{"id":"e6ca9d18-1f87-43c2-b151-23c8fab2d736","arxiv_id":"2505.20090","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A receding-horizon optimizer over two funnel parameters is layered on top of a finite-time funnel feedback law, with proofs of recursive feasibility and asymptotic stabilization.","lead":"This paper proposes a control scheme that lets a predictive optimizer tune the shape of a funnel-shaped error boundary, while a separate high-gain feedback law keeps the output inside that boundary. The scheme solves a two-variable optimization at each step, and the paper proves feasibility, bounded cost, and convergence to the equilibrium.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 8's contradiction assumes a uniform-in-c bound on alpha_c(4/9); the stated hypotheses allow this quantity to blow up on a bounded c-interval, so the theorem is not proved as stated.","rationale":"The reader's weakest-assumption choice was Lemma 3, the external finite-time funnel-control result from [8]. That is certainly load-bearing: if it failed, the equilibrium endpoint argument and the control-bound argument would both collapse. However, Lemma 3 is a published external theorem, and nothing internal to this paper gives a concrete reason to doubt it. The more immediately actionable concern is the unstated uniformity of the family alpha_c in Theorem 8. The proof of Theorem 8 requires a finite upper bound on alpha_c evaluated at a fixed argument (4/9 or 64/81) uniformly over all active parameters that can occur. Corollary 7 gives only an upper bound on c_i^*, and the assumptions on alpha_c are per-c bijections with no continuity or equicontinuity in c. Since the active parameters can in principle accumulate at any point of the allowed interval, the required uniform bound is not guaranteed by the stated hypotheses. I constructed an allowed family where alpha_c(4/9) is unbounded on a bounded c-interval, so the theorem as stated is incomplete. The concrete family alpha_c(s) = 2c/(1-s) used in the numerical example does satisfy the needed uniform bound, and the rest of the proof is largely coherent modulo this extra hypothesis and minor typographical issues. Thus the paper merits revision rather than rejection: add the regularity assumption to Theorem 8 and Theorem 5's setup, and the central stabilization claim should hold for the intended class of alpha-families. This does not change the reader's CONDITIONAL verdict, but it sharpens the condition under which the paper should be accepted.","tokens_in":11153,"tokens_out":15977,"duration_ms":174784,"concrete_test":"Verify the gap analytically by taking alpha_c(s) = 2c + (s/(1-s)) / |c-1| for c != 1 and alpha_1(s) = 2 + s/(1-s); this family satisfies every condition stated in Section 3.1, but sup_{c in [1/2,3/2]} alpha_c(4/9) = infinity. Rerun the proof of Theorem 8 for this family: the bound in Step 3 becomes infinite for active parameters approaching c = 1, so the contradiction does not follow. Then amend Theorem 8 with the explicit assumption sup_{c in [c_min,c_max]} alpha_c(4/9) < infinity (or joint continuity of alpha_c in c) and check that the paper's example alpha_c(s) = 2c/(1-s) satisfies this condition. If the amended proof goes through, the original theorem statement still needs this added hypothesis.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The decisive step in the stability proof is the control bound in Theorem 8, Step 3 and Step 4. To contradict the unboundedness of the input, the proof uses bounds of the form ||u(tcrit)|| <= Nhat(alpha_c(4/9)) * 2/3 and ||u(\\hat tcrit)|| <= Nhat(alpha_c(64/81)) * 8/9, where c is the currently active funnel parameter. Corollary 7 only shows that the set {c_i^*} is bounded above; it gives no lower bound, and the stated assumptions on alpha_c are made pointwise in c: each alpha_c is merely a continuous bijection [0,1) -> [2c,infty), with no joint continuity or uniform local boundedness in c. The stated hypotheses therefore do not imply sup_{c in (0,C]} alpha_c(4/9) < infinity. An explicit allowed family is alpha_c(s) = 2c + (s/(1-s)) / |c-1| for c != 1 and alpha_1(s) = 2 + s/(1-s); each alpha_c is a continuous bijection with the required range, but alpha_c(4/9) tends to infinity as c -> 1. If a run of Algorithm 4 had active parameters accumulating near such a blow-up point, the purported contradiction in Theorem 8 would disappear. This is a statement-level gap in the theorem as written, not a counterexample to the concrete scheme with alpha_c(s) = 2c/(1-s), for which the needed uniform bound does hold. The fix is to add an explicit regularity assumption on the family (alpha_c), for example joint continuity of (c,s) |-> alpha_c(s) on compact c-intervals, or directly sup_{c in [c_min,c_max]} alpha_c(4/9) < infinity.