{"id":"f37e85ef-4bd0-4f5d-82a5-6c2a27429b27","arxiv_id":"2411.18681","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"After denoising, temporal classes V, VII, and VIII of the black hole X-ray binary IGR J17091-3624 show signs of determinism, contradicting the earlier all-stochastic classification.","lead":"This paper re-analyzes X-ray lightcurves of the black hole binary IGR J17091-3624 after applying several noise filters, and reports that some of its temporal classes show deterministic, non-random structure, unlike earlier claims. The result would support the 'twin' relationship between IGR J17091-3624 and GRS 1915+105, but the analysis lacks a control test on pure noise.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing noise-only control: filtering may imprint deterministic structure, so all NS classifications in Tables 4–6 are unverified artifacts.","rationale":"The paper's entire reversal of the previous all-stochastic finding comes from denoising: Table 1 shows no NS class in the original data, and NS labels appear only after applying the four filters. Therefore the validity of the filters is the pivot on which the central claim turns. Section 3 explicitly lists 'avoid introducing additional local correlation' as a required property, but this is asserted, not verified. Convolution filters are mathematically guaranteed to induce autocorrelation over the kernel width, so the requirement is not obviously satisfied; whether that autocorrelation changes the GP dimension estimate, Betti numbers, or PCA eigenvalue ratios in a way that mimics NS is an empirical question. The paper does not include any noise-only control, so the false-positive rate of the full pipeline is unknown. This is distinct from the scientific question of whether IGR J17091-3624 is deterministic: it concerns the measurement device (the denoising plus classification chain). The concern is load-bearing because if any of the four filters converts Poisson noise into NS-looking signals, every NS classification in Tables 4–6 is an artifact and the claimed similarity to GRS 1915+105 collapses. We agree with the reader's assessment; this is the same weakest assumption they identified. Other issues (method disagreements, missing error bars, autoencoder trained on GRS 1915+105 data) are secondary: they matter only after the filter-independence question is resolved. Consequently the reader's REJECT verdict is unchanged: the paper cannot support its central claim without the control experiment we propose. If the control later shows negligible false positives, the claim would be supported; until then, rejection or at least 'unverified' is the honest reading.","tokens_in":8950,"tokens_out":4175,"duration_ms":37808,"concrete_test":"For each ObsID, simulate 100 pure-Poisson lightcurves matched to the observed count rate and length, plus 100 AR(1) linear-stochastic surrogates matched to the lag-1 autocorrelation. Apply the identical denoising and classification pipeline of §5 (BOX/GAU/NLM/ADA; CI nmsd with D2 saturation, SVD Betti, PCA+DBSCAN, autoencoder DS with threshold 1.5) and measure the fraction classified NS. If this false-positive rate exceeds the quoted 5% for any filter/method combination, then the NS labels for classes V–VIII in Table 6 cannot be distinguished from filter-imprinted structure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that IGR J17091-3624 classes V, VII, and VIII are primarily NS after denoising—rests on the assumption stated in §3, requirement 2: the filters 'should avoid introducing additional local correlation.' This assumption is never tested. All NS evidence appears only after denoising: the unfiltered CI analysis in Table 1 gives nmsd < 3 for every class, while Table 2 reports nmsd up to 11.492 for class VIII after ADA. The BOX and GAU filters are convolution kernels of width 2k+1 (Eqs. in §3.1); applied to white noise they introduce triangular autocorrelation at lags ≤ k, lengthening the correlation time and flattening the D2(M) curve toward saturation. NLM and ADA similarly replace each point by a weighted combination of neighbors. If such filtering is applied to a purely stochastic lightcurve, CI can yield D2 saturation and nmsd > 3, SVD can produce loops in the E1–E2 plane, and PCA eigenvalue ratios can exceed the S threshold—exactly the signatures used to label classes V–VIII as NS. The paper provides no control on pure Poisson noise or a linear stochastic process with matched count rate, duration, and power spectrum to establish the false-positive rate of the pipeline after denoising. Because the methods disagree in several cases (Tables 3, 5, 6) and the final labels rely on majority voting, the only way to distinguish filter-induced artifacts from real determinism is such a control. Without it, the paper's central conclusion is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper analyzes RXTE PCA lightcurves of the black hole X-ray binary IGR J17091-3624 in nine temporal classes. The authors apply four denoising filters (boxcar and Gaussian convolution, non-local means, and an adaptive denoising algorithm) and then classify each class as stochastic (S) or non-stochastic (NS) using four methods: correlation integral with surrogate analysis (nmsd), singular value decomposition