{"id":"4ae43db8-88d5-45b2-a834-dc89ba62dd2c","arxiv_id":"2509.06188","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A driftless linear approximation at the initial state, re-solved on progressively shorter intervals, is claimed to drive constrained nonlinear systems arbitrarily close to any reachable target; the proof rests on a false differential inequality.","lead":"This paper proposes a control method that ignores nonlinear drift and steers a system to a target using only simple piecewise-constant inputs built from the input matrix at the starting point. A generalist should care because the controller is unusually simple, but the convergence proof contains a mathematical error.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (21) is false at t=0: the constant ||xtg-x0|| was dropped, and the corrected Gronwall bound does not converge to zero.","rationale":"The reader's verdict is REJECT, and the weakest assumption correctly identifies a false differential inequality. My independent derivation confirms that (21) is not merely a gap but a concrete algebraic error: the term ||xtg-x0|| appears as an additive constant in the bound on dot_Q1, and the paper's transformation to c t is invalid. The proof of Theorem 1 and Corollary 1 depends entirely on this inequality to obtain the exponentially decaying v_n. When the correct inequality is used, the bound on the error at the end of an interval does not vanish as the interval length goes to zero; it retains a term proportional to the error at the start of the interval. Thus the claimed monotone convergence to zero is not established. The scalar counterexample confirms the claimed bound (15) is false even for a simple stable linear system. Since the main theoretical guarantee is unsupported, the REJECT verdict stands. The simulations are illustrative but cannot replace the missing proof.","tokens_in":14868,"tokens_out":7572,"duration_ms":73022,"concrete_test":"Independently re-derive Eq. (21) from (19) and (20) on the interval [0,t*]. The correct inequality is dot_Q1(t) <= ||e0|| + A t + D_S Q1(t) with e0=xtg-x0 and A=||f(x0)||+D_S||e0||, not c t + D_S Q1(t). Then test the theorem's bound (15) on the scalar linear system xdot=-x+u, x0=2, xtg=0, t*=0.1. Compute u1=-x0/t*=-20 and the resulting state x(0.1)=2e^{-0.1}-20(1-e^{-0.1})≈3.71. Compare to the claimed bound v1=6(e^{0.1}-1)≈0.63. The bound is violated, confirming the false inequality is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central convergence proof (Theorem 1, Corollary 1) relies on the differential inequality (21): dot_Q1(t) <= c t + D_S Q1(t) on [0,t*] with Q1(0)=0. At t=0 this reads 0 <= ||xtg-x0||, which is false whenever x0 != xtg. The derivation error is concrete: from (19) and (20) the correct inequality is dot_Q1(t) <= ||xtg-x0|| + (||f(x0)|| + D_S||xtg-x0||) t + D_S Q1(t). The constant term ||xtg-x0|| was incorrectly absorbed into c t, since c includes (D_S+1)||xtg-x0|| but is multiplied by t. The same error propagates to (32) at every interval start. With the correct inequality, Gronwall gives E_{k+1} <= E_k e^{D_S dt} + A(e^{D_S dt}-1)/D_S, which has a limit as dt->0 equal to E_k, not zero; hence the monotone 1/n shrinkage in Lemma 1 and the guarantee in Corollary 1 are unsupported. A simple scalar test shows the claimed bound (15) is numerically false: for xdot = -x + u, x0=2, xtg=0, t*=0.1, one has D_S=1, c=6, so v1=6(e^{0.1}-1)≈0.63, but the proposed u1=-x0/t* yields |x(0.1)|≈3.7, exceeding v1. The simulations in Section 5 do not repair this fatal gap in the proof.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a controller for constrained nonlinear systems (Problem 1) using a sequence of piecewise-constant inputs derived from the linear driftless approximation xdot = g(x0)u. The time horizon is partitioned into intervals of length t*/n, and at each interval the control is the minimum-energy input for the driftless model. The authors claim a monotone bound on the endpoint error, conditions guaranteeing ||u||_infinity <= 1, and convergence to any neighborhood of the target. Section 5 reports simulations on ADMIRE, van der Pol, wing rock, and a two-link robot, with code available.","tokens_in":15299,"tokens_out":5457,"duration_ms":60107,"significance":"The idea is simple and computationally attractive: it uses only g(x0), ignores the drift f, and produces explicitly constructed piecewise-constant inputs. If the theorem were valid, the explicit bound (15) and the input-constraint condition (14) would be a useful contribution. The paper also provides reproducible code and openly discusses assumption violations. However, the main proof is invalid at two load-bearing points, Eqs. (21) and (32), so the central convergence guarantee is not established. The paper cannot be accepted without a substantially different proof.","major_comments":[{"comment":"At