{"id":"09f5eb92-ccbb-4b76-bd3e-3c380670ff16","arxiv_id":"2411.17810","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"After denoising RXTE light curves, several temporal classes of IGR J17091-3624 show non-stochastic signatures, supporting a dynamical resemblance to GRS 1915+105.","lead":"This paper applies noise-reduction techniques to X-ray light curves of the black hole binary IGR J17091-3624 and reports that several temporal classes show deterministic, non-stochastic signatures. If correct, it supports the twin relationship with GRS 1915+105 and suggests earlier stochastic classifications were dominated by Poisson noise.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validation gap: no synthetic ground-truth test shows that denoising cannot imprint deterministic structure; NLM/ADA and even BOX/GAU can turn stochastic input into locally smooth series that the subsequent CI/SVD/PCA tests may read as NS.","rationale":"The paper is honest about its heuristics and the agreement among methods is mildly supportive; noting that SVD on the original data already shows structure for classes VI, VII, and VIII indicates that not all signal is filter-generated. However, the strong causal statement in Section 6, that the previously reported stochasticity is 'an artifact of Poisson noise domination,' requires the filter to be faithful, not merely effective at smoothing. Section 3.3 acknowledges the lack of ground truth and sets NLM's key parameter from an RMS knee in the observed data itself. The absence of synthetic benchmarks is therefore the load-bearing gap, exactly as the Reader identified. The paper's own negative controls—classes I, II, III, IX remaining S—are valuable but asymmetric: they demonstrate that the pipeline does not label every filtered series NS, not that the NS labels are correct. The proposed synthetic test would settle whether the false-positive rate is controlled for Poisson noise, red noise, and known chaos, and would also verify that the methods retain sensitivity to true deterministic dynamics. Until then, CONDITIONAL is appropriate; the Reader's verdict stands unchanged.","tokens_in":19916,"tokens_out":4408,"duration_ms":46073,"concrete_test":"Run the exact pipeline (BOX, GAU, ADA, NLM; then CI/NMSD, SVD/Betti, PCA/DBSCAN) on synthetic stochastic light curves matched to each IGR observation: (i) pure Poisson counts with the observed mean rate and length; (ii) linearly filtered Gaussian or red-noise processes with the observed PSD; and (iii) a known chaotic process, e.g., Lorenz, sampled as Poisson counts, as a positive control. For each class, compute the false-positive rate = fraction of stochastic realizations labeled NS by at least three filter/test combinations. If the rate exceeds about 5%, or if the positive control is not recovered, the denoising-induced-determinism concern is confirmed and the headline claim must be downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the premise that the four filters in Sections 3.1–3.3 remove Poisson noise without imprinting low-dimensional structure. This premise is not validated. BOX and GAU are low-pass convolutions; ADA fits local polynomials; NLM averages self-similar patches. Applied to a pure stochastic process, these operations produce correlated, locally smooth series. The paper's main discriminator for CI is the IAAFT surrogate test in Section 4.1.1, but surrogates are generated from the filtered series itself. If a filter imprints nonlinear smoothness, especially NLM whose weights and sensitivity h are fit to the observed data, the surrogates preserve the second-order power spectrum but not the specific patch-redundancy structure, so the NMSD statistic can reject the null for stochastic input. The SVD/Betti and PCA/DBSCAN tests likewise have no calibrated false-positive rate on filtered noise. Class IV is labeled NS partly because one SVD/NLM portrait 'looks like' an attractor, and class VI is S by CI/PCA but NS by SVD. The fact that several classes remain S after filtering is a useful control, but it only shows that not all stochastic data are converted to NS; it does not show that the four claimed NS classes are genuine.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes nine RXTE (and one Chandra) light curves of IGR J17091–3624, applies four denoising filters (BOX, GAU, ADA, NLM), and uses three classification diagnostics (correlation integral with IAAFT surrogates, SVD with Betti numbers, and PCA with DBSCAN) to distinguish stochastic from non-stochastic behavior. The authors report that temporal classes IV, V, VII, and VIII, which were previously classified as stochastic by Adegoke et al. (2020), become non-stochastic after denoising under a majority rule over methods. They interpret this as evidence that IGR J17091–3624 is dynamically similar to GRS 1915+105 and that the earlier stochastic conclusion was an artifact of Poisson noise contamination.","tokens_in":20243,"tokens_out":5311,"duration_ms":48257,"significance":"If the central claim were established, it would overturn a published classification result and materially strengthen the twin-source interpretation of IGR J17091–3624 and GRS 1915+105. The paper also contributes a practical multi-filter, multi-diagnostic pipeline and is transparent about its parameter choices and about disagreements among methods; the use of public archival data and the