with Betti numbers, an autoencoder-based deviation-from-stochasticity metric (D_S), and PCA eigenvalue ratios clustered with DBSCAN. They find that classes V, VII, and VIII are primarily NS, in contrast to a previous study that found all classes stochastic, and interpret this as evidence for a dynamical twin nature with GRS 1915+105. The paper includes a summary table of consensus classifications and a discussion of future work combining spectral and timing properties.","tokens_in":9292,"tokens_out":7695,"duration_ms":72209,"significance":"The result, if valid, would overturn the prior all-stochastic classification of IGR J17091-3624 and strengthen the analogy with GRS 1915+105. The paper has strengths: it applies multiple independent classification techniques, provides a link to the autoencoder code, and explicitly acknowledges classes with conflicting evidence (labeled S/NS). However, the central claim hinges entirely on the ability of the denoising filters to remove Poisson noise without imprinting deterministic structure; this premise is stated as a requirement but never calibrated on noise-only data. The conclusions are therefore not yet supported, but the gap is addressable with a controlled false-positive study.","major_comments":[{"comment":"Requirement 2 of §3 states that the filtering should avoid introducing additional local correlation, yet the paper never tests this on noise-only data. The unfiltered CI analysis gives nmsd < 3 for every class (Table 1), while the filtered analyses produce nmsd up to 11.492 for class VIII after ADA (Table 2). Because BOX and GAU are convolution kernels and NLM and ADA replace each point with weighted combinations of neighbors, applying these filters to white noise will lengthen the correlation time and can make D2(M) saturate, which is exactly the signature used to label a signal NS. The authors must run their full pipeline on pure Poisson noise and on linear stochastic processes with matched count rate, duration, and power spectrum, and report the false-positive rate for each method and for the consensus rule. Without such a control, the NS classifications in Tables 4-6 are unverified artifacts.","section":"§3, Tables 1-2"},{"comment":"The four methods disagree for several classes; for example, class VII is NS by CI, SVD, and PCA but S by the autoencoder, and class VIII is NS by CI, SVD, and PCA but S by the autoencoder. The paper's 'primary NS' conclusion for classes V, VII, and VIII rests on a ≥3-of-4 majority rule, but the paper does not justify that this rule has a low false-positive rate after denoising. A control experiment that counts how often three or more methods simultaneously produce NS on filtered white noise is necessary; the current consensus rule is otherwise an uncalibrated post-hoc construction.","section":"Tables 3, 5, 6"},{"comment":"No uncertainties are reported for nmsd or D2, despite the fact that nmsd is computed from only 19 surrogates and its null distribution is not quantified. Borderline cases such as class IV (nmsd = 2.234 original, 3.623 GAU) could cross the threshold with sampling noise. Additionally, the paper performs 45 classification tests (9 classes × 5 data versions) without any multiple-comparison correction; the probability of at least one spurious NS call is non-negligible. The authors should provide error bars or bootstrap intervals and either apply a multiple-comparison correction or explicitly argue why the consensus rule protects against multiplicity.","section":"§4.1, Tables 1-2"}],"minor_comments":[{"comment":"The caption of Fig. 1 describes class VIII as 'determined to be S unanimously by all filtering techniques', but Table 2 and §6 state that class VIII is NS (e.g., nmsd = 11.492 after ADA). If the figure shows a different class (e.g., class I or II), the caption must be corrected; as written it contradicts the paper's central result.","section":"Fig. 1"},{"comment":"The behavior column lists 'S/NS*' for class IX, but the asterisk is never defined in the table caption or the text; please clarify what it indicates.","section":"Table 2"},{"comment":"The heading 'Principle Component Analysis' should be 'Principal Component Analysis'.","section":"§4.4"},{"comment":"The text refers to 'IGR J17091-362-I' in the second paragraph; this appears to be a typo for 'IGR J17091-3624-I'.","section":"§5.2"},{"comment":"The text says a class is labeled S or NS based on a consensus of at least three filtering techniques, but Table 6 reports classes IV and VI as 'S/NS' overall. Please clarify whether 'S/NS' means no consensus and how the labeling rule applies to these entries.","section":"§5.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a proceedings contribution, and the missing noise-only control may stem from space constraints. That said, the central claim is strong enough that a revision with a false-positive calibration is essential before publication, even in a proceedings volume. I do not see a need to verify code integrity; the code link is a positive sign."