{"id":"210d0e5e-f8aa-40eb-800c-66e978a65a50","arxiv_id":"2606.24641","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A suboptimal MPC scheme on a reduced-order model is proven to keep the full plant stable and feasible, provided the sampling interval and the fast unmodeled dynamics scale together to make the time-scale separation wide enough.","lead":"SMART-MPC combines two common shortcuts in real-time control—computing the control only approximately and using a simplified model of the plant—and proves that, if you sample fast enough and the neglected dynamics are fast enough, the closed-loop system still reaches its target. The paper gives a Lyapunov-based stability theorem and a robot-arm simulation that illustrates when the shortcuts work and when they break.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof mismatches the boundary-layer Lyapunov function with the equilibrium manifold used in Theorem 2; the fast-subsystem decrease is proven for deviations from ξ_eq(x,Π(z)), not from ξ_eq(x,Π(z*)).","rationale":"Stress-testing the main theorem, I looked for the least secure condition required for the central claim. The strongest competitor was Assumption 3's global contraction of a single SQP iteration, which is indeed not verified and is a serious applicability gap. But that concern attacks whether the theorem applies to the experiments; it does not attack the theorem under its stated assumptions. The coordinate mismatch in the proof of Theorem 1 attacks the theorem itself. In Section 4.2 the authors introduce ξ̃=ξ−ξ_eq(x,Π(z)) and prove a Lyapunov decrease for G(ξ̃)+κL(z̃). In Section 4.4 they then invoke Theorem 2 with the natural equilibrium manifold h(x)=(ξ_eq(x,Π(z*)),z*). For Theorem 2, ψ=y−h(x), so the corresponding U would be G(ξ−ξ_eq(x,Π(z*)))+κL(z−z*). These two functions are different; Lemma 2 does not prove the decrease of the latter. A reader cannot follow the proof as written. The gap is patchable — one can apply Theorem 2 after the coordinate change with h=0 — but the patch requires re-checking the Lipschitz, contraction, and forward-invariance conditions in the new coordinates, including a local version of (A.8). Since the paper does not do this, the correct recommendation remains CONDITIONAL: the theorem may be true, but the proof needs revision. The reader's identified weakest assumption (Assumption 3) is real but different; hence agreement_with_reader is 'disagree'.","tokens_in":27332,"tokens_out":22366,"duration_ms":206249,"concrete_test":"Re-derive the proof of Theorem 1 in the transformed state (x,ξ̃,z̃), applying Theorem 2 with h=0. Check: (i) Lemma 2's U is a Lyapunov function for the boundary layer y⁺=G(x,y,δ) with y=(ξ̃,z̃); (ii) Lemma 1 implies the required δ-Lipschitz bound on F(x,y,δ) after substituting ξ=ξ̃+ξ_eq(x,Π(z̃+z*)), with any additional constant absorbed; (iii) a localized version of (A.8) on Ω_V(γ), rather than X_0×Y, supports the Lyapunov decrease. If all three hold, Theorem 1 survives with a revised proof; if any fails, the central claim lacks a valid proof.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.4 applies the abstract Theorem 2 to the original closed loop (13) with fast equilibrium h(x)=(ξ_eq(x,Π(z*(x))), z*(x)). The boundary-layer variable in Theorem 2 is ψ=y−h(x), so the Lyapunov function should be G(ξ−ξ_eq(x,Π(z*)))+κL(z−z*). Lemma 2 instead proves decrease of U=G(ξ̃)+κL(z̃) for ξ̃=ξ−ξ_eq(x,Π(z)), i.e., the deviation from a z-dependent equilibrium. These differ whenever z̃=z−z*(x)≠0 because ξ_eq(x,Π(z̃+z*))≠ξ_eq(x,Π(z*)). Thus condition (A.6) is not actually established for the system to which Theorem 2 is applied. The gap is repairable by reformulating the full system in coordinates (x,ξ̃,z̃) with h=0 and re-verifying Lemma 1's bounds after the coordinate change, but this is not what the paper does. Independently, (A.8) is checked only on Ω_V(γ), not on the product X_0×Y required by Theorem 2. As written, the proof of Theorem 1 is incomplete at its central step.