{"id":"4fed1b70-9e0f-47fd-bae9-efed6f287985","arxiv_id":"1909.01725","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"A monotone transformation of a Gaussian process alters its autocorrelation through a distribution-specific function, and the power-law exponent survives when the first-order term of that function dominates.","lead":"This paper analyzes how correlations change when a Gaussian time series is rescaled to follow a different distribution, and uses that analysis to generate non-Gaussian time series with long-range power-law correlations. It is useful for anyone who needs surrogate data for testing signals in physics, finance, or biology.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (36) prescribes the inverse of the fGn power-law spectrum; as written, FFM with H > 0.5 produces high-frequency-dominated noise, not the intended l^{-(2-2H)} correlations.","rationale":"The theoretical derivation in Secs. II-III is internally sound: Eq. (8) is the correct NORTA-type integral, the Hermite expansion in Eq. (17) follows from the Mehler expansion, and Eq. (31) is exact for genuine stationary Gaussian time series. The reader's weakest assumption about finite-N FFM outputs not being exactly bivariate normal is a legitimate but secondary approximation, and the paper's own numerical checks mitigate it. The more serious issue is a concrete self-inconsistency in the FFM prescription: Eq. (36) uses the reciprocal of the spectral exponent required by Eq. (35). Because the entire application section instructs the reader to use this spectrum, the algorithm as written would not generate the claimed power-law correlated Gaussian input, and the subsequent transformation would not produce the advertised output. All displayed results (e.g., Fig. 10 and the IBM example) are consistent with the correct spectral exponent, so the fix is likely a sign change in Eq. (36); nevertheless, the manuscript should not be accepted without that correction and a verification that the figures were produced with the corrected spectrum. Hence the verdict should move from ACCEPT to CONDITIONAL.","tokens_in":25014,"tokens_out":15153,"duration_ms":162250,"concrete_test":"Implement FFM exactly as stated in Eq. (36): S(f_j) = f_j^{2H-1} for j=1,...,N/2, random uniform phases, inverse FFT, with N=2^14 and H=0.85. Estimate the autocorrelation CG(l) and compare with Eq. (34). If CG(1) is negative and the large-l decay exponent is near -1 rather than 2H-2 = -0.3, Eq. (36) is wrong. Then rerun with S(f_j) = f_j^{1-2H} and confirm that the measured CG(l) matches Eq. (35), which would settle that the correction is a sign flip in the written spectrum.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's application rests on FFM generating a Gaussian series with autocorrelation CG(l) as in Eqs. (34)-(35). For H > 0.5, CG(l) is positive and decays as l^{2H-2}; by the Wiener-Khinchin theorem, the low-frequency power spectrum of such a process must scale as f^{1-2H}. Equation (36), however, states S(f_j) proportional to f_j^{2H-1}, which is the reciprocal exponent. For H=0.85 this gives S(f) proportional to f^{0.7}, i.e., high frequencies dominate and the low-frequency divergence that produces long-range power-law correlations is absent. A reader implementing Eq. (36) literally would obtain a noise with roughly white or anti-correlated high-frequency character, not the l^{-0.3} decay shown in Fig. 10. The empirical figures appear consistent with the correct sign (1-2H), suggesting a typographical sign error rather than a conceptual failure, but as written the proposed generalization of FFM is not reproducible. This is load-bearing because the algorithm in Application I is the central practical deliverable.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies how the Pearson correlation between two Gaussian variables changes when both are transformed by the inverse-CDF map z = F^{-1}(Phi(z_G)) to a destination marginal distribution. It derives the pair-correlation function C(C_G) from the bivariate Gaussian integral (Eq. 8), obtains a power-series expansion with coefficients given by one-dimensional integrals (Eqs. 16-17), and establishes general properties such as oddness for symmetric marginals and the existence of a negative lower bound C_min for non-symmetric marginals. Closed forms are given for the uniform and lognormal cases, and numerical results are presented for logistic, Laplace, arcsine, symmetric Pareto, exponential, Weibull, and Pareto distributions. The paper then extends the pair result to time series via C(l) = C(C_G(l)), and uses this to propose a generalization of the Fourier Filtering Method (FFM) for generating power-law correlated time series with an arbitrary marginal distribution. The method is illustrated by synthesizing a series that mimics the marginal distribution and autocorrelation power-law tail of