{"id":"3bdac20b-a367-43f4-ac1b-2e536a146661","arxiv_id":"2505.04954","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A deterministic multi-trajectory probing scheme plus ridge regression learns the linearization of a noisy nonlinear system, achieving O(N^{-1/4}) worst-case error from N experiments under initialization constraints.","lead":"This paper gives a finite-sample guarantee for learning a linearized model of an unknown nonlinear system when experiments can only start inside a restricted region. It uses many one-step trajectories and regularized least squares, with an error bound that balances noise against nonlinearity.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's printed noise term contains an extra 1/q factor that makes the stated O(N^{-1/4}) consistency false as written; the proof's own decomposition yields a noise term without this factor.","rationale":"The most load-bearing issue is not the initialization-noise idealization but a direct mismatch between Theorem 1 and the proof that supports it. The theorem's noise term has an extra 1/q factor; taken literally, it prevents the stated q proportional to N^{-1/4} choice from driving the error to zero. The proof, however, clearly shows the correct noise term lacks this factor, so the advertised O(N^{-1/4}) rate is recoverable once the display is corrected. The reader's rationale also flagged this scaling error, although the reader's weakest-assumption field emphasized exact resetting; hence partial agreement. The conditional verdict remains appropriate: the core idea and proof strategy are plausible, but the formal statement as printed cannot be certified and must be corrected before acceptance.","tokens_in":14343,"tokens_out":8670,"duration_ms":86590,"concrete_test":"Independently re-derive the noise term in Eq. (8) from Eq. (9), Lemma 2, and Eq. (10), keeping all constants. If the result is 3*sigma_w * sqrt(log(...)) / sqrt(N q^2/(n+p) + lambda) with no 1/q in front, then the displayed theorem is inconsistent with its own proof and the claimed q proportional to N^{-1/4} consistency rate is unsupported; if a missing argument produces the 1/q, that argument must be supplied and checked.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Eq. (8), the 'Error due to noise' term is displayed as (5*sigma_w/q) * sqrt(log(...)) / sqrt(N q^2/(n+p) + lambda). Substituting the proof's Eq. (9)-(10) together with Lemma 2 gives ||W Z' (Z Z' + lambda I)^{-1}|| <= 3*sigma_w * sqrt(log(...)) / sqrt(lambda_min(Z Z') + lambda) <= 3*sigma_w * sqrt(log(...)) / sqrt(N q^2/(n+p) + lambda). No 1/q factor appears. With m=0 and lambda=0, the printed noise term scales as 1/(q^2 sqrt(N)); choosing q proportional to N^{-1/4} then makes it O(1), not O(N^{-1/4}), so the advertised consistency rate does not follow from the displayed theorem. The correct scaling 1/(q sqrt(N)) balances the nonlinearity term, which is O(beta q), at q proportional to N^{-1/4}. This is an internal inconsistency between the theorem statement and its proof, not a modeling dispute. The exact-initialization caveat identified by the reader is legitimate but secondary: this scaling issue is load-bearing even in the ideal-reset case.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies finite-sample identification of the linearized dynamics Theta = [A B] of a discrete-time nonlinear system x_{k+1} = f(x_k,u_k) + w_k around the origin, under the constraint that experiments can only be initialized inside a feasible set S. It proposes a deterministic multi-trajectory data acquisition scheme (Algorithm 1) that cyclically initializes the state-input vector at m +/- q e_i within S, followed by a regularized least-squares estimator (Algorithm 2). The main result (Theorem 1) bounds ||hat Theta - Theta|| by three terms: a noise term, a nonlinearity term, and a regularization term. The paper claims that when S is an open set containing the origin, choosing q proportional to N^{-1/4} gives consistency at rate O(N^{-1/4}), and that when S does not contain the origin a small error is still possible if S is close enough to the origin. Numerical experiments on a pendulum-like system and on a strongly nonlinear system illustrate the trade-offs and show better behavior than single-trajectory Gaussian-input identification. The paper explicitly notes that exact resetting is assumed and that initialization perturbations are only tested numerically.","tokens_in":14534,"tokens_out":25061,"duration_ms":229821,"significance":"If