{"id":"7f701940-31e7-48c4-aa60-d526c9a760ec","arxiv_id":"2501.13270","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A discrete-time version of REAP model predictive control selects an adaptive KKT step size at each iteration to preserve feasibility under early termination.","lead":"This paper adapts the step size of an optimization routine inside a model predictive controller so that, even if computation stops early, the chosen control input still respects safety constraints. It would help drones and robots with weak onboard computers run constrained control with a safety guarantee.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's proof relies on inequality (22), which is false because B is affine in λ; the convergence and boundedness guarantees are therefore unsupported as written.","rationale":"REAP's value proposition is exactly that a feasible suboptimal solution is always available and that the virtual dynamics converge to the tightened optimum. Theorem 2 is the only support for convergence, and it also supplies the boundedness used in the proof. The false inequality (22) therefore attacks the core mathematical claim. The counterexample is not contrived: B is affine in λ by construction, so any joint strong-monotonicity constant must be zero. This is a structural property, not a numerical artifact. The experimental and simulation results are useful evidence that the scheme works in practice, and the adaptive σ idea is sensible; those parts deserve credit. However, empirical success cannot replace a proof that relies on a false inequality. A corrected proof might exist using a different Lyapunov function or by adding regularization, but as written the central theorem fails. The normalization issue in Theorem 1 is also legitimate and would need fixing; it is a second distinct theoretical gap, but the convergence gap is the one that most directly blocks the headline guarantee. Consistent with the reader, I would not reject the manuscript outright; the appropriate outcome is conditional acceptance pending a correct convergence proof or a weakened claim.","tokens_in":18798,"tokens_out":13914,"duration_ms":155206,"concrete_test":"Verify inequality (22) on the one-constraint example. Let β>0, B(u,λ)=0.5u² - λ log(-βu), and choose u†=-1/β, λ†=1/β². For z=(0,δ), compute the left side of (22): the first component of the field is βδ, multiplied by z_u=0; the second component is log(-βu†)=0, multiplied by δ; hence LHS=0, whereas μ||z||²=μδ²>0. This analytical calculation settles that (22) is false as stated; if the authors instead intend μ||u-u†||² on the right, then the claimed dual convergence does not follow.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing gap is in Theorem 2. Its proof requires inequality (22): ([∇uB; -∇λB])^T ([u-u†; λ-λ†]) ≥ μ ||[u-u†; λ-λ†]||² for some μ>0. For the modified barrier (9), ∇λB = -log(-β(η_i^T u+γ_i)) is independent of λ, so the vector field is affine in the dual variables and cannot be strongly monotone in the λ-direction. Concretely, for the scalar problem min 0.5u² subject to u+1/β≤0, take u†=-1/β, λ†=1/β², and perturb λ by δ with u=u†. At this point ∇uB=βδ but z_u=0, and ∇λB=0, so the left side of (22) is 0 while the right side is μδ²; no μ>0 exists. Thus the Lyapunov difference bound (28), the invariance of T in (21), and the claimed convergence to (u†,λ†) are unsupported. This is independent of the cross-evaluation issue the reader noted; even evaluating (22) at the same step does not save it. The Theorem 1 normalization concern is a separate scale-dependence gap; the convergence proof is the more fundamental failure for the paper's central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the discrete-time (Euler) implementation of the Robust-to-Early Termination (REAP) MPC scheme proposed by Hosseinzadeh et al. The authors introduce an adaptive Karush-Kuhn-Tucker parameter sigma(tau|t), chosen at each computation step from the distance to the constraint boundaries, and claim that this choice preserves the invariance of the feasible sets D_hatu and D_hatlambda, maintains anytime feasibility, and guarantees convergence to the tightened optimal solution. The theoretical results are stated as Theorems 1-3. The paper also reports simulation studies for a Parrot Bebop 2 model, a comparison of fixed