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes model predictive funnel control (MPFC), a two-layer control scheme in which an outer MPC layer updates, at each sampling instant, the parameters (c,T) of a time-varying funnel boundary, while an inner finite-time funnel controller (from [8]) drives the output to zero within that boundary. The main theoretical results are Theorem 5 (initial and recursive feasibility, plus funnel constraint satisfaction), Lemma 6 (boundedness of the closed-loop infinite-horizon cost), Corollary 7 (boundedness of the optimized c-values), and Theorem 8 (asymptotic stability under positive definite Q). The proofs use the finite-time funnel control lemma from [8] and assume an optimization oracle that returns a strict upper bound on the optimal value at every step.","tokens_in":11485,"tokens_out":13957,"duration_ms":134905,"significance":"If the gap identified below in Theorem 8 is repaired, the paper makes a useful contribution: it combines model-free funnel control with model-based predictive optimization while keeping the number of optimization variables independent of the horizon, and it provides rigorous feasibility, cost-bound, and stability arguments. The proof of Theorem 8 is an intricate contradiction argument based on crossing times, and Lemma 6's telescoping of the cost is clean and correct. The main computational premise, the strict upper-bound oracle, is stated explicitly in Algorithm 4 rather than hidden, which is a strength. The numerical example illustrates the proposed scheme, though it is not a systematic comparison.","major_comments":[{"comment":"The contradiction in Theorem 8 requires a bound on N-hat(alpha_c(4/9)) and N-hat(alpha_c(64/81)) that is uniform over the active parameters c = c_i^*. The stated hypotheses only require each alpha_c to be a continuous bijection [0,1) -> [2c,infty) pointwise in c, and Corollary 7 only establishes that {c_i^*} is bounded above, with no lower bound. An allowed family is alpha_c(s) = 2c + (s/(1-s))/|c-1| for c != 1 and alpha_1(s) = 2 + s/(1-s); each alpha_c is a continuous bijection with the required range, but alpha_c(4/9) tends to infinity as c -> 1, so the bound in Step 3 can blow up along a sequence of active parameters. The proof therefore does not establish the claimed contradiction as stated. I recommend adding an explicit uniformity assumption on the family (alpha_c), for example joint continuity of (c,s) |-> alpha_c(s) on compact c-intervals, or directly sup_{c in [c_min,c_max]} alpha_c(4/9) < infinity. This uniformity holds for the concrete choice alpha_c(s) = 2c/(1-s) used in the numerical section.","section":"Theorem 8, Steps 3–4"},{"comment":"The candidate construction in Step 1 divides by ||dot_psi|_{[t-hat, t-hat+H]}||_infty, which is zero whenever psi is constant on that interval (for example, psi == const > 0). In that case the expression for T-hat is undefined, so the proof does not cover all admissible outer funnel functions. The feasibility claim itself remains true, but the proof needs a separate case for ||dot_psi||_infty = 0, or an alternative construction such as T-hat = H and c-hat = (psi(t-hat) + ||y-hat||)/(2H).","section":"Theorem 5, Step 1"}],"minor_comments":[{"comment":"There are two notation typos in the proof: the derivative in the mean-value theorem step should be applied to ||y_cl(t-hat_crit)||, not to ||y_cl(t_crit)||, and the expression 82/92 should be read as (8/9)^2 = 64/81.","section":"Theorem 8, Step 4"},{"comment":"The sums defining y_cl, u_cl, and phi_cl start at i=1, which omits the first sampling interval [t_0,t_1) = [0,h). They should start at i=0, or the interval [0,h) should be included separately, to match the definition of y_i on [t_i,t_{i+1}] for i in N_0.","section":"Section 4, closed-loop definitions"},{"comment":"The algorithm assumes that at every step an oracle returns a strict upper bound V-hat_H(t_i,y_i) > V_H(t_i,y_i). This is a genuine computational premise; the paper could usefully comment on how such a bound might be obtained in practice, for example by evaluating a feasible suboptimal candidate, and on what happens if the oracle only provides an approximate bound.","section":"Algorithm 4"}],"recommendation":"major_revision","confidential_remarks":"The main theorem is currently conditional on an unstated uniformity condition on the family (alpha_c). This is fixable with an additional assumption, and the concrete scheme in the numerical example satisfies it. The paper's reliance on [8, Thm. 3.1] is appropriate but means the novelty lies mainly in the MPFC architecture and the stability proof rather than in the finite-time funnel lemma itself."