{"id":"fc502a98-c7df-4873-aa0c-b907c6551998","arxiv_id":"2505.08715","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Invariant tori of dimension two and three can be computed from single short trajectories using Birkhoff RRE, Bayesian MAP rotation-vector inference, and KZ lattice reduction.","lead":"This paper introduces a numerical method to compute invariant tori of dimension two or higher from a single short trajectory of a symplectic map, without continuation or initial guesses. It combines Birkhoff reduced-rank extrapolation, Bayesian inference of the rotation vector, and lattice reduction to choose the best torus coordinates.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing assumption is the unproven convergence of Birkhoff RRE frequencies to true Fourier frequencies for d>1: Sections 3.1 and 5 state this is open, yet the full method and its validation depend on it. The reported residuals are integrated checks and do not isolate whether Ω_j → ω·k_j.","rationale":"The reader's weakest assumption identifies the same load-bearing issue: the method's success depends on Birkhoff RRE frequency convergence for d>1, which the paper explicitly leaves unproven. My read agrees and adds that the final residuals are not a direct test of this convergence, because the same estimated rotation vector is used both to build the torus and to evaluate the KAM residual. This makes the empirical evidence suggestive rather than conclusive for the general claim. I do not see an internal inconsistency or a simpler fatal flaw: the parameterization step is standard least-squares, the KZ choice is well motivated, and the reported examples include a plausible genuine a posteriori residual. The paper also has real strengths: the code is public, the experiments are reproducible in structure, and the failure-mode analysis is honest. The conditional verdict is therefore appropriate: the method is a valuable heuristic with strong empirical support, but the central mathematical assumption needs either proof, a rigorous bound, or a clear statement of its failure regions before the claim can be accepted unconditionally.","tokens_in":28830,"tokens_out":9517,"duration_ms":111965,"concrete_test":"Construct a synthetic d=2 invariant torus with a known analytic embedding and known Diophantine rotation vector, e.g. h(S(θ)) = Σ_{|k|≤K} c_k e^{2πik·θ} with prescribed exponentially decaying c_k. Generate exact trajectory samples h(F^t x) on this torus. Run Birkhoff RRE with T=2J for J=250, 500, 1000, 2000, and 4000, and compute the maximum error of the J0=30 strongest recovered frequencies Ω_j matched to the true ω·k_j. Then run the full pipeline (MAP, KZ, parameterization) and compare the recovered rotation vector to the true ω up to the GL(2,Z) and sign ambiguities. If the matched frequency error does not decrease systematically with J or is not below about 1e-8 at J=4000, the unproven convergence is a practical failure; if it does decrease, the concern is mitigated but still lacks a proof for generic d.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the frequencies Ω_j returned by Birkhoff RRE (Sec. 3.1) converge to the true temporal frequencies ω·k_j for invariant tori of dimension d>1. The paper explicitly says: \"There is currently no theory for the convergence of the frequencies Ω_j to true values ω·k_j as J→∞\" (Sec. 3.1), and the conclusion repeats that \"a proof of the convergence of the Birkhoff RRE roots to the frequencies ω·k is still an open question.\" This is not a peripheral technicality: the MAP step (Sec. 3.2, Eqs. 13 and 25) takes the measured Ω_j as input, and the final parameterization (Sec. 2) assumes the recovered rotation vector. If Ω_j does not converge for some d>1 case, the method can fail even when the trajectory is on a perfectly smooth torus. The empirical evidence, while encouraging, is an integrated check: Table 2 shows recovered frequencies for one 2D torus, and the reported RKAM residuals in Section 4.2 are computed with the same estimated frequencies used to construct S. A wrong but nearby frequency estimate could still yield small residuals on a finite trajectory and a 25×25 grid, so these numbers do not isolate frequency convergence. The two admitted failure modes in Sec. 4.1—nearly resonant rotation vectors and filamentary tori—are exactly the regimes where the unproven convergence assumption is least secure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a fully data-driven method for computing invariant tori of dimension d > 1 from a single, relatively short trajectory, without continuation or an initial guess. The pipeline has four stages: (i) Birkhoff reduced rank extrapolation (RRE) to extract many temporal frequencies and their amplitudes from the trajectory; (ii) a Bayesian maximum a posteriori (MAP) problem, Eq. (25), that selects a valid rotation vector from among the measured frequencies; (iii) a Korkine-Zolotarev lattice reduction that