{"id":"fb5c9f64-140a-439c-ae9a-4cc5f8bbc8d9","arxiv_id":"2502.00243","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"FPW is a weighted phase-binning periodicity statistic, derived from a Gaussian Process model, that detects periodic signals of arbitrary waveform in irregularly sampled time series.","lead":"This paper introduces FPW, a fast period-finding algorithm that folds unevenly sampled light curves into phase bins and sums weighted bin signals, with CPU and GPU code. The authors say it is being applied to 1.5 billion objects in ZTF data because it detects arbitrary waveform shapes, from smooth waves to sharp eclipses.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Printed C++ code computes half of Eq. 1 (Minv=1/(2*VtCinvV), no 1/alpha^2 denominator term), so the chi-square(M-1) calibration and the 75%/3x sensitivity ratios do not describe the deployed statistic; a Monte Carlo null test is required before those claims can be trusted.","rationale":"The reader's weakest assumption correctly identifies the most load-bearing condition: the implemented statistic's null distribution. I agree that the Appendix D code returns half of Eq. 1 and drops the 1/alpha^2 term, and that even with that discrepancy corrected the chi-square(M-1) assumption is unverified for irregular sampling and the chosen weighting. This is not an objection to the GP derivation itself, which is a reasonable route to a phase-binning statistic; it is an internal mismatch among Eq. 1, Appendix A.4, Appendix B, and the code listing, compounded by the paper's own statement that statistical evaluation is deferred to Paper II. Because all the quantitative sensitivity claims in the abstract and Sections 3.1-3.3 depend on this calibration, the paper needs a corrected code listing and a null-distribution validation section before acceptance. The current CONDITIONAL verdict remains appropriate; no additional adjustment is needed.","tokens_in":17093,"tokens_out":7541,"duration_ms":80279,"concrete_test":"Simulate null light curves with the same irregular sampling, heteroscedastic errors, frequency grid, and bin counts (M=4, 5, 10, 20) used in the paper. For each M, draw at least 1e5 Gaussian realizations, run both the printed C++ deltaChi2 and an exact implementation of Eq. 1 (with and without the 1/alpha^2 term), and compare the empirical distributions and tail probabilities to chi-square_M and chi-square_{M-1}. If the printed-code statistic has mean about (M-1)/2 rather than M-1, the factor-two discrepancy is confirmed; if the exact implementation's tails differ beyond Monte Carlo error, the chi-square(M-1) significance assumption fails and the sensitivity claims in Figures 1 and 4 must be recomputed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claims (75% of Lomb-Scargle significance for sinusoids, 3x significance for narrow eclipses, 2e-9 false-alarm level) all assume that the statistic actually computed is Eq. 1 and that it is chi-square distributed with M-1 degrees of freedom under the null. Both legs are insecure. In Appendix D, deltaChi2 computes Minv = 1.0/(2*VtCinvV[i]) and accumulates ytCinvV[i]*Minv*ytCinvV[i], with no +1/alpha^2 denominator term. This returns exactly half of the alpha->infinity limit of Eq. 1 for every trial frequency. Thus either the code is not the algorithm described by Eq. 1, or Eq. 1 is not what the code implements; in either case the printed code does not reproduce the statistic whose significance is claimed. Independently of that factor, the paper itself states in Section 1 that Paper II will evaluate the statistics of the algorithm, and no Monte Carlo null test appears in the present text. A retained 1/alpha^2 term makes each bin term a scaled chi-square_1 with scale V_m/(V_m+1/alpha^2), so the sum is not chi-square unless that term is negligible; and subtracting a mean estimated from heteroscedastic data does not automatically reduce M degrees of freedom to M-1. These are not cosmetic issues: the false-alarm thresholds, the relative-sensitivity curves in Figures 1 and 4, and the practical claims about ZTF-scale searches all inherit the null distribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Fast Periodicity Weighting (FPW), a phase-binning statistic for detecting periodicity in irregularly sampled time series. For each trial frequency, data are folded into M phase bins and the statistic is the sum over bins of the squared inverse-variance-weighted bin sum, normalized by the sum of inverse variances plus a