{"id":"376e24d2-a482-43fa-b5f5-bd5ac6703b88","arxiv_id":"1908.11736","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Residual GNSS time series are modeled as white noise plus colored noise plus a Levy process, with a heuristic that classifies the Levy component as Gaussian, fractional, or stable.","lead":"This paper proposes that the leftover noise in GPS station position time series can be split into white noise, colored noise, and a third random component that follows a Levy process. The authors use simulations and three real GPS stations to sort that third component into Gaussian, fractional, or stable Levy classes based on how the fitted noise parameters change when the time window is expanded.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The N-step classifier is never validated against simulations with a known Levy component: the supporting simulations inject only white+colored noise, so the three-class claim rests on arbitrary thresholds and an acknowledged noise-model dependence.","rationale":"Read in good faith, the paper proposes an exploratory tool: fit a three-component model (white + colored + Levy) to residual GNSS series and use the stability of estimated functional/stochastic parameters under length extension to label the Levy component. The idea is plausible, and the paper is appropriately hedged in §3.4 and the conclusions, explicitly acknowledging noise-model dependence and the difficulty of separating fractional from stable Levy. However, the central claim in the abstract is stronger than the evidence. The simulations in §3.3.1 are generated from a white+power-law model with no injected Levy process, and the classification rule of Table 1 uses 3%/20% thresholds without calibration. Consequently the labels 'Gaussian Levy', 'Fractional Levy', and 'Stable Levy' are inferred from parameter variation and ARMA/FARIMA fits rather than from any known Levy component. This is not an internal inconsistency, but it is a correctness risk: the same large parameter variations could be produced by colored-noise amplitude, functional-model absorption, or unmodeled offsets, all acknowledged in the text. Since the reader's CONDITIONAL verdict already requires sensitivity analysis and validation, no verdict change is needed; the proposed ground-truth simulation would directly settle whether the classifier identifies Levy type.","tokens_in":17465,"tokens_out":4662,"duration_ms":50854,"concrete_test":"Run the N-step classifier on simulated residual series with a known injected third component: (i) standard Brownian motion (Gaussian Levy), (ii) fractional Levy stable motion with H > 1/α, and (iii) symmetric α-stable Levy motion with α < 2, each added to the same white+power-law noise used in §3.3.1. Repeat over at least 100 realizations, using both PL+WN and FN+WN as estimation models and thresholds varied over, e.g., 1%, 5%, 10%, 15%, 25%, and 30%. If the classifier labels do not match the injected Levy class substantially above chance, or if labels are strongly threshold- or noise-model-dependent, the claimed three-class identification fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's three-class identification rests entirely on the N-step classifier in §3.2/Table 1, but nowhere is that classifier tested against series with a known Levy component. The simulations in §3.3.1 are generated as white noise plus power-law noise only; the 'third r.v.' is not simulated, so the labels Gaussian/fractional/stable Levy are never compared with ground truth. The rule itself uses arbitrary thresholds (3%, 20%) on the percentage variation of estimated parameters, and the text acknowledges two confounds: the results depend strongly on the chosen stochastic noise model (PL+WN vs FN+WN, §3.3.2 and Fig. 2), and part of the noise is absorbed into the functional model (§3.3.1, §3.4). Indeed §3.3.1 states it is 'much more difficult to discriminate between the fractional Levy and the stable Levy.' Thus large parameter variation is not a demonstrated indicator of a stable Levy process; it could equally reflect colored-noise amplitude, unmodeled offsets, or functional-model misfit, as the paper itself concedes. The load-bearing assumption is therefore that percentage variation is a faithful proxy for the Levy type of the third r.v., and this assumption is unvalidated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper models residual GNSS position time series as the sum of three random variables: white noise, coloured noise, and a third variable assumed to belong to the family of Levy processes. To identify the type of the third variable, the manuscript proposes an N-step method in which the length of the time series is extended in steps of 0, 0.3, 0.5, 0.7, 0.8, and 