{"id":"bf1de329-2d98-4128-9afc-c8333da3eda5","arxiv_id":"2505.19454","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"This paper introduces a modal integral-collocation transcription for optimal control in which each coordinate is represented by one set of Chebyshev or Legendre coefficients for its highest derivative, and demonstrates it on benchmarks, orbit raising, and a rocket landing flip.","lead":"This paper presents a way to convert optimal control problems into numerical optimization by approximating only the highest derivative of each state with Chebyshev or Legendre polynomials, then integrating to obtain the states. It tests the method on benchmark problems, orbit transfers, and a Starship-style landing flip, showing it works and that the choice of polynomial basis affects speed and reliability.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The CGL quadrature weights in Eqs. (32a)-(32b) are internally inconsistent and sum to 17/6 for n=2, so the printed transcription cannot reproduce the reported cost functionals.","rationale":"The reader's weakest assumption was the global-polynomial limitation and node-only constraint enforcement. Those are legitimate scope limitations, but they are not internally inconsistent with the method's stated design. I found a more concrete and checkable problem: the CGL quadrature formulas in Section II.B are algebraically wrong as printed. For n=2 they do not integrate a constant, which means the cost functional in Eq. (68) is not faithfully discretized for that node set. Since the central claim is that DOPIC is a valid direct pseudospectral transcription and the numerical examples validate the method by matching analytical costs, an error in the quadrature formulas directly undermines reproducibility of the evidence. The fix is elementary, and the numerical results might still be correct if the authors used the standard Clenshaw-Curtis weights in their code; but the paper must be corrected before the validation can be accepted. I therefore keep the CONDITIONAL verdict but shift the condition: correct and re-report the CGL quadrature weights, or provide code confirming the weights actually used.","tokens_in":25979,"tokens_out":18010,"duration_ms":160530,"concrete_test":"For n=2, construct the CGL nodes tau=[-1,0,1], compute the weights from Eqs. (32a)-(32b), and check whether the sum equals 2 (necessary for exact integration of constants). If the sum is not 2, the published quadrature is internally inconsistent; then recompute the minimum-fuel problem at n=4 using both the printed weights and the corrected Clenshaw-Curtis weights, and compare J to the analytical value 14.142. If the printed-weights cost does not match to the reported accuracy, the validation figures cannot be reproduced from the paper as written.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The CGL quadrature weights printed in Eqs. (32a)-(32b) do not define a valid quadrature rule. For n=2 (nodes -1, 0, 1), Eq. (32a) gives w1 = (2/2)(1 - [cos(0)/(0^2-1) + cos(2*pi/2)/(4*1^2-1)]) = 1*(1 - (-1 - 1/3)) = 7/3, while Eq. (32b) gives w0 = w2 = 1/2^2 = 1/4. The weights sum to 17/6, not 2, so the rule fails the basic test of integrating f(tau)=1 exactly over [-1,1]. The correct Clenshaw-Curtis weights for these nodes are w0 = w2 = 1/3 and w1 = 4/3. Because Eq. (68) uses these weights to evaluate the Bolza cost, an independent implementation following the paper will compute incorrect performance indices. The claimed convergence of J to analytical values (e.g., 14.142 for the minimum-fuel problem) therefore either relies on unstated corrected weights or is numerically suspect. This is a direct inconsistency in the transcription as published and blocks reproduction of the central validation.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Direct Orthogonal Polynomial Integral Collocation (DOPIC), a modal direct pseudospectral transcription of optimal control problems. The highest derivative of each state is represented by a single Chebyshev or Legendre polynomial series, and all lower derivative levels are obtained by repeated exact integration of that series, so that only one vector of modal coefficients is needed per coordinate. The Bolza cost is evaluated by quadrature, controls are represented nodally, and the resulting NLP is solved with fmincon. The method is demonstrated on a second-order minimum-fuel problem, the Breakwell bounded-state problem, planar circular orbit raising (minimum-time and maximum-radius variants), and a simplified Starship landing flip maneuver. The authors conclude that the single-coefficient-set representation reduces the parameter space and that the examples validate the method.","tokens_in":26210,"tokens_out":9556,"duration_ms":89461,"significance":"If