is not established as written.","section":"§V-A, Theorem 1; Appendix B, Lemma 6"},{"comment":"The displayed regret bound in Theorem 2 does not transparently follow from the stated lemmas with the stated exponents. In the proof of Lemma 4, Part i gives O(T mn H^2(1−γ)^H), not the O(T√(m^2 n^3)H^2(1−γ)^H) term displayed in Theorem 2. Lemma 1 gives O(T√(m^2 n^2)H^3(ϵ1+ϵ3+ϵ)/ϵ*), which with ϵ1(H)=c1H(1−γ)^H yields T mn H^4(1−γ)^H/ϵ*, not the term T√(m^2 n^3)H^5(1−γ)^H/ϵ* displayed in Theorem 2. These differences change the H-dependence and the m,n exponents in the final bound; although the O~(√T) rate may survive after a careful re-derivation, the theorem as printed is not a consequence of the supplied lemmas and should be corrected.","section":"§V-B, Theorem 2, Lemmas 1, 3, and 4"}],"minor_comments":[{"comment":"There are small typos: \"indicater\" in the Notations paragraph, and the class K is written as K ⊆ R^{n×m} while the controller ut = −Kxt requires K ∈ R^{m×n}.","section":"§I-A and §II"},{"comment":"Definition 3 defines ϵ-strict safety only for ϵ > 0, but Corollary 2 states the result for \"ϵ0 ≤ 0\" and later applies it with ϵ0 = 0; please clarify the convention for non-positive buffer parameters.","section":"Definition 3 and Corollary 2"},{"comment":"The symbol w̄ appears in the display for Gf without definition; it appears to mean the disturbance bound w introduced in Section II.","section":"Appendix D, Lemma 10"},{"comment":"The numerical experiment parameters H=⌊log T⌋, ϵ=log(T)/√T, and η=1/(√T log T) are not the choices prescribed by Corollary 1, and for T=30 the condition H ≥ log(2κ²)/log((1−γ)^{-1}) may fail; please state explicitly whether the simulation is intended as a heuristic illustration or as a realization of the theorem's parameter selection.","section":"§VI"},{"comment":"The phrase \"the hidden constant coefficients is the polynomial\" should be \"the hidden constants are polynomials\", and \"This finishs the proof\" in Lemma 4 should be corrected to \"finishes\".","section":"§V-B, Theorem 2"}],"recommendation":"major_revision","confidential_remarks":"The mismatches in Theorem 1 appear to be typographical/index errors rather than a fundamental flaw in the proof strategy, because the appendix's own lemmas provide larger bounds and the set-shrinkage argument is coherent. However, the displayed exponents in Theorem 2 and Corollary 1 need a full re-derivation after the constants are corrected, and the safety guarantee as written is currently unproven. Given the novelty of the convex buffer-zone analysis and the likely fixability, I recommend a major revision rather than rejection. The paper should also recheck whether the claimed O~(m^{3/2}n^{3/2}√T) bound survives the corrected definitions or whether a modestly worse polynomial dependence in m,n is the honest statement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper before citing it: the central idea is good, and the result is probably true, but Theorem 1 as written is not proven. The authors extend the affine-constraint result of [3] to general convex safety sets using a buffer-zone defined by set shrinkage and expansion, and they achieve deterministic safety with O~(sqrt(T)) regret that does not depend on the number of constraints. That is a genuine gap in the literature, and the convex-analysis machinery in Lemma 2 is the right tool for it. The paper is also honest: it cites the relevant prior work, acknowledges in footnote 3 that Assumption 2 is stronger than the finite-horizon strict-safety assumption in [3], and does not hide the missing projection implementation.\n\nThe soft spot is load-bearing but likely fixable. Theorem 1 states eps1(H) = c1 H (1-gamma)^H and eps2(eta,H) = c2 sqrt(m n^3) eta H^2. The appendix proves different bounds: Lemma 5 gives an extra sqrt(mn) factor in eps1, and Lemma 6 gives an extra sqrt(m) factor in eps2. Since Corollary 1 chooses H, eta, and epsilon using the stated formulas, an implementer following the paper may select parameters that violate the actual buffer inequality eps1 + eps2 <= epsilon <= eps* - eps1 - eps3. The safety proof's containment (X_eps)_{-eps1-eps2} subset X then does not close. This is not a cosmetic typo; it affects the paper's main claim. My read is that all the missing powers can be absorbed into the constants without changing the regret rate, so the result survives, but the theorem needs to be restated and re-proven before the paper is dependable.\n\nOther concerns are minor. The projection oracle onto Omega_eps is never specified, and the numerical section is a two-dimensional toy with T=30. Neither undermines the theory. The regret analysis follows the standard OGD/DAC program, but that is the correct framework, and the constraint-count-free bound is a genuine improvement.