{"id":"ee843611-c24c-4d4d-8468-5b4e0ea54fe6","arxiv_id":"2507.02772","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A hybrid systems framework treats suboptimal MIQP solver output in feedback control as a perturbation and claims practical stability under small suboptimality.","lead":"This paper models a mixed-integer quadratic program solver running in a feedback loop as a hybrid dynamical system, treating solver suboptimality as a bounded perturbation. It claims that if the suboptimality is small enough the closed loop remains practically stable, and it illustrates the idea with spacecraft rendezvous model predictive control simulations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2 is not established: Lemma 3's flow decrease fails for suboptimal y, and ϵ-suboptimality is not shown to be a ρ-perturbation in the sense of Proposition 1.","rationale":"The paper's advertised contribution is Theorem 2, a practical-stability bound for suboptimal MIQP feedback. That theorem rests on Lemma 3 and on Proposition 1. The proof of Lemma 3 is a one-line citation with three assertions, and the flow-decrease assertion is false for the Lyapunov function in Eq. (7) whenever y is suboptimal and x=0: L is constant, but |χ|_A is positive. The example above satisfies the qualitative stability content of Assumption 4, and if Assumption 3 is granted, the stated assumptions still do not imply the required decrease; thus the assumption set is not sufficient. If Assumption 3 is not granted, it is itself unjustified for this L, because no argument shows that the optimality gap controls distance to Y. Independently, even granting Lemma 3, Theorem 2's identification of solver suboptimality with the additive perturbations of [13, §3.3.3] is not established. Lemma 1's upper semicontinuity is topological, not quantitative; it does not imply that ψ_ϵ lies within ρB of ψ_0 for ρ controlled by ϵ. The theorem's unspecified σ therefore does not provide a meaningful quantitative guarantee. The reader's weakest-assumption pick, Assumption 4, is related: it simply assumes the nominal closed loop is stable, so the framework does not prove stability of optimal MIQP-MPC from first principles, only analyzes perturbations of an already stable loop. I agree this is a serious limitation, but the more specific, checkable flaw is the invalid Lyapunov decrease and the unverified perturbation identification. These are internal to the argument, not a matter of disagreeing with consensus, so the reader's REJECT verdict remains appropriate.","tokens_in":13577,"tokens_out":10424,"duration_ms":131837,"concrete_test":"Construct the scalar HMIQP with f(x,u)=−x, κ(y)=y, Problem 1 as min y² subject to y≥0 (so Y={0}), Δt=1. At χ=(1,0,0)∈C, L=1 and |χ|_A=1, but during flow ˙x=−x gives ˙L=−x²=0 at x=0 while |χ|_A stays 1, so Lemma 3's decrease condition fails. This one-state example disproves the claimed UGpAS without extra assumptions. Separately, solve a one-integer-variable MIQP with ϵ>0 and compare ψ_ϵ with ψ_0+ρB to test the perturbation identification; if ψ_ϵ is not contained in ψ_0+ρB for ρ→0 as ϵ→0, Theorem 2's use of Proposition 1 is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Theorem 2) is not derived. To invoke Proposition 1 ([13, Thm 3.26]) for robustness, one needs (i) a uniform global pre-asymptotic stability certificate for the nominal HMIQP and (ii) a certified ρ-perturbation model of HMIQP that contains H_ϵ. Neither is established. For (i), Lemma 3 asserts L-decrease along flows and jumps under Assumptions 3–4. But L in Eq. (7) is (ϕ_ϵ−V)+½|x|²; it has no term for τ and only penalizes y through the optimality gap. During flows y is fixed, so the flow decrease inherits only ˙x=f(x,κ(y)) from Assumption 4; this bounds the x-component of L, not the y-distance to Y. The claim ˙L ≤ −ρ(|χ|_A) therefore requires an error bound linking (ϕ_ϵ−V) to dist(y,Y), which is not stated or proved; Assumption 3 merely asserts the resulting K∞ bounds exist. Without that, the KL estimate required by Proposition 1 is unavailable. For (ii), Theorem 2 asserts that H_ϵ is 'the same as' a system with disturbances and that 'all conditions from Proposition 1 hold.' A ρ-perturbation in [13, §3.3.3] perturbs states or inputs by d with |d|≤ρ; an ϵ-optimal solution y∈ψ_ϵ(x,η,ϵ) need not lie within any ρ-ball of ψ_0(x,η) unless a quantitative upper semicontinuity modulus for ψ is proven. No such modulus is given, and the set-valued flow ˙τ=1+eτ is not checked against the hybrid basic conditions. The theorem also never relates σ to ϵ, so the bound is vacuous as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces a hybrid dynamical system HMIQP for feedback control in which a parametric