the formulas as printed were correct, DOPIC would be a useful and reasonably general modal alternative to the established nodal pseudospectral methods: it applies to Chebyshev and Legendre bases with their standard grids, avoids rewriting dynamics in first-order form, and the benchmark results against analytical values and published orbit-raising transfer times are encouraging. The paper also documents several nontrivial limitations honestly, including the even/odd Breakwell artifact and the restriction to global collocation. However, the printed quadrature weights for the Chebyshev and CP2K node sets are not valid integration rules, and because Eq. (68) evaluates the cost with those weights, the reported performance indices cannot be reproduced from the manuscript as written. This defect is load-bearing for the central validation and must be fixed. The absence of any convergence theorem or error analysis also means the method's validity currently rests entirely on numerical evidence.","major_comments":[{"comment":"The quadrature weight formulas for the Chebyshev-Gauss, Chebyshev-Gauss-Lobatto, and CP2K nodes are internally inconsistent and do not define valid quadrature rules. For n=2, Eqs. (32a)–(32b) give w0=w2=1/4 and w1=7/3, which sum to 17/6 instead of 2; the rule therefore does not integrate f(tau)=1 exactly over [-1,1]. The correct Clenshaw-Curtis weights for these three nodes are w0=w2=1/3 and w1=4/3. The same defect appears in Eq. (31): for n=1 the two CG nodes each receive weight 2, again summing to 4 instead of 2, and Eq. (33) similarly gives weight 2 at each of the two CP2K nodes for n=1. Since Eq. (68) uses these weights to approximate the Bolza cost, an independent implementation following the printed transcription will compute incorrect performance indices and cannot reproduce the reported values such as J=14.142 in Fig. 2 or J=2.240 in Fig. 4. Please correct the weight formulas (or clearly state which different weights the actual code used) and re-run the affected examples.","section":"II.B (Eqs. 31–33) and III.C (Eq. 68)"},{"comment":"The Breakwell results for odd approximation orders are presented as valid solutions even though the continuous state constraint x(t) <= l is enforced only at the collocation nodes. As the paper itself notes, odd orders place no node at the constraint peak, so the constraint is inactive at every node and the reported cost underestimates the analytical value; in fact the polynomial state history violates the continuous constraint between nodes. This is not merely a cosmetic artifact: it means the odd-order performance indices in Fig. 4 correspond to infeasible trajectories. Please either enforce the constraint on a denser set of points, add a quantitative statement of maximum inter-node constraint violation, or explicitly exclude odd-order solutions from the accuracy claims.","section":"IV.B (Eq. 78)"},{"comment":"The paper offers no convergence theorem or error bound for the DOPIC transcription, and it explicitly restricts the presentation to global collocation (III.C). For a new transcription method, the central claim that DOPIC is a valid and accurate pseudospectral method would be materially strengthened by at least a convergence statement for smooth problems and a discussion of how the polynomial order relates to the NLP solution error. In the current manuscript, after the quadrature issue is fixed, the only support for the central claim is the numerical evidence; the Starship problem is additionally verified only by internal consistency across node sets, since no external solution is provided. Please add a formal or at least quantitative convergence discussion, and state clearly which of the numerical claims are validated externally and which are not.","section":"III.C and V"}],"minor_comments":[{"comment":"The recurrence for Chebyshev polynomials of the second kind is misprinted: U_{n+1}(tau) = 2 tau U_n(tau) - U_n(tau) should read U_{n+1}(tau) = 2 tau U_n(tau) - U_{n-1}(tau).","section":"II.A (Eq. 14)"},{"comment":"There is a spelling error in 'Chebyshev-Guass-Lobatto' (should be 'Gauss'), and Section IV contains the duplicated phrase 'All computations are are run'.","section":"II.A"},{"comment":"The caption of Fig. 2 refers to the 'second order minimum control problem'; the problem is the second-order minimum-fuel problem. This should be made consistent.","section":"Fig. 2"},{"comment":"The claim that 'no additional coefficients are needed for each integration' is slightly misleading because the integration matrices B_j grow in dimension at each integration level; the statement would be clearer as 'no additional independent coefficient vectors are introduced'.","section":"III.A"}],"recommendation":"major_revision","confidential_remarks":"The decisive