chooses a homology basis adapted to the shape of the torus; and (iv) an adaptive least-squares Fourier parameterization, Eq. (7) and Sec. 2.2. The method is tested on a coupled standard map (1000 random initial conditions, with 851 classified as integrable), on island chains, and on three three-dimensional tori in the cislunar elliptic restricted three-body problem. The authors report KAM residuals RKAM between 3.63e-6 and 1.63e-5 for the 3D examples and a 95% success rate for the standard-map ensemble, and they discuss two failure modes: nearly resonant rotation vectors and highly anisotropic (filamentary) tori.","tokens_in":29156,"tokens_out":3219,"duration_ms":32613,"significance":"If the central claims hold, this would be a practically valuable tool for astrodynamics and plasma physics, since it removes the need for continuation or a good initial guess and works from a single trajectory, including for tori of dimension three. The paper's strengths are its honest validation methodology: the KAM residual in Eq. (29) is evaluated on a uniform grid rather than on the fitting trajectory, and the validation error Rh in Eq. (9) uses held-out trajectory points. The authors also explicitly acknowledge the main theoretical gap, namely that convergence of Birkhoff RRE frequencies to the true Fourier frequencies omega·k is unproven for d > 1. The numerical experiments are extensive and the code is publicly available in SymplecticMapTools.jl. The main risk is that the entire method depends on that unproven convergence, and the reported residuals, while encouraging, are integrated checks that do not isolate whether the frequency estimates converge.","major_comments":[{"comment":"The load-bearing assumption of the method is that the Birkhoff RRE frequencies Omega_j converge to the true temporal frequencies omega·k_j for d > 1. The paper explicitly states in Sec. 3.1 that 'There is currently no theory for the convergence of the frequencies Omega_j to true values omega·k_j as J→∞' and repeats in the Conclusion that this is an open question. The MAP step in Sec. 3.2, Eq. (25), takes the measured Omega_j as input and restricts candidate rotation vectors to subsets of Omega, so if this convergence fails, the method collapses even on a perfectly smooth torus. The empirical evidence in Sec. 4 is integrated: the reported RKAM residuals are computed with the same estimated frequencies used to construct the parameterization, and a wrong but nearby frequency could still yield small residuals on a finite trajectory sampled on a 25×25 grid. I therefore ask for a targeted numerical test that isolates frequency convergence, for example comparing Birkhoff RRE frequencies against frequencies obtained from a high-accuracy parameterization-method torus for one or two d = 2 and d = 3 examples, with controlled J, T, and analytic or high-precision ground truth. Without such a test, the central methodological claim is not fully supported.","section":"Sec. 3.1 and Sec. 5"},{"comment":"The prior hyperparameters C and r in sigma(k) = C e^{-r||k||} are fit by least squares to the same trajectory's sorted Fourier magnitudes H_j, Eq. (20), and then used as the prior in the MAP objective (25). This is an empirical-Bayes procedure that uses the data twice: the same trajectory both determines the prior and is the evidence in the posterior. This double use can overstate posterior confidence and may bias the wavenumber assignment toward the fitted smoothness model. The paper does not report sensitivity of the final rotation vector or RKAM to changes in C and r, nor does it compare with a fixed, cross-validated choice. I request a sensitivity study that perturbs the fitted C and r (or holds out a fraction of the H_j when fitting them) and reports how often the rotation vector and the final RKAM change. The order-statistic model in Eq. (22) also assumes no measurement error in the H_j, which is an additional idealization that should be stated and tested.","section":"Sec. 3.2, Eq. (20) and Eq. (25)"},{"comment":"The paper claims in the Introduction that 'we successfully parameterize 95% of the initialized tori for the weakly coupled standard map,' but the success criterion is not defined: no threshold on RKAM (or Rh) is given, and it is unclear whether the denominator is 1000 initialized trajectories, the 851 classified as integrable, or the 853 appearing in the Fig. 7 caption. The count inconsistency (851 in the text versus 853 in the caption) also needs correction. Since this 95% claim is a headline result, the definition of success and the exact counts must be made precise and reproducible.","section":"Sec. 4.1, Fig. 7 caption and Introduction"}],"minor_comments":[{"comment":"The name 'Korkine-Zolatarev' is a misspelling; the standard spelling is Korkine-Zolotarev.","section":"Abstract and Sec. 3.3"},{"comment":"The definition of R_h^2 writes R2_h = 1/R2_h0 sum ||h(F^t(x)) - hat h(theta + omega t)||^2, but the second