{"id":"efb37176-3770-40d4-b601-67928e1fb0ed","arxiv_id":"2411.12899","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"An adaptive control barrier function with an error-bound margin that is nonincreasing and, under persistency of excitation, vanishing is derived and demonstrated on two numerical examples.","lead":"This paper proposes an adaptive safety controller that learns unknown system parameters online while keeping the system inside a safe region. Under a persistence of excitation condition, the safety margin is claimed to shrink to zero, so the controller becomes as good as if it knew the true parameters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 2(c) is false as stated: eventual full-rank Omega_k does not imply theta_k -> theta*, so the advertised vanishing-conservativeness result is unproven; a uniform PE lower bound is needed.","rationale":"The paper's central advertised feature is vanishing conservativeness under persistency of excitation, and the formal support is Proposition 2(c) (theta_k -> theta*) together with Proposition 3(d) (nu_k -> 0), used in Proposition 4(c) to conclude that the adaptive constraint converges to the ideal constraint. Reading the proof of Proposition 2(c) carefully, it infers theta_k -> theta* solely from Delta V_k < 0, which is a classic invalid inference for time-varying systems; the contraction ratio can approach 1. This is not a matter of disagreement with a community consensus but an internal correctness failure, because an explicit admissible sequence Phi_k = 2^{-k} makes Omega_k full rank for every k while the parameter error does not vanish. The safety lower-bound construction in Proposition 4(a) is correct, and the recursive error bound is a useful contribution, but the vanishing-conservativeness result, the main novelty, is unsupported. The reader's weakest_assumption identifies the same issue, and I agree with the reader's verdict of REJECT; adding a uniform PE lower bound and supplying the omitted proofs of Theorems 1 and 2 would be a plausible repair, but the current version does not establish the advertised claim.","tokens_in":13528,"tokens_out":10033,"duration_ms":94848,"concrete_test":"Run the scalar RLS recursion that satisfies the paper's assumptions: p = 1, k_n = 0, sigma_k = 1, Phi_k = 2^{-k}, theta* = 1, theta_0 = 0. The update is theta_{k+1} = (Phi_k^2 theta* + theta_k) / (1 + Phi_k^2), and the error evolves as e_{k+1} = (1 + 2^{-2k})^{-1} e_k. Since sum_{k=0}^{infinity} 2^{-2k} < infinity, the product prod_{k=0}^{infinity} (1 + 2^{-2k})^{-1} converges to a positive constant, so e_k does not approach 0. Print e_k for k = 0,...,50 and observe that it flatlines at a positive value; this directly contradicts Proposition 2(c) under its stated full-rank assumption (Omega_k = 2^{-2k} > 0 for all k) and, by extension, Proposition 4(c).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 2(c), the basis for the advertised vanishing-conservativeness result (Proposition 4(c)), asserts that if Omega_k is full rank for all k >= k_i, then the parameter error tilde_theta_k converges to 0. The proof only establishes Delta V_k < 0, i.e., V_{k+1} < V_k, and a strictly decreasing nonnegative sequence need not converge to 0. The underlying recursion is tilde_theta_{k+1} = sigma_k P_k tilde_theta_k, whose contraction ratio is rho_k = sigma_k lambda_max(P_k) = sigma_k / (sigma_k + lambda_min(Omega_k)) < 1 for each k. Full rank gives no uniform bound rho_k <= c < 1; if lambda_min(Omega_k) goes to 0 fast enough, the product of ratios converges to a positive constant. Concrete counterexample: p = 1, k_n = 0, sigma_k = 1, Phi_k = 2^{-k}; then Omega_k = 2^{-2k} > 0, so Omega_k is full rank for every k, yet the error recursion is tilde_theta_{k+1} = (1 + 2^{-2k})^{-1} tilde_theta_k, and tilde_theta_k converges to tilde_theta_0 * prod_{j=0}^{infinity} (1 + 2^{-2j})^{-1} > 0, not 0. Thus theta_k does not converge to theta*, and Proposition 4(c) is unsupported. Proposition 3(d) also contains a circular step: it asserts that ||tilde_theta_{k+1}|| -> 0 to conclude nu_k -> 0, which is exactly the fact that needs proof. The missing ingredient is a persistency-of-excitation condition with a uniform lower bound, for example lambda_min(Omega_k) >= epsilon > 0 for all large k, which would make rho_k <= sigma/(sigma+epsilon) < 1 and yield convergence; pointwise full rank is