{"id":"90ee251b-4014-408c-99ff-59487eb10049","arxiv_id":"2412.19683","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A machine learning model trained on recurrence quantifiers of a standard map can detect resonances in other 2D systems, but its generalization to a 4D map requires matching embedding conditions and yields unclear peaks.","lead":"This paper combines recurrence analysis with a long short-term memory neural network to automatically detect orbital resonances in weakly nonintegrable dynamical systems. The method is demonstrated on a standard map, a second map, and a deformed Kerr spacetime, but the higher-dimensional case works only under embedding.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'regardless of dimensionality' claim rests on a single embedded-train/embedded-test configuration in the 4D map; without a threshold-normalized full-state test or quantitative resonance labels, the success may be an embedding or threshold artifact rather than a universal RQA signature.","rationale":"The reader identified the feature-manifold transfer assumption as the weakest load-bearing premise, and the paper's own Section III B supports this: the basic network fails on the 4D map, and only the embedded network on embedded data produces recognizable peaks. My stress test sharpens the same concern: the failure may be caused by the use of absolute recurrence thresholds rather than a fundamental dimensional limitation, and the absence of quantitative ground-truth resonance labels for the 4D map means one cannot distinguish a true universal signature from an embedding or threshold artifact. The proposed fixed-RR retraining test directly isolates the threshold issue while keeping the full-state representation, which is the regime the abstract claims. I do not recommend changing the verdict from CONDITIONAL: the method shows promise on 2D systems, but the higher-dimensional evidence is too thin to justify the strong wording 'regardless of dimensionality.' The suggested conditional acceptance with requests for code, data, and improved 4D validation remains the appropriate outcome.","tokens_in":12105,"tokens_out":4405,"duration_ms":53375,"concrete_test":"Retrain the same basic LSTM architecture on the standard map, but compute all RQA input features using per-trajectory thresholds chosen to achieve a fixed recurrence rate (e.g., RR = 0.05, as in Fig. 2) instead of the fixed absolute epsilon values used in the paper. Then evaluate this fixed-RR basic network on the full 4D state of the 4D map (i.e., without embedding). If resonances become visible in the network output aligned with known resonance locations, the original failure was a threshold-normalization artifact and the dimensionality claim gains support. If the network still fails on the full 4D state, the 'regardless of dimensionality' claim is not supported, because the only successful configuration requires embedding on both training and test sides.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that recurrence quantifiers carry imprints of resonant behavior regardless of dimensionality and that the trained LSTM automates resonance detection—is only as strong as the 4D evidence. Section III B states that the only working 4D configuration is the embedded network on embedded 4D data; mismatched combinations fail. This is explicitly load-bearing because the abstract's 'regardless of the system's dimensionality' must be supported by a higher-dimensional demonstration. The current demonstration is vulnerable in two ways. First, the RQA features are computed with fixed absolute thresholds (epsilon = 0.001,...,1) after rescaling to [0,2π], not with a fixed recurrence rate. In 4D, and especially after embedding, the distribution of pairwise distances changes, so the feature vectors may lie outside the training manifold. Second, no independent ground truth for resonance identity in the 4D map is provided: APLE only distinguishes regular from chaotic motion, not resonant from non-resonant tori. The peaks in Fig. 10 could therefore be threshold-induced or embedding-induced periodicities rather than universal resonance signatures. The conclusion 'we conclude that this method is effective' is thus supported mainly by visual inspection of two 2D cases and one partial 4D case. If the 4D success is an artifact, the method's value for EMRIs—where more than two degrees of freedom matter—is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a resonance detection method that combines recurrence quantification analysis (RQA) with a bidirectional LSTM network. The network is trained on hand-labeled orbits of the standard map and then applied to other systems: the de Vogeleare map, geodesic motion in the