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is the first paper to claim non-stochastic (NS) behavior in IGR J17091-3624 after denoising, specifically classes V, VII, VIII, contradicting the earlier all-stochastic classification [7]. That claim is new and worth taking seriously, but the evidence as presented does not support it. Every NS classification appears only after filtering, and there is no noise-only control to show the filters are not manufacturing the structure. The stress-test note lands; in fact the paper's own §3 lists 'avoid introducing additional local correlation' as a requirement, and the paper never tests it. The 'circularity' concern is better framed as an uncalibrated instrument: using a chaos-detector on data does not presuppose the answer, but using it without noise-only false-positive testing does.\n\nWhat the paper does well: the writing is clear, the data are public, the prior null result [7] is handled fairly, and the authors honestly label disagreements as S/NS. The combination of four filters and four classifiers is new for this source, and the autoencoder and SVD give independent looks.\n\nSoft spots. First, no control. BOX, GAU, NLM, ADA all smooth by weighted averaging of neighbors; smoothing white noise lengthens correlation time, flattens D2(M), creates SVD loops, raises PCA eigenvalue ratios, exactly the claimed NS signatures. Second, methods disagree on headline classes: autoencoder labels VII/VIII as S, while CI/SVD/PCA label NS; the >=3-of-4 consensus hides that. Third, free parameters are underreported: embedding dimension, delay, ADA window order, NLM bandwidth, PCA cutoff, DBSCAN epsilon. No uncertainties on nmsd/D2, no multiple-comparison correction across 9 classes and 5 filters. Fourth, several figures are missing from the arXiv version, so SVD topology claims cannot be checked.\n\nBottom line: plausible but unestablished. A matched control, pure Poisson noise and linear stochastic lightcurves with same count rate, duration, and PSD through the same pipeline, would settle false positives. Not a desk-reject; deserves a referee who can ask for that experiment. I would not cite the NS result until the control is in.","headline":"Plausible and new, but the NS classifications appear only after denoising and the missing noise-only control leaves filter artifacts as a live alternative.","tokens_in":779,"tokens_out":1692,"would_cite":false,"duration_ms":152574,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that after removing Poisson noise, several temporal classes of the black hole X-ray binary IGR J17091-3624 show deterministic (non-stochastic) dynamics, contradicting the earlier all-stochastic finding and strengthening…","keywords":["black hole X-ray binary","IGR J17091-3624","GRS 1915+105","Poisson noise","denoising","stochastic versus deterministic dynamics","correlation integral","autoencoder"],"falsifier":"Take synthetic lightcurves that are known to be pure Poisson or linear stochastic processes, match them to IGR J17091-3624 in length and count rate, pass them through the same four filters and four classifiers, and check whether any filter yields $\\mathrm{nmsd}\\ge 3$ or $D_S\\ge 1.5$ for a known-stochastic input; if so, the paper's NS classifications are artifacts.","tokens_in":8718,"feed_emoji":"🕳️","tokens_out":6396,"duration_ms":49278,"temperature":0.7,"pith_summary":"The paper asks whether the black hole X-ray binary IGR J17091-3624 is truly stochastic or whether earlier classifications were spoiled by Poisson noise. It applies four denoising filters and four independent nonlinearity tests, and reports that temporal classes V, VII, and VIII are primarily non-stochastic after denoising while classes I, II, and IX remain stochastic. The paper's stated aim is to overturn the earlier all-stochastic verdict and to show that IGR J17091-3624, like GRS 1915+105, switches between stochastic and deterministic states. A sympathetic reader would care because the result would make the two sources twins not only in timing behavior but in underlying dynamics.","feed_headline":"Denoising reveals deterministic states in black hole IGR J17091","feed_subtitle":"Classes V, VII, and VIII look stochastic in raw data but non-stochastic after noise filtering, matching GRS 1915+105.","key_machinery":"The machinery is a two-stage pipeline. First, four denoising filters — adaptive denoising (ADA), non-local means (NLM), Gaussian convolution (GAU), and boxcar convolution (BOX) — reduce Poisson noise, with the stated requirement that filtering not introduce additional local correlation. Second, four independent classifiers label each temporal class as S or NS: the correlation integral with surrogate analysis and the criterion $\\mathrm{nmsd}>3$; singular value decomposition with Betti-number topology; a multi-scale autoencoder whose deviation-from-stochasticity metric $D_S$ is thresholded at 1.5; and PCA eigenvalue-ratio features clustered with DBSCAN. A class is declared S or NS when at least three of the four methods agree, and conflicting cases are labeled S/NS.","core_discovery":"The paper claims that the temporal classes V, VII, and VIII of IGR J17091-3624 are primarily NS (non-stochastic or deterministic-looking) once Poisson noise is suppressed, whereas classes I, II, and IX are S, and classes IV and VI remain S/NS. This directly contradicts the earlier hypothesis, cited as [7], that all IGR J17091-3624 classes are stochastic. The paper further claims that this S/NS switching mirrors GRS 