versus adaptive sigma, suboptimality estimates, experimental results on the drone, and a MATLAB package called DiscreteREAP.","tokens_in":19063,"tokens_out":6798,"duration_ms":70455,"significance":"If correct, the paper would provide a practically useful discretization guideline for REAP: a computable step-size selection rule that preserves the anytime feasible and convergent behavior of the continuous-time scheme. The adaptive-sigma idea is sensible, and the paper contains a notable empirical component: a comparison across five sigma strategies, a study with state and input constraints, an assessment of degradation under reduced available computation time, and real-time experiments on a Parrot Bebop 2. The release of the DiscreteREAP MATLAB package is also a positive contribution that lowers the barrier for other researchers. However, the theoretical core contains load-bearing gaps: Theorem 1 is scale-dependent unless constraint normals are normalized, and the convergence proof of Theorem 2 relies on a monotonicity inequality that is false for the modified barrier used in the paper.","major_comments":[{"comment":"The invariance proof compares a Euclidean distance in u-space with a constraint-value change. Equation (18) bounds |eta_i^T (hatu(tau)-hatu(tau-1))| by ||eta_i|| ||hatu(tau)-hatu(tau-1)||, but equation (16a) then bounds ||eta_i|| times the step size by the Euclidean distance delta_i^u to the shifted hyperplane eta_i^T theta + gamma_i + epsilon = 0. For a point with constraint margin d_i = -eta_i^T hatu - gamma_i - epsilon, the Euclidean distance is d_i/||eta_i||. The proof therefore only guarantees |eta_i^T (hatu(tau)-hatu(tau-1))| <= d_i/||eta_i||, which is larger than the allowed constraint-value margin d_i whenever ||eta_i||<1. Hence Theorem 1, and consequently the anytime feasibility statement in Theorem 3, is not proved for general constraint normalizations. The fix is either to state explicitly that all normals are normalized to unit length, or to define delta_i^u as the constraint-value margin -eta_i^T hatu - gamma_i - epsilon rather than the Euclidean distance.","section":""},{"comment":"Inequality (22) is asserted as a strong-monotonicity property of the vector field [nabla_u B; -nabla_lambda B], but the modified barrier function in (9) is affine in lambda, so nabla_lambda B = -log(-beta(eta_i^T u + gamma_i + 1/beta)+1) is independent of lambda. The vector field therefore cannot be strongly monotone in the lambda direction. A concrete counterexample is the scalar problem min 0.5 u^2 subject to u <= 0, with tightened constraint u + 1/beta <= 0 and modified barrier B(u,lambda)=0.5u^2 - lambda log(-beta u). At u = -1/beta, lambda = 1/beta^2 + delta, the field value is (beta delta, 0) while (u-u^dagger, lambda-lambda^dagger) = (0, delta), so the left side of (22) is 0 and no mu>0 can satisfy the inequality. This invalidates the Lyapunov difference bound (28), the invariance of the set T in (21), and the claimed convergence to (u^dagger, lambda^dagger). Additionally, the proof applies (22) with gradients evaluated at tau-1 paired with the error at tau, a cross-evaluation that is not justified by the stated inequality even if strong monotonicity held.","section":""},{"comment":"There is an internal inconsistency between the convergence claim and the stated behavior of the adaptive parameter. Remark 4 states that if delta_i^hatu(tau-1|t)=0 or delta_i^hatlambda(tau-1|t)=0, then sigma(tau|t)=0 and consequently sigma(tau+theta|t)=0 for all theta=0,1,...; the iterates stall for the rest of the computation interval at time t. Under this mechanism, sigma(0|t)>0 but sigma(1|t)=0 already implies the trajectory never converges. Theorem 2's hypothesis only excludes \"sigma(tau|t) equivalently equal to zero at all tau\", which is insufficient; the proof needs either a strict positivity assumption on sigma for all tau or a mechanism that restarts the evolution after a zero step. As written, the convergence result is not supported for trajectories that touch the shifted constraint boundary before reaching u^dagger.","section":""}],"minor_comments":[{"comment":"The sentence \"its not clear if its properties are maintained\" contains a grammatical error: it should be \"it is not clear whether its properties are maintained\".","section":""},{"comment":"The