the main theorem is corrected, the paper makes a useful contribution: it provides one of the first finite-sample guarantees for learning a local linear model of a nonlinear stochastic system with a constrained initialization region, without requiring global Lipschitzness of the unmodeled dynamics. The decomposition of the error into noise, nonlinearity, and regularization terms is clean, the data acquisition scheme is simple and deterministic, and the numerical experiments support the qualitative predictions. The proofs are self-contained in the appendix, including persistent excitation, sub-Gaussian concentration, and Taylor remainder bounds. The exact-initialization limitation is disclosed honestly, and robustness to initialization jitter is offered only as empirical evidence. The claimed O(N^{-1/4}) rate, if supported, would be a useful improvement over the existing practice for nonlinear system identification.","major_comments":[{"comment":"The displayed \"Error due to noise\" term contains a spurious factor 1/q and an inconsistent denominator. In the proof, Eqs. (9)-(10) combined with Lemma 2 and Lemma 1 give the noise contribution bounded by 3 sigma_w sqrt(log(9n/delta) + (n+p) log(1 + ...)) / sqrt(N q^2/(2(n+p)) + lambda), with no 1/q factor. For m = 0 and lambda = 0, the printed term scales as 1/(q^2 sqrt(N)), so the recommended choice q proportional to N^{-1/4} leaves this term O(1), contradicting the claimed O(N^{-1/4}) consistency in Section 4.2. The proof's own bound balances the O(beta q) nonlinearity term exactly at q proportional to N^{-1/4}, so Eq. (8) and the related rate statements in Section 4.2 must be corrected. I recommend replacing the noise term by 3 sigma_w sqrt(L) / sqrt(N q^2/(2(n+p)) + lambda) or an equivalent conservative upper bound without the 1/q factor, and re-deriving the displayed constants in the theorem.","section":"Theorem 1, Eq. (8)"}],"minor_comments":[{"comment":"The square-root expression following Eq. (24) is ambiguous as typeset; it should read sqrt(2(n+p)/(N q^2 + 2 lambda (n+p))) + 2 sqrt(lambda)(n+p)/(N q^2 + 2 lambda (n+p)). The current typesetting suggests that the second summand is inside the radical, which is dimensionally inconsistent.","section":"Section 7.4, proof of Lemma 3"},{"comment":"The logarithmic factor in Theorem 1 omits the term zeta = 4 lambda (n+p)/N that appears in the denominator of Lemma 2's log argument. This omission is a valid loosening of the bound, but it should be explicitly acknowledged so that the two displays are visibly consistent.","section":"Theorem 1, Eq. (8) and Lemma 2"},{"comment":"The numerical robustness experiment adds Gaussian perturbations to the designed initial conditions, but the formal analysis assumes exact resetting. The paper's own caveat at the end of Section 5.1 correctly states that a detailed analysis is left for future work; I recommend making this limitation explicit in the contributions or abstract as well, since the formal guarantee does not cover inexact initialization.","section":"Section 5.1"},{"comment":"The expression sqrt(lambda N n beta^2 b^2 q^4) in the regularization term should be parenthesized or written as sqrt(lambda N n) beta b q^2 to avoid ambiguity about what is inside the radical.","section":"Theorem 1, Eq. (8)"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the scaling problem in Theorem 1 appears to be a theorem-statement/proof mismatch rather than a fundamental flaw; the proof's own decomposition actually supports the advertised O(N^{-1/4}) rate once the noise term is corrected. I therefore see this as a corrigible major revision, not a rejection. Please ask the authors to reconcile Eq. (8) with Lemmas 1 and 2 and to re-check all derived rates in Section 4.2. The exact-initialization assumption is acknowledged by the authors, so I do not treat it as a novelty or overclaim issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: the paper attacks a real gap—finite-sample guarantees for identifying the Jacobian of a noisy nonlinear system when you can only probe from a feasible region. The deterministic multi-trajectory design, the ridge regression analysis, and the explicit trade-off between bias from nonlinearity and variance from noise are genuinely new compared with the authors' prior CDC paper, and the numerics back the qualitative storyline. The writing is clear and the proof structure is honest.