factor R2_h0 is defined immediately after; please make the dependence on the theta initialization explicit, since the expression as written is ambiguous.","section":"Sec. 2.1, Eq. (9)"},{"comment":"In the sentence 'Typical values of these constants are pmax = 10 and epsilon_ada = 10^-8,' the symbol epsilon_ada appears to be a typo for epsilon_isl, which is the island tolerance defined earlier in the same paragraph.","section":"Sec. 3.4"},{"comment":"There is a duplicated word in 'a resolution of (K1, K2) = (24, 18) would be needed to capture all of the same Fourier modes for for the sheared torus.'","section":"Sec. 2.3"},{"comment":"The statement that 'the weighted Birkhoff classification rate approximately matches the Birkhoff RRE classification rate in the right panel' is made without a quantitative comparison; reporting the actual classification counts for the weighted Birkhoff average at the same trajectory lengths would make the comparison in Fig. 6 more informative.","section":"Sec. 4.1, Fig. 6"}],"recommendation":"major_revision","confidential_remarks":"The paper is well within the scope of a dynamical-systems journal and the code availability is a plus. The main risk is the unproven convergence of Birkhoff RRE frequencies for d > 1; I would encourage the editor to ask for a dedicated numerical convergence study or a proof sketch before accepting. The empirical-Bayes circularity in Sec. 3.2 is also a concern but is fixable with sensitivity analysis. The '95%' success-rate claim needs tight definition; otherwise it is not reproducible from the text."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should read this if you care about practical computation of invariant tori. The paper delivers a real advance: a fully automatic pipeline that computes d>1 tori from a single short trajectory, with no initial guess and no continuation. The core combination—Birkhoff RRE for frequencies, Bayesian MAP for wavenumber labeling, KZ lattice reduction for homology choice, and adaptive least-squares parameterization—is genuinely new and is validated on coupled standard maps and the elliptic restricted three-body problem, including 3D tori. The code is public, and the authors report KAM residuals around 1e-5 to 1e-6 for the 3D examples, with a 95% success rate on 1000 random trajectories in the 2D standard map case.\n\nThe paper is also honest about its main soft spot. Section 3.1 and the conclusion both state there is no proof that the Birkhoff RRE frequencies converge to the true Fourier frequencies for d>1. This is not a peripheral technicality: the MAP step takes those frequencies as input, and if they are wrong the whole method collapses. The stress-test note is right that the reported residuals are integrated checks, so they do not isolate frequency convergence. A wrong but nearby frequency could still yield small residuals on a finite trajectory and a 25x25 grid. That said, the authors do not overclaim. They call the convergence an open question and treat the method as a numerical heuristic with strong empirical support. The two failure modes they identify—nearly resonant vectors and filamentary tori—are exactly where the unproven convergence is least secure, and they discuss mitigations. The empirical-Bayes prior is fit to the same trajectory data used in the MAP step, which is a mild circularity but not a fatal one; the final torus is independently checked by a uniform-grid KAM residual and held-out points.\n\nIs the lack of a convergence proof disqualifying? I do not think so for a computational paper. The method is a practical engineering contribution, and the authors are explicit about the theory gap. But it does mean the method's reliability is not yet well characterized beyond the tested examples. A serious referee should push for either a partial convergence result, a cleaner numerical study of frequency convergence on a known torus, or a more detailed mapping of failure regions. The paper is well-written, the numerical experiments are carefully designed, and the adaptive least-squares parameterization alone is a useful contribution.\n\nBottom line: this is a solid, important subfield advance. It deserves a serious peer review. I would bring it to reading group and I would cite it if I worked on computational tori. The theoretical gap should be acknowledged in the final version, but it should not block publication.","headline":"A practically valuable pipeline for computing higher-dimensional invariant tori from single short trajectories, with an honest but real theoretical gap in the frequency-convergence step.","tokens_in":29736,"tokens_out":1662,"would_cite":true,"duration_ms":18915,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37J40","37M10","65P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"A single short trajectory can determine a higher-dimensional invariant torus without initial guesses or continuation.","keywords":["invariant tori","Birkhoff