Johannsen-Psaltis spacetime, and a 4D symplectic map. The authors report clear resonance detection in the two 2D test cases and a partial success in the 4D case when both training and test RQA features are computed from time-delay embedded data. The stated motivation is detecting resonances in EMRI systems, where higher-dimensional phase spaces prevent the use of rotation-number methods.","tokens_in":12523,"tokens_out":3528,"duration_ms":35852,"significance":"Resonance detection in higher-dimensional near-integrable systems is an open problem with direct relevance to EMRI waveform modeling, and the proposed RQA+ML pipeline is a plausible and potentially useful approach. The 2D demonstrations, especially the transfer from the standard map to the de Vogeleare map and to a black-hole spacetime, are visually convincing and illustrate a practical workflow. The paper also gives a clear description of the network architecture, training data, and recurrence-analysis parameters, and it builds on standard, widely used tools. However, the central claim that recurrence quantifiers carry imprints of resonant behavior 'regardless of the system's dimensionality' rests on a single configuration in the 4D map and lacks an independent quantitative ground truth for resonance identity in that test case. As a result, the significance for EMRI applications, where more than two degrees of freedom matter, is not yet firmly established.","major_comments":[{"comment":"The claim in the abstract and conclusion that recurrence quantifiers carry imprints of resonant behavior 'regardless of the system's dimensionality' is not supported by the 4D experiment as presented. The authors state in Section III B that only the embedded network applied to embedded 4D map data produces recognizable resonance peaks; the basic network on full-state 4D data fails (Fig. 9), and the cross combinations also fail. This makes the higher-dimensional success conditional on a specific embedding preprocessing pipeline, not a demonstration of a dimensionality-independent RQA signature. The conclusion 'we conclude that this method is effective' therefore overstates the strength of the 4D evidence.","section":"Section III B and Section IV"},{"comment":"There is no independent ground truth for resonance identity in the 4D map. The APLE indicator defined in Eq. (20) distinguishes regular from chaotic trajectories but does not label a regular torus as resonant or non-resonant. The statement in Section III B that 'the network is reacting to resonances in their true locations' is based on visual comparison with the APLE geography in Fig. 5, not on a quantitative resonance classification. Without an independent method to identify resonant tori in the 4D map, the peaks in Fig. 10 could be artifacts of the embedding or of the recurrence thresholds rather than universal resonance signatures.","section":"Section II F 3 and Section III B"},{"comment":"The RQA features are computed with fixed absolute thresholds epsilon = 0.001, ..., 1 after rescaling the phase space to [0, 2*pi]^n, rather than with a fixed recurrence rate. Because the distribution of pairwise distances changes with dimensionality and after time-delay embedding, the 4D/embedded feature vectors may lie outside the feature manifold of the standard-map training data. The authors themselves show in Fig. 3 that RQA values depend strongly on the choice of epsilon. The paper should either demonstrate that the results are robust to threshold normalization, for example by using a fixed recurrence rate, or justify why absolute thresholds transfer across systems and dimensions.","section":"Section II D and Section III B"}],"minor_comments":[{"comment":"The sentence 'Although the tool that we are presenting here can be used is quite generic' is ungrammatical and should be revised.","section":"Abstract"},{"comment":"In the definition of P(epsilon, v), the factors (1 - R_{i,j}) and R_{i,j+k} omit the explicit epsilon dependence that is written in Eq. (5a); the notation should be made consistent.","section":"Section II B, Eq. (5b)"},{"comment":"The caption refers to 'the 1/3 resonance is highlighted in red', while the text in Section II F 2 and Section III A 2 identifies the resonance as omega_r/omega_theta = 2/3. This inconsistency should be corrected.","section":"Fig. 8 caption"},{"comment":"There is a typo in 'thetorch.nn.LSTM module'; it should read 'the torch.nn.LSTM module'.","section":"Section II D"},{"comment":"The statement that a code release is 'under preparation' limits reproducibility; the authors should consider providing the code with the submission or at least a complete list of hyperparameters and data-generation scripts.