{"id":"f11295ff-5d3b-4f89-9aa2-b046756a2904","arxiv_id":"2412.08772","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A perturbation-based optimal control method for learning is proposed, but proof errors and missing baselines undermine the generalization claim.","lead":"The paper presents an optimal-control view of learning where a small control term nudges gradient descent, then uses perturbation theory to compute approximate solutions. The mathematical proof has errors, and the claim that this improves generalization is not backed by rigorous analysis or proper baseline comparisons.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The perturbation expansions, even if repaired, only show that θ* nearly minimizes the validation cost; the paper neither proves nor experimentally demonstrates that this improves generalization to unseen data, so the central claim is unsupported.","rationale":"The reader's verdict is REJECT, and I agree that the paper is not acceptable in its current form. However, my primary load-bearing concern is not the same as the reader's weakest_assumption. The reader emphasizes the unproved convergence of the perturbation expansions; that is a real technical gap, especially because the bang-bang control determined by Eq. (2.5) has switching times whose shift in ε must be controlled before an O(ε²) expansion of uϵ is justified. But I see the deepest problem elsewhere: the paper's stated goal is improving generalization, while Propositions 2.1 and 2.2 only provide perturbation estimates for the validation-cost functional. They do not establish any property of test-set performance, and the numerical experiments do not include the controlled comparisons or error bars needed to support such a claim. This is a mismatch between the central claim and the mathematical content, independent of whether the perturbation proofs are corrected. I checked the main perturbation argument: with the corrected adjoint equation ṗ0=∇²J0(θ0)p0, the claimed O(ε²) convergence rate in Proposition 2.2 is plausibly correct, so the proof defects flagged by the reader are serious but likely fixable. The generalization claim, by contrast, is not a matter of a few missing terms; it requires new theoretical content or a proper empirical demonstration. Hence the verdict should remain REJECT, but for the reason that the central claim is unsupported rather than solely because of the adjoint-equation typo. My proposed test directly checks whether the claimed generalization improvement is empirically real; if it fails, the paper's central contribution is falsified, and if it passes, the paper still needs a rigorous statement of what 'generalization' means and a proof that the decomposition achieves it.","tokens_in":8047,"tokens_out":10011,"duration_ms":107889,"concrete_test":"On the saturated-water regression problem, hold out a test set disjoint from Z1 and Z2 (or use repeated k-fold cross-validation); for several values of ε and noise levels, compute the test loss of θ*=θ0(T)+εθ1(T), θ0(T), and the ordinary least-squares fit, with means and standard deviations over repeated train/validation/noise splits. If θ* does not outperform both baselines on held-out test data, the generalization claim fails; if it does, the claim still requires a proof relating test loss to the expansion.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the small-parameter method improves generalization in the considered learning class. The main mathematical results (Propositions 2.1 and 2.2) are conditional perturbation estimates: assuming the expansions in (2.1) converge, they give an O(ε²) expansion of the optimal value and show J^ε[uϵ]−J^ε[u0]=O(ε²). Even if these results are granted and the evident typo in Eq. (2.4) is corrected—it should read ṗ0=∇²J0(θ0)p0, not the dimensionally inconsistent ∇²J0(θ0)—the theorems only imply that the algorithm approximately minimizes the validation cost Φ(·,Z2). No theorem compares test-set loss, no definition of generalization is provided, and Eqs. (2.8)–(2.9) are merely finite-difference identities with no sign, no bound, and no expectation over data. The numerical section plots training versus testing loss, but it lacks a baseline comparison (e.g., θ0(T) or ordinary least squares), error bars, repeated train/validation/noise splits, and code. It therefore cannot establish that the method improves generalization to unseen data. The central claim thus exceeds what the mathematics and numerics actually support.