notation in equation (16b) contains a misplaced comma: \"nabla_lambda B(x(t), r, hatu, (tau-1|t), hatlambda(tau-1|t))\" should read \"nabla_lambda B(x(t), r, hatu(tau-1|t), hatlambda(tau-1|t))\".","section":""},{"comment":"The table caption and text use \"ceases\" where \"cases\" is intended, and the phrase \"for all ceases\" should be corrected.","section":""},{"comment":"The claim that B and its gradients are \"bounded and definite\" is unclear; the intended meaning appears to be that the logarithm argument remains positive so that B is well defined and finite on the iterates. The wording should be revised.","section":""},{"comment":"Several reference titles contain typos: reference [26] reads \"onvex Optimization\" and reference [34] reads \"gradient synamics\". These should be corrected before publication.","section":""},{"comment":"The text \"See Figure 2 and (3) for geometric illustrations\" appears to refer to Figure 3, not equation (3); the cross-reference should be fixed.","section":""}],"recommendation":"major_revision","confidential_remarks":"The empirical and experimental sections are substantial and the MATLAB package is a useful contribution, but the theoretical guarantees are the paper's central claim and the current proof of Theorem 2 is invalid. The scale-dependence issue in Theorem 1 is local and fixable, but the convergence proof needs a fundamentally different argument or a restriction to a setting where a valid Lyapunov function exists. I would recommend a major revision and a re-review of the revised proofs."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper addresses a real practical gap: continuous-time REAP needs a discrete-time implementation, and the existing advice to use a sufficiently small Euler step leaves the user with no way to choose that step without either stalling the virtual dynamics or risking constraint violations. The authors' adaptive KKT parameter, chosen from distance-to-constraint bounds, is a sensible and novel extension of [25]. The simulation study and the Parrot Bebop 2 experiment are credible, and the zero-violation results with the adaptive rule in the tested cases are worth attention. The MATLAB package is a useful addition.\n\nThe theory, however, has load-bearing problems. The proof of Theorem 2 relies on inequality (22), which claims strong monotonicity of the stacked vector field [∇_u B; −∇_λ B]. But the modified barrier is affine in λ: ∇_λ B = −log(−β(η_i^T u+γ_i)) has no λ-dependence. A concrete counterexample is the scalar problem min 0.5u² subject to u+1/β≤0. At u = u† = −1/β and λ = λ†+δ, the u-gradient changes by βδ while the λ-gradient is unchanged, so the left side of (22) is 0 and the right side is μδ²; no μ>0 exists. This invalidates the Lyapunov difference bound (28), the invariance of T, and the claimed convergence to (u†,λ†). This is not a minor gap; the convergence theorem is the core of the paper.\n\nTheorem 1 has a separate scaling error. δ_i^u is the Euclidean distance to the shifted hyperplane, whereas η_i^T Δu is a change in the constraint value. The comparison in the proof is valid only for unit normals; with σ_i^u as in (16a), a non-normalized constraint with ||η_i||<1 can be crossed by the allowed step. This is fixable by normalizing the constraint normal or rescaling σ_i^u.\n\nThe experiments remain useful and the adaptive scheme may well work, but the theoretical guarantees as stated are not proven. I would not cite this paper for a convergence result. I would, however, send it to peer review: the practical extension is valuable, the proof failures are concrete and likely repairable, and a referee can point to exactly what needs to be fixed. As it stands it is a conditional reject.","headline":"The adaptive KKT step-size scheme is a practical and well-tested idea, but the central convergence proof relies on a false strong-monotonicity inequality, so the theory as written does not support the claims.","tokens_in":19636,"tokens_out":8336,"would_cite":false,"duration_ms":79026,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C55","90C25","90C30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that discrete-time Euler REAP with a per-step adapted KKT parameter preserves anytime feasibility and convergence.","keywords":["Model Predictive Control","Robust-to-Early