1 year; the percentage variations in estimated functional and stochastic model parameters are then compared with thresholds of 3% and 20%. Table 1 labels the third variable as Gaussian Levy, fractional Levy, or stable Levy according to these thresholds. The method is applied to simulated white-plus-power-law noise time series at three coloured-noise amplitudes and to three real GNSS stations (DRAO, ASCO, ALBH). The paper concludes that fractional Levy processes offer an alternative to fractional Brownian motion and that stable Levy processes appear when heavy-tailed residuals arise from functional-model misfit.","tokens_in":17749,"tokens_out":4140,"duration_ms":40653,"significance":"If the proposed classification were validated, the N-step method would give geodesy practitioners a practical heuristic for selecting among noise models. The paper is clearly structured, connects Levy processes to established geodetic noise models, provides closed-form variance derivations in the appendices, and uses standard software (Hector). However, the central three-class claim is not yet established: the classifier is never tested against time series with a known Levy component, and its thresholds are introduced without justification or sensitivity analysis. As a result, the reported identification of Gaussian, fractional, and stable Levy processes is currently a plausible hypothesis rather than a demonstrated result.","major_comments":[{"comment":"The N-step classifier is never validated against time series with a known Levy component. The simulations in §3.3.1 generate only white noise plus power-law noise (PL+WN); the third random variable is absent by construction. Consequently, the labels 'Gaussian Levy', 'fractional Levy' and 'stable Levy' assigned in Fig. 1 and Tables 2–3 are not compared with ground truth. The statement in §3.3.1 that 'it is much more difficult to discriminate between the fractional Levy and the stable Levy' is an acknowledgement that the load-bearing assumption—that percentage variation of estimated parameters is a faithful proxy for the Levy type—is unvalidated. The authors should simulate series with injected Levy processes (e.g., Brownian motion, fractional Levy stable motion, and alpha-stable motion) and quantify classification accuracy.","section":"§3.3.1 and Table 1"},{"comment":"The classification thresholds of 3% and 20% are introduced without justification or sensitivity analysis. Because the definition of the three classes in Table 1 is itself based on these thresholds, the subsequent identification of Gaussian, fractional, and stable Levy in simulations is partly a restatement of the definitions. The thresholds should be derived from a loss function, calibrated on simulations, or at minimum varied in a sensitivity analysis to show that the conclusions in §3.3 are not threshold-dependent. As it stands, changing these two numbers to, say, 5% and 15% could change the labels assigned to real stations.","section":"§3.2, Table 1"},{"comment":"The reported correlations do not provide evidence for the three-way distinction. For all scenarios, Corr. Normal and Corr. Levy are within mutual uncertainties (e.g., case C, beta=1.1: 0.89±0.50 vs 0.96±0.18; case B, beta=1.1: 0.92±0.21 vs 0.94±0.14), so the residual distribution fits both families almost equally well. The paper's own text concedes the difficulty of separating fractional from stable Levy. Without a statistical test (e.g., a likelihood-ratio test that accounts for the Gaussian case being nested in the alpha-stable family) or classification-error rates, the claim that three classes are identifiable is not supported.","section":"§3.3.1, Table 3"},{"comment":"The classification of the real stations is not robust to the choice of stochastic noise model. The authors state that 'there is a strong dependence with the selected noise model' and that the results depend on whether PL+WN or FN+WN is used. Because the paper selects the noise model a priori using information criteria and does not show that the resulting Levy class is invariant to that choice, the assignment of DRAO and ASCO to the fractional Levy class is conditional on an unverified modeling decision. A sensitivity check over both noise models, and ideally over the Up coordinate, is needed.","section":"§3.3.2, Fig. 2"}],"minor_comments":[{"comment":"The phrase 'imply potential anxiety' is unclear; consider 'may indicate uncertainty' or 'may signal model inadequacy'. The term 'anxiety' is used in several places to mean functional-model misfit, but it is not standard statistical terminology and should be defined or replaced.","section":"Abstract and §3.4"},{"comment":"The symbols ≜, ≃, ∼, and ⁄= are defined by reference to 3% and 20% variation, but