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formulates a supervised-learning parameter-estimation problem as an optimal control problem for a gradient-flow ODE in which a small parameter ε multiplies a control input and a squared-gradient term. The terminal cost is the validation loss. Under the assumption that the optimal state, adjoint, and control admit regular perturbation expansions in ε, the paper derives zeroth- and first-order optimality systems, proposes an algorithm that returns θ0(T)+εθ1(T), and claims an O(ε²) estimate for Jε[uε]−Jε[u0]. Numerical results for polynomial fits of thermophysical properties of water are reported.","tokens_in":8358,"tokens_out":8572,"duration_ms":85537,"significance":"If the advertised claim were established, the decomposition would be attractive: it replaces a single weakly controlled optimal control problem by a sequence of easier zeroth- and first-order problems. The paper also makes the useful observation that the zeroth-order solution is obtained by solving the unperturbed gradient flow and an adjoint equation, and the numerical section treats a concrete regression example. However, the central assertion that the method improves generalization is neither mathematically defined nor demonstrated; the main propositions are conditional perturbation estimates, and the numerical experiments lack baselines and repeated splits. The paper therefore does not currently establish its title claim.","major_comments":[{"comment":"The proof of Proposition 2.1 contains an incorrect first-order adjoint expansion. In the displayed expansion of the p-equation, the term −ϵu0(t)(∇B(θ0(t),˜Z(1)))^T p1(t) should have p0(t), because p(t)=p0(t)+ϵp1(t)+O(ϵ²). The expansion also omits the third-derivative contribution ϵ D(∇²J0)(θ0(t))[θ1(t)] p0(t) that arises when expanding ∇²J0(θ^ϵ(t)) p^ϵ(t). The zeroth-order conclusion (2.3)–(2.4) survives because it only requires the O(1) terms, but the displayed derivation does not validly establish the first-order adjoint equation.","section":"Appendix A / Proposition 2.1"},{"comment":"The proof of Proposition 2.2 as written is dimensionally inconsistent. It concludes J^ϵ[u^ϵ] − J^ϵ[u0] = ϵ(u0(t)−u^ϵ(t))B(θ0(t),˜Z(1)) + O(ϵ²), where the left side is a scalar while the right side is a vector (or an unintegrated integrand). The missing integral and inner product, ϵ∫_0^T ⟨p0(t), (u0(t)−u^ϵ(t))B(θ0(t),˜Z(1))⟩ dt, are essential for the cancellation argument. In addition, the step (u0−u^ϵ)=O(ϵ) is imported from the assumed expansion (2.1), so the O(ϵ²) rate is conditional on that expansion.","section":"Appendix A / Proposition 2.2"},{"comment":"The paper never formalizes 'generalization.' The optimization objective (1.1) is the validation loss Φ(θ^ϵ(T),Z(2)), and Proposition 2.1 expands that loss; no test-set or population risk, no train/validation split distribution, and no expectation are introduced. Equations (2.8) and (2.9) are finite-difference identities for the training and validation losses and contain no sign, no upper bound, and no statistical statement, so they do not support the phrase 'improvements in the model training loss and that of the model validating loss.' Consequently the abstract's claim that the framework improves generalization is not a consequence of the theorems.","section":"Sections 1–2 / generalization claim"},{"comment":"The numerical experiments do not provide evidence for the generalization claim. Figures 3.1–3.3 plot training loss versus testing loss but include no baseline such as θ0(T), ordinary least squares, or the unperturbed gradient flow; no error bars, no repeated train/validation/noise splits, and no code are reported. The validation set has only m2=6 points. Tables 3.1 and 3.2 give fitted parameters and residual standard deviations on the original 22-point dataset, again without any comparator, so the figures cannot be read as demonstrating improved performance on unseen data.","section":"Section 3"},{"comment":"The perturbation analysis is conditional by the paper's own statement. Proposition 2.1 assumes that the expansions in (2.1) converge and satisfy the Euler–Lagrange equations, and Section 1(d)(iii) merely assumes existence and boundedness of θ^ϵ(t) for all admissible controls and ε∈(0,εmax). No conditions on J0, Φ, the loss, or the noise variance are given that would imply this regularity. The paper thus provides a conditional asymptotic decomposition rather than a theorem that the method applies to a specified class of learning problems.","section":"Section 1(d) and (2.1)"}],"minor_comments":[{"comment":"The adjoint equation is missing the factor p0(t); it should read ˙p0(t)=∇²J0(θ0(t),Z(1)) p0(t).","section":"Eq. (2.4) and Algorithm Step 