{"id":"b904c5c8-038c-43e4-ab4f-3e4b275ddc0f","arxiv_id":"2412.00922","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An OCO-plus-reference-governor framework gives constrained nonlinear control systems a dynamic regret bound linear in the optimizer's regret and path length, and linear only in cost variation when the optimizer converges Q-linearly.","lead":"This paper builds a modular controller for nonlinear systems with unknown, time-varying costs and hard safety constraints, combining an online optimizer with a reference governor. It proves regret bounds that scale linearly with the optimizer's regret and path length, and shows the linear dependence on the optimizer regret cannot be improved.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 2's lower-bound proof is vacuous: the adversary's cost functions make η_t = rt, so ROCO_T = 0 and the claimed optimality of linear dependence on OCO regret is unproven.","rationale":"The reader's concern about Assumption 5 is legitimate: the paper asserts, without proof, that the scalar RG and almost all RGs satisfy condition (7), and Lemma 3 and Theorem 1 need that property. That is a real applicability gap. However, the more load-bearing issue is Proposition 2's vacuous proof, because it directly underpins the abstract's optimality claim. A vacuous lower bound is an internal correctness issue, not merely an unproven hypothesis about existing RGs. The main regret bound (Theorem 1, Corollary 1) is conditional and may survive, so I keep the CONDITIONAL verdict but add a condition that Proposition 2 must be repaired or the optimality claim softened. Agreement with the reader is partial: we agree Assumption 5 is a gap, but the primary weakness is different.","tokens_in":24382,"tokens_out":17931,"duration_ms":163396,"concrete_test":"Recompute ROCO_T for the cost functions defined in Appendix F; verify that Ls_t(rt)=0 and η_t=rt, giving ROCO_T=0 for all t and reducing Proposition 2's inequality to 0 ≤ RT. To test whether the optimality claim can be salvaged, fix any η* in S_v, set Lt(x,u)=||x-h(η*)||^2+||u-g(h(η*),η*)||^2, and check whether RT ≥ c·ROCO_T holds for some universal c>0 along the closed-loop trajectory; if no such c exists, the claim should be withdrawn or reproven.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 2 (Section 4) purports to show that the linear dependence of RT on ROCO_T in Theorem 1 cannot be improved. The proof in Appendix F chooses Lt(x,u)=||(x-h(rt), u-g(h(rt),rt))||^2. Then the steady-state cost is Ls_t(v)=||(h(v)-h(rt), g(h(v),v)-g(h(rt),rt))||^2, which attains its minimum 0 at v=rt. By Definition 1, η_t=rt and hence ROCO_T = Σ_t (Ls_t(rt)-Ls_t(η_t)) = 0 for every t. The displayed inequality RT ≥ ROCO_T therefore reduces to RT ≥ 0; it is a tautology and supplies no lower bound with positive OCO regret. Consequently, the conclusion that a linear bound is optimal, i.e., cannot be improved upon, does not follow from the proof. The flaw is internal: the adversary defines the benchmark to be the algorithm's own output, forcing the OCO regret to zero. If one instead fixes a nontrivial benchmark η*, the term Σ(Lt(xt,ut)-Ls_t(rt)) is no longer guaranteed nonnegative, so the claimed inequality RT ≥ ROCO_T is not established by this argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a modular online convex optimization (OCO) framework for constrained nonlinear discrete-time systems with time-varying unknown costs: an OCO algorithm generates a reference, a reference governor (RG) enforces state and input constraints, and a stabilizing feedback tracks the reference. The main result (Theorem 1, Eq. (22)) bounds the closed-loop dynamic regret by a constant plus the OCO algorithm's dynamic regret and path length, under Lipschitz regularity, exponential stability, and an abstract RG progress condition (Assumption 5). Corollary 1 specializes to Q-linearly convergent OCO algorithms and gives a bound linear in the variation of the optimal references. Proposition 2 claims that the linear dependence on the OCO regret cannot be improved. A numerical CSTR example compares four OCO/RG combinations and reports normalized regret and computation times.","tokens_in":24684,"tokens_out":13351,"duration_ms":124351,"significance":"The modular separation of OCO from reference governing is conceptually useful and broader than existing linear-system OCO-control results: it provides per-step constraint satisfaction for nonlinear systems while retaining a regret