{"id":"ae66f8df-4022-41c5-8e7b-4a20ef8c04aa","arxiv_id":"2507.12257","paper_version":4,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"PLaCy detects causal links by applying Granger causality to time-varying power-law spectral features (log-log slope and intercept) instead of raw signals.","lead":"This paper introduces PLaCy, a causal discovery method that tracks the power-law slope and intercept of each time series' frequency spectrum in sliding windows and performs Granger causality tests on those spectral features. The authors claim this approach is more robust to noise and non-stationarity than standard time-domain causal discovery, and they evaluate it on synthetic and two real-world benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.1 is unproven and its key assumption restates the conclusion; Appendix C shows the fitted λ is frequency-dependent, and Table 5 contradicts the 'same graph' claim in the paper's own low-noise setting.","rationale":"The reader's weakest-assumption assessment is on target, and my stress-test converges on the same point. I find the theorem's assumption 5 circular and Appendix C's own formula showing frequency dependence of the fitted λ to be the sharpest technical reason to doubt the method's theoretical foundation. I would not escalate to REJECT because the paper includes real supporting evidence for the method as a heuristic: public code, negative-control Petersen tests (Appendix E.6), FDR robustness (E.8), and a limitations section that honestly notes failure modes such as slowly varying spectra and short series. These are independent empirical checks, and the method does outperform in many nonstationary and noisy settings, e.g., the dOU(σ_m^g>0) rows of Table 1. The problem is that the claims are stated more strongly than the evidence: the theorem is not proven, and the abstract's 'consistently outperforms' is contradicted by the low-noise rows of Table 5, where Granger is much better at C=0.2. A CONDITIONAL verdict that requires (i) proving or explicitly restricting Theorem 3.1 and (ii) replacing 'consistently' with a qualified claim is the right outcome. No change to the reader's verdict is needed.","tokens_in":55741,"tokens_out":10032,"duration_ms":109354,"concrete_test":"Using the released code, reproduce the Table 5 cell OU(σ_m^g=0), N=5, σ_a^g=1.0, C=0.2, σ_b=0 and compare PLaCy's F1 against time-domain Granger's F1 over the same 100 runs. If the gap (0.13 vs 0.85) reproduces, Theorem 3.1's 'same graph' claim is contradicted in a configuration that satisfies the theorem's linearity, stationarity, and additive-noise assumptions. Independently, re-derive the theorem without using assumption (5) as a premise; a successful derivation must show that the frequency-dependent correction (log c + a3 − a1)/log f in Appendix C either vanishes in the relevant limit or is provably captured by the OLS fit in a way that preserves the graph.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is Theorem 3.1: Granger causality on the spectral features (a_i, λ_i) recovers the same graph G* as time-domain Granger analysis. This theorem is not established. Its assumption (5), 'causal dependencies are identifiable from amplitude-spectral dynamics,' is essentially the conclusion, and the appendix does not derive it from the other assumptions. Theorem B.3 only shows that, under a signal-dominance condition, (a_i, λ_i) 'retain information' about a causal influence; it does not prove that no spurious edges are introduced, so step (B)'s 'no spurious edges introduced' claim is unsupported. More concretely, Appendix C's own calculation for the additive model used to generate the synthetic data gives λ1 ≈ λ3 + (log c + a3 − a1)/log f. This depends on frequency, so the single OLS slope fitted in Eq. (2) is not a well-defined power-law exponent: it is an average over the fitted frequency band, and the assumption that each component admits a local power-law spectral approximation fails for the additive causal injection xj ← xj + C·xi(t−τ). The contradiction is visible in the paper's own Table 5: for OU(σ_m^g=0), C=0.2, σ_b=0, Granger attains F1 = 0.85±0.20 while PLaCy attains 0.13±0.21, so PLaCy is not recovering the same graph as time-domain Granger. This also undercuts the abstract's 'consistently outperforms.' Unless the theorem is restricted to a specific asymptotic limit and to causal strengths above the spectral-superposition threshold of Appendix C Eq. (6), it is either false or vacuous for the primary experimental setting.