1"},{"comment":"There are grammatical errors such as 'whose the time-evolution,' and the notation ε is used both for the small parameter in (1.2) and for the noise variables ε_i in item (a), which invites confusion.","section":"Abstract and Section 1"},{"comment":"The line J^ϵ[u] = Φ(θ0(T),Z(2)) − ϵ⟨p0(T),θ1_u(T)⟩ uses θ1_u before its u-dependence is defined; the definition should be given before this display.","section":"Appendix A / Proposition 2.2"},{"comment":"The table has incomplete and inconsistent entries, including missing β values for several rows and a stray '0' in the last row, and the header uses 'µ × 106' without stating the exponent convention.","section":"Table 3.3"},{"comment":"The paper should state explicitly what Figures 3.1–3.3 plot as the horizontal axis and specify the fixed final time T used in the experiment; otherwise the figures cannot be reproduced.","section":"Section 3"},{"comment":"Reference [3] is a self-citation used only as background; the dependence of the central derivation on [3] should be clarified or removed.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript overclaims 'generalization' without defining it. The mathematical core is a conditional perturbation expansion of a validation-loss optimal control problem, and the appendix contains errors in the first-order adjoint expansion and in Proposition 2.2's final expression. If the authors were to rescope the paper as an approximation method for the validation-cost problem, fix the proofs, and add a serious empirical comparison, a future submission could be viable; as submitted, the central claim is not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper applies a textbook regular perturbation expansion to a weakly-controlled gradient system and claims that the resulting two-term approximate optimal parameter improves generalization. The math is mostly a straightforward exercise, but the paper's central claim is unsupported, and the proof of the main convergence estimate contains real errors.\n\nWhat is genuinely there: the idea of splitting the Pontryagin optimality conditions into zeroth and first orders and solving them sequentially for a learning problem is a reasonable exercise, and the author correctly identifies that the zeroth-order solution is just the gradient flow on the training loss. The decomposition itself is not new—it is Bender–Orszag level—but applying it to this particular cost with the squared-gradient term and a validation-time penalty is a concrete setting that could be worked out. The paper is also honest about assuming the expansions converge, though that assumption is never justified.\n\nThe soft spots are serious. First, the proof of Proposition 2.2 in the appendix is wrong. The adjoint equation there is written with p1 multiplying the ∇B term even at zeroth order, and it drops the third-derivative term that should appear from expanding ∇B. The step where p0 dt is set equal to ∇²J0 is dimensionally inconsistent—vector equals matrix. Those are not typos; they break the proof of the O(ε²) estimate. Second, the paper never defines generalization, and no theorem compares the output parameter's loss on unseen data. Even with correct algebra, the results only show that the approximate parameter approximately minimizes the validation cost. Equations (2.8)–(2.9) are finite-difference identities with no sign or bound. The numerical section plots training versus testing loss, but there is no baseline, no error bars, no repeated splits, and no code, so it cannot support the claim.\n\nWho gets value from this? Possibly a reader who wants to see how perturbation theory can be applied to a control-theoretic learning problem, but they would need to fix the errors themselves. The paper is not suitable for publication in its current form, and I would desk-reject it. If the author repaired the proof, added a rigorous statement about what the method does and does not imply for generalization, and ran a proper comparison, it might become a minor contribution.