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of the journal and addresses a timely problem in EMRI resonance modeling. The main concern is the unsupported dimensionality claim: the 4D demonstration is the load-bearing piece, but it lacks a quantitative ground truth and a robustness check against threshold choices. If the authors can provide an independent 4D resonance label (e.g., via frequency analysis) and show that the result is not an embedding or threshold artifact, the paper would be much stronger. I would also encourage releasing the code at the revision stage, as the current 'under preparation' note hinders independent verification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the pipeline: multi-threshold RQA quantifiers fed as a sequence into a bidirectional LSTM, trained on hand-labeled standard map islands and transferred to other systems. I don't know of that exact combination in the literature, and the 2D demonstrations are convincing. The de Vogeleare map and the black hole Poincaré section show clean output peaks at the resonant islands, and the training/validation split is sensible.\n\nThe paper is also honest about its limitations. The authors explicitly report that the basic network fails on the 4D map and that only the embedded network on embedded data works. They acknowledge the threshold sensitivity of RQA and the absence of rotation numbers in higher dimensions. That transparency buys goodwill.\n\nThe soft spots are real, though. The abstract says the method works \"regardless of the system's dimensionality,\" and the 4D evidence does not support that. There is exactly one working configuration: embedded training data, embedded test data, with embedding dimension and delay chosen separately. The mismatched combinations fail, which raises the possibility that the network is keying on embedding-induced periodicities rather than a universal resonance signature. On top of that, there is no independent ground truth for resonance identity in the 4D map. APLE only separates regular from chaotic motion, not resonant from non-resonant tori, so the peaks in Fig. 10 could be threshold or embedding artifacts. No code or data is released yet, and there are no quantitative metrics anywhere. Everything is visual, and the only number reported is validation loss. For a machine-learning paper, that is a significant gap.\n\nThe citation pattern is fine. The paper builds on the right RQA and recurrence-plot literature, and the astronomical context is well sourced.\n\nThe likely reader is someone working on resonance detection in near-integrable systems, particularly for EMRI modeling. They will get a clear description of a useful 2D tool and a fair warning about how fragile the higher-dimensional generalization currently is.\n\nRecommendation: send it to peer review. The core idea deserves referee time, and the 2D results justify it. But the authors should be pushed to either strengthen the 4D evidence or temper the dimensionality claim, and to release code and data so the transfer result can be independently checked.","headline":"Promising 2D resonance detector whose \"regardless of dimensionality\" claim outruns the evidence; the 4D success is one embedded-only configuration with no independent ground truth.","tokens_in":12965,"tokens_out":1757,"would_cite":false,"duration_ms":19989,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37M10","68T07","83C57"],"pacs":[],"model":"deepseek-v4-flash","headline":"Recurrence analysis plus a standard-map-trained LSTM can locate resonances even in a four-dimensional map where conventional visualization tools do not exist.","keywords":["recurrence quantification analysis","resonance detection","long short-term memory","standard map","time-delay embedding","Johannsen-Psaltis spacetime","extreme mass ratio inspirals","4D map"],"falsifier":"Compute recurrence-quantification features for the 4D map under several embedding dimensions and delays, and compare the embedded network's output peaks with the APLE resonance map: if the peak locations shift whenever $m$ or $\\tau$ changes, the detections are artifacts of the embedding rather than signatures of the resonances.","tokens_in":1926,"feed_emoji":"🧠","tokens_out":2853,"duration_ms":393223,"temperature":0.7,"pith_summary":"This paper proposes that recurrence plots contain enough information about orbital resonances, the periodic islands in weakly non-integrable systems, that a machine-learning network can learn to spot them from the patterns alone without needing the system's equations or frequencies. To test this, the authors train a bidirectional LSTM on sweeps of initial conditions of the standard map, using 70 recurrence-quantification indicators per orbit, and then apply it to other systems. The