the set of parameters over which the variation is computed is not specified, and it is unclear how absolute differences are normalized when parameters have different units. Please state the formula for the percentage variation explicitly.","section":"§3.2"},{"comment":"The text refers to 'Figure 1a, 1b, 1c' while the caption uses (A), (B), and (C); please make the cross-referencing consistent.","section":"§3.3.1"},{"comment":"The derivation assumes uncorrelatedness between the residual trend and the noise, and the authors later state that the resulting sigma^2 should be seen as an upper bound. This caveat should appear at the point of the derivation rather than only after Eq. (9).","section":"§3.4, Eq. (7)"},{"comment":"The approximation sign in Eq. (14) is introduced without showing how the cross terms are handled; a short derivation step for the cross term would improve readability.","section":"Appendix B, Eq. (14)"}],"recommendation":"major_revision","confidential_remarks":"This manuscript is an arXiv preprint from 2019; if this is a resubmission, the authors should clarify what has been updated relative to the original posting. The paper draws heavily on the authors' prior work, which is not itself a problem, but the novelty relative to Montillet and Yu (2015) should be stated explicitly. The main concern for the editor is that the central three-class identification is currently supported only by an unvalidated threshold rule; I believe this is fixable within the manuscript's scope by adding simulations with known Levy components and a threshold sensitivity analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes a decomposition of residual GNSS time series into white noise, colored noise, and a third random variable belonging to the Levy family, plus an N-step length-variation heuristic to classify that third component. That is genuinely new, and the heuristic is practical enough that geodesists may want to try it. The real-data analysis with three stations is a nice touch, and the paper is honest about some limitations: it acknowledges noise-model dependence and the absorption of noise into the functional model. The literature on Levy processes and FARIMA is covered adequately.\n\nThe soft spots are substantial, though. The stress-test note is on target: the classifier is never validated against simulations with a known Levy component. The supporting simulations inject only white plus power-law noise; they do not simulate a Gaussian, fractional, or stable Levy third variable. So the labels in the simulations are never compared with ground truth. The 3% and 20% thresholds in Table 1 are arbitrary, and there is no sensitivity analysis around them. The correlations in Table 3 show the Normal and Levy fits are close in all scenarios, and the authors themselves admit it is much more difficult to discriminate between fractional and stable Levy. The real-station classifications in Section 3.3.2 are therefore speculative, especially since the results change with the chosen noise model (PL+WN vs FN+WN). The abstract overstates what is actually established.\n\nThe variance derivations in Section 3.4 are reasonable but tangential, and they do not rescue the classifier. Still, the paper is coherent on its own terms and shows clear thinking; the limitations are at least partially acknowledged in the discussion.\n\nWho should read this? People working on GNSS noise models who want a framework for thinking about heavy tails, model misspecification, and the limits of Gaussian/FARIMA descriptions. It deserves a serious referee because the question is important and the proposed heuristic is novel, but the paper needs major revision: validate the classifier on simulated Levy processes, add sensitivity analysis for the thresholds, and tone down the abstract. As is, I would not cite it for the claim of three identifiable Levy classes, but I might cite it as an early attempt to broaden the stochastic toolkit for geodetic residuals. My recommendation: send it to peer review, but flag the missing validation as a load-bearing issue.","headline":"The decomposition is new and worth discussing, but the abstract's clean three-class claim outruns the evidence.","tokens_in":18238,"tokens_out":1763,"would_cite":false,"duration_ms":21166,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M10","60G51","60G52"],"pacs":[],"model":"deepseek-v4-flash","headline":"Residual GNSS time series decompose into white noise, colored noise, and one of three Lévy processes, identifiable by a one-year length-variation rule.","keywords":["Lévy processes","mixed spectrum","coloured noise","GNSS time series","fractional Brownian motion","alpha-stable distribution","length-variation method"],"falsifier":"Simulate