\n\nWho should read this: anyone working on constrained online nonstochastic control, and anyone who needs a concrete example of how set operations can replace affine constraint structure. It deserves a serious referee, but only after the authors correct the Theorem 1 constants, reconcile them with the appendix, and say something about how the projection is computed.","headline":"A real and likely correct extension of online nonstochastic control to convex safety constraints, but Theorem 1 as stated is not proven because the constants in the statement do not match the appendix bounds.","tokens_in":25889,"tokens_out":1808,"would_cite":true,"duration_ms":23588,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C05","90C25"],"pacs":[],"model":"deepseek-v4-flash","headline":"A single online gradient-descent controller can enforce general convex safety constraints at every time step under adversarial bounded disturbances while still achieving $\\tilde{O}(\\sqrt{T})$ regret against the best safe linear policy in…","keywords":["online nonstochastic control","convex safety constraints","disturbance-action control","regret bounds","projected online gradient descent","buffer zone","linear time-invariant systems","deterministic safety"],"falsifier":"One concrete adversarial test: for an LTI instance satisfying Assumptions 1-2 and a parameter choice satisfying Theorem 1, compute the worst-case disturbance sequence that maximizes, over $t$, the distance of the OGD-BZC state from the boundary of $X$ and the input from the boundary of $U$. If any bounded disturbance sequence drives the state out of $X$ or the input out of $U$, the safety theorem is false.","tokens_in":94,"feed_emoji":"🛡️","tokens_out":9464,"duration_ms":153016,"temperature":0.7,"pith_summary":"The paper addresses online control of a linear time-invariant system whose state and input must stay inside fixed convex sets at every time step, while disturbances and costs are chosen adversarially. It proposes OGD-BZC, a projected online gradient-descent algorithm over the class of disturbance-action controllers, and proves that with suitable memory, buffer size, and step size, the system remains safe under every bounded disturbance sequence. The regret of OGD-BZC versus the best safe linear policy in hindsight is shown to be $\\tilde{O}(\\sqrt{T})$, the same order as the unconstrained online nonstochastic control benchmark. If correct, this means general convex safety constraints can be enforced deterministically without sacrificing the sublinear regret rate, and without the regret depending on the number of constraints.","feed_headline":"One algorithm enforces convex safety every step at ~√T regret","feed_subtitle":"A single projected gradient rule obeys general convex state and input constraints at all times under adversarial noise.","key_machinery":"Disturbance-action controllers with a safe-policy set defined by set shrinkage and expansion. The safe set is $\\Omega_\\epsilon = \\{M \\in \\mathcal{M}: \\mathring{h}_x(M) \\bar{B}(w) \\subseteq X_\\epsilon \\text{ and } \\mathring{h}_u(M) \\bar{B}(w) \\subseteq U_\\epsilon\\}$, where $X_\\epsilon$ and $U_\\epsilon$ are the $\\epsilon$-shrunk constraint sets and $\\mathring{h}_x, \\mathring{h}_u$ map a fixed policy to worst-case reachable sets for state and input under bounded disturbances. Projecting the online gradient step onto $\\Omega_\\epsilon$ keeps the actual state and input inside $X$ and $U$ once $\\epsilon$ is chosen to cover the exponentially decaying horizon-truncation error and the slowly-changing-policy approximation error. Lemma 2, which contains $\\Omega_\\epsilon$ between two convex interpolations of the policy set and the strictly safe linear policy $M(K_{ss})$, is the mechanism that turns convexity into distance bounds and removes the constraint-count dependence.","core_discovery":"The central claim is that a single online algorithm, OGD-BZC, can satisfy general closed convex state and input constraints at all times for an LTI system under adversarial bounded disturbances, and still match the usual unconstrained regret bound. The algorithm maintains a disturbance-action controller and runs projected online gradient descent on an approximate convex cost, projecting at each step onto a set $\\Omega_\\epsilon$ of policies that keep the surrogate state and input inside the $\\epsilon$-shrunk safety sets for all possible disturbances. By choosing the memory $H$, buffer size $\\epsilon$, and step size $\\eta$ according to the theorem conditions, the paper proves deterministic safety and, for sufficiently large $T$, regret bounded by $\\tilde{O}(m^{1.5} n^{1.5} \\sqrt{T})$ in Corollary 1. The proof avoids the affine structure used in earlier work by bounding the gap between the safe policy set and the best linear policy through convex set-containment lemmas, so the regret does not scale with the number of constraints.","pith_inferences":["Because the regret rate matches the unconstrained setting, the paper suggests that the price of deterministic convex safety is hidden in constants and logarithmic factors rather than in the $T$-dependence; a natural next question is whether similar $\\tilde{O}(\\sqrt{T})$ guarantees hold for time-varying or partially observable systems, where the strict-margin assumption is harder to satisfy.","The buffer construction relies on knowing the strict safety margin $\\epsilon^*$ of some linear policy, which may be unknown in practice; an adaptive scheme that estimates $\\epsilon^*$ online or