network detects resonances in the de Vogeleare map and in the Poincare section of a test particle in the deformed Kerr spacetime known as the Johannsen-Psaltis metric, where the target is the $\\omega_r/\\omega_\\theta = 2/3$ resonance relevant for extreme-mass-ratio inspiral modeling. The key demonstration is the 4D map: a plain network trained on full phase-space data fails, but a network trained on time-delay-embedded data and applied to embedded 4D data produces peaks at the true resonance locations. The paper concludes that recurrence quantifiers encode resonance structure even where conventional rotation-number and visual methods are unavailable.","feed_headline":"An LSTM trained on the standard map catches resonances in a 4D map","feed_subtitle":"Recurrence statistics plus a single trained network automate resonance detection, including black-hole EMRI orbits.","key_machinery":"The load-bearing object is the 70-dimensional recurrence-quantification input vector: for each orbit, seven RQA measures, $RR$, $DET$, $LAM$, $L$, $L_{\\rm entr}$, $DIV$, and $V_{\\max}$, computed at ten recurrence thresholds $\\epsilon$. These vectors, arranged as a sweep across initial conditions, are fed into a bidirectional LSTM whose cells are stacked along the initial-condition direction rather than along time. Embedding is the second mechanism: when only a scalar time series is available, time-delay embedding with dimension $m$ and delay $\\tau$ reconstructs the phase space, and the paper shows that matching the embedding between training and test data is what makes the 4D detection work.","core_discovery":"The paper's central claim is that recurrence quantifiers, the statistical measures of the diagonal and vertical line structures in a recurrence plot, carry detectable imprints of resonant islands, and that an LSTM trained once on the standard map can serve as a resonance detector for other dynamical systems, including a four-dimensional map where rotation-number and visual methods are unavailable. The basic network, trained on full phase-space coordinates, cleanly localizes resonances in two two-dimensional maps and in the Poincare section of a test particle orbiting a deformed Kerr black hole. In the 4D map, the same architecture detects resonances only when both training and test data pass through time-delay embedding: the embedded network applied to embedded data produces output peaks that align with the resonance geography charted by the APLE indicator, while the two mismatched combinations show nothing. The paper thus claims a proof of concept for a dimensionality-agnostic, automated resonance-detection pipeline, with the preprocessing convention required to match between training and application.","pith_inferences":["An implication the authors leave implicit is that the 4D success may owe more to embedding-induced features of the recurrence statistics than to a truly system-independent resonance signature; a decisive check would be to train on embedded standard-map data and test on a different embedded 4D system whose resonances are independently mapped.","If the preprocessing-match requirement generalizes, then each new application domain will need its own embedding calibration, which weakens the 'train once, use anywhere' reading of the method.","A testable extension for EMRI work is to apply the embedded network to synthetic gravitational-wave snapshots of inspiraling small bodies crossing a resonance and ask whether the output peaks track the crossing time; if they do, the method could become a resonance-crossing detector in waveform data.","The hand-labeling protocol, which marks an island as resonant only if at least two neighboring initial conditions fall in it, shapes the learned target; changing the sweep spacing or labeling single-point islands could change which resonances are detected and is worth quantifying."],"forward_implications":["Resonance localization can be automated for two-dimensional systems without manual rotation-number curves: the network trained on the standard map detects islands in the de Vogeleare map and in the black-hole Poincare section.","For higher-dimensional systems, the practical recipe is to embed the available scalar observables and apply a network trained on embedded data; the 4D map test shows that this combination recovers the known resonance geography.","In extreme-mass-ratio inspiral modeling, the method offers a way to identify extended resonances such as the $\\omega_r/\\omega_\\theta = 2/3$ resonance in the Johannsen-Psaltis spacetime, helping to decide which perturbation parameters matter for waveform modeling.","Because recurrence quantifiers are computable from any trajectory and do not require a Poincare