MIQP is solved recursively; the hybrid state is (y,x,τ), the jump map selects y from the ε-optimal set ψε, and the flow map is ẋ=f(x,κ(y)), τ̇=1. Under a rationality assumption on Q and A, Theorem 1 claims HMIQP satisfies the hybrid basic conditions. Lemma 3 claims uniform global pre-asymptotic stability of the stabilizing set A = Y×{0}×[0,Δt] under Assumptions 3 and 4 when Problem 1 is solved optimally, and Theorem 2 claims a semiglobal practical robust stability bound for suboptimal solver parameters modeled as bounded perturbations. The paper closes with MIQP-MPC simulations on a spacecraft rendezvous problem using time-limit, memory-limit, and iteration-limit parameters.","tokens_in":13938,"tokens_out":6962,"duration_ms":77759,"significance":"The proposed framework addresses an interesting and timely problem: giving a hybrid-systems account of suboptimal MIQP solvers in feedback loops. If Theorem 2 were valid, the paper would provide a useful bridge between parametric optimization and hybrid systems, and the practical robustness statement would be valuable. The explicit definition of solver iterates and the classification of solver parameters are also conceptually useful. However, the paper's central stability claims are not established by the presented proofs; several assumptions carry the entire load and the key perturbation argument is missing. As a result, the theoretical contribution is currently not reliable.","major_comments":[{"comment":"The proof of Lemma 3 asserts 'the flow dynamics are asymptotically stable, i.e. ˙L(χ) ≤ −ρ(|χ|A) ∀χ ∈ C, which is true under Assumption 4,' but this does not follow from Assumption 4. During flows, y and the sampled parameters c,η are fixed, so the term (ϕε−V) in L (Eq. (7)) is constant; the only evolving part is x, and Assumption 4 does not imply that 1/2|x|^2 is a Lyapunov function for the nonlinear dynamics ẋ=f(x,κ(y)). Moreover, L contains no term that bounds dist(y,Y) for suboptimal y, so the required decrease ˙L ≤ −ρ(|χ|A) cannot be concluded without an additional quantitative relationship between the optimality gap and the distance of y to Y. Thus the KL estimate needed for Theorem 2 is not derived.","section":"Section IV, Lemma 3"},{"comment":"Assumption 3 postulates the existence of K∞ functions α1, α2 with α1(|χ|A) ≤ L(χ) ≤ α2(|χ|A), and Assumption 4 postulates that κ(y) renders ẋ=f(x,κ(y)) asymptotically stable. These assumptions essentially assert the Lyapunov-theoretic properties that Lemma 3 is supposed to prove; in particular, Assumption 3 pre-assumes that L behaves as a proper indicator of A, and Assumption 4 pre-assumes nominal closed-loop stability. The paper provides no evidence or construction showing that these assumptions are consequences of the MIQP structure or of warm starting.","section":"Section IV, Assumptions 3 and 4"},{"comment":"Theorem 2 invokes Proposition 1 ([13, Thm 3.26]) but does not verify two of its key hypotheses. First, to identify He with a hybrid system with disturbances in [13, Sec. 3.3.3], one must show that each point y ∈ ψε(x,η,ε) lies within a bounded distance ρ of the nominal set ψ0(x,η); this requires a quantitative upper-semicontinuity modulus for ψε, which is not stated or proved. Second, the flow perturbation eτ ∈ [−1,∞) is unbounded above, so it cannot be represented as a bounded ρ-perturbation with finite ρ. Consequently, the robust-stability KL bound is not established.","section":"Section IV, Theorem 2"},{"comment":"The bound |π(t,j)|A ≤ β(|π(0,0)|A, t+j)+σ never relates σ to the suboptimality parameter ε or to the perturbation bound ρ. As stated, σ is an unspecified positive number that can always be taken large enough to make the inequality trivially true, so the theorem does not yield a quantitative guarantee that better solver parameters imply smaller steady-state error.","section":"Section IV, Theorem 2"}],"minor_comments":[{"comment":"In Eq. (7), the term 1/2 x^2 should be 1/2 |x|^2 for a vector state.","section":"Section II.A, Eq. (7)"},{"comment":"V is defined as a set-valued mapping V: R^n × R^m ⇒ R, but the value function is single-valued; this notation is confusing.","section":"Section II.A"},{"comment":"The caption of Fig. 6 says 'As expected in Figure 2' but should refer to Figure 6; several figure cross-references are mismatched.","section":"Section V, Figures"},{"comment":"The simulations consist of single runs without error bars or statistical repetition; the unexplained instability at 0.07 s in Fig. 2 is reported without a supporting numerical investigation, which weakens the empirical validation.","section":"Section V"},{"comment":"There are minor typos, including 'Nestrov' for 'Nesterov', and the notation