uses conservative bounds on $H$ and $\\eta$ would make the algorithm more directly implementable.","The projection step onto $\\Omega_\\epsilon$ may be computationally heavy for high-dimensional convex constraints, a point the paper itself flags as future work; an approximate projection that preserves membership up to a smaller buffer might keep deterministic safety while reducing per-step cost.","The numerical regret can be negative relative to the best safe linear policy because disturbance-action controllers form a larger policy class, which suggests that measuring regret against the best safe DAC rather than the best linear policy would be a harder and possibly more informative benchmark."],"forward_implications":["For any LTI system satisfying the strict safe linear policy assumption, OGD-BZC keeps $x_t$ in $X$ and $u_t$ in $U$ for every $t$ and every bounded disturbance sequence, so safety is deterministic rather than probabilistic or cumulative.","The regret bound is $\\tilde{O}(\\sqrt{T})$ under the Corollary 1 parameter choice, matching the unconstrained convex-cost rate up to log factors and dimension and system constants.","The result covers general closed convex constraints, including nonlinear constraint sets, and the regret expression has no factor counting the number of constraints, improving on the affine-only predecessor.","For sufficiently large $T$, the theorem's inequalities leave room to choose $H$, $\\epsilon$, and $\\eta$ whenever an $\\epsilon^*$-strictly safe linear policy exists, so the safety and regret guarantees are simultaneously attainable."],"supporting_citations":[{"why":"Supplies the disturbance-action controller framework, the state/control decomposition in Proposition 1, and the reduction to online convex optimization with memory that underlies the algorithm.","marker":"[1]"},{"why":"Provides the standard OGD analysis and online nonstochastic control background used for the performance-gap bound in Lemma 4.","marker":"[2]"},{"why":"Establishes the affine-constraint buffer-zone approach and the policy-mapping lemmas that OGD-BZC generalizes to convex constraints.","marker":"[3]"},{"why":"Defines the shrinkage and expansion operations used to build the convex buffer zone and the safe policy set.","marker":"[27]"},{"why":"Provides the convex-analysis facts, including support-function and Minkowski-operation properties, used in Proposition 2 and Lemma 2.","marker":"[28]"}],"fun_headline_variants":["Convex safety at every step, regret still √T","One controller meets all convex limits, √T regret","Always-safe convex control, √T regret under attack","Never violate convex constraints, get √T regret"],"cache_read_input_tokens":28032,"weakest_assumption_plain":"There exists a fixed linear controller that, under every possible disturbance sequence, keeps the state and input strictly inside the constraints by at least a fixed positive margin for the entire infinite horizon.","fun_headline_variants_meta":{"raw":{"variants":["Convex safety at every step, regret still √T","One controller meets all convex limits, √T regret","Always-safe convex control, √T regret under attack","Never violate convex constraints, get √T regret"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000521,"raw_usage":{"total_tokens":2484,"prompt_tokens":871,"completion_tokens":1613,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":1550}},"tokens_in":487,"tokens_out":1613,"duration_ms":16803,"temperature":1.0,"reasoning_tokens":1550,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T00:54:02.861292+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One concrete adversarial test: for an LTI instance satisfying Assumptions 1-2 and a parameter choice satisfying Theorem 1, compute the worst-case disturbance sequence that maximizes, over $t$, the distance of the OGD-BZC state from the boundary of $X$ and the input from the boundary of $U$. If any bounded disturbance sequence drives the state out of $X$ or the input out of $U$, the safety theorem is false.","supporting_citations":[{"cited_title":"Online control with adversarial disturbances,","cited_arxiv_id":null,"evidence_quote":"Supplies the disturbance-action controller framework, the state/control decomposition in Proposition 1, and the reduction to online convex optimization with memory that underlies the algorithm."},{"cited_title":"Online optimal control with affine constraints,","cited_arxiv_id":null,"evidence_quote":"Establishes the affine-constraint buffer-zone approach and the policy-mapping lemmas that OGD-BZC generalizes to convex constraints."},{"cited_title":"The impact of the geo- metric properties of the constraint set in safe optimization with bandit feedback,","cited_arxiv_id":null,"evidence_quote":"Defines the shrinkage and expansion operations used to build the convex buffer zone and the safe policy set."},{"cited_title":"Schneider, Convex bodies: the Brunn-Minkowski theory , second expanded edition ed., ser","cited_arxiv_id":null,"evidence_quote":"Provides the convex-analysis facts, including support-function and Minkowski-operation properties, used in Proposition 2 and Lemma 2."}],"review_version":1}