not sufficient. The safety lower-bound property (Proposition 4(a,b)) is sound, and the numerical examples are illustrative, but the central vanishing-conservativeness claim does not follow. Theorem 1 and Theorem 2 are also stated without proof (omitted for space), though the convergence gap alone is enough to invalidate the advertised result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a sampled-data adaptive control-barrier-function (CBF) scheme for nonlinear systems with linearly parameterized uncertainty. A recursive-least-squares estimator with a sliding window and regularization produces a parameter estimate theta_k and a computable nonincreasing upper bound nu_k on the parameter estimation error. These are interpolated continuously in time and used inside a higher-order CBF constraint, together with a slack variable and a closed-form optimal control law, to guarantee constraint satisfaction and to make the constraint no more conservative over time. The paper further claims that, under a persistency-of-excitation condition, theta_k converges to theta*, nu_k converges to zero, and the adaptive CBF constraint converges to the ideal constraint that would be used if theta* were known. Two numerical examples, a pendulum and a nonholonomic robot, illustrate the approach.","tokens_in":13869,"tokens_out":8003,"duration_ms":84976,"significance":"If the convergence claims were established, the paper would make a useful contribution: it gives a closed-form adaptive CBF construction, a computable nonincreasing error bound, and a systematic way to reduce conservativeness while maintaining safety. The safety part of the argument, in particular the fact that psi is a lower bound on the ideal CBF when nu bounds the estimation error (Proposition 4(a)-(b)), is sound and is a genuine strength. However, the central asymptotic claims about vanishing conservativeness are not proven under the assumptions stated in the manuscript. The current 'full rank' assumption is not persistency of excitation, and the proofs of parameter convergence and nu_k -> 0 contain a load-bearing gap. The significance of the paper therefore depends on a correction of the excitation assumption and the convergence proofs.","major_comments":[{"comment":"The proof of Proposition 2(c) is invalid: strict decrease of the Lyapunov-like function V_k does not imply V_k -> 0 when the contraction ratio tends to 1. From equation (16), tilde_theta_{k+1} = sigma_k P_k tilde_theta_k, so the contraction factor is sigma_k lambda_max(P_k) = sigma_k / (sigma_k + lambda_min(Omega_k)). The assumption that Omega_k is full rank for all k >= k_i only gives lambda_min(Omega_k) > 0 pointwise; if lambda_min(Omega_k) decays to zero sufficiently fast, the product of contraction factors has a positive limit. Concretely, take p = 1, k_n = 0, sigma_k = 1, and Phi_k = 2^{-k}; then Omega_k = 2^{-2k} is full rank for every k, yet tilde_theta_{k+1} = (1 + 2^{-2k})^{-1} tilde_theta_k, so tilde_theta_k converges to tilde_theta_0 * prod_{j=0}^infty (1 + 2^{-2j})^{-1} > 0 rather than to zero. Thus theta_k does not converge to theta* under the stated assumption, and Proposition 4(c), which relies on this result for the advertised vanishing-conservativeness property, is unsupported. A uniform lower bound such as lambda_min(Omega_k) >= epsilon > 0 for all large k is needed to obtain a uniform contraction factor less than 1.","section":"Section IV, Proposition 3(d)"}],"minor_comments":[{"comment":"The phrase \"2 two numerical examples\" should be corrected to \"two numerical examples.\"","section":"Abstract"},{"comment":"The word \"Lipshchitz\" should be \"Lipschitz.\"","section":"Section III"},{"comment":"The heading \"SAFE AND OPTIMAL CONTROL\" is misspelled in the manuscript as \"safe an optimal control.