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PLaCy, a causal discovery method for multivariate time series whose spectra approximately follow power laws. Each series is segmented into sliding windows, a line is fitted to log spectral amplitude versus log frequency in each window, and multivariate Granger tests are applied to the resulting sequences of intercepts and exponents. The central theoretical claim, Theorem 3.1, is that Granger causality performed on these spectral features recovers the same causal graph as time-domain Granger analysis. The paper supports this with a broad synthetic benchmark based on Ornstein-Uhlenbeck processes under several additive/multiplicative noise regimes, two real-world datasets, negative-control experiments, FDR robustness checks, and a complexity analysis.","tokens_in":56109,"tokens_out":5419,"duration_ms":68389,"significance":"The idea of summarizing each window by its power-law exponent and intercept is attractive, and the empirical campaign is unusually thorough: it includes multiple noise types, N=5 and N=10, negative controls, FDR correction, and publicly available code. If the invariance theorem were valid, the method would justify replacing time-domain Granger inference with spectral-feature inference for scale-free processes, with potentially broad applications in finance, neuroscience, and climate science. However, the theorem is not established as stated, and the abstract's claim of consistent outperformance is contradicted by the paper's own tables. The contribution therefore currently rests on heuristic empirical evidence rather than on the advertised theoretical guarantee.","major_comments":[{"comment":"The theorem is not proven, and as stated it is circular. Its fifth assumption, 'causal dependencies are identifiable from amplitude-spectral dynamics,' is essentially the graph-preservation claim that the theorem is supposed to establish; no derivation of that assumption from the earlier, more primitive assumptions is given. Step (B) of Appendix B relies on Theorem B.3, which only shows that, under a signal-dominance condition, the features (a_i, λ_i) 'retain information' about a causal influence. It does not prove that no spurious edges are introduced, nor that the graph learned from (a, λ) equals G*. Thus the 'complete formal proof' promised after Theorem 3.1 is not supplied.","section":"§3.1, Theorem 3.1"},{"comment":"For the additive data-generation mechanism used in Section 5.1, the paper's own derivation yields λ1 ≈ λ3 + (log c + a3 − a1)/log f. This is frequency-dependent, so no single exponent λ exists for the summed process; the OLS slope fitted in Eq. (2) is an average over a frequency band. Consequently the local power-law approximation assumed in Theorem 3.1 is violated by the very model used to generate the synthetic data, and the theorem does not apply to the experiments it is invoked to justify.","section":"Appendix C, Eqs. (5)-(8)"},{"comment":"The claim that the method 'consistently outperforms state-of-the-art alternatives' is contradicted by the paper's own results. In Table 5, in the row OU(σm_g=0), C=0.2, σ_b=0.0, PLaCy attains F1=0.13±0.21 whereas Granger attains F1=0.85±0.20; Table 1 shows the same pattern (0.14±0.22 vs 0.58±0.26). This is the low-noise setting in which time-domain Granger analysis should be most reliable, and it directly contradicts Theorem 3.1's assertion that the two analyses recover the same graph. The empirical claims must be restricted to the noisy or non-stationary regimes where the spectral averaging is helpful, and the abstract should be revised accordingly.","section":"Abstract; Tables 5 and 1"},{"comment":"Step (A1) is also not established to the standard required for valid VAR-based inference. Theorem B.1 gives a CLT-style argument for approximate Gaussianity of (a, λ), but the proof acknowledges that overlapping windows induce temporal dependence in the noise and does not prove that the residual process satisfies the assumptions needed for the Granger Wald tests, such as absence of serial correlation and correct model specification. The claim that the feature series 'satisfies the classical conditions' of Lütkepohl is therefore unsupported.","section":"Appendix B, Theorem B.1 and step (A)"}],"minor_comments":[{"comment":"The legend lists 'PCMCI' twice; the second entry appears to be PCMCIΩ and should be relabeled.","section":"Figure 2"},{"comment":"The algorithm title is typeset as 'PL ACY'; this should