reference to the NoLiTSA codebase are helpful for reproducibility. However, the claim is currently conditioned on an unvalidated assumption about the denoising filters and on test-specific threshold choices, so the significance is high but not yet secure.","major_comments":[{"comment":"Section 3 states as a design criterion that the filters should not introduce additional local correlation, yet the convolution filters in Eqs. (1)–(5) are explicitly low-pass smoothers and ADA fits overlapping local polynomials (Eq. 6). No synthetic ground-truth control is reported: Poisson-only realizations matched to the observed count rates and lengths are never passed through the four filters and then through the CI/IAAFT, SVD/Betti, and PCA/DBSCAN pipeline. Because these operations can convert white noise into a locally smooth, correlated series, the central claim that classes IV, V, VII, and VIII are non-stochastic after denoising is not yet supported. The authors should add false-positive rates for each filter–test combination on filtered pure noise and on known chaotic and stochastic benchmark systems.","section":"Secs. 3.1–3.3, Tables 3 and 6"},{"comment":"The IAAFT surrogate test is applied to the filtered series rather than to the original observation, and for NLM the filtered series is constructed from patch redundancy with a sensitivity parameter h tuned per lightcurve (Eq. 9, Fig. 1). IAAFT surrogates preserve the power spectrum and amplitude distribution of this already-filtered series, but they do not preserve the patch-redundancy structure that NLM specifically builds on; hence nmsd > 3 can reject the null for stochastic input if the filter imprints such structure. The Table 3 classifications therefore need to be accompanied by a surrogate or synthetic test applied to the unfiltered data, or by a surrogate scheme consistent with the filtering model.","section":"Sec. 4.1.1, Eq. (12)"},{"comment":"The PCA/DBSCAN hyperparameters r_cut = 124 and epsilon = 40 are selected by maximizing the silhouette score on the very set of 45 signals that is subsequently labeled S/NS. This makes the Figure 12 split partly a product of thresholds fitted to the target data rather than an independent classification. With only nine source classes and multiple filtering variants, the reported silhouette score of 0.693 is not, by itself, evidence of robustness. A synthetic benchmark, cross-validation, or thresholds fixed a priori would be needed to show that the S/NS separation is not an artifact of threshold optimization.","section":"Sec. 5.3, Fig. 12"},{"comment":"The statement in Sec. 5.1 that classes IV, V, VII, and VIII are determined to be NS 'unanimously by our filtering methods' is contradicted by the paper's own Table 6, where SVD gives S/NS for class IV and where class VI is listed as S/NS overall. The final assignment of IV = NS in Sec. 6 is justified by the NLM SVD portrait, even though the stated formal criterion ('if three or more methods agree') is not met. The headline claim should be revised to distinguish robustly NS classes from ambiguous ones, and the internal inconsistency between the narrative and Table 6 should be resolved.","section":"Secs. 5.1 and 6, Tables 3 and 6"}],"minor_comments":[{"comment":"The sentence 'We move on to more involved methods that address these issues. explicitly preserved during boxcar denoising.' appears unfinished; please revise it.","section":"Sec. 3.1.1"},{"comment":"The source name is misspelled as 'IGR J17091–362' in two places; it should be 'IGR J17091–3624'.","section":"Secs. 1 and 4.1"},{"comment":"The heading uses 'Principle Component Analysis'; the standard term is 'Principal Component Analysis'.","section":"Sec. 4.3"},{"comment":"The manuscript does not state the embedding dimension range Mmin–Mmax, the delay used for each class, or the number of surrogates per lightcurve; these values are needed to interpret the nmsd statistics and D2 saturations.","section":"Table 3 and Eq. (12)"},{"comment":"For full reproducibility, the authors should provide the custom code for the filters and for the Betti-number and DBSCAN classification, rather than only citing the NoLiTSA package.","section":"Secs. 3 and 4"}],"recommendation":"major_revision","confidential_remarks":"The missing synthetic-control experiment is the key issue: it is fixable within the scope of the paper and should not require new observations. The paper is within the journal's scope, and the authors are honest about disagreements among methods, but the current abstract overstates the evidence relative to Table 6. I saw no concerns about attribution or overlap beyond the acknowledged prior work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper does something new and reports it clearly, but the central claim rests on a validation gap. The authors show that four temporal classes of IGR J17091-3624 look non-stochastic after denoising, contradicting Adegoke et al. (2020), and they support it with CI/surrogate, SVD/Betti, and PCA/DBSCAN. NLM applied to astrophysical time series, and the quantitative spin on SVD/PCA classification, are real contributions.\n\nCredit where due: the internal control is useful — classes I, II, and IX remain stochastic under every filter, which shows the pipeline does not blindly label everything non-stochastic. The authors are honest about several limitations, including the heuristic SNR measure and the knee-based choice of NLM sensitivity. The agreement among very different filters (boxcar and Gaussian as well as NLM and ADA) for the NS classes is reassuring.