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SMART-MPC, a framework for suboptimal and reduced-order nonlinear model predictive control of continuous-time interconnected systems. The plant is a two-timescale system (1) with target state x and fast extra state ξ; the MPC optimizer is modeled as a discrete-time dynamical system (12) that performs a generic optimization algorithm for a finite number of iterations. The controller uses a reduced-order prediction model that constrains the extra dynamics to its equilibrium manifold. The main result, Theorem 1, claims that under Assumptions 1–4, for every r>0 and γ∈(0,γmax), there exist δ̄,κ̄ such that for all sufficiently small sampling intervals δ with ε=rδ and κ≥κ̄, the closed-loop system (13) is recursively feasible and exponentially stable at (x⋆, ξ_eq(x⋆,Π(z⋆(x⋆))), z⋆(x⋆)). The proof uses a singular-perturbation decomposition: Lemma 1 establishes Lipschitz and sampling bounds, Lemma 2 treats the boundary-layer system, Lemma 3 treats the reduced system, and Theorem 2 from Appendix A is invoked to combine them. The theoretical claims are complemented by MuJoCo experiments on an underactuated PenduBot.","tokens_in":27767,"tokens_out":7305,"duration_ms":71565,"significance":"If the result holds, the paper provides a useful unifying framework for two practically important MPC approximations—suboptimal optimization and reduced-order prediction—by treating the sampling time as a tunable parameter and the optimizer as part of the closed-loop dynamics. The proof strategy is conceptually appealing and the assumptions are stated explicitly rather than hidden. The paper also benefits from a disclosed preliminary version [36], a re-proved and generalized prior result [41], and extensive virtual experiments. However, several load-bearing gaps in the current manuscript prevent the theorem from being accepted as proven. The most serious is a coordinate mismatch in the application of Theorem 2: Lemma 2 proves decrease in coordinates relative to the z-dependent equilibrium ξ_eq(x,Π(z)), while Theorem 2 requires a boundary-layer deviation from the manifold h(x)=(ξ_eq(x,Π(z⋆(x))), z⋆(x)). There is also an unstated identification between the MPC internal prediction model F^R_MPC and the exact sampled reduced map f^R, and a locality gap in the verification of condition (A.8). These issues are substantial but appear repairable, so the manuscript is not beyond the scope o","major_comments":[{"comment":"The application of Theorem 2 uses the wrong boundary-layer coordinate. Theorem 2 (Appendix A) requires the equilibrium manifold h(x)=(ξ_eq(x,Π(z⋆(x))), z⋆(x)) and the boundary-layer variable ψ=y−h(x); condition (A.6) must hold for U(ψ). However, Lemma 2 proves decrease of U(ξ̃,z̃) with ξ̃=ξ−ξ_eq(x,Π(z)) (Eqs. (25), (28)), i.e., the deviation from a z-dependent equilibrium, not from ξ_eq(x,Π(z⋆(x))). These two deviations differ whenever z−z⋆(x)≠0. Consequently, the claimed decrease of V in Eq. (19), stated in Eq. (34), does not follow from Theorem 2 as written. This is the central step of the proof of Theorem 1 and must be repaired, for example by carrying out the singular-perturbation argument in coordinates (x, ξ̃, z̃) with the equilibrium at the origin and re-deriving the cross-coupling bounds, or by proving a modified theorem with a state-dependent Lyapunov function.","section":"Section 4.4 and Appendix A"},{"comment":"The paper identifies the MPC internal prediction model F^R_MPC in Eq. (11) with the exact sampled reduced dynamics f^R in Eq. (23), but these are different objects. The feasible set X_T in Eq. (14) is defined through F^R_MPC, a discrete-time approximation of the reduced continuous-time dynamics, while Lemma 3 and the reduced system (30) use f^R, the exact sample map. Since the optimizer state z⋆(x) and the algorithm A in Assumption 3 solve problem (11) using F^R_MPC, the reduced closed-loop dynamics are not f^R(x,Π(z⋆(x)),δ) unless F^R_MPC≡f^R. No consistency condition or approximation-error bound is provided. The simulations use RK4 with δ_MPC=0.07 s, so this is not merely a formal distinction. This gap must be addressed, either by assuming exact discretization or by including the discretization mismatch explicitly in the analysis.","section":"Section 3.1 vs. Section 4.3"},{"comment":"Condition (A.8) of Theorem 2 requires f(x,y,δ)∈X_0 for all (x,y)∈X_0×Y. In the paper, the verification in Eqs. (31)–(33) is performed only on Ω_V(γ), not on the full product X_T×Ξ×Z required by the theorem. The argument then invokes Theorem 2 to obtain forward invariance of Ω_V(γ), but forward invariance is exactly what is needed to justify restricting the verification to Ω_V(γ). As written, the proof is circular. A local version of Theorem 2 that establishes invariance and the Lyapunov decrease simultaneously on a sublevel set would be needed, or the condition (A.8) must be verified on the full domain.","section":"Section 4.4, condition (A.8)"},{"comment":"Assumption 