{"id":"38a2caaf-5318-4b2c-9773-e340ad40c7c5","arxiv_id":"2504.16328","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"Parameterizing positive-definite penalty matrices through eigenvalues and rotation angles and optimizing them with particle swarm optimization reduced control cost by up to 65% versus diagonal matrices in tested aerospace examples.","lead":"This paper tunes full, non-diagonal penalty matrices for controllers using eigendecomposition, and compares them with standard diagonal matrices in three aerospace control problems. The full matrices improved the optimized control cost by up to 65% in the tested cases, though the comparison depends on how well the tuning optimizer ran.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The empirical 65% claim depends on PSO reaching near-optimal diagonal and full solutions with equal effort; the paper's admitted limited PSO runs and absent statistical tests leave the improvement as a possible search artifact.","rationale":"I read the paper as claiming two things: (i) a practical parameterization of full SPD penalty matrices that enforces definiteness and spans a broader set than diagonal matrices, and (ii) numerical evidence that this broader set yields up to 65% improvement in aerospace control examples. Part (i) is technically sound and independently reproducible: Eq. (2) with positive eigenvalues and an orthogonal Q parameterized via spherical coordinates/Gram-Schmidt indeed produces SPD matrices, and diagonal SPD matrices are included when Q=I. The provided GitHub code and Appendix A matrices support reproducibility. Part (ii) is where the argument must carry the paper's significance. The evidence consists of PSO comparisons with 5-10 stochastic runs, no statistical tests, and explicitly arbitrary PSO settings (Sections 4.1 and 6). Because the full parameterization changes both the feasible set and the optimization landscape, lower average costs are confounded with PSO search behavior. The modified Case E benchmark for Q-law adds another confound. This is the same soft spot the reader identified; my read strengthens it only by adding that the single-initial-condition setup makes overfitting a plausible alternative explanation. A controlled equal-budget statistical comparison would settle whether the improvement is real. If it survives, the paper merits acceptance as a useful engineering contribution; if not, the paper's contribution reduces to the (correct but less surprising) observation that a superset parameterization weakly dominates.","tokens_in":19804,"tokens_out":12123,"duration_ms":124662,"concrete_test":"Re-run the Section 4.1 LQR detumbling case (source of the 65% claim) with a fixed computational budget: identical number of objective evaluations and same random seed stream for both diagonal and full parameterizations, over 100 PSO restarts each, reporting full distributions and a Mann-Whitney test. If the full-matrix distribution does not stochastically dominate the diagonal distribution at the 95% level, or if the best-of-100 gap is inconsistent with the reported 65%, the headline improvement is an optimization artifact rather than evidence for the parameterization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two layers. The mathematical layer—that eigendecomposition with positive eigenvalues implicitly parameterizes the full SPD cone and therefore contains the diagonal parameterization as a subset—is sound; Eq. (2) plus the GEAGSP construction is a legitimate parameterization. The load-bearing layer is empirical: the abstract's 'improvements up to 65%' requires that the reported diagonal results are near the diagonal-constrained optimum, and that the full results beat them for reasons related to cross-coupling rather than search artifacts. This is exactly where the paper is weakest. Section 4.1 states swarm size and iteration limits were 'arbitrarily chosen such that solutions we deemed to be optimal ... could be obtained', and Section 6 concedes 'solutions are by no means optimal due to limited runs of PSO'. Tables 1, 2, 4, and 5 report 5-10 stochastic PSO runs with no confidence intervals or hypothesis tests; Case E for Q-law is run on a modified problem (E*) because the original boundary conditions did not converge. Since the full parameterization has a different, higher-dimensional search space (e.g., 27 vs. 9 variables for the 6-state LQR), unequal local-minimum trapping or more effective exploration could produce the lower costs even if the true diagonal-constrained optimum is equal or better. Overfitting is also plausible: each problem uses one initial condition, so the extra off-diagonal degrees of freedom can tune to that scenario. The paper does not establish that the improvement is caused by the enlarged SPD cone rather than by how PSO explores it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes parameterizing positive-definite penalty matrices as K = Q Λ Q^T, where Λ is a diagonal matrix of positive eigenvalues and Q is an orthogonal matrix built with a