\n\nRecommendation: do not send to peer review as is; the load-bearing proof error and the missing generalization argument make it a waste of referee time. But it is not incoherent nonsense—the author is thinking about a real problem, just not solving it yet.","headline":"A textbook perturbation expansion applied to a weakly-controlled gradient system, but the central claim about improving generalization is unsupported and the main convergence proof has errors.","tokens_in":8859,"tokens_out":3694,"would_cite":false,"duration_ms":35745,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49K15","34E10","68T05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A weakly-controlled gradient system with a small control parameter splits the learning problem into zeroth- and first-order ODEs whose aggregate parameter estimate is accurate to $O(\\epsilon^2)$ in the validation cost.","keywords":["aggregation","decomposition","generalization","Hamiltonian function","learning problem","optimal control problem","perturbation theory","Pontryagin's maximum principle"],"falsifier":"Pick a low-dimensional regression problem with a non-convex validation loss and a fixed small $\\epsilon$; solve the full optimal control problem numerically by shooting or dynamic programming, and compare the true optimal cost with the paper's first-order prediction $\\Phi(\\theta_0(T), Z^{(2)}) + \\epsilon\\langle\\nabla\\Phi(\\theta_0(T), Z^{(2)}), \\theta_1(T)\\rangle$. If the difference fails to shrink like $O(\\epsilon^2)$ even while the expansions (2.1) appear to converge numerically, the central claim is falsified.","tokens_in":7849,"feed_emoji":"🎯","tokens_out":8676,"duration_ms":76202,"temperature":0.7,"pith_summary":"The paper studies learning as an optimal control problem: model parameters evolve by gradient descent on the training loss, plus a small control term that scales a noise-dithered squared-gradient vector. The cost is the validation loss at the final time, and the control is chosen to minimize it. The central result is that the optimal value and the optimal parameter admit expansions in the small parameter: the zeroth-order term is the plain training flow, and the first-order correction is obtained from a backward adjoint equation plus a linearized forward equation. This yields an approximate optimum, $\\theta_0(T) + \\epsilon\\theta_1(T)$, whose validation cost matches the true optimum up to $O(\\epsilon^2)$, and the same second-order rate separates the true optimal control from the zeroth-order one. The paper also presents a three-step algorithm and numerical demonstrations on thermophysical-property regression.","feed_headline":"First-order correction, second-order accuracy for learned models","feed_subtitle":"Splitting a small-control learning problem into zeroth- and first-order steps yields a provably near-optimal parameter.","key_machinery":"The central object is the small-parameter expansion of the Pontryagin maximum principle for the Hamiltonian $H^\\epsilon(\\theta, p, u) = \\langle p, -\\nabla J_0(\\theta, Z^{(1)}) + \\epsilon u B(\\theta, \\tilde{Z}^{(1)})\\rangle$. Equating equal powers of $\\epsilon$ in the Euler-Lagrange equations yields the decomposed system: the uncontrolled training gradient flow (2.3), the backward adjoint (2.4) whose terminal condition is the negative validation gradient, the zeroth-order control $u_0(t) = \\arg\\max \\langle p_0(t), u B(\\theta_0(t), \\tilde{Z}^{(1)})\\rangle$ (a bang-bang selection when $U = [-1,1]$), and the first-order linearized ODE (2.6) for $\\theta_1$. The machinery converts a single high-dimensional optimal control problem into two forward and one backward ODE solves, aggregating the zeroth- and first-order solutions.","core_discovery":"In the paper's own terms, Proposition 2.1 states that if the optimal state, adjoint, and control expand as $\\theta^\\epsilon = \\theta_0 + \\epsilon\\theta_1 + O(\\epsilon^2)$, $p^\\epsilon = p_0 + \\epsilon p_1 + O(\\epsilon^2)$, $u^\\epsilon = u_0 + \\epsilon u_1 + O(\\epsilon^2)$, then the optimal value satisfies $J^\\epsilon[u^\\epsilon] = \\Phi(\\theta_0(T), Z^{(2)}) + \\epsilon\\langle\\nabla\\Phi(\\theta_0(T), Z^{(2)}), \\theta_1(T)\\rangle + O(\\epsilon^2)$, with $\\theta^* = \\theta_0(T) + \\epsilon\\theta_1(T) + O(\\epsilon^2)$. The zeroth-order pair $(\\theta_0, p_0)$ obeys a forward gradient flow on the training loss and a backward adjoint equation with terminal condition $-\\nabla\\Phi(\\theta_0(T), Z^{(2)})$; the first-order correction $\\theta_1$ is driven by $u_0 B(\\theta_0, \\tilde{Z}^{(1)})$ along a linearization of the training flow. Proposition 2.2 adds the convergence-rate estimate $J^\\epsilon[u^\\epsilon] - J^\\epsilon[u_0] = O(\\epsilon^2)$, showing that the zeroth-order control alone is nearly optimal.","pith_inferences":["If the expansion converges, the same recipe should apply to any gradient-flow-based learning rule: add a small control along a dithering direction, compute the adjoint at the training solution, and take one linearized correction step; the paper does not test this generality, but it follows directly from the derivation.","The role of the added noise in $\\tilde{Z}^{(1)}$ is to give