reduced rank extrapolation","rotation vector","Bayesian maximum a posteriori estimation","Korkine-Zolatarev lattice reduction","coupled standard map","restricted three-body problem","quasiperiodic motion"],"falsifier":"Run the frequency-extraction step on a trajectory of a known smooth two-dimensional torus with a prescribed rotation vector, increasing the filter size $J$, and check whether the computed frequencies approach the exact values $\\omega\\cdot k$; the central claim fails if the top frequencies plateau away from machine precision or never contain a pair of wavenumbers with determinant $\\pm1$.","tokens_in":28517,"feed_emoji":"🪐","tokens_out":10736,"duration_ms":99381,"temperature":0.7,"pith_summary":"The paper presents a frequency-based method for computing invariant tori of dimension greater than one from a single short trajectory of a dynamical system. No continuation, no initial guess, and no preferred coordinate system are required, which is what makes the method practical when little is known about the system in advance. The pipeline first extracts the trajectory's dominant frequencies with Birkhoff reduced rank extrapolation, then infers a valid rotation vector by Bayesian maximum a posteriori estimation, then selects short, nearly orthogonal loops around the torus by Korkine-Zolatarev lattice reduction, and finally fits Fourier coefficients by least squares. On a coupled standard map the authors report successfully parameterizing 95 percent of the initialized tori, and on three-dimensional Earth-Moon restricted three-body tori they report KAM residuals near $10^{-5}$ to $10^{-6}$ from trajectories of a few thousand iterates. If these results hold, a short trajectory is enough to recover both the geometry and the rotation vector of a higher-dimensional invariant torus.","feed_headline":"One trajectory computes an invariant torus, no guesses needed","feed_subtitle":"A four-step frequency pipeline recovers 2D and 3D tori with residuals near 1e-5.","key_machinery":"The load-bearing object is the Birkhoff reduced rank extrapolation (Birkhoff RRE) filter: a palindromic set of coefficients $c_j$ that extrapolates the trajectory, whose residual $R_{\\mathrm{RRE}}$ classifies the trajectory as integrable or chaotic, and whose filter roots $\\psi_j$ produce high-accuracy frequency estimates $\\Omega_j\\approx\\omega\\cdot k_j$. Around this sit two further mechanisms: a Bayesian maximum a posteriori estimator that assigns wavenumbers to the measured frequencies and recovers a valid rotation vector, and a Korkine-Zolatarev lattice reduction on the averaged metric $G=\\sum_k \\|h_k\\|^2 kk^\\top$ that selects the homology generators giving the most compact Fourier representation. The final least-squares coefficient fit is what converts the trajectory and rotation vector into an explicit torus parameterization.","core_discovery":"The central claim is that the rotation vector and Fourier parameterization of a $d$-dimensional invariant torus can be recovered from a single finite trajectory, without continuation or initial guesses, by combining temporal frequency estimation with a Bayesian labeling of wavenumbers. Specifically, Birkhoff RRE returns high-precision frequencies $\\Omega_j$ and magnitudes $H_j$; the MAP step identifies which lattice wavenumbers $k_j$ produced those frequencies and thereby finds a valid rotation vector $\\omega$; the KZ reduction replaces $\\omega$ by an equivalent rotation vector whose loops on the torus are short and nearly orthogonal; and a least-squares fit of the trajectory against modes $e^{2\\pi i k\\cdot\\theta}$ yields the torus Fourier coefficients. The authors demonstrate the pipeline on random initial conditions of a weakly coupled standard map, on island chains, and on three-dimensional tori in the Earth-Moon restricted three-body problem, reporting KAM residuals as low as $3.63\\times10^{-6}$ with trajectory lengths on the order of $10^3$--$10^4$.","pith_inferences":["If a convergence proof for the frequency roots were supplied, the Bayesian uncertainty $\\sigma_\\omega$ could be set from theory, which would turn the pipeline into a method with end-to-end error estimates for the recovered torus.","Because only injectivity of the observable $h(S)$ is required, the same steps should work with delay embeddings or other observables, extending the method to systems whose natural state space is not Euclidean.","Since trajectory generation is often the dominant cost, the short-trajectory property suggests this approach could map out many tori in parameter scans where continuation-based methods struggle.","The KZ step is a lattice shortest-vector problem on the torus's average metric; the same reduction could be used to identify slow and fast directions in higher-dimensional systems for model reduction."],"forward_implications":["For the weakly coupled standard map with 1000 random initial conditions, 851 trajectories are classified