\n\nThe soft spot is the missing synthetic ground-truth test. All four filters smooth the data, and the IAAFT surrogates are generated from the filtered series itself. That means the CI test can reject the null for filtered stochastic data if the filter imprints smooth, low-dimensional structure. The stress-test note is right: there is no calibrated false-positive rate for the SVD and PCA steps on filtered noise. On top of that, the PCA cutoff (r_cut=124) and DBSCAN epsilon (40) are optimized via silhouette score on the very signals being classified, so the S/NS labels are partly functions of fitted choices. The paper's own claim in Section 6 — that the earlier stochastic classification is 'an artifact of Poisson noise domination' — is stronger than the evidence supports. What the data show is consistency with determinism after denoising, not proof that the previous result was wrong.\n\nNone of this makes the paper incoherent. The methods are described in enough detail to reproduce, and the negative controls help. But the load-bearing premise — that denoising removes noise without creating structure — needs to be tested with simulated light curves of known chaotic and stochastic origin.\n\nFor a reader working on nonlinear timing analysis of accreting systems, this is worth engaging with, and it deserves a serious referee. I would send it to peer review, with the synthetic validation as the main requested revision.","headline":"Careful multi-method analysis suggests determinism in IGR J17091-3624 after denoising, but the missing synthetic ground-truth control leaves the central claim unproven.","tokens_in":20757,"tokens_out":3074,"would_cite":false,"duration_ms":26501,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["97.60.Lf","95.75.Wx","05.45.-a"],"model":"deepseek-v4-flash","headline":"The black hole X-ray binary IGR J17091-3624, previously classified as purely stochastic, shows four temporal classes that are deterministic after removing Poisson noise, strengthening its status as a dynamical twin of GRS 1915+105.","keywords":["black hole X-ray binary","IGR J17091-3624","GRS 1915+105","Poisson noise","denoising","correlation integral","singular value decomposition","principal component analysis"],"falsifier":"Apply the same four filters to synthetic lightcurves generated from a stationary stochastic process with the same length, mean count, and RMS as each IGR class, then run the same CI/surrogate, SVD, and PCA pipeline; if these pure-noise controls are classified as non-stochastic, the denoising is imprinting determinism and the paper's conclusion fails.","tokens_in":19716,"feed_emoji":"🕳️","tokens_out":5925,"duration_ms":45045,"temperature":0.7,"pith_summary":"This paper argues that IGR J17091-3624, long thought to be an outlier because its X-ray flickering looked purely random, actually hides deterministic dynamics. The apparent stochasticity, the authors claim, is an artifact of Poisson photon noise swamping a faint source. Applying four different denoising procedures (moving average, Gaussian, adaptive polynomial, and non-local means) and then testing with three independent statistical probes, they find that four of the nine temporal classes are non-stochastic after cleaning. That would make IGR J17091-3624 a dynamical twin of GRS 1915+105, and it would mean the earlier stochastic-only classification needs to be revised.","feed_headline":"Black hole's 'random' flickers found to hide deterministic order","feed_subtitle":"Faint source's stochastic label traced to Poisson noise; four temporal classes match GRS 1915+105.","key_machinery":"The argument runs on the conjunction of four denoising filters and three classification tests. The filters are the boxcar (moving average), Gaussian convolution, an adaptive denoising algorithm (ADA) that fits overlapping low-order polynomials to windows of the lightcurve, and non-local means (NLM), which averages similar short patches wherever they occur in time. The tests are the Grassberger-Procaccia correlation integral with surrogate analysis (quantified by the normalized mean sigma deviation, nmsd, with nmsd > 3 rejecting the stochastic null), singular value decomposition of the delay-embedded data matrix read through Betti numbers (β0 + β1 > 1 means structure), and principal component analysis of the eigenvalue-ratio curve clustered with DBSCAN. Agreement among at least three of the four filters is required to label a class stochastic or non-stochastic.","core_discovery":"The central claim is that temporal classes IV, V, VII, and VIII of IGR J17091-3624 are deterministic (non-stochastic) once Poisson noise is removed; classes I, II, III, and IX remain stochastic, and class VI is ambiguous. The authors show that in every unfiltered lightcurve the correlation-integral dimension matches the embedding dimension and the surrogate test says 'stochastic', reproducing the earlier conclusion. After any of the four filters, the same four classes produce correlation dimensions that saturate below the embedding dimension, SVD phase portraits with multiple topological features, and PCA eigenvalue-ratio outliers, with all three tests agreeing, while the stochastic classes stay stochastic. They conclude