eτ = [−1,∞) should be clarified as an interval rather than a vector.","section":"Introduction"}],"recommendation":"reject","confidential_remarks":"The paper's high-level idea has merit, but the central stability theorem is not proven; the gaps are load-bearing and would require substantial new analysis (quantitative perturbation bounds and a Lyapunov construction) beyond the current scope. I do not recommend major revision; the results as presented do not support the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper gives a hybrid-systems perspective on recursive MIQP feedback control, modeling solver suboptimality as a perturbation from the optimal solution set. That framing is genuinely new and worth attention: it connects parametric MIQP regularity (Bank–Hansel) to hybrid robust stability (Sanfelice), and it targets a real problem—setting time and memory limits for MIQP MPC. The first result, Theorem 1, is a routine check that upper semicontinuity of the ϵ-optimal mapping under rationality implies the hybrid basic conditions; it's fine.\n\nThe trouble is the central stability theorem. Lemma 3 claims a Lyapunov decrease along flows and jumps, but the proof just asserts the inequalities hold under Assumptions 3 and 4. They don't. The Lyapunov function L = (φ_ε−V)+½|x|² does not decrease in the gap term during flow, because y and the parameters are held constant; its flow derivative depends only on x·f(x,κ(y)). Assumption 4 gives asymptotic stability of the closed-loop system, but not that ½|x|² is a Lyapunov function for it, nor that the decrease is uniform in y over Y. And Assumption 3 already assumes L is proper and positive definite with respect to A—essentially the conclusion you want. So Lemma 3 does not provide the KL estimate needed.\n\nTheorem 2 then calls on a robustness result from Sanfelice's book. The stress-test note is on target: to apply a ρ-perturbation theorem you need to show the ϵ-optimal set ψ_ε lies within a bounded distance of ψ_0. The paper never proves that, and the bound σ is never related to ϵ. Without such a modulus, H_ε may not be a small perturbation of the nominal system at all. So the advertised guarantee is not established.\n\nThe simulations are illustrative, but they are single runs without error bars, and the instability at the 0.07 s soft time limit is dismissed as 'poor numerics' without evidence. That's a minor point compared to the proof gap, but worth noting.\n\nWho should read this? Control practitioners interested in a hybrid interpretation of suboptimal MIQP may find the framework thought-provoking. But I would not rely on the stability theorem as written. The paper deserves a serious referee because the question is important and the hybrid framing is novel—it's exactly the kind of work that might mature into a solid contribution after the perturbation assumption is made precise and the Lyapunov argument is fixed. In its current form it should not be accepted; I'd lean toward major revision or reject, depending on how the authors respond.\n\nRecommendation: send to peer review only if the editor expects the authors to actually close the gap. If I were the editor, I'd be inclined to give it a chance with heavy revision rather than desk-reject.","headline":"New hybrid framing of suboptimal MIQP feedback control, but the main stability theorem relies on assumptions that assume the conclusion and a perturbation model that is not verified.","tokens_in":14459,"tokens_out":5518,"would_cite":false,"duration_ms":61317,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49M37","93C30","90C11","93D15"],"pacs":[],"model":"deepseek-v4-flash","headline":"Suboptimal MIQP solvers in feedback loops can still yield stable control, with error bounded by the solver's suboptimality.","keywords":["mixed-integer quadratic programming","hybrid systems","model predictive control","suboptimal solver parameters","feedback control","parametric optimization","practical stability"],"falsifier":"A concrete test: run the rendezvous MPC simulation from Section V with a soft time limit that frequently cuts off solver progress before any feasible iterate is produced, but keep the perturbation small enough to satisfy Assumption 5. If the observed trajectory violates the bound $|\\pi(t,j)|_\\mathcal{A} \\leq \\beta(|\\pi(0,0)|_\\mathcal{A}, t+j) + \\sigma$ (e.g., the objective function diverges for a time limit between the stable 0.05 s and the unstable 0.04 s), then Theorem 2's sufficient conditions fail to hold for that regime.","tokens_in":13314,"feed_emoji":"🎛️","tokens_out":1658,"duration_ms":21538,"temperature":0.7,"pith_summary":"This paper aims to prove that running a mixed-integer quadratic programming (MIQP) solver to only moderate accuracy, not to optimality, still produces a stable feedback controller for a dynamical system, provided the suboptimality is small. The authors frame the recursive MIQP solver and the controlled plant together as one hybrid system — flowing continuously while the plant evolves and jumping whenever a new control is computed. Their central result, Theorem 2, gives a bound on how far any trajectory of this hybrid system can wander from the ideal stabilizing set: it stays within a disturbance term sigma of a class-KL decay, so practical stability is preserved. The paper validates this with spacecraft rendezvous simulations using realistic solver limits (time, memory, iteration caps).","feed_headline":"Suboptimal MIQP solvers still stabilize feedback control","feed_subtitle":"A hybrid-systems proof bounds the error from imperfect solver output, so faster approximate solves can be used safely in MPC.","key_machinery":"The central object is the hybrid system $\\mathcal{H}_{\\mathrm{MIQP}}$ with state $\\chi = (y, x, \\tau)$: during flows, the plant $\\dot{x} = f(x, \\kappa(y))$ evolves and the timer $\\tau$ increases; at jumps, the solver returns $y^+ \\in \\psi_\\epsilon(x(\\tau), \\eta(x(\\tau)), \\epsilon)$, the $\\epsilon$-optimal set of the parametric MIQP. The key identity is the Lyapunov function $L = (\\phi_\\epsilon - V) + \\tfrac{1}{2} x^2$, which encodes both suboptimality and plant-state deviation, and the theorem that small $\\epsilon$ (suboptimal solver parameters) perturbs the hybrid system only slightly, so the semiglobal practical robustness result applies.","core_discovery":"The paper establishes that a recursive MIQP-based feedback controller, solved suboptimally by a black-box solver, can be modeled as a hybrid dynamical system with bounded perturbations. Their main theorem states that under regularity conditions (rational problem data, a closed stabilizing set, and a Lyapunov function bounded by class-K-infinity functions), warm-started suboptimal MIQP feedback keeps the closed-loop trajectories uniformly globally practically pre-asymptotically stable: every solution satisfies $|\\pi(t,j)|_\\mathcal{A} \\leq \\beta(|\\pi(0,0)|_\\mathcal{A}, t+j) + \\sigma$ for all $(t,j)$, where $\\sigma$ is the perturbation bound induced by the solver's suboptimality. The proof builds a Lyapunov function $L = (\\phi_\\epsilon - V) + \\tfrac{1}{2}x^2$ that combines deviation from optimality with deviation from the equilibrium, and then applies a robustness result for hybrid systems with small perturbations.","pith_inferences":["The paper leaves implicit a practical tuning rule: the threshold on solver suboptimality is not a fixed number but depends on the plant's stability margin, so a user could in principle estimate how much suboptimality the closed loop tolerates by probing the nominal Lyapunov decay rate.","The authors mention time-varying solver parameters only in passing; a direct extension would be to prove stability under an additional assumption that the suboptimality $\\epsilon_k$ decays along the trajectory, which would connect this framework to step-size-like scheduling for MIQP solvers.","The theory is developed for constant suboptimality $\\epsilon$; a testable prediction is that if a solver is run with a monotonically increasing iteration budget, the perturbation term $\\sigma$ in Theorem 2 should shrink to zero, which the simulations in Section V appear to support but do not formally verify.","Because the stability result relies on warm starting being feasible at every step, a practical extension would be to characterize how often warm-start infeasibility occurs in practice and how that would enter the perturbation bound."],"forward_implications":["If Theorem 2 is correct, practitioners can intentionally trade solution accuracy for computation speed: any solver parameter that degrades optimality only mildly (soft time limits, memory caps, iteration limits) will not break closed-loop stability, as long as the suboptimality stays below a threshold.","The hybrid-systems framing implies that analyses of MIQP feedback controllers can reuse the large toolbox of hybrid stability and robustness results, not just continuous-time or discrete-time tools.","The framework applies not only to MIQPs but, as the authors note, extends to recursive mixed-integer linear programs and indefinite quadratic programs with minor adjustments, expanding the class of optimization-based controllers that can be analyzed this way.","The simulation results suggest that even very tight solver budgets (e.g. a 0.05-second time limit or a 4 MB memory cap) can yield stable spacecraft rendezvous control, except in cases where