1915+105, whose classes are already known to be a mix of stochastic and deterministic behavior; one class pair, III and $\\nu$, is noted to differ. The claim is stated in Section 6 as 'IGR J17091-3624 classes V, VII, and VIII are primarily NS' and 'we have uncovered potential complex dynamics in IGR J17091-3624 using multiple denoising methods.'","pith_inferences":["Editorial inference: the paper does not report a control in which the four filters are applied to synthetic pure Poisson or linear stochastic series; such a control is the natural way to test whether any filter manufactures NS-looking structure.","Editorial inference: because the autoencoder was trained on signals that include GRS 1915+105 lightcurves, its $D_S$ threshold may be tuned to a brighter source; checking whether classifications depend on count rate would clarify transferability.","Editorial inference: the transient S/NS claim predicts that repeated observations of the same temporal class could flip classification depending on the Poisson realization, which a re-observation campaign could test."],"forward_implications":["The earlier all-stochastic classification of IGR J17091-3624 would be a noise artifact rather than a property of the source.","IGR J17091-3624 and GRS 1915+105 would share transient S/NS switching, strengthening the twin-source hypothesis.","Classes V, VII, and VIII become the promising targets for follow-up studies of deterministic accretion-flow dynamics in this source.","The same filter-and-classify pipeline could be applied to other faint X-ray binaries where Poisson noise dominates."],"supporting_citations":[{"why":"The prior study that classified all IGR J17091-3624 temporal classes as stochastic; this paper's target.","marker":"[7]"},{"why":"The GRS 1915+105 classification into stochastic and deterministic classes that this paper extends to IGR J17091-3624.","marker":"[4]"},{"why":"Supplies the autoencoder-based deviation-from-stochasticity metric and the $D_S \\ge 1.5$ threshold used as one classifier.","marker":"[15]"},{"why":"The adaptive denoising algorithm for detecting determinism in heavy noise.","marker":"[10]"},{"why":"Non-local means denoising adapted to one-dimensional lightcurves.","marker":"[8, 9]"},{"why":"The correlation integral method for estimating correlation dimension $D_2$.","marker":"[11]"},{"why":"Surrogate analysis used to test the linear stochastic null hypothesis.","marker":"[12]"},{"why":"The $\\mathrm{nmsd}>3$ criterion for declaring a timeseries non-stochastic.","marker":"[13]"},{"why":"PCA and SVD matrix-based methods for classifying S versus NS timeseries.","marker":"[14]"}],"fun_headline_variants":["Denoising reveals determinism in black hole IGR J17091","IGR J17091: chaos to order via denoising","Noise hides determinism in black hole IGR J17091","Stochastic label for IGR J17091 challenged by denoising"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The conclusion stands on the assumption that the four denoising filters remove Poisson noise without imprinting deterministic structure onto an underlying stochastic signal; the paper lists this requirement but never validates it with a control on pure noise.","fun_headline_variants_meta":{"raw":{"variants":["Denoising reveals determinism in black hole IGR J17091","IGR J17091: chaos to order via denoising","Noise hides determinism in black hole IGR J17091","Stochastic label for IGR J17091 challenged by denoising"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000414,"raw_usage":{"total_tokens":2149,"prompt_tokens":965,"completion_tokens":1184,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":1116}},"tokens_in":581,"tokens_out":1184,"duration_ms":9536,"temperature":1.0,"reasoning_tokens":1116,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:58:51.119329+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take synthetic lightcurves that are known to be pure Poisson or linear stochastic processes, match them to IGR J17091-3624 in length and count rate, pass them through the same four filters and four classifiers, and check whether any filter yields $\\mathrm{nmsd}\\ge 3$ or $D_S\\ge 1.5$ for a known-stochastic input; if so, the paper's NS classifications are artifacts.","supporting_citations":[{"cited_title":"Misra, aff, K.P","cited_arxiv_id":null,"evidence_quote":"Surrogate analysis used to test the linear stochastic null hypothesis."},{"cited_title":"Harikrishnan, R","cited_arxiv_id":null,"evidence_quote":"The $\\mathrm{nmsd}>3$ criterion for declaring a timeseries non-stochastic."},{"cited_title":"Identification of Stochasticity by Matrix-decomposition: Applied on Black Hole Data","cited_arxiv_id":"2307.07703","evidence_quote":"PCA and SVD matrix-based methods for classifying S versus NS timeseries."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The adaptive denoising algorithm for detecting determinism in heavy noise."},{"cited_title":"Adegoke, B","cited_arxiv_id":null,"evidence_quote":"The prior study that classified all IGR J17091-3624 temporal classes as stochastic; this paper's target."},{"cited_title":"Adegoke, P","cited_arxiv_id":null,"evidence_quote":"The GRS 1915+105 classification into stochastic and deterministic classes that this paper extends to IGR J17091-3624."}],"review_version":1}