\"","section":"Section V heading"},{"comment":"The proofs of Theorem 1 and Theorem 2 are omitted. If the authors rely on standard CBF arguments and on [24], they should state the precise theorem numbers and provide a short proof sketch that accounts for the sampled-data construction of theta and nu; otherwise the central safety result is not self-contained.","section":"Section V, Theorems 1 and 2"},{"comment":"The examples do not report any measure such as lambda_min(Omega_k) along the trajectories, so the reader cannot check whether the excitation condition that would support the convergence claims is actually satisfied in the simulations.","section":"Section VI and VII"}],"recommendation":"major_revision","confidential_remarks":"The main issue is that the paper advertises persistency of excitation but only assumes pointwise full rank of Omega_k. This is a correctable, but central, gap: the authors need to introduce a uniform PE condition and re-prove Propositions 2(c), 3(d), and 4(c) under that condition. If the authors cannot establish convergence under a standard PE assumption, the vanishing-conservativeness claim should be withdrawn and the paper reframed around the nonincreasing-conservativeness result."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the Gutierrez-Hoagg preprint. The safety-constraint machinery is real: using the RLS estimate plus a computable nonincreasing error bound as a margin in the CBF is a sensible extension of Taylor-Ames and Lopez et al., and the lower-bound property (Prop 4(a)-(b)) is correct. The closed-form control is a standard QP solution, and the two examples illustrate the intended performance benefit. That part earns its keep.\n\nThe problem is the headline: vanishing conservativeness under persistency of excitation. Proposition 2(c) is false as stated. The proof infers θ_k → θ* from ΔV_k < 0, but a strictly decreasing nonnegative sequence need not go to zero. Concretely: p=1, σ_k=1, Φ_k=2^{-k}. Then Ω_k = 2^{-2k} > 0 for every k, so the full-rank assumption holds, but the error recursion (16) gives e_{k+1} = (1+2^{-2k})^{-1} e_k, so e_k converges to a positive limit, not 0. Eventual full rank is too weak; you need a uniform lower bound λ_min(Ω_k) ≥ ε > 0 to make the contraction ratio stay below 1. Proposition 3(d) repeats the same gap: it uses ||tilde_θ_{k+1}|| → 0 to conclude ν_k → 0, which is the very fact in question. Consequently Prop 4(c), the advertised convergence to the ideal CBF constraint, is unsupported. Without it, the title feature doesn't follow.\n\nThe theorems on optimality and closed-loop safety are stated without proofs, 'omitted for space.' That is not acceptable for a journal version, though if the proofs are truly standard it's a completeness issue rather than a substantive one.\n\nThe abstract promises a persistency-of-excitation condition, but the formal assumption is only eventual full rank of Ω_k; no uniform PE bound appears anywhere.\n\nAll that said, the flaws are repairable. Add a proper PE condition with a uniform lower bound on Ω_k, fix the convergence proof, supply the missing proofs, and the safety result plus the nonincreasing-conservativeness result stand. I would send the current version to peer review anyway, because a serious referee should weigh in on whether the uniform PE fix is enough and whether the examples satisfy it. But I wouldn't cite the vanishing-conservativeness claim as it stands.","headline":"Sound safety construction with a nonincreasing error margin, but the headline vanishing-conservativeness theorem is false as stated: eventual full-rank Ω_k does not imply θ_k→θ*, and a uniform PE bound is needed.","tokens_in":14580,"tokens_out":6145,"would_cite":false,"duration_ms":53896,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C40","93C10","93D30"],"pacs":[],"model":"deepseek-v4-flash","headline":"An adaptive control-barrier constraint that estimates unknown model parameters online is shown to converge to the ideal safety constraint under persistency of excitation, preserving safety throughout.","keywords":["control barrier functions","adaptive control","recursive least squares","parameter estimation","persistency of excitation","state constraints","safety-critical control","vanishing conservativeness"],"falsifier":"Take the least-squares recursion (7)–(11) with a regressor sequence $\\Phi_k$ that is full rank for every $k$ but whose information matrix $\\Omega_k$ has smallest eigenvalue tending to zero while $\\sigma_k$ is held fixed, so that $\\sigma_k\\lambda_{\\max}(P_k)\\to 1$; if $\\|\\theta_k-\\theta^*\\|$ and $\\nu_k$ do not converge to zero, then the convergence claims in Propositions 2(c) and 3(d) are false.","tokens_in":13203,"feed_emoji":"🛡️","tokens_out":8805,"duration_ms":78739,"temperature":0.7,"pith_summary":"This paper shows how to enforce state constraints (safety) in control systems with unknown parameters without paying a permanent conservativeness penalty. It couples a sampled-data