the solver makes insufficient progress, which matches the theoretical requirement of small perturbation."],"supporting_citations":[{"why":"Supplies the parametric MIQP stability and upper semi-continuity results (rational matrix assumptions) that underpin Lemma 1 and Proposition 2.","marker":"[19]"},{"why":"Provides the hybrid systems framework, the basic conditions, the robustness theorem (Proposition 1), and the KL-stability definitions used in Lemma 3 and Theorem 2.","marker":"[13]"},{"why":"Defines the hybrid basic conditions and the relationship between outer semicontinuity and upper semicontinuity used in Lemma 2 and Theorem 1.","marker":"[12]"},{"why":"Establishes that feasible warm-started suboptimal MPC is asymptotically stable, used to justify Assumption 4.","marker":"[28]"},{"why":"Provides the suboptimal MPC simulation baseline and the observation that a too-short time limit destabilizes MPC, which the paper reproduces and explains.","marker":"[29]"},{"why":"Documents the specific solver parameters (soft time limit, soft memory limit, iteration limit) that the paper analyzes and tests.","marker":"[24]"},{"why":"Defines the $\\epsilon$-optimal set mapping and the parametric optimization regularity used in Eq. (1) and Lemma 1.","marker":"[21]"}],"fun_headline_variants":["Suboptimal MIQP solves yield practically stable MPC","Hybrid proof shows suboptimal MIQP control is stable","Warm-started suboptimal MIQP: hybrid stability for MPC","Suboptimal MIQP feedback stable via hybrid systems theory"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Assumption 4 states that the closed-loop system with the exact (optimal) MPC feedback is already asymptotically stable — without that, the theorem gives no guarantee for the suboptimal case, so the proof's force is in extending an existing stability property, not in creating stability from nothing.","fun_headline_variants_meta":{"raw":{"variants":["Suboptimal MIQP solves yield practically stable MPC","Hybrid proof shows suboptimal MIQP control is stable","Warm-started suboptimal MIQP: hybrid stability for MPC","Suboptimal MIQP feedback stable via hybrid systems theory"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001061,"raw_usage":{"total_tokens":4382,"prompt_tokens":807,"completion_tokens":3575,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":423,"completion_tokens_details":{"reasoning_tokens":3509}},"tokens_in":423,"tokens_out":3575,"duration_ms":31008,"temperature":1.0,"reasoning_tokens":3509,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:21:42.191692+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: run the rendezvous MPC simulation from Section V with a soft time limit that frequently cuts off solver progress before any feasible iterate is produced, but keep the perturbation small enough to satisfy Assumption 5. If the observed trajectory violates the bound $|\\pi(t,j)|_\\mathcal{A} \\leq \\beta(|\\pi(0,0)|_\\mathcal{A}, t+j) + \\sigma$ (e.g., the objective function diverges for a time limit between the stable 0.05 s and the unstable 0.04 s), then Theorem 2's sufficient conditions fail to hold for that regime.","supporting_citations":[{"cited_title":"Stability of mixed-integer quadratic programming problems","cited_arxiv_id":null,"evidence_quote":"Supplies the parametric MIQP stability and upper semi-continuity results (rational matrix assumptions) that underpin Lemma 1 and Proposition 2."},{"cited_title":"Hybrid feedback control","cited_arxiv_id":null,"evidence_quote":"Provides the hybrid systems framework, the basic conditions, the robustness theorem (Proposition 1), and the KL-stability definitions used in Lemma 3 and Theorem 2."},{"cited_title":"Hybrid dynamical systems","cited_arxiv_id":null,"evidence_quote":"Defines the hybrid basic conditions and the relationship between outer semicontinuity and upper semicontinuity used in Lemma 2 and Theorem 1."},{"cited_title":"Suboptimal model predictive control (feasibility implies stability)","cited_arxiv_id":null,"evidence_quote":"Establishes that feasible warm-started suboptimal MPC is asymptotically stable, used to justify Assumption 4."},{"cited_title":"Autonomous satellite rendezvous and proximity operations with time- constrained sub-optimal model predictive control","cited_arxiv_id":null,"evidence_quote":"Provides the suboptimal MPC simulation baseline and the observation that a too-short time limit destabilizes MPC, which the paper reproduces and explains."},{"cited_title":"Non-linear parametric optimization , volume 58","cited_arxiv_id":null,"evidence_quote":"Defines the $\\epsilon$-optimal set mapping and the parametric optimization regularity used in Eq. (1) and Lemma 1."}],"review_version":1}