bound expressed in terms of the chosen OCO algorithm. The proof structure in Lemmas 1-3 and Theorem 1 is transparent, and the decomposition into transient, RG-tracking, and OCO-regret terms is a genuine strength. The numerical study with code availability also helps the reader assess the design choices. However, two issues affect the advertised claims: the proof of the optimality statement (Proposition 2) is vacuous as written, and Assumption 5 is an existence assertion whose satisfaction by the exemplar RGs is not demonstrated. With Proposition 2 repaired or removed and Assumption 5 made constructive, the remaining framework is a solid contribution.","major_comments":[{"comment":"The lower-bound proof is vacuous. The adversary chooses L_t(x,u)=||(x-h(r_t), u-g(h(r_t),r_t))||^2, so the steady-state cost L^s_t(v)=||(h(v)-h(r_t), g(h(v),v)-g(h(r_t),r_t))||^2 attains its minimum at v=r_t. Hence η_t=r_t by Definition 1 and R^OCO_T=Σ(L^s_t(r_t)-L^s_t(η_t))=0 for every t. The displayed inequality R_T≥R^OCO_T then reduces to R_T≥0, which is already true for these nonnegative costs. This construction supplies no lower bound with positive OCO regret, and therefore does not establish the claimed optimality of a linear dependence on R^OCO_T. A valid lower bound must exhibit an adversary with large R^OCO_T and compare R_T to a positive multiple of R^OCO_T; as written, the proof forces the benchmark to equal the algorithm's own output.","section":"Section 4, Proposition 2 (Appendix F)"},{"comment":"Assumption 5 is load-bearing: Lemma 3 and Theorem 1 require existence of a class-K function ρ, an integer M, and ε>0 satisfying (20). The paper asserts without proof that the scalar RG (8) and 'almost all RGs' satisfy the progress condition (7), but no ρ is constructed for the scalar RG, the command governor, or any concrete class of RGs. This is not a purely technical formality: Remark 1 explicitly lists command governors with weighting that violate the stated condition (7), so Assumption 5 is not automatically satisfied by standard designs. Please provide a proof for at least the scalar RG (8) and the standard command governor, or replace Assumption 5 with constructive sufficient conditions that can be checked for a given RG. Without this, the main regret bound is conditional on an unverified abstract property.","section":"Section 3.2, Assumption 5 and Lemma 3"},{"comment":"The text states that Q-linear convergence (14) was 'verified numerically' with κ≈0.986 for all q_t∈[50,250] and c_t∈[0.25,0.65]. A finite simulation cannot verify a property over continuous parameter ranges, and no proof is supplied. Since the numerical section uses this to claim that the assumptions of Corollary 1 are satisfied, this claim should be rephrased as empirical evidence for the specific tested instances, or supported by a theoretical verification using the structure of the example.","section":"Section 5.4, numerical verification of Q-linear convergence"}],"minor_comments":[{"comment":"The definition α_t=ρ^{-1}(ε) presumes that ε lies in the range of ρ, but Assumption 5 only states ρ:R_{\\ge0}\\to[0,1] is class K, which does not imply surjectivity. Choose ε from the range of ρ, or define α_t as any value satisfying ρ(α_t)=ε.","section":"Lemma 3, Eq. (21)"},{"comment":"The claim that replacing (7) with the R-weighted inequality 'only leads to additional constant factors' is not demonstrated; a brief derivation or a reference would make the modification precise.","section":"Remark 1"},{"comment":"The cross-section notation switches between O_x(v) and O_v(x) for the same object; standardizing the notation would improve readability.","section":"Notation, Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The main regret decomposition and Theorem 1 appear sound, and I would not reject the paper on those grounds. The serious issue is Proposition 2: as written, the lower-bound proof is a tautology and does not support the abstract's optimality claim. The authors should either provide a genuine lower-bound construction with positive OCO regret or rescope the claim. Assumption 5 also needs concrete verification for the presented RGs before the framework can be regarded as generally applicable. These are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe modular OCO-plus-reference-governor framework is the real contribution here. Theorem 1 is a genuine composition result: per-step constraint satisfaction plus dynamic regret linear in the OCO algorithm's regret and path length, for nonlinear systems under time-varying unknown costs. The proof structure (Lemmas 1-3) is clean and the numerical comparison is concrete and reproducible. That part deserves a serious look.