generalized Euler-angle/Gram-Schmidt parameterization (GEAGSP). The parameters are tuned with particle swarm optimization (PSO), and the method is compared against conventional diagonal penalty matrices in three aerospace control settings: a Lyapunov-stabilized Zermelo navigation problem, LQR and Lyapunov-based spacecraft attitude control, and low-thrust trajectory design using both a generic Lyapunov law and Q-law. The reported results show lower control effort, time of flight, or higher final mass for the full parameterization, with the abstract citing improvements up to 65%. The paper also provides numerical penalty matrices in Appendix A and a public code repository.","tokens_in":20138,"tokens_out":6693,"duration_ms":68052,"significance":"The mathematical core is sound: with Λ > 0 and Q ranging over O(N), the factorization K = Q Λ Q^T is a legitimate parameterization of the positive-definite cone, and it contains the diagonal parameterization as a subset. The manuscript is careful about reproducibility details (integration tolerances, PSO settings, event-detection thresholds, and concrete penalty matrices in Appendix A), and the presentation of the GEAGSP algorithm is a useful practical contribution. If the reported improvements are genuine, the method gives control practitioners a simple way to enlarge the searchable design space for LQR and Lyapunov-type controllers. However, the central empirical claim currently rests on a small number of stochastic PSO runs with no statistical analysis, on an admitted lack of optimality, and on one benchmark that was modified to force convergence; these weaknesses are load-bearing for the stated conclusions.","major_comments":[{"comment":"The abstract's 'improvements up to 65%' is an empirical claim that assumes PSO reaches comparably converged optima for the diagonal and full parameterizations. The paper itself states that the PSO swarm size and iteration limits were 'arbitrarily chosen such that solutions we deemed to be optimal ... could be obtained' (§4.1) and that 'solutions are by no means optimal due to limited runs of PSO' (§6). Tables 1, 2, 4, and 5 report only 5–10 runs per case with no confidence intervals or hypothesis tests, and the full-parameterization runs exhibit large variance (e.g., Table 1 rest-to-rest Q2,R2 values range from 0.06420 to 0.38302, and Table 2 detumbling Kp,2,Kd,2 values range from 0.05594 to 0.10420). Section 3 reports only a single pair of costs for Zermelo, despite PSO being stochastic. Because the full parameterization has a different and higher-dimensional search space, unequal local-minimum trapping could produce lower reported costs even if the true diagonal-constrained optimum is no worse. Please provide evidence that the diagonal results are at or near the diagonal-constrained optimum (e.g., many more multi-start runs, convergence diagnostics, or comparison with a reference solution), and add a statistical comparison of the two parameterizations.","section":"§3–§5, Tables 1, 2, 4, 5; §4.1; §6"},{"comment":"The Q-law Case E is not solved on the original benchmark from Ref. [6]; Table 3 shows that the 'Case E*' boundary conditions change the initial inclination from 0.06° to 30.06° and the target inclination from 116° to 86° because the original case 'did not converge' with the authors' Q-law implementation. This modification is not a minor numerical detail: it changes the difficulty and shape of the transfer, and the improvement reported for Case E may be specific to the modified case rather than to the penalty-matrix parameterization. Please either solve the original Case E or justify explicitly why Case E* is a legitimate proxy, and report the convergence failure of the original case separately from the performance comparison.","section":"§5, Table 3, Case E*"},{"comment":"Each example is optimized for a single initial condition or a single transfer scenario, and the extra off-diagonal degrees of freedom can simply tune to that one scenario. This is a concrete overfitting risk for the 65% LQR improvement in particular, since the full parameterization has many more free parameters than the diagonal one while the validation is performed on the same trajectory used for tuning. Please test robustness by re-optimizing over several initial conditions or transfer scenarios and reporting the distribution of improvements, or otherwise demonstrate that the gains obtained with the full parameterization are not scenario-specific.","section":"§4 and §5 (LQR attitude, Lyapunov attitude, low-thrust transfers)"},{"comment":"The paper asserts that the Cayley transform and Givens rotation parameterizations are 'comparable in performance' to GEAGSP but provides no numerical comparison or supporting data for this claim. If these alternatives are not used in the results, the statement should be either supported with a small comparative experiment or removed.","section":"§2.1–§2.3"}],"minor_comments":[{"comment":"The indexing in the displayed symmetric matrix does not correctly define an M = (N^2+N)/2 