be fixed to 'PLaCy'.","section":"Algorithm 1"},{"comment":"The text describes a p-value-based adaptive procedure for choosing the window length, while the experiments fix l=50; the relation between the adaptive procedure and the fixed default should be clarified.","section":"Section 3 and Table 11"},{"comment":"The description 'p-value of 0.05' for selecting the window length is ambiguous: it should specify which null hypothesis is tested and over which frequency range.","section":"Appendix D.1"}],"recommendation":"major_revision","confidential_remarks":"The empirical core of the paper is serious, but the two headline claims — the invariance theorem and consistent outperformance — are both load-bearing and currently unsupported or contradicted by the paper's own data. I recommend major revision rather than rejection because the method's empirical value in non-stationary settings could survive a substantial repositioning: removing or heavily restricting Theorem 3.1, explicitly identifying regimes where PLaCy fails, and replacing 'consistently outperforms' with qualified comparisons. If the authors intend to retain the invariance theorem as a central contribution, the revision would need to be much deeper than a local fix."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"PLaCy is a plausible and well-tested idea: instead of running Granger causality on raw time series, extract a windowed log-log spectral slope and intercept and run multivariate Granger on those features. The specific combination is new relative to the spectral-causality literature (Geweke, DTF, BCGeweke), and the empirical campaign is broad—four OU scenarios, N=5 and 10, negative controls, two real datasets, released code. Credit where due: they include Petersen negative-control tests, report FDR-corrected results, and have an honest limitations section. In the noisy, multiplicative-noise regimes that motivate the paper, PLaCy beats Granger and the other baselines by a wide margin.\n\nThe soft spot is the theory. Theorem 3.1 claims causal discovery on (a,λ) recovers the same graph as time-domain Granger. The appendix doesn't prove it. Step (B1) assumes 'causal dependencies are identifiable from amplitude-spectral dynamics'—that is the conclusion packaged as an assumption. The supporting lemmas show approximate Gaussianity and 'retaining information,' but nothing rules out spurious edges. Worse, Appendix C's own calculation for the additive injection used in the synthetic data gives λ1 ≈ λ3 + (log c + a3 − a1)/log f, which depends on frequency; the single OLS slope over the band is not a well-defined exponent, so the clean power-law carrier assumption fails exactly in the primary experimental setting.\n\nThe paper's own Table 5 makes the problem concrete. In OU(σm_g=0) with C=0.2 and σb=0, Granger gets F1=0.85±0.20 while PLaCy gets 0.13±0.21. In that low-noise stationary case, PLaCy is not recovering the time-domain graph at all. That directly contradicts the abstract's 'consistently outperforms' and the theorem's 'same graph' guarantee. The method is genuinely robust in noisy settings, but the theoretical claim overstates its range of validity.\n\nThis is still a paper worth refereeing. The empirical contribution is substantial, the method is new, and the flaw is fixable: either restrict the theorem to an asymptotic regime where it can be proven, or reframe PLaCy as a robust heuristic with empirical justification. As written, the paper should not be accepted with the theory intact. A serious referee would demand the theorem be corrected or removed. I'd send it to review, but expect major revision.","headline":"A solid empirical heuristic whose central theorem is unproven; the paper's own low-noise results contradict the 'consistently outperforms' claim.","tokens_in":56674,"tokens_out":3001,"would_cite":true,"duration_ms":29959,"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":"Power-law spectral slopes preserve causal structure under noise.","keywords":["causal discovery","time series","power-law spectra","Granger causality","spectral feature extraction","non-stationary noise","Ornstein-Uhlenbeck process","scale-free processes"],"falsifier":"Generate two independent power-law noise series and make y a lagged copy of x, y(t) = x(t − τ), so the two series have identical spectra and hence identical fitted (a, λ) features: a time-domain Granger test would detect the causal lag, while a method operating only on (a, λ) cannot, revealing a boundary of the