that the source hosts a complex underlying dynamical system similar to GRS 1915+105, and that prior findings of pure stochasticity were noise artifacts.","pith_inferences":["The same pipeline could be applied to other faint X-ray binaries whose 'stochastic' classifications come from noise-dominated RXTE or NICER lightcurves, potentially overturning several published results.","Because the NLM filter preserves stochasticity when the signal is genuinely random (the S classes stay S), the method itself carries a built-in control; a natural next test is to benchmark all four filters on simulated chaotic signals embedded in Poisson noise to quantify detection thresholds.","If the determinism is real, the underlying dynamical system should be identifiable in higher-cadence future observations, for example in the phase-space topology of the attractor or in dimension estimates that match one of the GRS 1915+105 classes."],"forward_implications":["IGR J17091-3624 displays at least as much dynamical complexity as GRS 1915+105, so the two sources can be studied as the same physical class of accreting black holes.","The four non-stochastic classes (IV, V, VII, VIII) can now be mapped onto specific accretion-flow states: a Keplerian disk for VII and VIII, a transitional advection-dominated flow for IV, and a mixed disk/advection flow for V.","The stochastic classes I, II, III, and IX all correspond to power-law-dominated, general advective accretion flow states, tightening the coupling between spectral state and temporal dynamics.","Poisson-noise removal should become a standard pre-processing step before non-linear analysis of any faint X-ray source, not just this object."],"supporting_citations":[{"why":"Classified all IGR J17091-3624 temporal classes as stochastic or noise-dominated; the claim this paper directly refutes.","marker":"Adegoke et al. (2020)"},{"why":"Supplies the correlation-integral method and correlation dimension for detecting determinism in time series.","marker":"Grassberger & Procaccia (1983)"},{"why":"Defines the nmsd statistic and the nmsd > 3 threshold used to reject the stochastic null hypothesis.","marker":"Harikrishnan et al. (2006)"},{"why":"Provides the adaptive denoising algorithm (ADA) designed to detect chaos in heavy noise.","marker":"Tung et al. (2011)"},{"why":"Introduces non-local means denoising, adapted here with a Poisson-noise-scaled decay parameter to lightcurves.","marker":"Buades et al. (2005)"},{"why":"Gives the combined spectral/temporal classification mapping C and S behaviors to four accretion-flow states, which the paper extends.","marker":"Adegoke et al. (2018)"},{"why":"Establishes SVD-based phase portraits for GRS 1915+105 classes, the approach this paper makes quantitative with Betti numbers.","marker":"Misra et al. (2006)"},{"why":"Supplies the PCA eigenvalue-ratio method and its features (MER, VAR, Area) for classifying stochastic versus deterministic series.","marker":"Chakka & Sinha (2024)"},{"why":"Defines the nine temporal classes of IGR J17091-3624 used throughout the analysis.","marker":"Altamirano et al. (2011)"}],"fun_headline_variants":["Denoising reveals deterministic order in black hole twin","IGR J17091 shows deterministic classes like GRS 1915","Noise removal unmasks determinism in black hole flickers","Black hole twin's 'stochastic' flecks are deterministic","Poisson noise hides deterministic pattern in IGR J17091"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole classification rests on the assumption that the denoising filters remove Poisson noise without creating artificial low-dimensional structure in a signal that is in fact purely random.","fun_headline_variants_meta":{"raw":{"variants":["Denoising reveals deterministic order in black hole twin","IGR J17091 shows deterministic classes like GRS 1915","Noise removal unmasks determinism in black hole flickers","Black hole twin's 'stochastic' flecks are deterministic","Poisson noise hides deterministic pattern in IGR J17091"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000436,"raw_usage":{"total_tokens":2264,"prompt_tokens":1037,"completion_tokens":1227,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":1140}},"tokens_in":653,"tokens_out":1227,"duration_ms":8927,"temperature":1.0,"reasoning_tokens":1140,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:49:03.550073+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the same four filters to synthetic lightcurves generated from a stationary stochastic process with the same length, mean count, and RMS as each IGR class, then run the same CI/surrogate, SVD, and PCA pipeline; if these pure-noise controls are classified as non-stochastic, the denoising is imprinting determinism and the paper's conclusion fails.","supporting_citations":[{"cited_title":"K., Mukhopadhyay, B., & Misra, R","cited_arxiv_id":null,"evidence_quote":"Classified all IGR J17091-3624 temporal classes as stochastic or noise-dominated; the claim this paper directly refutes."},{"cited_title":"2006, Physica D: Nonlinear Phenomena, 215, 137","cited_arxiv_id":null,"evidence_quote":"Defines the nmsd statistic and the nmsd > 3 threshold used to reject the stochastic null hypothesis."},{"cited_title":"2011, Phys","cited_arxiv_id":null,"evidence_quote":"Provides the adaptive denoising algorithm (ADA) designed to detect chaos in heavy noise."}],"review_version":1}