\n\nThe problem: the printed Theorem 1 is internally inconsistent. The 'error due to noise' term carries a factor 1/q in the numerator in addition to the 1/q that comes from the singular-value lower bound. Plugging the proof's own equations (9)-(10) and Lemma 2 into the decomposition gives a noise term of order (1/sqrt(N q^2)) sqrt(log), i.e., no extra 1/q. As printed, with q ~ N^{-1/4} the noise term is O(1), so the consistency claim does not follow from the statement. The corrected scaling, 1/(q sqrt N), balances the nonlinearity term O(beta q) at q ~ N^{-1/4}, giving the advertised rate. So this is very likely a typographical slip, but it is load-bearing: a referee has to be able to verify the bound as stated, and right now the theorem as written is false.\n\nTwo secondary points. First, the theory assumes perfect resets to deterministic initial points z0 = m ± q e_i. Lemma 1 and Lemma 3 depend on that exactly. The robustness experiment in Section 5.1 only adds small Gaussian perturbations and shows the error stays small; that's a legitimate scope limitation, clearly flagged by the authors, not a hidden flaw. Second, the numerical comparison with a single trajectory under i.i.d. Gaussian inputs convincingly demonstrates the failure mode the paper is motivated by.\n\nThe citation pattern is fine: self-citations to the conference version and to [33] for matrix concentration lemmas are standard and don't carry the result's weight. I found no circular parameter fitting.\n\nRecommendation: send it to review. A competent referee can check the corrected bound in an afternoon. The problem is well-motivated, the algorithm is practical, and the corrected result is likely to stand. My verdict is conditional: fix the statement of Theorem 1, double-check Lemma 3's algebra, and either extend the initialization-robustness analysis or state the exact-reset scope more prominently.","headline":"Useful and well-scoped paper on learning linearized models under initialization constraints, but Theorem 1's printed noise bound has a 1/q scaling slip that makes the advertised O(N^{-1/4}) consistency false as written; the fix is straightforward and the underlying idea is sound.","tokens_in":15105,"tokens_out":3895,"would_cite":false,"duration_ms":37724,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that restart experiments with deterministic initial perturbations can consistently recover the linear part of a nonlinear system, at a worst-case rate of $O(N^{-1/4})$ when the feasible region contains the origin.","keywords":["system identification","linearized model","nonlinear dynamics","finite-sample guarantees","multiple trajectories","regularized least squares","initialization constraints","persistent excitation"],"falsifier":"Run the scalar system $x_{k+1} = x_k + x_k^2 + w_k$ with $\\Theta = 1$, small sub-Gaussian noise variance, $m=0$, $\\lambda=0$, and $q = c_0 N^{-1/4}$. If the empirical $\\|\\hat{\\Theta} - 1\\|$ across many independent trials fails to shrink at the predicted $O(N^{-1/4})$ rate as $N$ grows, or fails to shrink to zero in the noiseless case, then the claimed consistency rate and bound are disproved.","tokens_in":14094,"feed_emoji":"🎛️","tokens_out":9836,"duration_ms":94355,"temperature":0.7,"pith_summary":"This paper asks a practical question: when a physical system is genuinely nonlinear but a control engineer only needs its linearized model, how much data and what kind of experiments are needed to recover that model with a guarantee? It proposes collecting $N$ one-step trajectories from restart experiments whose initial state-plus-input points are chosen deterministically as small perturbations $\\pm q e_i$ around a center $m$ inside a feasible region, then fitting $\\Theta = [A\\;B]$ by regularized least squares. The central result is a finite-sample bound on $\\|\\hat{\\Theta} - \\Theta\\|$ that splits into noise, nonlinearity, and regularization terms and exhibits a bias-variance-like trade-off in the perturbation size $q$. When the feasible region contains the origin and $q$ is chosen to shrink as $N^{-1/4}$, the bound gives worst-case consistency at rate $O(N^{-1/4})$; for a truly linear system it recovers the familiar $O(N^{-1/2})$ rate. This matters because existing finite-sample identification guarantees mostly assume the dynamics are exactly linear or require a single long trajectory with random inputs and globally Lipschitz nonlinearities, both of which can fail in mild nonlinear settings.","feed_headline":"Restarting