\n\nThe soft spot is Proposition 2. The stress-test note is correct: the adversary in the proof chooses L_t(x,u) = ||(x-h(r_t), u-g(h(r_t),r_t))||^2, which makes the steady-state cost zero at v = r_t, so one minimizer is η_t = r_t and R^OCO_T = 0. The displayed inequality R_T ≥ R^OCO_T reduces to R_T ≥ 0, which is a tautology. It does not establish that linear dependence on R^OCO_T is optimal, because the constructed instance has no positive OCO regret. To prove optimality you would need an instance with R^OCO_T > 0 and R_T ≥ c R^OCO_T. The text's interpretation after the proof overclaims what the argument shows. This is a load-bearing flaw in the \"optimality\" claim, but not in Theorem 1 itself.\n\nSecond, Assumption 5 is genuinely an abstract property. The paper asserts the scalar RG (8) and \"almost all\" RGs satisfy it without constructing ρ or proving the contraction. This is an addressable gap, but as written Lemma 3 rests on an unproven existence claim. A referee should ask for a proof or a reference that establishes it for scalar RGs.\n\nMinor: the Q-linear condition for OGD is verified numerically (κ ≈ 0.986), not analytically; that is fine for a numerical study, but the paper should be explicit that the Corollary 1 bound is conditional on a numerically verified property.\n\nWho is this for? People working on OCO-based control, reference governors, and safe online optimization. The upper-bound framework is a useful modular template. It deserves peer review, but Proposition 2 needs to be fixed or demoted to a remark. I would send it out with major-revision expectations.","headline":"Solid modular OCO+RG upper-bound result for nonlinear constrained control, but the claimed optimality of linear dependence on the OCO regret rests on a vacuous lower-bound proof.","tokens_in":25195,"tokens_out":4543,"would_cite":true,"duration_ms":37254,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A modular framework combines online convex optimization with reference governors to control nonlinear systems under time-varying unknown costs, guaranteeing constraint satisfaction and a dynamic regret bound that is linear in the…","keywords":["online convex optimization","reference governor","nonlinear systems","constraint satisfaction","dynamic regret","path length","modular control"],"falsifier":"Pick a stabilizing feedback and safe-set pair satisfying Assumptions 1–4, and a reference governor that keeps the closed loop safe but whose progress $\\nu_t$ toward an infeasible desired reference is zero for some reachable state—for example, a scalar governor on a safe set whose cross-section at that state is a single point. If that governor still yields good closed-loop performance but inequality (7) fails for every class-$\\mathcal{K}$ function $\\rho$, then the proof mechanism of Lemma 3 and Theorem 1 is violated; conversely, exhibiting such a case with regret growing faster than the claimed linear-in-path-length bound would refute the theorem's reach.","tokens_in":24189,"feed_emoji":"⚙️","tokens_out":8643,"duration_ms":68497,"temperature":0.7,"pith_summary":"This paper proposes a modular control framework that separates the task of minimizing time-varying, a priori unknown cost functions from the task of keeping a nonlinear dynamical system inside its state and input constraints. An online convex optimization (OCO) algorithm proposes a reference signal based on previous costs; a reference governor modifies that reference only when needed to guarantee the closed loop stays in a safe set; a stabilizing feedback tracks the modified reference. The paper proves that the resulting closed-loop dynamic regret is bounded by a constant plus the OCO algorithm's own dynamic regret plus a term proportional to how often and how far the optimizer moves its reference. It further shows that a linear dependence on the optimizer's dynamic regret is unavoidable, and that for Q-linearly convergent optimizers the regret is linear only in the variation of the cost functions, matching known optimal bounds for linear systems. If correct, the framework yields per-time-step constraint satisfaction with a performance loss that is as small as the underlying online optimizer allows.","feed_headline":"Constrained nonlinear control gets