Termination","Discrete-Time Implementation","Anytime Feasibility","Primal-Dual Gradient Flow","Modified Barrier Function","KKT Parameter Adaptation","Limited Computing Capacity"],"falsifier":"Run a scalar numerical test with constraint $u\\le 0$, the modified barrier (9), and update (14), starting inside the $\\epsilon$-band such as $u=-\\epsilon/2$ while setting $\\sigma(\\tau|t)$ equal to the upper bound in (15); if $u(\\tau|t)>0$ appears for any $\\tau$, then Theorem 1's invariance claim is false.","tokens_in":18513,"feed_emoji":"⚙️","tokens_out":10756,"duration_ms":108593,"temperature":0.7,"pith_summary":"Model Predictive Control normally requires an optimization to finish before the next sampling instant, which is hard on limited hardware. The Robust-to-Early Termination (REAP) approach instead turns the MPC problem into a virtual dynamical system whose state is always a feasible, suboptimal control sequence, so computation can stop at any moment. The original REAP was analyzed in continuous time. This paper shows that the same guarantees survive a practical Euler discretization when the KKT parameter, the gain that controls how fast the virtual system moves, is chosen at every computation step based on the distance to the active constraints. Under that adaptive choice, the feasible sets stay invariant, the iterates converge to the tightened optimum, and every truncated computation still yields a constraint-satisfying control.","feed_headline":"Adaptive gain keeps early-terminated MPC feasible","feed_subtitle":"Discrete-time REAP preserves anytime feasibility when the KKT parameter is chosen per step.","key_machinery":"The central machinery is the virtual primal-dual gradient-flow system (10) driven by the modified barrier function (9), together with a per-step adaptive gain $\\sigma(\\tau|t)$. In continuous time a fixed gain works because infinitesimal steps cannot overshoot; in discrete time the paper makes the gain shrink exactly when a step would threaten a constraint. The upper bound (15)-(16) measures, for each constraint, the Euclidean distance $\\delta_i^{\\hat u}(\\tau-1|t)$ to the shifted hyperplane $\\eta_i^\\top\\theta+\\gamma_i+\\epsilon=0$ and the distance $\\delta_i^{\\hat\\lambda}(\\tau-1|t)$ to the dual nonnegativity boundary, then divides by the local gradient norm with a small smoothing factor to obtain the largest safe gain. This converts the fixed-step Euler update into an adaptive-step algorithm that is slow near constraints and fast away from them.","core_discovery":"The paper's claim is that discrete-time REAP, implemented as Euler updates in (14), keeps both central properties of the continuous-time scheme: anytime feasibility and convergence. The enabling device is a state-dependent KKT parameter. At computation step $\\tau$, the gain $\\sigma(\\tau|t)$ is capped by the Euclidean distance from the current iterate to each shifted constraint hyperplane, and from the dual iterate to its nonnegativity boundary, divided by the local gradient magnitude, producing the upper bound in (15) and (16). Theorem 1 proves these per-step caps keep the iterates inside $D_{\\hat u}$ and $D_{\\hat\\lambda}$; Theorem 2 proves boundedness and convergence to the tightened optimum using a Lyapunov function; Theorem 3 uses the growth of the modified barrier near the boundary to conclude constraint satisfaction at every step. The authors read this as making REAP implementable in discrete time without sacrificing its central guarantee that a feasible suboptimal control is always available when computation is cut short.","pith_inferences":["The paper does not develop this, but the adaptive $\\sigma$ rule is functionally a per-step backtracking line search; coupling it with a projected or predictor-corrector step could reduce the conservatism noted in Remark 3.","If the guarantee holds as stated, REAP becomes a candidate for fixed-point processors with hard deadlines, since no offline bound on the iteration count is needed, only the per-step distance computation.","The $\\epsilon$-band subtlety in the distance-to-shifted-constraint bound suggests a practical hardening: compute the distance to the original constraint boundary rather than the shifted one, or keep iterates outside the $\\epsilon$-band, before relying on Theorem 1's step bound.","A testable