parameterization of a symmetric matrix: the row patterns shown (e.g., row 2 begins with k_{N−1}, row 3 begins with k_3) are inconsistent, and the placement of diagonal entries is unclear. Please rewrite the display with standard symmetric indexing or with an explicit small example such as N = 3.","section":"§2, Eq. (1)"},{"comment":"The caption of Figure 10(b) reads 'Case E trajectory: minimum-fuel Q-law with K2', but Section 5.1 describes these panels as the generic Lyapunov control-law trajectories for Case E. Please correct the caption so that it matches the content.","section":"§5.1, Figure 10 caption"},{"comment":"The expression Q_{1,2} = (1 + W_P P) (SND)^T K_{1,2} (ND) is ambiguous because the parentheses around SND and ND are not placed. Please write the expression as (S N D)^T K (N D) and state the dimensions of S, N, and D.","section":"§5.2, Eq. (30)"},{"comment":"The input line says j = 1,...,N−i, which is empty for i = N, and the Gram–Schmidt step in line 6 is described only verbally. A short clarifying sentence about the N = i case and the construction of the basis B would improve reproducibility.","section":"§2.3, Algorithm 1"},{"comment":"Section 4.1 reports an average decrease of 65.4697% for the detumbling LQR case, while Section 6 says 'in the most extreme case, the improvement in performance was 65%.' Because 65% is the average improvement for one maneuver, not an extreme-case value, please clarify whether the quoted number refers to a mean or to a single best run.","section":"§4.1 and §6"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about its limitations, and the core parameterization is mathematically standard and correctly presented. The main risk is that the headline empirical claim (65% improvement) may not survive a more careful optimization comparison; however, the issue is fixable within the manuscript's scope by adding statistical validation, more runs, and a defensible treatment of Case E. I would encourage the editor to request a major revision rather than rejection, because the practical tuning method could be valuable if validated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a follow-up to the authors' earlier arXiv paper [17], applying an eigendecomposition parameterization of full positive-definite penalty matrices to three aerospace control settings. The genuinely new part is the numerical study—Zermelo navigation, LQR and Lyapunov attitude control, and generic Lyapunov plus Q-law low-thrust transfers—comparing full matrices against diagonal ones. The parameterization itself is standard (eigendecomposition plus an orthogonal-matrix parameterization), but it is clearly explained and the GEAGSP variant is implemented with code on GitHub.\n\nWhat it does well: the mathematics is sound. Eq. (2) with positive eigenvalues and a full orthogonal Q parameterizes the SPD cone, so diagonal matrices are a subset; the claim that full matrices broaden the achievable set is true at the feasible-set level. The paper is transparent about matrices in the appendix, gives tolerances, bounds, swarm sizes, and iteration limits, and explicitly concedes in Section 6 that the PSO solutions are not optimal. That honesty helps.\n\nThe soft spot is the empirical comparison. The 65% figure comes from 5–10 stochastic PSO runs with no confidence intervals or hypothesis tests. Section 4.1 says swarm sizes and iteration limits were 'arbitrarily chosen' to get solutions the authors deemed optimal, and Section 6 admits limited runs. Since the full parameterization has more decision variables (27 vs 9 for the 6-state LQR), the lower cost could come from a more flexible search space finding better local optima, not necessarily from cross-coupling. The modified Case E* for Q-law is a second red flag, though disclosed. Single initial conditions per problem also leave room for overfitting to that scenario.\n\nI don't think this sinks the paper. The improvements are consistent across independent cases, and in several low-thrust runs the diagonal cost is nearly constant while the full cost varies and drops, which looks like a real difference rather than pure noise. Still, the headline number needs a controlled comparison: equal function-evaluation budgets, more restarts, and statistical summaries. This paper is for practitioners who tune LQR, CLF, or Q-law gains and want a practical way to explore full penalty matrices. I'd send it to review, with the controlled comparison as the main requested revision.","headline":"Useful numerical extension of the authors' own GEAGSP parameterization, but the headline 65% improvement is not yet supported by a controlled optimization comparison.","tokens_in":20699,"tokens_out":3129,"would_cite":false,"duration_ms":31214,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15B48","49N10","93D05","70M20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Replacing diagonal penalty matrices with full positive-definite matrices built from eigenvalues and eigenvector rotations improves control performance by up to 