experiments reveals the linear core of nonlinear systems","feed_subtitle":"A finite-sample proof shows one-quarter-power consistency when restarts start near the origin.","key_machinery":"The load-bearing object is the Taylor decomposition of the nonlinear map around the origin, $x_{k+1} = \\Theta z_k + w_k + r_k$ with $\\Theta = [A\\;B]$ and remainder $r_k = h(z_k)$ whose components are bounded by $\\beta\\|z_k\\|_1^2$ (Assumption 1). Algorithm 1's sign-alternating deterministic perturbations $\\pm q e_i$ around $m$ are the second piece: they make the design covariance $ZZ'$ well conditioned, with $\\lambda_{\\min}(ZZ') \\ge Nq^2/[2(n+p)]$ (Lemma 1), while the sign changes cancel the center $m$ in cross terms; this is what lets small $q$ keep the nonlinearity bias low without losing persistent excitation. Algorithm 2's regularized least squares with ridge parameter $\\lambda$ produces the closed-form estimator, and the proof decomposes the error via the triangle inequality into a bias term from $\\lambda$, a nonlinearity term from $RZ'(ZZ'+\\lambda I)^{-1}$, and a noise term bounded through a sub-Gaussian self-normalized martingale inequality. The regularization parameter is not merely numerical: it damps the noise and nonlinearity terms at the cost of an irreducible $\\|\\Theta\\|$ bias.","core_discovery":"The paper's central claim is Theorem 1: under Assumption 2, meaning that the deterministic points $m \\pm q e_i$ lie in the feasible region $S$, and under Assumption 1, meaning that the Taylor remainder $r_k$ satisfies $|r_{i,k}| \\le \\beta \\|z_k\\|_1^2$ near the origin, the regularized least-squares estimate $\\hat{\\Theta} = XZ'(ZZ' + \\lambda I)^{-1}$ produced by Algorithms 1 and 2 satisfies, with probability at least $1-\\delta$, a three-term error bound. The noise term scales roughly like $\\sigma_w/(q\\sqrt{N})$, the nonlinearity term like $\\beta q$, and the regularization term like $\\lambda\\|\\Theta\\|$ divided by the covariance size. The author's conclusion is that consistent learning of the linearized dynamics is possible with finitely many short trajectories even when the nonlinearity is not globally Lipschitz: when $S$ is open and contains $0$, taking $m=0$, $\\lambda=0$, and $q = c_0 N^{-1/4}$ yields $\\|\\hat{\\Theta} - \\Theta\\| = O(N^{-1/4})$. The same framework yields small error for convex feasible regions that do not contain the origin, provided the center $m$ is close enough to $0$ and $N$ is large, and for perfectly linear systems the bound reduces to an $O(1/\\sqrt{N})$ rate comparable with existing random-input results.","pith_inferences":["Beyond the paper: the same alternating-perturbation design should extend to estimating the quadratic term of the Taylor expansion, since finite-difference combinations of the one-step outputs isolate second-order coefficients; this would give a direct data-driven check of Assumption 1's constant $\\beta$.","Beyond the paper: a fully adaptive experiment could choose $q$ per batch by estimating the noise variance and nonlinearity curvature from earlier batches, potentially achieving faster than worst-case convergence; Theorem 1's trade-off suggests this is feasible but does not analyze it.","Beyond the paper: with inexact initialization modeled as an extra bounded disturbance, the two proofs that fail (Lemma 1 and Lemma 3) could be patched by adding an initialization-error term to the bound; the numerical robustness shown in Section 5.1 is evidence this patch would work, but the paper leaves the proof open."],"forward_implications":["Engineers can obtain a certified linear model for controller design using many one-step restart experiments, and the guarantee holds regardless of the spectral radius of $A$, so even unstable nonlinear systems are in scope.","When initialization can reach a neighborhood of the reference point, the perturbation size $q$ should be chosen to balance noise and nonlinearity; a default of $q \\propto N^{-1/4}$ gives guaranteed worst-case consistency at rate $O(N^{-1/4})$.","In the special case of perfectly linear dynamics, the bound reproduces the $O(1/\\sqrt{N})$ sample complexity of established multiple- and single-trajectory linear identification methods, so the approach does not sacrifice the linear benchmark.","If the feasible region excludes the origin, the center $m$ should be as close to the origin as constraints allow, and a larger regularization $\\lambda$ becomes useful when