3 requires a unique globally exponentially stable equilibrium z⋆(x) for the optimizer dynamics A(z,x) for every x∈X_T, with a known Lyapunov function L satisfying (16). This is a very strong property, especially for the nonconvex OCP (11). The experiments in Section 5 use a single SQP/RTI iteration per sampling instant, but the paper provides no verification—analytical or numerical—that this algorithm satisfies Assumption 3, nor even that the relevant fixed point and contraction exist over the operating region. Since real-time SQP schemes are typically only locally contractive, the experimental demonstration does not validate the theorem in the regime claimed. The paper should either prove or numerically check the contraction property for the RTI scheme, or state the experimental section as an illustration rather than a validation of Assumption 3.","section":"Assumption 3 vs. Section 5"}],"minor_comments":[{"comment":"The notation f^R_MPC and f^R is easy to confuse. Consider renaming the exact sampled reduced map, e.g., f^R_exact, to emphasize that Eq. (23) is not the MPC prediction model used in Eq. (11).","section":"Notation, Eq. (23)"},{"comment":"The statement 'we retain only the zeroth-order term in the Taylor expansion and bound the remainder linearly in δ' is imprecise. For a rigorous bound one should write the integral form of the remainder and give the explicit bound in terms of δ and the Lipschitz constants.","section":"Appendix B, Eq. (B.4)"},{"comment":"The recursive feasibility conclusion says Ω_V(γ) is forward invariant, implying x_t∈X_T. This is correct because V≥W and γ<γ_max, but the implication is not explicitly justified; a one-sentence clarification would help.","section":"Theorem 1 statement"},{"comment":"The experimental plots are informative, but the figures do not show the sampling instants or the actual MPC update times. Given that the paper emphasizes the sampling time δ as a design parameter, indicating the discretization grid would improve readability.","section":"Figures 7–10"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising and well-motivated framework, and the proof is detailed, but the central application of Theorem 2 suffers from a coordinate mismatch that breaks the claimed derivation of Eq. (34). The F^R_MPC versus f^R identification and the local verification of (A.8) are additional load-bearing gaps. These issues appear repairable, but they require substantial reworking of Section 4.4 and the related statements. The experimental section also needs to address whether the single-iteration SQP scheme actually satisfies Assumption 3. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What to know: this paper proposes SMART-MPC, a unified timescale-separation analysis for suboptimal MPC on reduced-order models. Treating the sampling time and the fast-dynamics scale through the tuning relation epsilon = r*delta is a genuinely neat idea, and the combination of suboptimal optimization and model reduction in one Lyapunov-based guarantee is new. The writing is clear and the authors disclose the preliminary version and the prior theorem they build on.\n\nThe central theorem is plausible but the proof as written does not go through. The stress-test note holds up on reading. Lemma 2 establishes decrease of U in coordinates (xi_tilde, z_tilde) where xi_tilde = xi - xi_eq(x, Pi(z)) and z_tilde = z - z*(x). But Theorem 2, as applied, requires a Lyapunov function on the deviation from the fast equilibrium h(x) = (xi_eq(x, Pi(z*(x))), z*(x)), i.e. on (xi - xi_eq(x, Pi(z*)), z - z*). These are different whenever z differs from z*(x). So condition (A.6) has not been verified for the boundary-layer system to which Theorem 2 is actually applied. The gap is likely repairable by rewriting the system in coordinates with h=0 and re-checking the necessary Lipschitz bounds, but that rewriting is not in the paper. A second related issue: condition (A.8) is checked only on Omega_V(gamma), not on the full product set required by the abstract theorem.\n\nThe assumptions are also very strong. Assumption 3 demands a globally exponentially stable optimizer iteration with known Lyapunov function for every x in X_T; the experiments use a single SQP iteration per sample, and no evidence is given that it satisfies this contraction property. Similarly, the analysis identifies the MPC prediction model with the exact sampled reduced map f^R, while the simulations use RK4 with delta_MPC = 0.07 s; the mismatch is unquantified. These are patchable, but they mean the theorem is conditional and the simulations do not verify its hypotheses.