provably tight regret bound","feed_subtitle":"A modular controller plus reference governor keeps the system safe and makes regret grow only with cost variation.","key_machinery":"The mechanism that carries the argument is the modular composition of three components: an OCO algorithm producing a desired reference $r_t$, a reference governor that outputs $v_t = \\mathrm{RG}(x_t, r_t)$ so that $(x_t, v_t)$ lies in a forward-invariant safe set $\\mathcal{O}$, and a stabilizing feedback $u_t = g(x_t, v_t)$. The load-bearing identity is the regret decomposition in the proof of Theorem 1, which splits the closed-loop cost into the OCO algorithm's steady-state regret, a transient cost proportional to $\\|x_t - h(v_t)\\|$, and a reference-tracking cost proportional to $\\|v_t - r_t\\|$. The transient term is controlled by an exponential-stability Lyapunov function (Lemma 1) and the tracking term by the governor's progress condition (Assumption 5): when $r_t$ is not yet feasible, the governor must step at least distance $\\nu_t$ toward $r_t$ and reduce the remaining distance by a factor $1 - \\rho(\\nu_t)$, with $\\rho$ a class-$\\mathcal{K}$ function. Lemma 3 turns that per-step progress into an average contraction over a fixed horizon $M$, which yields the linear dependence of the regret on the optimizer's path length.","core_discovery":"The central claim, formalized as Theorem 1, is that under mild regularity assumptions—a steady-state map, exponential stability of the stabilizing feedback, a forward-invariant safe set with positive margin around each admissible steady state, and a reference governor that always moves toward the desired reference when it is infeasible—the closed loop satisfies $(x_t, u_t) \\in \\mathcal{Z}$ for all $t$ and its dynamic regret $R_T$ satisfies $R_T \\leq c_0 + R_T^{\\mathrm{OCO}} + c_{PL} R_T^{PL}$. Here $R_T^{\\mathrm{OCO}}$ is the dynamic regret of the OCO algorithm against the sequence of optimal steady-state references and $R_T^{PL}$ is its path length; the constants depend on the Lyapunov data, the safety margin, and the governor's progress rate. Proposition 2 proves that the linear dependence on $R_T^{\\mathrm{OCO}}$ cannot be improved. For OCO algorithms that converge Q-linearly on constant costs, Corollary 1 sharpens the bound to a constant plus a linear term in the cumulative movement of the optimal references, which is known to be the best possible order. The same results recover the OCO-with-memory setting as a special case.","pith_inferences":["The proof identifies the governor's worst-case progress rate as the true safety–performance trade-off; measuring $\\rho$ for a given safe set on a given system would let a practitioner compute the regret constants before deployment.","Since Assumption 3 only needs exponential stability rather than global stability, the same modular argument should extend to locally stabilizing controllers if the safe set is chosen inside the region of attraction, giving a route to model-predictive or tube-based inner loops.","The authors' stated limitation of exact dynamics suggests a natural next test: replace the steady-state map $h$ with a set-valued uncertainty model and re-derive the regret bound as a function of the uncertainty diameter; the framework's modularity makes that perturbation analysis independent of the OCO algorithm choice.","For slowly time-varying cost functions, the bound predicts that the closed-loop cost should stay within a constant of the optimal steady-state cost; this could be checked in experiments or simulations by feeding the framework a sequence of slowly varying references and recording regret per window."],"forward_implications":["Every instance of the framework—any OCO algorithm and any reference governor satisfying Assumptions 4–5—guarantees $(x_t, u_t) \\in \\mathcal{Z}$ at each time $t$, not merely asymptotically.","The closed-loop dynamic regret scales at most linearly with the underlying OCO algorithm's dynamic regret and its path length; choosing a Q-linearly convergent algorithm makes the regret linear only in the cumulative variation of the optimal steady-state references.","No control algorithm can beat a linear dependence on the OCO algorithm's dynamic regret, so the framework's loss relative to the optimizer is the best possible order.","The same analysis covers online convex optimization with memory (switching costs), so regret guarantees for that setting follow as a special case.","Larger safe sets for the