IBM daily absolute returns.","tokens_in":25250,"tokens_out":9942,"duration_ms":111707,"significance":"The central theoretical result is sound and useful: the relation C(l) = C(C_G(l)) is derived without fitting to the target autocorrelation, and the expansion in Hermite polynomials gives a principled, parameter-free way to quantify how different marginals attenuate or distort Gaussian correlations. The closed-form uniform and lognormal results are clean, and the numerical checks in Figs. 7 and 10 support the analytical formulas. The systematic comparison across symmetric, bounded, and heavy-tailed distributions is a useful practical guide. The proposed FFM generalization is potentially valuable, provided the algorithmic spectrum is stated correctly. The main weakness is that the FFM application rests on approximate finite-N Gaussianity, and, more seriously, the spectrum written in Eq. (36) has the wrong exponent and is not reproducible as stated.","major_comments":[{"comment":"The exponent in the FFM spectrum is inverted. For a fractional Gaussian noise with autocorrelation (34)-(35), C_G(l) ~ H(2H-1) l^{2H-2}, so by the Wiener-Khinchin theorem the low-frequency spectrum must scale as S(f) ~ f^{1-2H}. Equation (36) instead prescribes S(f_j) proportional to f_j^{2H-1}. For H=0.85 this gives S(f) ~ f^{0.7}, which is high-frequency-dominated and would not produce the l^{-0.3} decay reported in Fig. 10; a reader implementing Eq. (36) literally would not obtain a long-range correlated series. The figures suggest the correct sign was used in the computations, but as written the algorithm is not reproducible. This is load-bearing because the generalized FFM is the central practical deliverable. Please correct the exponent and the sentence following Eq. (36), and re-verify the resulting spectra.","section":"IV.A, Eq. (36)"},{"comment":"The identity C(l) = C(C_G(l)) requires that for every lag l the pair (z_G,i, z_G,i+l) is jointly bivariate normal with correlation C_G(l). This holds exactly for Gaussian processes such as AR(1) in stationarity and for true fGn, but FFM outputs are finite Fourier series with deterministic amplitudes and random phases, so they are only approximately Gaussian for finite N. The paper currently states without qualification that FFM outputs are Gaussian (Section IV.A, step 3). Please state this approximation explicitly and either add a quantitative check of the joint Gaussianity of FFM pairs or cite the standard large-N justification, since Eq. (31) is used to interpret all FFM-based examples, including Figs. 7 and 10.","section":"IV, Eq. (31)"}],"minor_comments":[{"comment":"The proportionality 'sign(H)' is a typo: H is positive throughout, so sign(H)=1. The intended sign is that of (2H-1), which is negative for H<1/2. Please replace sign(H) with sign(2H-1) or simply keep the explicit prefactor H(2H-1).","section":"IV.A, Eq. (35)"},{"comment":"The Hurst exponent H=0.87 is estimated from the same IBM autocorrelation whose power-law tail is then reproduced by the synthetic series. The agreement in Fig. 12 is therefore a self-consistency check, not an out-of-sample validation. Please state this explicitly in the text so that readers do not overinterpret the match.","section":"IV.B, Application II"},{"comment":"The derivative identity in Eq. (12) is correct but is introduced as 'not difficult to check.' Since it is the key to the whole expansion, a one-line reference to the Mehler expansion of the bivariate Gaussian density would make the derivation more transparent.","section":"II.A, Eq. (12)"},{"comment":"The noise level 2/sqrt(N-l) is a white-noise heuristic. For long-memory processes the fluctuations of autocorrelation estimates depend on H. The paper uses this only as a practical criterion, so this is not blocking, but acknowledging the heuristic nature would improve rigor.","section":"IV.A, Eqs. (38)-(39)"},{"comment":"The abstract and introduction promise a study of 'arbitrary' distributions, but the actual analysis is confined to a selected family of parametric distributions plus one empirical CDF. This is fine, but the scope should be worded as 'a broad class of distributions' rather than literally arbitrary.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":"The Eq. (36) sign error is the main technical blocker; once corrected and the FFM section rewritten, the paper is likely publishable. The novelty is moderate—the pair-correlation integral is essentially the NORTA/copula construction—but the systematic treatment of tail and symmetry effects and the connection to power-law time series generation provide sufficient added value for this journal. I would encourage the editor to ask for a careful revision of the algorithmic section and an explicit statement of the finite-N Gaussianity assumption."