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely new scheme, not a repackaging of funnel MPC. The authors optimize two funnel parameters (c,T) per receding-horizon step and let the inner funnel controller generate the input, which cuts the decision variables to two regardless of horizon length. That is a real departure from [18-20], which keep a static funnel and optimize control values.\n\nThe paper does solid work on the feasibility side. Theorem 5 gives explicit feasible candidates and the recursive argument via shifting (c,T) works. Lemma 6 telescopes correctly and gives the bounded closed-loop cost. The stability proof in Theorem 8 follows the standard terminal-equality route, and the contradiction structure is clever.\n\nThe soft spots are real but repairable. The main one is in Theorem 8, Steps 3 and 4: the control bounds use Nhat(alpha_c(4/9)) and Nhat(alpha_c(64/81)) assuming these are finite uniformly over the active parameter c. The stated assumptions only give pointwise continuous bijections alpha_c for each c; there is no joint continuity or uniform local boundedness in c. The stress-test example alpha_c(s) = 2c + (s/(1-s))/|c-1| (with a limit at c=1) satisfies every stated hypothesis but makes alpha_c(4/9) blow up as c->1. So the theorem as written is not proved. That said, the concrete family alpha_c(s)=2c/(1-s) satisfies the needed bound, so this is a gap in the statement, not a counterexample to the method. The fix is straightforward: add joint continuity on compact c-intervals, or directly assume sup over the relevant c-range of alpha_c(4/9) is finite.\n\nSecond, Algorithm 4 relies on an oracle that returns a strict upper bound Vhat_H > V_H and finds feasible points under it. No concrete optimizer is given. For a theory paper this is acceptable if flagged, but they should be clearer that the guarantees are conditional on that oracle.\n\nThird, the claimed computational advantage is not benchmarked; the numerical example is one trajectory, and the paper itself defers performance studies. That is fine for a brief paper, but it keeps significance from being higher than reasonable.\n\nMinor edge cases: in Theorem 5 Step 1, if psi_dot is identically zero, the candidate T divides by zero; use a max or separate case. There are also a few notation typos in Theorem 8 (t_crit vs hat t_crit).\n\nOverall: the central idea is sound and the gap is patchable. This paper is for readers working on funnel control or MPC for constrained nonlinear systems. It deserves a serious referee and, after a revision that fixes the alpha_c regularity assumption and discusses the oracle, I would support publication.","headline":"A genuinely new funnel-MPC hybrid with solid feasibility/cost results and a fixable regularity gap in the stability theorem.","tokens_in":12077,"tokens_out":2814,"would_cite":true,"duration_ms":25819,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C10","93D15","93C40"],"pacs":[],"model":"deepseek-v4-flash","headline":"Model predictive funnel control tunes its error tube online and proves that the closed-loop output converges to zero.","keywords":["model predictive control","funnel control","prescribed transient behavior","nonlinear output feedback","adaptive control","stability","recursive feasibility","equilibrium endpoint constraints"],"falsifier":"A direct check: apply the feedback law (4) with $\\alpha_c(s)=2c/(1-s)$ and $N(s)=s$ to the example system (18) from inside the funnel, and measure $\\|y(T)\\|$ and $\\|u\\|_\\infty$; a nonzero $\\|y(T)\\|$ or an unbounded input would falsify Lemma 3 and therefore Theorems 5 and 8. Running Algorithm 4 on the same example with $Q=I_2$ should give $\\|y_{\\mathrm{cl}}(t)\\to 0$; failure of that convergence would falsify Theorem 8.","tokens_in":10845,"feed_emoji":"🎛️","tokens_out":10566,"duration_ms":66586,"temperature":0.7,"pith_summary":"The paper introduces model predictive funnel control (MPFC), a two-layer controller in which a lower-level