section of a specific dimension, the same pipeline can in principle be applied to continuous-time systems by taking Poincare sections.","The failure of the cross-combinations in the 4D experiment implies that preprocessing must be considered part of the method: a detector trained on one embedding convention is not assumed to transfer to data prepared differently."],"supporting_citations":[{"why":"Defines recurrence plots and the RQA measures used to build the 70-dimensional input vectors.","marker":"[9]"},{"why":"Establishes time-delay embedding as a phase-space reconstruction, the preprocessing used for the 4D test.","marker":"[18]"},{"why":"Introduces the LSTM cell that the paper stacks along the initial-condition direction.","marker":"[19]"},{"why":"Defines the standard map on which the networks are trained and validated.","marker":"[21]"},{"why":"Introduces the Johannsen-Psaltis spacetime metric used in the black-hole test case.","marker":"[10]"},{"why":"Provides the geodesic setup and identifies the $\\omega_r/\\omega_\\theta = 2/3$ resonance targeted in the black-hole test.","marker":"[24]"},{"why":"Defines the 4D map used as the higher-dimensional test system.","marker":"[25]"},{"why":"Supplies the APLE indicator whose maps of 4D resonances are the comparison baseline.","marker":"[26]"},{"why":"Provides the algorithms used to select embedding dimension and delay for the training and test data.","marker":"[28]"}],"fun_headline_variants":["LSTM trained on standard map spots resonances in 4D maps","Recurrence + LSTM automates resonance detection across systems","One LSTM, many maps: resonance detection via recurrence","Machine learning reads recurrence plots to find resonances","From standard map to black holes: LSTM detects resonances"],"cache_read_input_tokens":14976,"weakest_assumption_plain":"The load-bearing premise is that recurrence-quantifier patterns are similar enough across different dynamical systems, when computed with the same thresholds and the same embedding choice, for a network trained on the standard map to recognize resonances in other systems; the 4D results show that this similarity holds only for matching embedding conventions.","fun_headline_variants_meta":{"raw":{"variants":["LSTM trained on standard map spots resonances in 4D maps","Recurrence + LSTM automates resonance detection across systems","One LSTM, many maps: resonance detection via recurrence","Machine learning reads recurrence plots to find resonances","From standard map to black holes: LSTM detects resonances"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000753,"raw_usage":{"total_tokens":3378,"prompt_tokens":998,"completion_tokens":2380,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":2296}},"tokens_in":614,"tokens_out":2380,"duration_ms":18290,"temperature":1.0,"reasoning_tokens":2296,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:57:35.312037+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute recurrence-quantification features for the 4D map under several embedding dimensions and delays, and compare the embedded network's output peaks with the APLE resonance map: if the peak locations shift whenever $m$ or $\\tau$ changes, the detections are artifacts of the embedding rather than signatures of the resonances.","supporting_citations":[{"cited_title":"Combining Machine Learning with Recurrence Analysis for resonance detection","cited_arxiv_id":"2412.19683","evidence_quote":"Defines recurrence plots and the RQA measures used to build the 70-dimensional input vectors."},{"cited_title":"Marwan, M","cited_arxiv_id":null,"evidence_quote":"Introduces the LSTM cell that the paper stacks along the initial-condition direction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the standard map on which the networks are trained and validated."},{"cited_title":"8 shows the result of the basic network applied to the Poincar´ e section of a test particle following a geodesic in the Johannsen-Psaltis spacetime metric (see Sec","cited_arxiv_id":null,"evidence_quote":"Introduces the Johannsen-Psaltis spacetime metric used in the black-hole test case."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the geodesic setup and identifies the $\\omega_r/\\omega_\\theta = 2/3$ resonance targeted in the black-hole test."},{"cited_title":"How to avoid potential pitfalls in recurrence plot based data analysis","cited_arxiv_id":"1007.2215","evidence_quote":"Defines the 4D map used as the higher-dimensional test system."},{"cited_title":"Takens, in Dynamical Systems and Turbulence, War- wick 1980, edited by D","cited_arxiv_id":null,"evidence_quote":"Provides the algorithms used to select embedding dimension and delay for the training and test data."}],"review_version":1}