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work with surrogate time series. The core result, C(l) = C(C_G(l)), is the standard Gaussian-copula/NORTA relation, and the authors properly cite Li-Hammond, Cario-Nelson, Chen, and Kugiumtzis rather than claiming it new. What is actually new is the systematic catalog of C(C_G) across symmetric and non-symmetric, bounded and unbounded distributions, plus the b1-based criterion in Eq. (39) for when power-law autocorrelation survives the transform and stays above the 2/sqrt(N) noise floor. That criterion is practical and, as far as I can check, correct. Fig. 7 is a nice validation that AR(1) and FFM series stay on the theoretical C(C_G) curve, and the uniform and lognormal cases are solved in closed form. The derivations in Eqs. (13)-(17) are clean.\n\nThe soft spots are real but manageable. Eq. (31) requires each pair of Gaussian variables to be jointly bivariate normal; this is exact for fGn and AR(1), only approximate for finite-N FFM output, and the paper does not push hard on that approximation. The IBM demonstration is a self-consistency check rather than an independent prediction, because the empirical CDF and the Hurst exponent are estimated from the same data whose autocorrelation is then matched. No code or data are shipped. None of these is fatal.\n\nOne thing is more serious. Eq. (36) states S(f_j) proportional to f_j^(2H-1). For H > 0.5 that gives a spectrum increasing with frequency, hence high-frequency-dominated noise, whereas fGn with H > 0.5 must have S(f) proportional to f^(1-2H), diverging at low frequency. The power-law tail in Eq. (35) and Wiener-Khinchin force the negative exponent. So a reader implementing the algorithm literally will not obtain the l^-0.3 decay shown in Fig. 10. The figures look correct, suggesting the authors used the right spectrum and the equation has a typo, but as written the central recipe is not reproducible. The stress-test note holds up. A referee should require the exponent be fixed, and ideally a short code snippet or pseudocode with the correct sign.\n\nBottom line: the theory and the b1 criterion are worth publishing, and the paper deserves serious peer review. I would send it to review, but with a required fix to Eq. (36) before acceptance.","headline":"Honest, well-grounded treatment of how Gaussian autocorrelations transform under marginal changes, but the FFM spectrum in Eq. (36) is typo-level wrong and must be fixed before the main algorithm is reproducible.","tokens_in":738,"tokens_out":852,"would_cite":false,"duration_ms":40634,"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":"Gaussian-to-any-distribution transforms preserve power-law autocorrelations when the mapping is nearly linear.","keywords":["Gaussian transformation","autocorrelation function","power-law correlations","Fourier Filtering Method","marginal distribution","Hurst exponent","long-range dependence","time series synthesis"],"falsifier":"Generate a long Gaussian AR(1) series with negative coefficient, transform it with a heavy-tailed lognormal marginal ($s=2$), and compare the empirical negative-lag autocorrelation $C(\\ell)$ with the analytic prediction $C(C_G(\\ell))$ from Eq. (29); a systematic deviation beyond the $2/\\sqrt{N-\\ell}$ noise level at lags where both are significant would show that the bivariate-normal pair assumption fails and Eq. (31) is not universal. A direct check is to estimate the joint density of each lagged pair and test whether, after standardizing the marginals, it is bivariate Gaussian.","tokens_in":24848,"feed_emoji":"📈","tokens_out":10276,"duration_ms":92221,"temperature":0.7,"pith_summary":"This paper establishes a transfer rule for correlations under monotone transformations: if a Gaussian time series with autocorrelation $C_G(\\ell)$ is mapped to any new marginal distribution by $z_i=F^{-1}(\\Phi(z_{G,i}))$, then the new autocorrelation is exactly $C(\\ell)=C(C_G(\\ell))$, where the function $C(C_G)$ is fixed by the destination distribution through a two-dimensional integral. It matters because standard generators such as the Fourier Filtering Method—a spectral method that imprints a prescribed power-law spectrum on Gaussian noise—produce Gaussian long-range correlated data, whereas real records are often non-Gaussian. The paper shows when and how the power-law exponent of the Gaussian series survives the transform: for symmetric destinations with bounded or fast-decaying tails, $C(C_G)$ is nearly linear, so $C(\\ell)\\simeq b_1 C_G(\\ell)$ and the Hurst exponent (the exponent controlling the power-law decay) is preserved. Heavy-tailed or strongly skewed destinations compress and distort correlations, and negative correlations may become unreachable. The paper closes with a practical prescription for producing synthetic series of arbitrary distribution with a controlled power-law autocorrelation tail, demonstrated on absolute stock returns.","feed_headline":"Same power-law exponent survives any distribution change","feed_subtitle":"A near-linear correlation map lets synthetic long-range series of any shape inherit the Hurst exponent.","key_machinery":"The load-bearing object is the function $C(C_G)$, the Pearson correlation of two variables obtained by sending Gaussian variables with correlation $C_G$ through $F^{-1}(\\Phi(\\cdot))$, where $F$ is the destination cumulative distribution. It is computed from the double integral in Eq. (8); when no closed form exists, the paper expands it as $C(C_G)=\\sum_{n=1}^\\infty b_n C_G^n$, with $b_n=(1/(n!