t=0, inequality (21) reads 0 <= ||xtg - x0||_infinity, which is false unless x0 = xtg; by continuity it fails for small positive t. The correct derivation from (19) and (20) gives \\dot Q1(t) <= ||xtg - x0|| + t(||f(x0)|| + D_S ||xtg - x0||) + D_S Q1(t). The constant ||xtg - x0|| cannot be absorbed into c t because c includes (D_S+1)||xtg - x0|| but is multiplied by t. Consequently the Gronwall step leading to (24), and hence the bound (15) for n=1, do not follow.","section":"Theorem 1, Eq. (21)"},{"comment":"The same omission occurs in the induction step: at t = t_k, the right-hand side of (32) is zero while the left-hand side is ||xtg - x_k||, which is not zero in general. Thus the induction step from v_k to v_{k+1} is invalid, and the claim that v_{n+1} <= n/(n+1) v_n applies to the actual error is unsupported. Corollary 1, including the endpoint guarantee and the inter-sample bound, relies on this step.","section":"Theorem 1, Eq. (32)"},{"comment":"With the corrected inequality, the endpoint error E_k = ||xtg - x_k|| would obey approximately E_{k+1} <= E_k e^{D_S Delta t_{k+1}} + A(e^{D_S Delta t_{k+1}} - 1)/D_S, whose limit as Delta t -> 0 is E_k, not zero. Hence the monotone 1/n contraction asserted in Lemma 1 cannot be concluded from the authors' Gronwall estimate; the argument requires a mechanism that forces E_k itself to decrease, which is absent. The central claim of convergence to the target is therefore not established.","section":"Lemma 1 and corrected recurrence"}],"minor_comments":[{"comment":"The sentence fragment 'The left-hand side Clearly,' is incomplete; it should read 'Clearly, ||u1(t)||_infinity <= 1 ...'.","section":"Theorem 1, Part 1"},{"comment":"The definition v_n = c/D_S (e^{Delta t_n D_S} - 1) divides by D_S; the case D_S = 0 should be handled separately or excluded.","section":"Lemma 1"},{"comment":"The symbol t is overloaded: it is both a variable and the upper bound of the interval for t*. Use a different symbol, e.g., t_max, for the solution of the transcendental equation.","section":"Eq. (14)"},{"comment":"In part (ii), t is used for a generic time while t* is the design parameter; this is confusing because t also appears in Eq. (14). Rename one of them.","section":"Corollary 1"},{"comment":"The abstract says the method completes the task even when assumptions are violated, but Example 5.3 explicitly reports a non-zero steady-state error. Suggest softening to 'bounded error' or 'satisfactory performance in some cases'.","section":"Abstract and Section 5.3"}],"recommendation":"reject","confidential_remarks":"The central theorem is not correct as stated; this is a proof error, not a disagreement with consensus or a circularity issue. The simulations are suggestive but do not repair the missing guarantee. I recommend rejection, though the underlying idea may be salvageable with a genuinely different analysis, e.g., one that does not rely on the false differential inequality."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a genuinely simple controller idea and honest simulations, but the central convergence theorem is not proven. Equation (21) is false at t=0 unless x0 = xtg, and the same error repeats at every interval boundary. The reader's analysis is right, and the stress-test note is accurate.\n\nWhat's new: using g(x0) as a driftless model, re-solving a pseudoinverse constant input over harmonic-length intervals is a clean combination. The controller needs almost no model, and the examples (ADMIRE, van der Pol, wing rock, two-link arm) show real behavior. The code is available. The paper is clearly written and the literature is relevant. The self-citation [32] supplies a standard minimum-norm fact; that's fine.\n\nThe problem: the proof of Theorem 1. In deriving (21), the constant term ||xtg - x0|| is dropped. Correct combination of (19) and (20) gives dot Q1 <= (||f(x0)|| + DS||xtg-x0||)t + DS Q1 + ||xtg-x0||, not c t + DS Q1. At t=0 the LHS is ||xtg-x0|| and the RHS is 0. The same flaw appears in (32). With the constant term retained, Gronwall yields a bound whose limit over shrinking intervals is the previous error, not zero, so Lemma 1 and Corollary 1 don't follow. The stress-test's scalar example is outside the theorem's assumptions (condition (14) fails because the input constraint isn't satisfied), but that does not rescue the proof: the differential inequality is false under the theorem's own conditions when x0 != xtg.\n\nThe simulation section is good supporting material, but simulations cannot patch a false inequality. The method might still work under the restrictive conditions (small drift relative to input authority), but the proof as written doesn't show it, and I'm not convinced the claimed monotone bound holds generally.