recursive-least-squares parameter estimator with a computable, nonincreasing upper bound on the estimation error, and feeds both into a control-barrier-function constraint. The central claim is that whenever the sampled regression data are persistently exciting—the relevant matrix $\\Omega_k$ is full rank for all large sampling instants—the parameter estimate converges to the true value, the error bound shrinks to zero, and the adaptive CBF constraint converges to the ideal constraint that a designer would write if the parameters were known. Safety is preserved at every time during this learning process. Two numerical examples, an inverted pendulum and a nonholonomic mobile robot, demonstrate that the adaptive constraint improves tracking relative to using a fixed conservative bound.","feed_headline":"Safety constraints converge to ideal ones under persistent excitation","feed_subtitle":"Adaptive control-barrier method estimates unknown parameters and shrinks its safety margin to zero under excitation.","key_machinery":"The load-bearing object is the computable error-bound recursion $\\nu_{k+1} = \\min\\{\\sigma_k\\lambda_{\\max}(P_k)\\nu_k, \\sqrt{\\nu_k^2+\\tau_k}\\}$, combined with the identity $\\tau_k=\\Delta V_k$ (Lemma 1). Here $\\Omega_k=\\sum_{i=k-k_n}^{k}\\Phi_i^T\\Phi_i$ is a sliding-window information matrix built from sampled regressors, $P_k=(\\sigma_k I_p+\\Omega_k)^{-1}$, and $\\tau_k$ is defined purely from measured data and the current estimate. The identity turns an unmeasurable Lyapunov-like decrease into a measurable quantity, and the min operation keeps $\\nu_k$ nonincreasing while always dominating the true error norm. The CBF constraint (27) uses $\\theta(t)$ and $\\nu(t)$ linearly, so whenever $\\nu(t)$ bounds the error, $\\psi$ is a lower bound of the ideal $\\psi^*$ (Proposition 4). Under full rank of $\\Omega_k$, both the parameter error and $\\nu_k$ are driven to zero, so $\\psi\\to\\psi^*$.","core_discovery":"Under a persistency-of-excitation condition, the adaptive CBF constraint $\\psi(x,\\theta(t),\\nu(t),\\hat u,\\hat\\delta)$ converges pointwise to the ideal constraint $\\psi^*(x,\\hat u,\\hat\\delta)$ that would be used with the true parameter vector $\\theta^*$, and this happens while the state remains inside the safe set for all time. The key is that the one-step decrease of the Lyapunov-like function for the regularized finite-horizon least-squares estimator is exactly computable from measured data: the quantity $\\tau_k$ defined in (19) equals $\\Delta V_k$, even though $\\theta^*$ is unknown. This identity yields a nonincreasing bound $\\nu_k$ on the parameter-error norm and, under the full-rank condition on $\\Omega_k$, forces both $\\|\\theta_k-\\theta^*\\|$ and $\\nu_k$ to zero. Consequently the CBF constraint begins conservative and asymptotically sheds its conservativeness, while the same estimate simultaneously improves the desired control law.","pith_inferences":["The same $\\tau_k=\\Delta V_k$ identity should transfer to any recursive-least-squares estimator whose regressors are sampled integrals of known functions, giving a ready-made nonincreasing error certificate for other adaptive safety filters; the paper does not develop this generality.","A quantitative excitation condition—a uniform lower bound on the smallest eigenvalue of $\\Omega_k$—would let one state convergence rates for $\\nu_k$ and hence a finite-time bound on the conservativeness gap; the paper only asserts asymptotic vanishing.","Since the safety constraint is valid for any $\\nu(t)$ that dominates the estimation error, one could replace the least-squares bound with any tighter set-membership or interval bound and keep the same $\\psi$ construction, potentially reducing conservativeness faster than the RLS bound does."],"forward_implications":["Safety is guaranteed for all time under the closed-form control $u^*$, regardless of whether the excitation condition holds.","Under the full-rank condition on $\\Omega_k$ for all large $k$, the parameter estimate converges to $\\theta^*$ and the error bound $\\nu_k$ tends to zero.","The enforced constraint converges to the ideal known-parameter constraint, so in the long run the adaptive filter behaves as if the model uncertainty were absent.","Because the error bound is nonincreasing, the desired control law $u_d(x,\\theta(t))$ improves monotonically