governor and faster stabilizing feedbacks shrink the constants in the regret bound, quantifying the design trade-off between safety margin and performance."],"supporting_citations":[{"why":"Supplies the survey of reference governors on which the general Assumption 5 characterization is based.","marker":"[14]"},{"why":"Establishes the explicit reference governor and Lyapunov-sublevel safe sets that satisfy the RG assumptions.","marker":"[15]"},{"why":"Provides the converse Lyapunov theorem used in Lemma 1 to derive the Lyapunov function from exponential stability.","marker":"[24]"},{"why":"Gives the optimal regret lower bound for linear systems that Corollary 1 is matched against.","marker":"[30]"},{"why":"Provides the earlier linear-systems version of the proposed framework, which this paper extends.","marker":"[41]"},{"why":"Defines online gradient descent, one of the OCO algorithms the framework is compatible with.","marker":"[54]"},{"why":"Introduces online convex optimization with memory, which Remark 3 recovers as a special case.","marker":"[3]"},{"why":"Provides the maximal output admissible set construction used as a safe set satisfying Assumption 4.","marker":"[17]"}],"fun_headline_variants":["Tight regret bound for constrained nonlinear control","Reference governors and OCO achieve optimal regret","Provably optimal regret for safe nonlinear control","Online convex optimization handles constrained nonlinear control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantees stand or fall on the requirement that whenever the optimizer's desired reference is not yet feasible, the reference governor must move a guaranteed positive step toward it and shrink the remaining gap by a factor that is bounded away from one.","fun_headline_variants_meta":{"raw":{"variants":["Tight regret bound for constrained nonlinear control","Reference governors and OCO achieve optimal regret","Provably optimal regret for safe nonlinear control","Online convex optimization handles constrained nonlinear control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000497,"raw_usage":{"total_tokens":2464,"prompt_tokens":1000,"completion_tokens":1464,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":1410}},"tokens_in":616,"tokens_out":1464,"duration_ms":10300,"temperature":1.0,"reasoning_tokens":1410,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:52:30.557518+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pick a stabilizing feedback and safe-set pair satisfying Assumptions 1–4, and a reference governor that keeps the closed loop safe but whose progress $\\nu_t$ toward an infeasible desired reference is zero for some reachable state—for example, a scalar governor on a safe set whose cross-section at that state is a single point. If that governor still yields good closed-loop performance but inequality (7) fails for every class-$\\mathcal{K}$ function $\\rho$, then the proof mechanism of Lemma 3 and Theorem 1 is violated; conversely, exhibiting such a case with regret growing faster than the claimed linear-in-path-length bound would refute the theorem's reach.","supporting_citations":[{"cited_title":"Garone, S","cited_arxiv_id":null,"evidence_quote":"Supplies the survey of reference governors on which the general Assumption 5 characterization is based."},{"cited_title":"Garone and M","cited_arxiv_id":null,"evidence_quote":"Establishes the explicit reference governor and Lyapunov-sublevel safe sets that satisfy the RG assumptions."},{"cited_title":"Jiang and Y","cited_arxiv_id":null,"evidence_quote":"Provides the converse Lyapunov theorem used in Lemma 1 to derive the Lyapunov function from exponential stability."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the optimal regret lower bound for linear systems that Corollary 1 is matched against."},{"cited_title":"Nonhoff, J","cited_arxiv_id":null,"evidence_quote":"Provides the earlier linear-systems version of the proposed framework, which this paper extends."},{"cited_title":"Zinkevich","cited_arxiv_id":null,"evidence_quote":"Defines online gradient descent, one of the OCO algorithms the framework is compatible with."},{"cited_title":"Anava, E","cited_arxiv_id":null,"evidence_quote":"Introduces online convex optimization with memory, which Remark 3 recovers as a special case."},{"cited_title":"Gilbert and K","cited_arxiv_id":null,"evidence_quote":"Provides the maximal output admissible set construction used as a safe set satisfying Assumption 4."}],"review_version":1}