extension would replace the Euler integrator with a Runge-Kutta or operator-splitting scheme and derive the analogous per-step gain bound from the local truncation error."],"forward_implications":["An Euler implementation of REAP with $\\sigma(\\tau|t)$ chosen by (15) keeps $\\hat u(\\tau|t)$ inside the tightened feasible set for every computation step, so early termination at any $\\tau$ still returns a feasible control.","The discrete iterates converge to the tightened optimum $u^\\dagger(t)$ as $\\tau\\to\\infty$, provided the adaptive gain does not become identically zero.","Fixed gains in discrete time can violate constraints; the paper reports violations in 48.9% to 100% of tested runs for fixed $\\sigma$, whereas the adaptive gain produced zero violations in the same tests.","When available computation time is only half the sampling period, the reported performance degradation stays below 0.3% relative to a powerful processor solving the MPC problem optimally."],"supporting_citations":[{"why":"Introduces continuous-time REAP and its anytime-feasibility and convergence properties, which this paper extends to discrete time.","marker":"[25]"},{"why":"Supplies the modified barrier function used to define the virtual dynamics and its logarithmic terms.","marker":"[31]"},{"why":"Alexandrov's theorem on convex polyhedra, used to show the barrier grows near the boundary so feasibility is enforced.","marker":"[39]"},{"why":"Primal-dual gradient-flow analysis whose convergence arguments are adapted into the discrete-time Lyapunov proof.","marker":"[28]"},{"why":"Monotone operator primer supporting the strong-convexity and monotonicity inequality used in Theorem 2.","marker":"[40]"}],"fun_headline_variants":["Distance-capped gain keeps discrete REAP always feasible","Per-step KKT caps let early-terminated MPC stay feasible","Discrete REAP: adaptive gain ensures anytime feasibility","Early-terminated MPC: distance-based gain preserves feasibility","REAP goes discrete: per-step gain sustains feasibility"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The invariance proof assumes that the distance to a slightly shifted copy of each constraint, divided by the local gradient norm, is a safe per-step movement bound that keeps iterates inside the original feasible set; near the true boundary this distance can exceed the remaining margin, so the bound may allow a step past the constraint.","fun_headline_variants_meta":{"raw":{"variants":["Distance-capped gain keeps discrete REAP always feasible","Per-step KKT caps let early-terminated MPC stay feasible","Discrete REAP: adaptive gain ensures anytime feasibility","Early-terminated MPC: distance-based gain preserves feasibility","REAP goes discrete: per-step gain sustains feasibility"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000699,"raw_usage":{"total_tokens":3141,"prompt_tokens":914,"completion_tokens":2227,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":2162}},"tokens_in":530,"tokens_out":2227,"duration_ms":17974,"temperature":1.0,"reasoning_tokens":2162,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:19:55.507163+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a scalar numerical test with constraint $u\\le 0$, the modified barrier (9), and update (14), starting inside the $\\epsilon$-band such as $u=-\\epsilon/2$ while setting $\\sigma(\\tau|t)$ equal to the upper bound in (15); if $u(\\tau|t)>0$ appears for any $\\tau$, then Theorem 1's invariance claim is false.","supporting_citations":[{"cited_title":"Polyak, Modified barrier functions (theory and methods), Mathemati- cal Programming 54 (1–3) (1992) 177–222","cited_arxiv_id":null,"evidence_quote":"Supplies the modified barrier function used to define the virtual dynamics and its logarithmic terms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Alexandrov's theorem on convex polyhedra, used to show the barrier grows near the boundary so feasibility is enforced."},{"cited_title":"Hosseinzadeh, B","cited_arxiv_id":null,"evidence_quote":"Primal-dual gradient-flow analysis whose convergence arguments are adapted into the discrete-time Lyapunov proof."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Monotone operator primer supporting the strong-convexity and monotonicity inequality used in Theorem 2."}],"review_version":1}