issue is the quadrature-weight defect in Eqs. (31)–(33): it blocks reproduction of the paper's central validation and should be fixed before acceptance. The novelty relative to the authors' earlier ICC work in [27,32] is moderate—the generalization to CGL, LG, LGL, and CP2K is real, but the core integration-matrix idea is already present in those references; a clearer novelty statement would help. With the formulas corrected and the Breakwell odd-order infeasibility addressed quantitatively, the paper would be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the modal integral-collocation transcription is a real and useful idea, but the printed quadrature weights for the Chebyshev grids are wrong, so as published the paper cannot reproduce its own cost numbers. A corrected version deserves a serious referee.\n\nWhat is actually new: the paper generalizes the successive-integration transcription to CGL nodes, Chebyshev polynomials of the second kind, and Legendre LG/LGL grids, with one modal coefficient set per coordinate. The integration-matrix derivation in Section III is coherent, and the benchmark problems match analytic solutions and published orbit-raising transfer times. The basis comparison and runtime data are genuinely informative.\n\nThe serious soft spot is the quadrature weights. Equations (31) and (32) include the i=0 term in the sum with denominator -1, which injects a spurious constant into every weight. I checked: for n=2, the CGL weights computed from Eq. (32) sum to 17/6 instead of 2, and the CG weights from Eq. (31) sum to 4 instead of 2. Since Eq. (68) uses these weights to evaluate the Bolza cost, anyone implementing the paper as written will compute wrong performance indices. This is almost certainly a typographical or derivation slip, because the reported numbers are consistent with the correct Clenshaw-Curtis/Fejér formulas, but it is a concrete reproducibility blocker.\n\nOther issues are proportionate and more minor. No code or data are provided, so the Starship flip result is unverifiable beyond internal consistency. Constraints are only enforced at the nodes, which explains the even/odd Breakwell artifact but also means inter-node constraint violations can go undetected. And the conclusion claims the method works with “any orthogonal polynomial set,” which oversells what was actually demonstrated for three families.\n\nBottom line: a genuinely useful transcription with a clear mathematical core, honest external benchmarking, and a fixable but important error in the printed formulas. The paper should go to peer review, with the authors asked to correct the quadrature weights, clarify the inter-node constraint limitation, and ideally release a minimal code implementation so the numerical claims can be independently checked.","headline":"The modal integral-collocation idea is sound, but the Chebyshev quadrature weights are misprinted, which blocks reproduction until fixed.","tokens_in":26740,"tokens_out":6658,"would_cite":false,"duration_ms":59261,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49M37","65L60","65D32"],"pacs":[],"model":"deepseek-v4-flash","headline":"A new direct pseudospectral transcription, orthogonal polynomial integral collocation, represents only the highest-order derivative of each coordinate and integrates it to recover all lower state levels, reducing the NLP decision-variable…","keywords":["direct pseudospectral methods","orthogonal polynomial integral collocation","Chebyshev polynomials","Legendre polynomials","trajectory optimization","nonlinear programming transcription","rocket landing flip maneuver"],"falsifier":"Take a bang-bang problem with a known switching time that does not coincide with any collocation node, compute the maximum deviation between the OPIC polynomial and the true highest-order derivative as $n$ increases, and also evaluate the state constraint at off-node points. If the error fails to decay spectrally or the constraint is violated between nodes, as the odd-order Breakwell solution already suggests can happen, the single-global-polynomial assumption is falsified.","tokens_in":25779,"feed_emoji":"🚀","tokens_out":6788,"duration_ms":53063,"temperature":0.7,"pith_summary":"This paper introduces a direct transcription method, orthogonal polynomial integral collocation (OPIC), for converting a continuous-time optimal control problem into a nonlinear program. Rather than representing each state level on a grid and using differentiation matrices to enforce the dynamics, OPIC expands only the highest-order derivative of each coordinate in an orthogonal polynomial series and obtains every lower derivative level by repeated exact integration. The result is a modal formulation in which one set of polynomial coefficients carries an