a residual time series whose third component is a genuine symmetric α-stable Lévy motion with known parameters (e.g., α=1.5), add it to white plus power-law noise, then run the N-step procedure with the PL+WN model and check whether the percentage variations exceed 20% in both the stochastic and functional models; if they do not, the rule mislabels a true stable Lévy process as fractional.","tokens_in":17279,"feed_emoji":"📡","tokens_out":8028,"duration_ms":73970,"temperature":0.7,"pith_summary":"After the usual geophysical signals are removed from daily GNSS position time series, what remains is usually treated as a mixture of white and colored noise. This paper argues that the residual should instead be modeled as a sum of three random variables: white noise, colored noise, and a third component that belongs to the family of Lévy processes. Using simulations and three real stations, the authors claim the third component falls into one of three identifiable classes: Gaussian Lévy (short-memory), fractional Lévy (long-memory and self-similar, an alternative to fractional Brownian motion), or stable Lévy (heavy-tailed, with large or infinite variance). They propose a practical N-step rule based on how estimated parameters change when the time-series length is extended over one year, with thresholds of 3% and 20%. If the rule works, geodesists would have a simple way to choose between ARMA and FARIMA noise models and would be alerted when missing geophysical signals are contaminating the residuals.","feed_headline":"GNSS residuals split into three Lévy noise classes","feed_subtitle":"A one-year length-variation rule tells Gaussian, fractional, and stable cases apart.","key_machinery":"The load-bearing device is the N-step length-variation method. The time series is truncated or extended in fractions of a year (0, 0.3, 0.5, 0.7, 0.8, 1), and at each step the stochastic model parameters (amplitudes of white and colored noise plus power-law index) and the functional model parameters (trend, seasonal harmonics) are re-estimated; the percentage variation between the first and N-th step is then compared with two thresholds. Variations below roughly 3% label the third random variable as Gaussian Lévy, variations between 3% and 20% label it fractional Lévy, and variations above 20% label it stable Lévy. The classification is anchored to the Lévy-stable machinery: the characteristic exponent α and Hurst parameter H determine which ARMA/FARIMA model should be used, and the variance formulas for residual trend and seasonal signals separate the finite-variance cases from the infinite-variance stable case.","core_discovery":"The central claim is that the residual time series obtained after subtracting the deterministic tectonic rate and seasonal signals from GNSS observations is not just white plus colored noise, but contains a third random variable that is a Lévy process. Depending on the memory and tail properties of the residual, this third variable is one of three kinds: a Gaussian Lévy process when short-memory ARMA dynamics dominate, a fractional Lévy process when the series is self-similar with long-range dependence (reducing to fractional Brownian motion at α=2), or a stable Lévy process when heavy tails and an effectively infinite variance point to a misfit between the chosen functional model and the observations. The paper tests this trichotomy on simulated mixed-spectrum time series and on real GNSS stations, reporting that the residual distribution is consistent with a Gaussian or fractional Lévy class for ordinary stations, while the station affected by slow-slip events shows the large parameter variation associated with the stable class.","pith_inferences":["The 3% and 20% thresholds are justified only by simulation experience, not by a formal statistical test; a natural next step is to derive a likelihood-ratio or information-criterion comparison between the three Lévy classes on the same one-year sliding windows.","If the functional model absorbs a substantial part of the colored noise, the N-step rule may systematically over-label fractional Lévy cases, so the method's false-positive rate should be measured on synthetic series where the true third r.v. is a known α-stable motion.","The same three-term decomposition could be useful for other long-memory geophysical or financial records with mixed spectra, such as river discharge or sea-level time series, where the stable-Lévy flag would signal missing forcing terms."],"forward_implications":["A geodesist can classify the residual noise of a real station by re-estimating the noise model on one-year-extended windows and applying the 3%/20% thresholds, without needing additional data types.","When the