high-gain funnel law keeps the output inside a shrinking error tube, and an upper-level model predictive optimizer reshapes the tube at every sampling step by choosing just two parameters, the funnel slope $c$ and the end time $T$. The authors prove that the optimization is always feasible—initially and at every subsequent step—and that with a positive-definite state cost the closed-loop output asymptotically converges to zero. The central trick is a funnel boundary that reaches zero in finite time: the exact finite-time tracking property of the lower-level law supplies the equilibrium endpoint constraint that MPC stability theory needs, without adding a hard constraint to the optimization. A reader should care because the decision-variable count stays independent of the prediction horizon, inter-sampling behavior is explicitly controlled, and the scheme links the model-free robustness of funnel control to the predictive quality of MPC.","feed_headline":"Funnel controller tunes its own error tube and provably converges","feed_subtitle":"Model predictive funnel control tunes two parameters per step, yet guarantees feasibility and stability.","key_machinery":"The load-bearing object is the finite-time funnel pair $(c,T)$ defining $\\varphi(t;c,T)=c(T-t)$, together with the feedback law $u(t)=(N\\circ\\alpha_c)(\\|y(t)\\|^2/\\varphi(t)^2)\\, y(t)/\\varphi(t)$, where $\\alpha_c:[0,1)\\to[2c,\\infty)$ is a bijection governing how quickly gain rises and $N:\\mathbb{R}_{\\ge0}\\to\\mathbb{R}$ is a surjection (the paper suggests $\\alpha_c(s)=2c/(1-s)$ and $N(s)=s\\cos s$). Lemma 3, taken from [8], guarantees that this law drives the output of every system in class $\\mathcal{S}$ to zero exactly at time $T$ with bounded input, starting from any initial output inside the funnel. That exact finite-time property plays the role of the terminal equality constraint in MPC stability theory [17]: the horizon can be extended at zero cost, which produces the inequality in Lemma 6 that bounds each sample-interval cost by the drop in the horizon cost. Corollary 7 turns the bounded closed-loop cost into boundedness of the sequence of active parameters $c_i$. Theorem 8 then argues by contradiction: if the output norm kept returning to a fixed $\\varepsilon$, the funnel law would have to generate unbounded control at the times when the output crosses $\\varepsilon/3$ and $2\\varepsilon/3$, but bounded $c_i$ and the distance-to-boundary structure of (4) give a uniform bound on the control, a contradiction.","core_discovery":"The paper's claim, stated on its own terms, is that receding-horizon optimization of the funnel parameters $(c,T)$ over a fixed horizon $H$ stabilizes every system in the class $\\mathcal{S}$. Theorem 8 makes this precise: under the assumptions of Theorem 5 and with $Q$ positive definite, $\\lim_{t\\to\\infty} y_{\\mathrm{cl}}(t) = 0$, so Algorithm 4 asymptotically stabilizes system (1). Theorem 5 establishes initial and recursive feasibility for every initial output with $\\|y(0)\\|<\\psi(0)$ and shows that the closed-loop output remains inside the funnel boundary on each sampling interval. Because the funnel boundary $\\varphi(t;c,T)=c(T-t)$ vanishes at $T$, Lemma 3 gives exact convergence to the equilibrium at the horizon end for free, which converts the classical MPC terminal-equality argument into a cost-decrease argument. The same proof yields bounded infinite-horizon closed-loop cost and bounded optimized parameters, and these bounds feed the contradiction argument that rules out any persistent deviation from zero.","pith_inferences":["One consequence the paper leaves implicit is that the two-variable parameterization makes the online optimization independent of the horizon length in a way that could scale to long horizons; a benchmark against classical MPC on computation time would test how much of that promise is realized in practice.","Because the terminal equilibrium is enforced by the feedback law rather than by an explicit endpoint constraint, the same architecture might extend to time-varying reference tracking if the finite-time exact-tracking result is replaced by an analogous moving-target version.","The stability proof uses only two structural ingredients—positive definiteness of $Q$ and a uniform control bound from bounded $c_i$ and distance-to-boundary—so the same argument should apply to any parameterized feedback family sharing those properties.","A testable extension is to relax the outer