noise, nonlinearity, or distance from the origin is large.","State predictions from the learned model are close to the true nonlinear map near the origin, with prediction error dominated by the identification error $\\|\\hat{\\Theta} - \\Theta\\|$ rather than by the nonlinear terms when $\\|z\\|$ is small (Proposition 1)."],"supporting_citations":[{"why":"The authors' earlier conference result that arbitrarily initialized short trajectories can recover the linearized model without Lipschitzness; this paper adds feasible-region constraints and explicit rates.","marker":"[32]"},{"why":"Baseline finite-sample guarantee for unmodeled dynamics that requires globally Lipschitz nonlinear terms; the paper's Assumption 1 is positioned as a milder replacement.","marker":"[27]"},{"why":"Finite-sample multiple-trajectory identification for truly linear systems, whose $O(1/\\sqrt{N})$ rate this paper matches in the linear special case.","marker":"[8]"},{"why":"Single-trajectory linear identification benchmark whose random-input rate the paper compares against.","marker":"[25]"},{"why":"Supplies the sub-Gaussian self-normalized martingale concentration lemma used to bound the noise term.","marker":"[33]"},{"why":"Supplies the matrix inequality used to compare inverse covariances in the noise bound.","marker":"[5]"}],"fun_headline_variants":["Finite data learns linear core of nonlinear systems","Multiple restarts reveal linear dynamics amid nonlinearity","Short trajectories beat single long runs for linear fits","Proof: finite samples suffice to linearize nonlinear models","Nonlinear systems leave linear fingerprints in short runs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee rests on resetting each experiment's initial state and input exactly to the deterministic points $m \\pm q e_i$; if initializations are only approximate, the proof's covariance lower bound and nonlinearity remainder bound no longer follow.","fun_headline_variants_meta":{"raw":{"variants":["Finite data learns linear core of nonlinear systems","Multiple restarts reveal linear dynamics amid nonlinearity","Short trajectories beat single long runs for linear fits","Proof: finite samples suffice to linearize nonlinear models","Nonlinear systems leave linear fingerprints in short runs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000503,"raw_usage":{"total_tokens":2497,"prompt_tokens":1024,"completion_tokens":1473,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":1401}},"tokens_in":640,"tokens_out":1473,"duration_ms":8726,"temperature":1.0,"reasoning_tokens":1401,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:17:48.479876+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the scalar system $x_{k+1} = x_k + x_k^2 + w_k$ with $\\Theta = 1$, small sub-Gaussian noise variance, $m=0$, $\\lambda=0$, and $q = c_0 N^{-1/4}$. If the empirical $\\|\\hat{\\Theta} - 1\\|$ across many independent trials fails to shrink at the predicted $O(N^{-1/4})$ rate as $N$ grows, or fails to shrink to zero in the noiseless case, then the claimed consistency rate and bound are disproved.","supporting_citations":[{"cited_title":"Learning linearized models from nonlinear systems with finite data","cited_arxiv_id":null,"evidence_quote":"The authors' earlier conference result that arbitrarily initialized short trajectories can recover the linearized model without Lipschitzness; this paper adds feasible-region constraints and explicit rates."},{"cited_title":"Accurate parameter estimation for safety-critical systems with unmodeled dynamics","cited_arxiv_id":null,"evidence_quote":"Baseline finite-sample guarantee for unmodeled dynamics that requires globally Lipschitz nonlinear terms; the paper's Assumption 1 is positioned as a milder replacement."},{"cited_title":"On the sample complexity of the linear quadratic regulator","cited_arxiv_id":null,"evidence_quote":"Finite-sample multiple-trajectory identification for truly linear systems, whose $O(1/\\sqrt{N})$ rate this paper matches in the linear special case."},{"cited_title":"Near optimal finite time identification of arbitrary linear dynamical systems","cited_arxiv_id":null,"evidence_quote":"Single-trajectory linear identification benchmark whose random-input rate the paper compares against."},{"cited_title":"Hermitian matrix inequalities and a conjecture","cited_arxiv_id":null,"evidence_quote":"Supplies the matrix inequality used to compare inverse covariances in the noise bound."}],"review_version":1}