\n\nWho gets value: researchers working on Lyapunov-based analysis of suboptimal or approximate MPC, and anyone using singular perturbation arguments in sampled-data control. The paper deserves a serious referee: the idea is significant and the proof can likely be fixed. It should not be desk-rejected, but it needs a substantive revision addressing the Theorem 2 application and either weakening or verifying Assumption 3.","headline":"Good idea, honest framework, but the proof of Theorem 1 has a real gap in applying the two-time-scale theorem, and the experiments don't verify the strong assumptions.","tokens_in":28194,"tokens_out":3841,"would_cite":false,"duration_ms":40633,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C70","93D30","93C55"],"pacs":[],"model":"deepseek-v4-flash","headline":"Sampling fast enough converts both optimization error and model mismatch into fast transients, so a suboptimal, reduced-order MPC keeps recursive feasibility and exponential stability.","keywords":["Model Predictive Control","suboptimal MPC","reduced-order model","timescale separation","singular perturbation","recursive feasibility","exponential stability","real-time iteration"],"falsifier":"Run the closed loop on the two-link arm with a fixed inductance L (so ε is fixed by hardware) while decreasing δ: if recursive feasibility and exponential stability continue to hold with no lower bound on δ, then the asserted ε=rδ mechanism is not necessary. Conversely, measure the one-step Lyapunov decrease of the optimizer alone on the nonconvex swing-up problem; if for some state in the claimed invariant set the function L(A(z,x)−z⋆(x))−L(z−z⋆(x)) is not negative, Assumption 3 fails and Theorem 1 cannot be the explanation for any observed stability.","tokens_in":27179,"feed_emoji":"⚙️","tokens_out":8059,"duration_ms":75044,"temperature":0.7,"pith_summary":"The paper's aim is to prove that a model predictive controller can operate with two cost-saving shortcuts at once — solving the underlying optimal control problem only approximately, and predicting with a reduced-order model that ignores part of the plant — and still keep the closed loop recursively feasible and exponentially stable. The key move is to treat the sampling interval not as an implementation detail but as a design parameter that induces timescale separation: the ignored fast dynamics are scaled so that both they and the optimizer's internal iterations settle much faster than the slow target state moves. The main theorem says that for any certified sublevel set of the closed-loop Lyapunov function, fast enough sampling and a sufficiently contractive optimizer give forward invariance (so feasibility never breaks) and exponential convergence to the combined equilibrium. The practical message is that model fidelity and optimization accuracy can be traded for sampling speed, with explicit constants describing how fast the sampling must be. Virtual experiments on an underactuated two-link arm support the predicted threshold behavior: fast sampling matches ideal benchmarks, slow sampling degrades into oscillations.","feed_headline":"Fast sampling makes suboptimal MPC provably stable","feed_subtitle":"A single optimizer step and a reduced model can suffice when the plant is sampled fast enough.","key_machinery":"The load-bearing object is the composite Lyapunov function V(x,ξ,z)=W(x)+G(ξ−ξ_eq(x,Π(z)))+κL(z−z⋆(x)), where W certifies stability of the reduced-order MPC closed loop, G certifies exponential contraction of the neglected dynamics to its equilibrium manifold, and L certifies contraction of the optimizer iterations to their fixed point; κ balances the fast and slow parts. Lemma 2 shows this combination is a Lyapunov function for the boundary-layer system, Lemma 3 gives a δ-scaled Lyapunov decrease for the reduced system, and a packaged two-timescale stability theorem in the appendix assembles them into a decrease of V along the full closed loop. The role of δ as a tunable small parameter is","core_discovery":"The central claim is Theorem 1: under four assumptions — Lipschitz plant, a globally stable equilibrium manifold for the neglected dynamics, a globally contractive optimizer with a known Lyapunov function, and a Lyapunov function for the reduced-model closed loop — for every r>0 and every sublevel set inside the feasible region there exist