as integrable by length 8001, and 95 percent of the initialized tori are successfully parameterized by the reported a posteriori measure.","The method computes three-dimensional tori of the Earth-Moon restricted three-body problem from trajectories of length 3335 to 8335, with KAM residuals of $3.63\\times10^{-6}$, $5.70\\times10^{-6}$, and $1.63\\times10^{-5}$.","Birkhoff reduced rank extrapolation classifies the standard-map trajectories as integrable with roughly an order of magnitude shorter trajectories than weighted Birkhoff averaging requires.","The main failure modes are nearly resonant rotation vectors and strongly filamentary tori; the authors show these are detectable through the KAM residual and resonance-order diagnostics.","Island-chain trajectories are handled by treating the chain period as an added rational frequency, so a single pipeline covers tori, islands, and chaotic classification."],"supporting_citations":[{"why":"Supplies the Birkhoff reduced rank extrapolation filter whose roots yield the frequency estimates that drive the entire method.","marker":"[30]"},{"why":"Supplies the ergodic projection and torus parameterization idea that the least-squares step generalizes from two dimensions.","marker":"[29]"},{"why":"Defines the weighted Birkhoff average whose convergence rates and residuals anchor the trajectory classification and comparison.","marker":"[24]"},{"why":"Provides the existing method for non-star-shaped one-dimensional torus curves that the new coordinate-agnostic approach extends to higher dimension.","marker":"[33]"},{"why":"Provides the Korkine-Zolatarev lattice reduction routine used to choose the optimized homology basis.","marker":"[43]"},{"why":"Represents the parameterization method whose need for initial guesses and continuation defines the bottleneck this method avoids.","marker":"[4]"}],"fun_headline_variants":["Single trajectory computes invariant tori","No guesses needed for higher-dimensional tori","One trajectory, robust tori recovery","Compute tori from one path, no initialization","Invariant tori from a single short run"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on an unproved assumption: the frequencies that the extrapolation step returns really converge to the true frequencies of the torus for tori of dimension two or higher, and the strongest recovered frequencies contain enough independent directions to form a valid rotation vector.","fun_headline_variants_meta":{"raw":{"variants":["Single trajectory computes invariant tori","No guesses needed for higher-dimensional tori","One trajectory, robust tori recovery","Compute tori from one path, no initialization","Invariant tori from a single short run"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1305,"prompt_tokens":891,"completion_tokens":414,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":349}},"tokens_in":507,"tokens_out":414,"duration_ms":4150,"temperature":1.0,"reasoning_tokens":349,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:47:42.644885+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the frequency-extraction step on a trajectory of a known smooth two-dimensional torus with a prescribed rotation vector, increasing the filter size $J$, and check whether the computed frequencies approach the exact values $\\omega\\cdot k$; the central claim fails if the top frequencies plateau away from machine precision or never contain a pair of wavenumbers with determinant $\\pm1$.","supporting_citations":[{"cited_title":"Ruth and D","cited_arxiv_id":null,"evidence_quote":"Supplies the Birkhoff reduced rank extrapolation filter whose roots yield the frequency estimates that drive the entire method."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ergodic projection and torus parameterization idea that the least-squares step generalizes from two dimensions."},{"cited_title":"Quantitative quasiperiodicity","cited_arxiv_id":null,"evidence_quote":"Defines the weighted Birkhoff average whose convergence rates and residuals anchor the trajectory classification and comparison."},{"cited_title":"Numerical computation of rotation numbers of quasi-periodic planar curves","cited_arxiv_id":null,"evidence_quote":"Provides the existing method for non-star-shaped one-dimensional torus curves that the new coordinate-agnostic approach extends to higher dimension."},{"cited_title":"HKZ and Minkowski Reduction Algorithms for Lattice- Reduction-Aided MIMO Detection","cited_arxiv_id":null,"evidence_quote":"Provides the Korkine-Zolatarev lattice reduction routine used to choose the optimized homology basis."},{"cited_title":"The Parameterization Method for Invariant Manifolds: From Rigorous Results to Effective Computations , volume 195 of Applied Mathematical Sciences","cited_arxiv_id":null,"evidence_quote":"Represents the parameterization method whose need for initial guesses and continuation defines the bottleneck this method avoids."}],"review_version":1}