graph-invariance claim. A second check is to compute the fitted λ for a sum of two power laws over a range of window lengths and measure how large the frequency-dependent correction (log c + a3 − a1)/log f is in practice.","tokens_in":2021,"feed_emoji":"📉","tokens_out":2323,"duration_ms":114581,"temperature":0.7,"pith_summary":"The paper argues that many real-world time series have power-law frequency spectra, and that the evolution of the two parameters describing that spectrum, the log-amplitude a and the spectral exponent λ, carries the causal signal. It proposes a method that fits the spectrum as A(f) = e^a $f^{{-λ}}$ in sliding windows and runs Granger causality tests on the resulting (a, λ) feature series. The central theorem states that this spectral feature mapping preserves the underlying causal graph, so discovery on the features is equivalent to discovery in the original time domain. If true, this would justify a robust preprocessing step that filters out multiplicative noise and non-stationary transients, improving causal discovery in finance, climate, neuroscience, and pollution data. The paper also reports higher F1 and true-negative rates than existing time- and frequency-domain methods on synthetic benchmarks and on real-world datasets with known causal structure, while acknowledging that the method is specialized to scale-free spectra.","feed_headline":"Power-law spectral slopes preserve causal structure","feed_subtitle":"Fit each window's spectrum to a power law and its slopes and amplitudes recover true causal links despite noise","key_machinery":"The central object is the power-law spectral feature pair (a, λ), obtained by an ordinary least squares fit of log A(f) = a − λ log f on the discrete Fourier transform amplitudes of each sliding window, where a is the log-amplitude scaling and λ is the spectral exponent measuring the decay rate. The construction turns each raw series into two feature series, and multivariate Granger causality tests with target λ_j and causing pair (a_i, λ_i) then reconstruct the graph. The invariance claim in Theorem 3.1 rests on assumptions including local power-law spectra, weak stationarity within windows, linear causal mechanisms, additive noise that does not dominate the signal, causal dependencies identifiable from amplitude-spectral dynamics, and the transformed features satisfying the standard assumptions for valid VAR-based Granger inference.","core_discovery":"For a linear structural causal process whose components have local power-law spectra, the sliding-window spectral fit (a, λ) preserves the causal graph. Theorem 3.1 claims that causal discovery performed on the transformed feature sequence (a, λ) recovers the same causal graph as causal discovery performed in the original time domain. The mechanism is that spectral exponents, rather than raw values, carry causal dependencies: when one variable causes another, perturbations in the slope or amplitude of the first variable's spectrum propagate into the fitted λ of the second, and a vector autoregressive Granger test can detect this while broadband noise and exogenous multiplicative fluctuations are damped by spectral averaging. The paper reports that this yields higher F1 and true-negative rates than baseline algorithms across four synthetic Ornstein-Uhlenbeck scenarios, including with multiplicative noise and non-equilibrium initialization, and on river hydrology and air pollution datasets with reference causal graphs.","pith_inferences":["The invariance theorem is explicitly conditional on causality being visible in amplitude-spectral dynamics, so a causal mechanism that shifts only the phase of the spectrum, such as a pure delayed copy of a signal, would be invisible to (a, λ) features; a phase-aware feature set might extend the method's reach.","The appendix derivation for a sum of two power laws gives the fitted λ_1 as approximately λ_3 plus a frequency-dependent correction, so slope estimates can vary with window length; this suggests a testable sensitivity probe on real data.","Because each raw series is compressed to two features, temporal resolution is traded for spectral resolution, and the stated minimum window of fifty points implies that causal lags shorter than the window length deserve direct benchmarking.","The positive results with a constraint-based algorithm run on spectral features suggest a generic spectral summarization layer could help other causal discovery paradigms, but the class of processes where this