65% in the reported aerospace examples.","keywords":["penalty matrix parameterization","eigendecomposition","positive-definite matrices","linear quadratic regulator","Lyapunov-based control","low-thrust trajectory optimization","Q-law","spacecraft attitude control"],"falsifier":"For the two-state Zermelo example, solve the diagonal parameterization $(k_1,k_2)$ and the full eigendecomposition parameterization $(\\lambda_1,\\lambda_2,\\theta)$ to verified global optimality with the same integration tolerances and compare the best costs; if the diagonal optimum equals the full optimum, the claimed advantage of the off-diagonal terms evaporates.","tokens_in":1715,"feed_emoji":"🛰️","tokens_out":1682,"duration_ms":90798,"temperature":0.7,"pith_summary":"The paper's thesis is that the usual diagonal penalty matrices in quadratic costs and Lyapunov functions are too restrictive: they omit cross-coupling terms that a full positive-definite matrix can provide, and full matrices can be searched just as easily if parameterized by eigendecomposition. It supplies an explicit parameterization, eigenvalues plus a minimally parameterized orthogonal eigenvector matrix, that makes positive-definiteness automatic, and optimizes those parameters with particle swarm optimization in three aerospace testbeds. Across LQR attitude control, Lyapunov attitude control, and low-thrust trajectory design, the full matrices reduce the performance objective, with the largest reported gain being 65%, and improve the mean objective in every reported case. A practitioner who adopts this parameterization can explore a larger class of feedback laws without changing the controller structure.","feed_headline":"Full penalty matrices cut control costs by up to 65%","feed_subtitle":"Eigendecomposition tuning keeps cross-coupling terms in quadratic costs, improving LQR, Lyapunov, and low-thrust results.","key_machinery":"The central object is the eigendecomposition $K = Q\\Lambda Q^{\\top}$ of a symmetric positive-definite penalty matrix. $\\Lambda$ holds positive eigenvalues, so sign-definiteness is automatic, and $Q$ is an orthogonal eigenvector matrix parameterized with $M = N(N-1)/2$ angular variables via a generalization of Euler angles combined with Gram-Schmidt orthogonalization. This converts the search for a positive-definite matrix into an unconstrained search over bounded angles and positive eigenvalues, while adding cross-coupling terms that a diagonal $K$ cannot represent. For full matrices, a diagonal binary selector matrix is premultiplied onto the error vector when certain states should not be penalized, avoiding rederivation of the cost.","core_discovery":"The paper claims that full, cross-coupled positive-definite penalty matrices are not merely feasible but beneficial. By writing a symmetric penalty matrix as $K = Q\\Lambda Q^{\\top}$ with positive eigenvalues $\\Lambda$ and an orthogonal eigenvector matrix $Q$, positive-definiteness is enforced implicitly, and the off-diagonal terms represent cross-coupling among states that a diagonal matrix cannot capture. Optimizing these matrices with particle swarm optimization gives lower costs than optimally tuned diagonal matrices in a two-state Lyapunov control problem, in LQR and Lyapunov-rate spacecraft attitude control, and in generic Lyapunov and Q-law low-thrust trajectory design; the largest reduction, about 65%, occurs in the LQR detumbling minimum-energy problem. The paper also presents three orthogonal-matrix parameterizations, Cayley transform, Givens rotations, and generalized Euler angles with Gram-Schmidt, and uses the last one for all results.","pith_inferences":["If the effect is genuinely due to cross-coupling, the same eigendecomposition parameterization should transfer to other quadratic weighting schemes such as model predictive control, control barrier functions, and sliding-mode weighting; the paper motivates this but does not demonstrate it.","A stricter test would hold the optimization budget fixed and compare the full parameterization against a globally optimal diagonal tuning, isolating the parameterization's contribution from the effects of a larger search space and stochastic search.","The long-duration low-thrust results show qualitatively different orbital-element histories between diagonal and full matrices, suggesting the full matrix reshapes the Lyapunov landscape rather than merely changing one gain; mapping the basin of attraction of each closed-loop law would test that interpretation.","A natural follow-up is gradient-based optimization on the same parameterization, since the paper notes the orthogonal parameterization admits gradients and could be refined locally after an initial heuristic search."],"forward_implications":["In LQR spacecraft attitude control, full $Q$ and $R$ matrices cut the mean minimum-energy cost from 0.423 to 0.146 in the detumbling maneuver and from 0.400 to 0.152 in the rest-to-rest maneuver, about 65% and 63% respectively.","In Lyapunov-based attitude control, full $K_p$ and $K_d$ matrices cut the mean cost from 0.107 to 0.076 in detumbling and from 0.091 to 0.068 in rest-to-rest.","In generic Lyapunov low-thrust control, the full penalty matrix reduced mean time of flight in all five transfer cases, with the largest mean savings being about 3.2 days in Case B and 2.2 days in Case E.","In Q-law low-thrust control, the full matrix reduced mean time of flight and increased mean final mass in every case, with Case E gaining about 10.5 days of flight time or close to 90 kg of final mass.","Because a diagonal matrix is a special case of the full parameterization, the method enlarges the achievable feedback-law class without adding constraints to the optimization and without changing the controller form."],"supporting_citations":[{"why":"Introduces the eigendecomposition idea for symmetric penalty matrices and the enforcement of sign-definiteness through eigenvalues.","marker":"[9]"},{"why":"Provides Cayley, Givens, and other minimal-variable parameterizations of orthogonal matrices against which the generalized Euler angle approach is compared.","marker":"[12]"},{"why":"The authors' earlier work that introduced the full-matrix parameterization used here and whose low-thrust results this paper extends.","marker":"[17]"},{"why":"Supplies the Q-law, the proximity quotient, the coasting mechanism, and the transfer-case benchmark set.","marker":"[6]"},{"why":"Provides the quaternion attitude dynamics and the Lyapunov attitude controller formulation used in the attitude examples.","marker":"[10]"},{"why":"Formulates the LQR problem and gain-tuning background that the quadratic-cost comparisons build on.","marker":"[22]"},{"why":"Particle swarm optimization, the metaheuristic used to tune the penalty-matrix parameters.","marker":"[25]"},{"why":"Shows extensive Q-law gain tuning and supplies benchmark low-thrust transfer solutions used for comparison.","marker":"[24]"}],"fun_headline_variants":["Eigendecomposition tuning cuts control costs up to 65%","Cross-coupled penalties improve aerospace control up to 65%","Full matrix tuning beats diagonal in LQR, Lyapunov, low-thrust","Unlocking off-diagonal penalties yields up to 65% better control","Aerospace control improves up to 65% with full penalty matrices"],"cache_read_input_tokens":22656,"weakest_assumption_plain":"The comparisons assume particle swarm optimization has searched both parameter spaces well enough that the remaining gap reflects the parameterization rather than the optimizer's luck or the larger number of free variables; the paper itself concedes that limited runs of PSO mean its solutions are not truly optimal.","fun_headline_variants_meta":{"raw":{"variants":["Eigendecomposition tuning cuts control costs up to 65%","Cross-coupled penalties improve aerospace control up to 65%","Full matrix tuning beats diagonal in LQR, Lyapunov, low-thrust","Unlocking off-diagonal penalties yields up to 65% better control","Aerospace control improves up to 65% with full penalty matrices"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001558,"raw_usage":{"total_tokens":6217,"prompt_tokens":930,"completion_tokens":5287,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":5193}},"tokens_in":546,"tokens_out":5287,"duration_ms":34284,"temperature":1.0,"reasoning_tokens":5193,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:06:27.933531+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the two-state Zermelo example, solve the diagonal parameterization $(k_1,k_2)$ and the full eigendecomposition parameterization $(\\lambda_1,\\lambda_2,\\theta)$ to verified global optimality with the same integration tolerances and compare the best costs; if the diagonal optimum equals the full optimum, the claimed advantage of the off-diagonal terms evaporates.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the eigendecomposition idea for symmetric penalty matrices and the enforcement of sign-definiteness through eigenvalues."},{"cited_title":"Shepard, S","cited_arxiv_id":null,"evidence_quote":"Provides Cayley, Givens, and other minimal-variable parameterizations of orthogonal matrices against which the generalized Euler angle approach is compared."},{"cited_title":"Expanding the Class of Quadratic Control-Lyapunov Functions for Low-Thrust Trajectory Optimization","cited_arxiv_id":"2408.14412","evidence_quote":"The authors' earlier work that introduced the full-matrix parameterization used here and whose low-thrust results this paper extends."},{"cited_title":"Schaub, J","cited_arxiv_id":null,"evidence_quote":"Provides the quaternion attitude dynamics and the Lyapunov attitude controller formulation used in the attitude examples."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Formulates the LQR problem and gain-tuning background that the quadratic-cost comparisons build on."},{"cited_title":"855–869.doi:10.1109/AERO.2005.1559377","cited_arxiv_id":null,"evidence_quote":"Shows extensive Q-law gain tuning and supplies benchmark low-thrust transfer solutions used for comparison."}],"review_version":1}