\n\nBottom line: reject the paper in its current form. The idea is worth a referee's attention, and a careful author could potentially repair or reframe the proof, but as submitted the central guarantee is unsupported. Would I cite it? No. Bring to reading group? Maybe, as a case study in why the constant term matters in Gronwall arguments.","headline":"The controller idea is simple and worth knowing, but the main convergence proof is wrong at Eq. (21) and the paper's central guarantee is unsupported.","tokens_in":15730,"tokens_out":6680,"would_cite":false,"duration_ms":71439,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C10","93C15","93B52"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a constrained nonlinear system can be driven arbitrarily close to a target state by piecewise constant inputs derived from a simple driftless linear approximation built only from the input map at the initial state, wit","keywords":["nonlinear control","driftless approximation","piecewise constant control","input constraints","Lipschitz continuous dynamics","minimum-energy control","pseudoinverse","arbitrary precision reachability"],"falsifier":"Take the scalar system x_dot = x + u with x0=0 and x_tg=1, which satisfies the paper's Lipschitz assumptions with D_f=1, D_g=0, and c=2. At t=0, the derivative Q1_dot(0) equals ||x_tg - x0||_inf = 1, while the right-hand side of (21) is 0. Since the inequality is asserted for all t in [0, t*] and both sides are continuous, it fails for small positive t as well; this blocks the derivation of the bound (15) and hence of Corollary 1.","tokens_in":14771,"feed_emoji":"🎯","tokens_out":9206,"duration_ms":107970,"temperature":0.7,"pith_summary":"This paper claims that a nonlinear control problem with input limits can be solved by a controller that ignores the drift term entirely. The controller uses only the input map at the starting state, g0 = g(x0), to form a linear driftless model, and over time intervals that shrink like t*/n it applies the minimum-energy constant input that would steer this driftless model to the target. The paper proves that the miss at each interval endpoint is bounded by a quantity v_n that decreases like 1/n, and that the input limit is respected when the interval scale t* satisfies a stated condition. If correct, this yields a simple alternative to feedback linearization and nonlinear model predictive control: no full model, no heavy computation, just pseudoinverse-based piecewise constant signals.","feed_headline":"Ignore the drift: piecewise constant inputs hit any target","feed_subtitle":"Minimum-energy controls over shrinking intervals shrink the error to zero while respecting input limits.","key_machinery":"The load-bearing object is the linear driftless approximation (5), x_dot = g0 u, together with the shrinking-interval schedule t_n = t* sum_{k=1}^n 1/k. The approximation turns the nonlinear problem into a trivial optimal-control problem: the optimal constant input is u_n = (1/Delta t_n) g0^dagger (x_tg - x_{n-1}), a pseudoinverse formula. The schedule makes Delta t_n = t*/n shrink monotonically to zero. The connecting quantity is the error bound v_n = (c/D_S)(exp(Delta t_n D_S) - 1): it is derived from the differential inequality (21) via a standard comparison inequality for differential equations, and Lemma 1 converts the shrinking intervals into the explicit decay v_n <= v_1/n, which driv","core_discovery":"Formally, the paper addresses Problem 1: for any epsilon > 0, reach and stay in the epsilon-ball around the target state x_tg while respecting the input limit ||u||_inf <= 1. The proposed solution is the linear driftless approximation (5), built from g0 = g(x0) alone. The time axis is split at t_n = t* times the harmonic sum from k=1 to n, so the interval lengths Delta t_n = t*/n shrink to zero, and the input on [t_{n-1}, t_n] is the constant u_n = (1/Delta t_n) g0^dagger (x_tg - x_{n-1}), the minimum-energy control that would make the driftless model hit the target at t_n. Theorem 1 asserts that if 2 c ||g0^dagger||_inf < 1 and t* lies between ||g0^dagger (x_tg - x0)||_inf and the unique po","pith_inferences":["Beyond the paper's proof, the key differential inequality (21) is asserted to hold from t=0 with Q1(0)=0, but at t=0 the left side is the initial distance to the target, which is positive unless x0 equals x_tg. A corrected argument would need to start the bound from that positive offset; the simulations may still be recoverable, but the theorem as written does not cover them.","The same shrinking-interval construction with a geometric partition, t_n = t_f (1 - a^{-n}), would give finite-horizon arrival and bounded input energy instead of the harmonic schedule's infinite total energy; testing whether the error-bound argument carries over is a direct next step.","For systems with more states than inputs, where g0 is not full row rank, the paper leaves the steady-state