entire degree of freedom, so the decision-variable vector of the NLP is much shorter than in standard nodal pseudospectral methods. The paper verifies the method on the minimum-fuel and Breakwell benchmark problems with analytical solutions, reproduces published planar circular orbit-raising transfer times, and solves a constrained Starship-style landing flip maneuver. A consistent finding across examples is that the choice of polynomial basis and associated grid materially affects runtime, accuracy, and convergence.","feed_headline":"Integrate, don't differentiate: a leaner way to solve optimal control","feed_subtitle":"The new transcription keeps one coefficient set per coordinate and matches published orbit-raising times.","key_machinery":"The mechanism is the integration-weight matrix chain. For each polynomial family, the indefinite integrals of the basis functions close onto adjacent basis functions, $\\int \\phi_i = b_i^+ \\phi_{i+1} + b_i^- \\phi_{i-1}$, so integrating the modal series $\\sum \\phi_i \\alpha_i$ is exactly a sparse matrix product $B_j$ with entries only on the super- and sub-diagonals. Repeatedly applying $B_j$ and subtracting the lower-bound evaluation vectors $v_j$ reconstructs every derivative of the state from the single vector $\\alpha$, while initial conditions enter through the polynomial terms $p_{m-j}(\\tau)$. This single-set-per-coordinate property is what replaces the state-ladder variables of nodal pseudospectral methods and drives the reduction in NLP dimension.","core_discovery":"The central claim is that a single $n$th-order orthogonal polynomial series can represent the highest-order derivative $\\ddot{y}$ (or $y^{(q)}$) of each coordinate, and that successive integrations of that series, carried out exactly through sparse integration-weight matrices $B_j$ built from the polynomial recurrences, yield all lower derivative levels with no additional coefficients. Collocation enforces the dynamics at the nodes, boundary conditions are imposed on the integrated state expressions, and controls stay nodal so that discontinuous or bang-bang control can still be represented. With this construction, the state part of the NLP decision vector contains only the $n+1$ modal coefficients $\\alpha$ per coordinate instead of a full ladder of nodal state values, which is the paper's main efficiency claim. The numerical experiments show that all tested bases, Chebyshev of the first and second kind and Legendre on their associated grids, reproduce the analytical benchmarks, match the orbit-raising transfer time of about $t_f \\simeq 3.32$ TU reported by earlier pseudospectral studies, and yield similar landing flip solutions, with second-kind Chebyshev often fastest and Legendre grids most reliable.","pith_inferences":["If domain segmentation were added, the modal coefficient vector per segment would become the natural adaptation target, allowing piecewise error near switching times to be reduced locally; nothing in the integral formulation prevents this extension.","The even-odd Breakwell discrepancy suggests a practical rule for users: when a state constraint may be active, prefer grids that include the constraint boundary or add off-node constraint checks, because node-only enforcement can leave the constraint violated between nodes.","The parameter-space advantage should be largest for high-order dynamics; a decisive head-to-head test would compare DOPIC with Gauss, Radau, and Legendre nodal methods on a fourth-order system at equal node counts, reporting decision-variable counts, accuracy, and solver time.","The paper leaves costate recovery open, but because the state is a global modal expansion, the NLP Lagrange multipliers could plausibly be mapped back to the indirect problem via a covector mapping theorem, turning DOPIC into a warm start for shooting methods."],"forward_implications":["For a second-order state, the state part of the NLP decision vector drops from roughly $2(n+1)$ nodal values to $n+1$ modal coefficients, with the gap growing for higher-order dynamics.","Because integration is a smoothing operation, derivative and state histories produced from the integrated series should be less error-amplifying than those obtained by differentiating nodal state values.","The method inherits spectral accuracy on smooth problems, so on trajectories whose highest derivative is smooth, increasing $n$ should drive the transcribed solution toward the continuous optimum quickly.","Control stays nodal, so the modal state representation does not prevent representing piecewise or bang-bang control histories.","Polynomial and grid choice is not neutral: on the tested problems, second-kind Chebyshev gave the lowest average runtime for the minimum-fuel case while Legendre LG and LGL