\\, \\sigma^2))\\left(\\int F^{-1}(\\Phi(x)) H_n(x)\\varphi(x)\\,dx\\right)^2$, where $H_n$ are Hermite polynomials and $\\sigma^2$ is the destination variance. For symmetric marginals the even coefficients vanish, leaving an odd series; $b_1$ is about 0.90-0.99 for uniform, logistic, Laplace, and arcsine distributions and decreases systematically as distribution tails lengthen. The identity $C(\\ell)=C(C_G(\\ell))$ turns this function into a transfer map between Gaussian and final autocorrelations, which is what makes the generalized Fourier Filtering Method possible.","core_discovery":"The central discovery is that the transformation $z_i=F^{-1}(\\Phi(z_{G,i}))$ converts Gaussian correlations through a deterministic, distribution-specific function: $C(\\ell)=C(C_G(\\ell))$, with $C(C_G)$ given by the double integral in Eq. (8) over the bivariate Gaussian density. The paper proves general properties: $C(0)=0$, $C(1)=1$, $C(C_G)$ is non-decreasing, and for symmetric destination marginals it is odd, so the feasible correlation range is $(-1,1)$; non-symmetric marginals have $C_{\\min}$ in $(-1,0)$, and heavier tails push $C_{\\min}$ toward $0$. It expands $C(C_G)=\\sum_{n=1}^\\infty b_n C_G^n$ with all coefficients nonnegative, and shows $b_1$ controls linearity: it is close to one for bounded, symmetric, and short-tailed distributions and shrinks toward zero as tails lengthen. Consequently, at sufficiently large lags, or whenever $b_1$ is large, $C(\\ell)\\simeq b_1 C_G(\\ell)$, so a power-law $C_G(\\ell)$ yields a power-law $C(\\ell)$ with the same Hurst exponent. The paper verifies this numerically on AR(1) (first-order autoregressive) and Fourier-filtered Gaussian series, then uses it to synthesize a series matching both the empirical marginal distribution and the power-law autocorrelation tail of a real financial volatility series.","pith_inferences":["A diagnostic the authors leave implicit: for any observed non-Gaussian long-memory series, the empirical autocorrelation should collapse onto $C(C_G(\\ell))$ for some Gaussian $C_G(\\ell)$; checking this collapse would test whether the data are consistent with a monotone Gaussian-copula model.","Because the transfer function depends only on the destination marginal, the recipe extends beyond the named distributions to any empirical distribution with a numerically computed inverse CDF; the financial example shows the route, and biomedical or climate records would be natural further tests.","The $b_n$ expansion predicts a quantitative rule: two destination marginals with the same $b_1$ value should show the same observable scaling range regardless of their higher-order coefficients, a comparison that could be made with heavy-tailed marginals matched in $b_1$ but differing in skewness.","Since the transform is monotone, rank correlations and mutual information between lagged values are unchanged, so the framework could be paired with rank-based dependence measures to separate the effect of the marginal from the effect of the copula in long-memory data."],"forward_implications":["The Fourier Filtering Method can be generalized: generate a Gaussian long-range correlated series, then apply $z_i=F^{-1}(\\Phi(z_{G,i}))$ to obtain any target marginal; when $b_1$ is large the final series keeps the same Hurst exponent.","The observable lag range of power-law behavior is set by solving $b_1 C_G(\\ell_{\\max})=2/\\sqrt{N-\\ell_{\\max}}$: long series or high $b_1$ give wide observable scaling, while short series with low $b_1$ may show no power-law at all.","For non-symmetric destination distributions, not every autocorrelation is realizable: the lower bound is $C_{\\min}$, which approaches 0 as tails become heavier, so strongly skewed long-tailed marginals are effectively incompatible with anticorrelated time series.","Because symmetric marginals have only odd terms in the expansion of $C(C_G)$, the linear approximation holds to larger values of $|C_G|$, making symmetric destinations the safer choice for synthesis.","The recipe reproduces both the empirical marginal distribution and the power-law autocorrelation tail of observed absolute-return series, with discrepancies confined to small lags."],"supporting_citations":[{"why":"It supplies the inverse cumulative