fractional Lévy class is identified, FARIMA becomes the appropriate residual model, and the Hurst exponent can be read off the fractional differencing parameter d via H = d + 0.5.","When the stable Lévy class is identified, it is a warning that the functional model is missing geophysical signals (or contains undetected offsets), and the residual variance is genuinely unbounded rather than a numerical artifact.","For ordinary GNSS stations whose residuals are Gaussian, the existing white-plus-power-law noise model remains sufficient, and the Lévy process reduces to the Brownian motion case.","The derived variance formulas imply that a residual linear trend grows quadratically with series length, which is why the stable Lévy case must be reserved for genuinely heavy-tailed residuals."],"supporting_citations":[{"why":"Supplies the mixed-spectrum noise model (white plus colored noise) and the simulation framework for GNSS position time series.","marker":"Williams et al. 2004"},{"why":"Provides the estimation software used to fit stochastic and functional model parameters in the N-step method.","marker":"Bos et al. 2013"},{"why":"Supplies the definition and parameterization of stable distributions that underpin the stable Lévy class.","marker":"Nolan 2009"},{"why":"Defines the fractional Lévy stable motion (fLsm), the object the fractional Lévy class is built on.","marker":"Weron et al. 2005"},{"why":"Established Lévy alpha-stable distribution and FARIMA for geodetic residuals, the direct predecessor this paper extends.","marker":"Montillet and Yu 2015"},{"why":"Introduced FARIMA, connecting the fractional parameter d to the Hurst exponent and motivating the ARMA/FARIMA dichotomy.","marker":"Granger and Joyeux 1980"},{"why":"Supplies the maximum-likelihood fitting of ARMA and FARIMA models used to characterize residuals.","marker":"Sowell 1991"},{"why":"Provides the information-criterion method used to select the stochastic noise model (FN+WN vs PL+WN) before applying the N-step classification.","marker":"He et al. 2019"}],"fun_headline_variants":["Lévy trichotomy explains GNSS residual noise","Mixed-spectrum residuals follow three Lévy classes","Lévy process family splits geodetic noise into three types","GNSS residuals: Gaussian, fractional, or stable Lévy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The classification rule depends on the untested premise that percentage changes in the estimated stochastic and functional model parameters over one-year extensions faithfully mirror the true type of the third random variable, with thresholds of 3% and 20%.","fun_headline_variants_meta":{"raw":{"variants":["Lévy trichotomy explains GNSS residual noise","Mixed-spectrum residuals follow three Lévy classes","Lévy process family splits geodetic noise into three types","GNSS residuals: Gaussian, fractional, or stable Lévy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000638,"raw_usage":{"total_tokens":2941,"prompt_tokens":949,"completion_tokens":1992,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":1925}},"tokens_in":565,"tokens_out":1992,"duration_ms":13516,"temperature":1.0,"reasoning_tokens":1925,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:24:22.118239+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a residual time series whose third component is a genuine symmetric α-stable Lévy motion with known parameters (e.g., α=1.5), add it to white plus power-law noise, then run the N-step procedure with the PL+WN model and check whether the percentage variations exceed 20% in both the stochastic and functional models; if they do not, the rule mislabels a true stable Lévy process as fractional.","supporting_citations":[{"cited_title":"(2013) Fast error analysis of continuous GNSS observations with missing data","cited_arxiv_id":null,"evidence_quote":"Provides the estimation software used to fit stochastic and functional model parameters in the N-step method."},{"cited_title":"Book online: http://www.math.ucla.edu/ biskup/275b.1.13w/PDFs/Nolan.pdf","cited_arxiv_id":null,"evidence_quote":"Supplies the definition and parameterization of stable distributions that underpin the stable Lévy class."},{"cited_title":"(2005) Complete description of all self-similar models driven by Levy stable noise, Phys","cited_arxiv_id":null,"evidence_quote":"Defines the fractional Lévy stable motion (fLsm), the object the fractional Lévy class is built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduced FARIMA, connecting the fractional parameter d to the Hurst exponent and motivating the ARMA/FARIMA dichotomy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the maximum-likelihood fitting of ARMA and FARIMA models used to characterize residuals."}],"review_version":1}