δ̄ and κ̄ such that with sampling period δ<δ̄, fast scale ε=rδ, and optimizer Lyapunov weight κ≥κ̄, the closed loop retains recursive feasibility and the equilibrium (target setpoint, optimizer fixed point, ignored state on its manifold) is exponentially stable. The proof works by writing the sampled-data closed loop as a two-timescale system in which the","pith_inferences":["Editorial inference: the same boundary-layer argument suggests a practical 'fastness ratio' — actuator bandwidth times sampling period — as a single number that determines whether suboptimal reduced-order MPC is safe; the paper does not phrase it this way, but it is a direct corollary of ε=rδ.","Editorial inference: Assumption 3 asks for a global Lyapunov contraction of the optimizer, but the proof only ever uses it along trajectories inside the certified invariant set; a local contraction in that set, verified numerically for the RTI/SQP scheme, would likely close the gap between theorem and experiment.","Editorial inference: the result could be turned into a self-testing MPC: the optimizer can estimate its own one-step contraction and the plant can estimate the fast dynamics' decay rate online, then adjust δ or κ to satisfy the key inequality, yielding an adaptive version of the scheme."],"forward_implications":["A controller that runs only one cheap optimizer iteration per sample and predicts with a model that ignores the motor/electrical dynamics can be certified stable provided the optimizer is contractive, by choosing the sampling period small enough; the experiments show parity with fully optimal MPC at δ=0.005 s.","Recursive feasibility becomes a provable property of the sampling rate: once the closed loop starts inside the certified sublevel set, the reduced MPC problem never becomes infeasible, even though the optimizer never solves the problem exactly.","The scaling ε=rδ gives an explicit design relation: the neglected physical dynamics must be at least as fast as the sampling mechanism for the guarantees to hold; when the fast time constant is fixed, there is an upper bound on δ before stability is lost.","The proof yields concrete Lyapunov-decrease inequalities that could be checked numerically offline, giving a certificate for a specific controller-plant pair before deployment.","For plants with genuinely slow parasitic dynamics, the theorem predicts — and the Monte Carlo results illustrate — that stability degrades gradually into oscillation rather than sudden failure, consistent with a shrinking boundary-layer margin."],"fun_headline_variants":["Fast sampling proves suboptimal MPC stable","Suboptimal MPC: fast sampling is the fix","Reduced-order MPC stabilized by fast sampling","Timescale separation enables suboptimal MPC","Suboptimal MPC: sample fast, stay stable"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the generic optimization algorithm has a globally exponentially stable fixed point with a known Lyapunov function satisfying a fixed quadratic decay (Assumption 3); for the experiments this is merely assumed for a single SQP iteration on a nonconvex problem, and the proof also needs the specific scaling ε=rδ rather than a fixed fast time constant.","fun_headline_variants_meta":{"raw":{"variants":["Fast sampling proves suboptimal MPC stable","Suboptimal MPC: fast sampling is the fix","Reduced-order MPC stabilized by fast sampling","Timescale separation enables suboptimal MPC","Suboptimal MPC: sample fast, stay stable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000964,"raw_usage":{"total_tokens":3931,"prompt_tokens":722,"completion_tokens":3209,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":3143}},"tokens_in":466,"tokens_out":3209,"duration_ms":21643,"temperature":1.0,"reasoning_tokens":3143,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T10:20:11.070211+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the closed loop on the two-link arm with a fixed inductance L (so ε is fixed by hardware) while decreasing δ: if recursive feasibility and exponential stability continue to hold with no lower bound on δ, then the asserted ε=rδ mechanism is not necessary. Conversely, measure the one-step Lyapunov decrease of the optimizer alone on the nonconvex swing-up problem; if for some state in the claimed invariant set the function L(A(z,x)−z⋆(x))−L(z−z⋆(x)) is not negative, Assumption 3 fails and Theorem 1 cannot be the explanation for any observed stability.","supporting_citations":[],"review_version":2}