error unquantified after the first interval; making that residual explicit would determine when the method is practically useful beyond Assumption 1.","Replacing g0 by g evaluated at the target or at an equilibrium might improve the approximation when drift is large; the paper raises this as a question, and the simulations do not yet resolve it."],"forward_implications":["For any epsilon > 0, after finitely many intervals the state at every sampling instant t_n is within epsilon of the target, and stays within epsilon for all later times.","The input constraint ||u||_inf <= 1 is guaranteed whenever the design parameter t* satisfies condition (14); the paper notes that the condition is sufficient, not necessary.","The controller can be implemented from a single evaluation of the input map g at the initial state, without knowledge of the drift term or a linearized model.","The error bound at interval endpoints shrinks monotonically, at least as fast as v_1/n, so accuracy improves as more intervals are used.","In the paper's simulation examples, the method completes the task even when global Lipschitz continuity or full row rank of g0 fails, and it outperforms linear state feedback in the compared case."],"supporting_citations":[{"why":"Supplies the minimum-energy optimal control for the driftless system that the method applies in each interval.","marker":"[32]"},{"why":"Supplies the comparison inequality used to convert the differential inequality (21) into the exponential error bound.","marker":"[30]"},{"why":"Used in Lemma 2 to guarantee existence and uniqueness of the positive solution of the transcendental equation defining condition (14).","marker":"[6]"},{"why":"Provides the pseudoinverse best-approximation property used in Remark 2 to describe the residual error when Assumption 1 fails.","marker":"[33]"},{"why":"Provides the inequality used to move the infinity norm inside the integral when bounding v(0,t).","marker":"[24]"},{"why":"Used to identify the derivative of Q1(t) with the instantaneous target error in the proof of Theorem 1.","marker":"[1]"}],"fun_headline_variants":["No drift, no problem: constant inputs hit any target","Shrinking steps, zero error: new control method","Driftless approximation outsmarts nonlinear control","Piecewise constant inputs break drift barrier","Even when assumptions fail, constant inputs rule"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The proof of the central theorem requires the differential inequality (21) to hold from t=0, but at t=0 it asserts that the initial distance to the target is at most zero, which only holds if the state already starts at the target.","fun_headline_variants_meta":{"raw":{"variants":["No drift, no problem: constant inputs hit any target","Shrinking steps, zero error: new control method","Driftless approximation outsmarts nonlinear control","Piecewise constant inputs break drift barrier","Even when assumptions fail, constant inputs rule"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00026,"raw_usage":{"total_tokens":1435,"prompt_tokens":763,"completion_tokens":672,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":614}},"tokens_in":507,"tokens_out":672,"duration_ms":8724,"temperature":1.0,"reasoning_tokens":614,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T23:59:00.150256+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the scalar system x_dot = x + u with x0=0 and x_tg=1, which satisfies the paper's Lipschitz assumptions with D_f=1, D_g=0, and c=2. At t=0, the derivative Q1_dot(0) equals ||x_tg - x0||_inf = 1, while the right-hand side of (21) is 0. Since the inequality is asserted for all t in [0, t*] and both sides are continuous, it fails for small positive t as well; this blocks the derivation of the bound (15) and hence of Corollary 1.","supporting_citations":[{"cited_title":"Padmanabhan and M","cited_arxiv_id":null,"evidence_quote":"Supplies the minimum-energy optimal control for the driftless system that the method applies in each interval."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the comparison inequality used to convert the differential inequality (21) into the exponential error bound."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Used in Lemma 2 to guarantee existence and uniqueness of the positive solution of the transcendental equation defining condition (14)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the pseudoinverse best-approximation property used in Remark 2 to describe the residual error when Assumption 1 fails."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the inequality used to move the infinity norm inside the integral when bounding v(0,t)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Used to identify the derivative of Q1(t) with the instantaneous target error in the proof of Theorem 1."}],"review_version":1}