helps beyond Ornstein-Uhlenbeck and the two real datasets remains to be mapped."],"forward_implications":["For processes with power-law spectra, time-domain Granger analysis can be replaced by Granger analysis on the (a, λ) features without changing the causal meaning of the recovered graph.","Multiplicative noise and non-equilibrium transients, which typically break stationarity assumptions, are suppressed by spectral averaging, so discovery remains accurate in regimes where constraint-based tests become unreliable.","Because the features are low-dimensional and estimated by linear regression, the preprocessing is cheap: the complexity is O(N L l log l) for the spectral step plus O(N^2 L) for the pairwise Granger tests.","The spectral feature extraction can be reused as a preprocessing layer for other discovery algorithms, as the supplementary experiments show improved F1 for a constraint-based algorithm run on spectral features.","The authors state that the method is specialized to scale-free processes with clear power-law spectra; for slowly varying spectra and very short series, time-domain methods such as standard Granger causality may be more appropriate."],"supporting_citations":[{"why":"Supplies the vector autoregressive framework whose conditions the transformed feature series must satisfy for valid Granger inference.","marker":"(Lütkepohl, 2005)"},{"why":"Gives the asymptotic complex-Gaussian distribution of DFT coefficients used to argue that fitted (a, λ) are approximately Gaussian and weakly stationary.","marker":"(Brillinger, 1981)"},{"why":"Derives power-law spectra for scale-invariant processes, motivating the A(f) = e^a f^{-λ} model and linking the spectral exponent to autocorrelation.","marker":"(Flandrin, 1989)"},{"why":"Provides the frequency-domain causal decomposition against which the paper positions its method, which instead tests on fitted spectral parameters.","marker":"(Geweke, 1982)"},{"why":"Documents self-organized criticality as a generative mechanism for ubiquitous 1/f and power-law noise in real-world systems.","marker":"(Bak et al., 1987)"},{"why":"Supplies a principal constraint-based baseline and the conditional-independence testing framework used for comparison.","marker":"(Runge et al., 2019)"},{"why":"Defines the Ornstein-Uhlenbeck process used to generate all synthetic benchmarks with power-law spectral characteristics.","marker":"(Uhlenbeck & Ornstein, 1930)"},{"why":"Supplies the AirQuality benchmark and its reference causal graph used for the real-world evaluation.","marker":"(Cheng et al., 2024)"}],"fun_headline_variants":["Power-law spectra keep causal discovery noise-robust","Spectral slopes unmask true causal links in noisy data","Causal discovery strengthened by power-law spectral fitting","Slope-based features resist noise in real-world time series"],"cache_read_input_tokens":58624,"weakest_assumption_plain":"The load-bearing premise is that causal influence always shows up as a measurable, stable change in the spectral slope of the affected series, which is essentially what Theorem 3.1 assumes as its condition that causal dependencies are identifiable from amplitude-spectral dynamics.","fun_headline_variants_meta":{"raw":{"variants":["Power-law spectra keep causal discovery noise-robust","Spectral slopes unmask true causal links in noisy data","Causal discovery strengthened by power-law spectral fitting","Slope-based features resist noise in real-world time series"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000136,"raw_usage":{"total_tokens":1096,"prompt_tokens":844,"completion_tokens":252,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":460,"completion_tokens_details":{"reasoning_tokens":188}},"tokens_in":460,"tokens_out":252,"duration_ms":3624,"temperature":1.0,"reasoning_tokens":188,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:50:19.836168+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate two independent power-law noise series and make y a lagged copy of x, y(t) = x(t − τ), so the two series have identical spectra and hence identical fitted (a, λ) features: a time-domain Granger test would detect the causal lag, while a method operating only on (a, λ) cannot, revealing a boundary of the graph-invariance claim. A second check is to compute the fitted λ for a sum of two power laws over a range of window lengths and measure how large the frequency-dependent correction (log c + a3 − a1)/log f is in practice.","supporting_citations":[],"review_version":1}