distribution technique used to map each Gaussian variable to the destination marginal.","marker":"[22]"},{"why":"It establishes that strictly increasing transformations preserve rank dependence, framing why only linear correlation changes.","marker":"[23]"},{"why":"It provides the Fourier Filtering Method that generates the Gaussian power-law correlated series used as the starting point.","marker":"[15, 16]"},{"why":"It first considered the numerical solution of the integral in Eq. (8) for generating variables with prescribed marginals and correlations.","marker":"[25]"},{"why":"It extends the correlation-transformation approach to random vectors with specified marginals.","marker":"[26]"},{"why":"It supplies the autoregressive-to-anything construction and feasibility results, including negative-correlation limits for non-symmetric marginals.","marker":"[29]"},{"why":"It previously discussed the validity of the time-series relation $C(\\ell)=C(C_G(\\ell))$ in the surrogate-data context.","marker":"[19]"},{"why":"It gives the fractional Gaussian noise autocorrelation function and the $2/\\sqrt{N}$ noise level used to define observable scaling lags.","marker":"[33]"},{"why":"It documents power-law autocorrelated stochastic processes for absolute returns, the real-world application target.","marker":"[46]"}],"fun_headline_variants":["Power-law correlations survive any distribution transform","Arbitrary distributions, same Hurst exponent","Transform Gaussian noise, keep power-law scaling","Any distribution, same long-range memory"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole calculation assumes that every pair $(z_{G,i}, z_{G,i+\\ell})$ of the Gaussian series is jointly bivariate normal with correlation $C_G(\\ell)$; this is exact for fractional Gaussian noise (the Gaussian long-memory process) and AR(1) processes but only approximate for finite-length Fourier-filtered outputs, and if the underlying dependence between lagged values is not the Gaussian copula, $C(\\ell)$ is not a function of $C_G(\\ell)$ alone.","fun_headline_variants_meta":{"raw":{"variants":["Power-law correlations survive any distribution transform","Arbitrary distributions, same Hurst exponent","Transform Gaussian noise, keep power-law scaling","Any distribution, same long-range memory"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000655,"raw_usage":{"total_tokens":3094,"prompt_tokens":1130,"completion_tokens":1964,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":746,"completion_tokens_details":{"reasoning_tokens":1910}},"tokens_in":746,"tokens_out":1964,"duration_ms":15071,"temperature":1.0,"reasoning_tokens":1910,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:10:03.214731+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a long Gaussian AR(1) series with negative coefficient, transform it with a heavy-tailed lognormal marginal ($s=2$), and compare the empirical negative-lag autocorrelation $C(\\ell)$ with the analytic prediction $C(C_G(\\ell))$ from Eq. (29); a systematic deviation beyond the $2/\\sqrt{N-\\ell}$ noise level at lags where both are significant would show that the bivariate-normal pair assumption fails and Eq. (31) is not universal. A direct check is to estimate the joint density of each lagged pair and test whether, after standardizing the marginals, it is bivariate Gaussian.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the inverse cumulative distribution technique used to map each Gaussian variable to the destination marginal."},{"cited_title":"Bernaola-Galvan, J","cited_arxiv_id":null,"evidence_quote":"It establishes that strictly increasing transformations preserve rank dependence, framing why only linear correlation changes."},{"cited_title":"Schreiber and A","cited_arxiv_id":null,"evidence_quote":"It first considered the numerical solution of the integral in Eq. (8) for generating variables with prescribed marginals and correlations."},{"cited_title":"Kugiumtzis","cited_arxiv_id":null,"evidence_quote":"It extends the correlation-transformation approach to random vectors with specified marginals."},{"cited_title":"Press, S.A","cited_arxiv_id":null,"evidence_quote":"It supplies the autoregressive-to-anything construction and feasibility results, including negative-correlation limits for non-symmetric marginals."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It previously discussed the validity of the time-series relation $C(\\ell)=C(C_G(\\ell))$ in the surrogate-data context."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It gives the fractional Gaussian noise autocorrelation function and the $2/\\sqrt{N}$ noise level used to define observable scaling lags."},{"cited_title":"Carpena, M","cited_arxiv_id":null,"evidence_quote":"It documents power-law autocorrelated stochastic processes for absolute returns, the real-world application target."}],"review_version":1}