the control a nonzero direction of action; an interesting testable extension is whether tuning the noise variance $\\sigma^2$ provides a practical way to control the size of the generalization improvement.","The $O(\\epsilon^2)$ rate suggests that higher-order corrections could be computed by continuing the asymptotic expansion, giving a systematic hierarchy of increasingly accurate training-aware validation optimizers, a direction not explored in the paper.","Because the theory assumes existence and convergence of the expansions, a natural stress test would be to run the algorithm on problems where gradient descent has bifurcations or non-unique minimizers; the paper's numerical examples are benign polynomial fits."],"forward_implications":["The approximate parameter $\\theta^* = \\theta_0(T) + \\epsilon\\theta_1(T)$ is computable by solving two forward ODEs and one backward ODE, avoiding a full nonlinear optimal-control solve.","The validation cost of this approximation is within $O(\\epsilon^2)$ of the true minimum, and the same $O(\\epsilon^2)$ gap separates the optimal control from the zeroth-order control $u_0$.","Generalization improvement is explicitly first-order: the $\\epsilon$ term in the expansion is the directional derivative of the validation loss along $\\theta_1(T)$, so the correction is a targeted step toward validation performance.","The decomposition is recursive in spirit: each order of the expansion can be solved independently and then aggregated, matching the paper's stated strategy.","For $U$ equal to the interval $[-1,1]$, the zeroth-order control is bang-bang, so the algorithm's control selection reduces to a switching rule along the adjoint direction."],"supporting_citations":[{"why":"Supplies Filippov's theorem and the reachable-set compactness argument used in assumption (iii) to ensure existence of the optimal control.","marker":"[1]"},{"why":"Provides Pontryagin's maximum principle, from which the Euler-Lagrange optimality conditions (1.3)-(1.5) are derived.","marker":"[2]"},{"why":"Supplies the perturbation-theory background for the asymptotic expansions of $\\theta$, $p$, and $u$ in the small parameter.","marker":"[4]"},{"why":"Supplies the asymptotic method used for expanding solutions of the differential equations in powers of $\\epsilon$.","marker":"[5]"}],"fun_headline_variants":["Small-parameter method sharpens learning generalization","Perturbation trick yields near-optimal learning","O(ε²) accuracy from a two-step learning split","Zeroth plus first order: near-optimal learning","Weak control, strong generalization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result depends on the assumption that the state, adjoint, and control have convergent power-series expansions in the small parameter $\\epsilon$ that satisfy the optimality equations; the paper states this as a condition rather than proving it.","fun_headline_variants_meta":{"raw":{"variants":["Small-parameter method sharpens learning generalization","Perturbation trick yields near-optimal learning","O(ε²) accuracy from a two-step learning split","Zeroth plus first order: near-optimal learning","Weak control, strong generalization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000539,"raw_usage":{"total_tokens":2641,"prompt_tokens":1058,"completion_tokens":1583,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":674,"completion_tokens_details":{"reasoning_tokens":1511}},"tokens_in":674,"tokens_out":1583,"duration_ms":12542,"temperature":1.0,"reasoning_tokens":1511,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:35:33.347044+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pick a low-dimensional regression problem with a non-convex validation loss and a fixed small $\\epsilon$; solve the full optimal control problem numerically by shooting or dynamic programming, and compare the true optimal cost with the paper's first-order prediction $\\Phi(\\theta_0(T), Z^{(2)}) + \\epsilon\\langle\\nabla\\Phi(\\theta_0(T), Z^{(2)}), \\theta_1(T)\\rangle$. If the difference fails to shrink like $O(\\epsilon^2)$ even while the expansions (2.1) appear to converge numerically, the central claim is falsified.","supporting_citations":[{"cited_title":"Bender & S.a","cited_arxiv_id":null,"evidence_quote":"Supplies the perturbation-theory background for the asymptotic expansions of $\\theta$, $p$, and $u$ in the small parameter."},{"cited_title":"Kevorkian & J.D","cited_arxiv_id":null,"evidence_quote":"Supplies the asymptotic method used for expanding solutions of the differential equations in powers of $\\epsilon$."}],"review_version":1}