funnel $\\psi$ or the strict upper-bound requirement $\\hat V_H > V_H$, for instance by penalizing the funnel parameters in the cost; the feasibility construction in Theorem 5 only needs $\\psi$ smooth and positive to build a candidate funnel."],"forward_implications":["Algorithm 4 is initially and recursively feasible for every system in class $\\mathcal{S}$ with initial output inside the outer funnel, so the optimization can never get stuck at a sampling instant.","With $Q$ positive definite, the closed-loop output converges to zero, so the scheme asymptotically stabilizes the system while staying inside the user-prescribed error boundary at all times.","The infinite-horizon closed-loop cost is bounded by the first horizon cost, and the optimized funnel parameters $c_i$ remain bounded, so the controller's effort stays under control.","The optimization has exactly two decision variables, $c$ and $T$, regardless of the horizon length $H$, so the per-step problem size does not grow when predictions are made further ahead.","Between sampling instants the funnel feedback law continues to act, so constraint satisfaction and inter-sampling behavior are explicitly handled rather than left open-loop."],"supporting_citations":[{"why":"Supplies the exact finite-time tracking result used as Lemma 3, making the equilibrium endpoint constraint automatic.","marker":"[8]"},{"why":"Provides the MPC terminal-equality stability theory that the cost-decrease and convergence argument is modeled on.","marker":"[17]"},{"why":"Defines the high-gain property and the system class $\\mathcal{S}$ used throughout the paper.","marker":"[2]"}],"fun_headline_variants":["Funnel MPC self-tunes error bounds with provable stability","Two-parameter MPC funnel that provably stabilizes its class","Receding-horizon funnel control: stability without horizon-sized optimization","Model predictive funnel: fewer variables, same guarantees, proven convergence","Self-tuned error funnel with recursive feasibility and exact convergence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument rests on the borrowed finite-time exact-tracking guarantee—that for every system in class $\\mathcal{S}$ the funnel law (4) drives the output to zero exactly at time $T$ with bounded input when the initial output lies inside the funnel—and if that guarantee failed for any admissible system, both the feasibility proof and the stability proof would collapse.","fun_headline_variants_meta":{"raw":{"variants":["Funnel MPC self-tunes error bounds with provable stability","Two-parameter MPC funnel that provably stabilizes its class","Receding-horizon funnel control: stability without horizon-sized optimization","Model predictive funnel: fewer variables, same guarantees, proven convergence","Self-tuned error funnel with recursive feasibility and exact convergence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000414,"raw_usage":{"total_tokens":2113,"prompt_tokens":890,"completion_tokens":1223,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":1137}},"tokens_in":506,"tokens_out":1223,"duration_ms":10003,"temperature":1.0,"reasoning_tokens":1137,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:06:01.772100+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct check: apply the feedback law (4) with $\\alpha_c(s)=2c/(1-s)$ and $N(s)=s$ to the example system (18) from inside the funnel, and measure $\\|y(T)\\|$ and $\\|u\\|_\\infty$; a nonzero $\\|y(T)\\|$ or an unbounded input would falsify Lemma 3 and therefore Theorems 5 and 8. Running Algorithm 4 on the same example with $Q=I_2$ should give $\\|y_{\\mathrm{cl}}(t)\\to 0$; failure of that convergence would falsify Theorem 8.","supporting_citations":[{"cited_title":"Exact output tracking in prescribed finite time via funnel control,","cited_arxiv_id":null,"evidence_quote":"Supplies the exact finite-time tracking result used as Lemma 3, making the equilibrium endpoint constraint automatic."},{"cited_title":"Optimal infinite-horizon feedback laws for a general class of constrained discrete-time systems: Stability and moving-horizon approximations,","cited_arxiv_id":null,"evidence_quote":"Provides the MPC terminal-equality stability theory that the cost-decrease and convergence argument is modeled on."},{"cited_title":"Funnel control of nonlinear systems,","cited_arxiv_id":null,"evidence_quote":"Defines the high-gain property and the system class $\\mathcal{S}$ used throughout the paper."}],"review_version":1}