grids gave the highest convergence rates on orbit raising."],"supporting_citations":[{"why":"Supplies the integral Chebyshev collocation foundation and the IVP/BVP solution framework that OPIC generalizes.","marker":"[27]"},{"why":"Presents the Gegenbauer integral transcription most similar to OPIC, giving the baseline for the integration-based approach.","marker":"[26]"},{"why":"Provides the discrete orthogonality relations, CP2K weights, and recurrence integrals used to build the collocation matrices.","marker":"[31]"},{"why":"Gives the spectral accuracy and minimax optimality results that justify the Chebyshev interpolation nodes.","marker":"[13]"},{"why":"Provides the analytical solutions for the minimum-fuel and Breakwell problems that serve as accuracy benchmarks.","marker":"[47]"},{"why":"Introduces the Legendre pseudospectral method whose orbit-raising results are used as the comparison target for the planar orbit transfer.","marker":"[17]"},{"why":"Gives a Chebyshev pseudospectral treatment of constrained optimal control used to further verify the orbit-raising solution.","marker":"[48]"},{"why":"Defines the pseudospectral optimal-control framework and notation within which OPIC is presented and compared.","marker":"[16]"},{"why":"Surveys the Gauss, Radau, and Legendre integral formulations that motivate the idea of enforcing dynamics through integration.","marker":"[22]"}],"fun_headline_variants":["Integral collocation: one coefficient set per coordinate for optimal control","Skip derivative states: integral collocation for optimal control","Integrating high-order derivatives for leaner optimal control","Pseudospectral optimal control by integrating, not differentiating"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that one global polynomial, fixed in advance and spanning the whole time interval, can faithfully capture the highest-order derivative of the trajectory, even when that derivative contains sharp switches or corners; the method does not adapt by splitting the domain.","fun_headline_variants_meta":{"raw":{"variants":["Integral collocation: one coefficient set per coordinate for optimal control","Skip derivative states: integral collocation for optimal control","Integrating high-order derivatives for leaner optimal control","Pseudospectral optimal control by integrating, not differentiating"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000974,"raw_usage":{"total_tokens":4141,"prompt_tokens":948,"completion_tokens":3193,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":3127}},"tokens_in":564,"tokens_out":3193,"duration_ms":19572,"temperature":1.0,"reasoning_tokens":3127,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:14:28.332734+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a bang-bang problem with a known switching time that does not coincide with any collocation node, compute the maximum deviation between the OPIC polynomial and the true highest-order derivative as $n$ increases, and also evaluate the state constraint at off-node points. If the error fails to decay spectrally or the constraint is violated between nodes, as the odd-order Breakwell solution already suggests can happen, the single-global-polynomial assumption is falsified.","supporting_citations":[{"cited_title":"Adaptive Collocation Methods Using Chebyshev Integration,","cited_arxiv_id":null,"evidence_quote":"Supplies the integral Chebyshev collocation foundation and the IVP/BVP solution framework that OPIC generalizes."},{"cited_title":"Fast, accurate, and small-scale direct trajectory optimization using a Gegenbauer transcription method,","cited_arxiv_id":null,"evidence_quote":"Presents the Gegenbauer integral transcription most similar to OPIC, giving the baseline for the integration-based approach."},{"cited_title":"B.,Chebyshev Polynomials in Numerical Analysis, Oxford University Press, 1968","cited_arxiv_id":null,"evidence_quote":"Gives the spectral accuracy and minimax optimality results that justify the Chebyshev interpolation nodes."},{"cited_title":"E., and Ho, Y","cited_arxiv_id":null,"evidence_quote":"Provides the analytical solutions for the minimum-fuel and Breakwell problems that serve as accuracy benchmarks."},{"cited_title":"Pseudospectral Chebyshev optimal control of constrained nonlinear dynamical systems,","cited_arxiv_id":null,"evidence_quote":"Gives a Chebyshev pseudospectral treatment of constrained optimal control used to further verify the orbit-raising solution."},{"cited_title":"An overview of three pseudospectral methods for the numerical solution of optimal control problems,","cited_arxiv_id":null,"evidence_quote":"Surveys the Gauss, Radau, and Legendre integral formulations that motivate the idea of enforcing dynamics through integration."}],"review_version":1}