even before parameter convergence is complete."],"supporting_citations":[{"why":"Supplies the recursive form and stability properties of the sliding-window variable-regularization least-squares estimator used to generate $\\theta_k$.","marker":"[32]"},{"why":"Provides the standard CBF quadratic-program formulation and forward-invariance analysis that Theorem 2 extends to the adaptive setting.","marker":"[11]"},{"why":"Introduces the baseline adaptive CBF approach that this paper extends by adding a computable error bound and vanishing conservativeness.","marker":"[27]"},{"why":"Shows how set-membership identification reduces conservativeness in adaptive CBFs, the line of work this paper continues with a least-squares bound.","marker":"[28]"},{"why":"Provides the closed-form safety-filter control result whose proof technique is cited for Theorem 1, the uniqueness of the minimizer.","marker":"[24]"},{"why":"Establishes the constructive CBF approach for forward invariance that underlies the safe-set construction $\\psi_0,\\ldots,\\psi_{d-1}$.","marker":"[10]"}],"fun_headline_variants":["Adaptive safety margins vanish under persistent excitation","CBF conservativeness shrinks to zero with excitation","Adaptive CBFs approach ideal safety under excitation","Excitation drives adaptive safety to exactness","Vanishing conservativeness in adaptive CBFs via excitation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The advertised vanishing of conservativeness depends on the parameter estimate actually converging to the true value, which the paper derives from strict decrease of a Lyapunov-like function under a full-rank condition on the sampled regression data; if strict decrease does not by itself force convergence, the central feature fails even though the safety guarantee survives.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive safety margins vanish under persistent excitation","CBF conservativeness shrinks to zero with excitation","Adaptive CBFs approach ideal safety under excitation","Excitation drives adaptive safety to exactness","Vanishing conservativeness in adaptive CBFs via excitation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000424,"raw_usage":{"total_tokens":2158,"prompt_tokens":910,"completion_tokens":1248,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":1173}},"tokens_in":526,"tokens_out":1248,"duration_ms":9235,"temperature":1.0,"reasoning_tokens":1173,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:05:37.244779+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the least-squares recursion (7)–(11) with a regressor sequence $\\Phi_k$ that is full rank for every $k$ but whose information matrix $\\Omega_k$ has smallest eigenvalue tending to zero while $\\sigma_k$ is held fixed, so that $\\sigma_k\\lambda_{\\max}(P_k)\\to 1$; if $\\|\\theta_k-\\theta^*\\|$ and $\\nu_k$ do not converge to zero, then the convergence claims in Propositions 2(c) and 3(d) are false.","supporting_citations":[{"cited_title":"On the stability and convergence of a sliding-window variable-regularization recursive-least-squares algorithm,","cited_arxiv_id":null,"evidence_quote":"Supplies the recursive form and stability properties of the sliding-window variable-regularization least-squares estimator used to generate $\\theta_k$."},{"cited_title":"Control barrier function based quadratic programs for safety critical systems,","cited_arxiv_id":null,"evidence_quote":"Provides the standard CBF quadratic-program formulation and forward-invariance analysis that Theorem 2 extends to the adaptive setting."},{"cited_title":"Adaptive safety with control barrier functions,","cited_arxiv_id":null,"evidence_quote":"Introduces the baseline adaptive CBF approach that this paper extends by adding a computable error bound and vanishing conservativeness."},{"cited_title":"Robust adaptive control barrier functions: An adaptive and data-driven approach to safety,","cited_arxiv_id":null,"evidence_quote":"Shows how set-membership identification reduces conservativeness in adaptive CBFs, the line of work this paper continues with a least-squares bound."},{"cited_title":"Constructive safety using control barrier functions,","cited_arxiv_id":null,"evidence_quote":"Establishes the constructive CBF approach for forward invariance that underlies the safe-set construction $\\psi_0,\\ldots,\\psi_{d-1}$."}],"review_version":1}