regularization term 1/alpha^2. The statistic is derived from a Gaussian Process model with a prior on per-bin signal amplitude (Appendix A). The paper claims FPW is waveform-agnostic, achieves ~75% of Lomb-Scargle's significance for pure sinusoids with M=4-5, and about 3x Lomb-Scargle's significance for narrow eclipses with M=20, while running fast enough for billion-source surveys. It includes C++/Python code, benchmarks against astropy Lomb-Scargle, and demonstrates the algorithm on two ZTF sources. The paper defers a full statistical evaluation and comparison to Paper II.","tokens_in":17433,"tokens_out":7882,"duration_ms":71808,"significance":"The algorithm is simple, computationally efficient, and the GP derivation in Appendix A is clean. If the sensitivity claims are correct, FPW would be a practical workhorse for periodicity searches in ZTF and similar surveys. However, the quantitative claims rest on the null distribution of the statistic and on the code implementing Eq. (1), both of which are problematic: the printed C++ code computes a different quantity (a factor of two and no alpha term), and the chi-square(M-1) null distribution is asserted without derivation or Monte Carlo verification. These issues are fixable but are essential to the paper's main conclusions.","major_comments":[{"comment":"The C++ function `deltaChi2` computes `Minv = 1.0/(2*VtCinvV[i])` and accumulates `ytCinvV[i]*Minv*ytCinvV[i]`. This corresponds to a denominator of `2*sum(1/sigma^2)` and omits the `1/alpha^2` term, so the code returns `0.5 * S_FPW` in the alpha->infinity limit of Eq. (1), not Eq. (1) as printed. Consequently, the chi-square(M-1) null distribution invoked in Sections 3.1 and 3.3, the 2e-9 false-alarm threshold, and the relative sensitivity values in Figures 1 and 4 do not apply to the provided code. The authors must reconcile the code with Eq. (1) and recalibrate the statistic's null distribution and sensitivity claims for the actual implementation.","section":"Appendix D, Section 4"},{"comment":"The paper asserts that S_FPW is distributed as chi-square with M-1 degrees of freedom under the null, but this is not derived or tested. The mean subtraction described in Section 3.5 removes a global unweighted mean, while the statistic is built from inverse-variance-weighted sums; the resulting constraint does not automatically reduce the effective degrees of freedom by one. In addition, the finite-alpha form of Eq. (1) makes each bin term a scaled chi-square variable rather than chi-square_1 unless 1/alpha^2 is negligible. The paper should provide either an explicit derivation of the null distribution for the actual pre-processing and weighting, or Monte Carlo simulations demonstrating that the empirical false-alarm rates match the claimed chi-square(M-1) thresholds on realistic ZTF-like sampling.","section":"Section 3.1, Appendix B"},{"comment":"The quantitative claims that FPW achieves ~75% of Lomb-Scargle's significance for sinusoids and ~3x for narrow eclipses are analytic predictions from Appendix B for idealized constant-error data and perfectly aligned phase bins. No simulations or injected-signal tests are presented to validate these ratios, and the two observational examples (Figures 2, 3, 5) are qualitative. Given that the code discrepancy and null-distribution issues affect the calibration, these numbers should be re-derived or replaced with empirical detection-efficiency measurements before the abstract's claims can be accepted.","section":"Sections 3.1 and 3.3, Figures 1 and 4"},{"comment":"The paper states that Paper II will 'evaluate the statistics of the algorithm' and provide quantitative comparisons. Since the present paper relies on specific significance values (e.g., 2e-9 false-alarm level and the 75%/3x ratios) to support its central claims, those statistical evaluations need to be included here, or the claims must be explicitly labeled as preliminary and the corresponding quantitative statements removed.","section":"Section 1, Section 5"}],"minor_comments":[{"comment":"The statement 'For exactly equal number of sources in each bin, Nm/Ntot = -1/M' should read '1/M' (positive).","section":"Section 3.5"},{"comment":"'ilustrate' should be 'illustrate'.","section":"Section 3.3"},{"comment":"The comment says 'Gets the fractional part' but `static_cast<int>(phase)` truncates; for positive inputs this is the floor, which is fine, but the comment could be more precise for negative phase values.","section":"Appendix D, makeIndices"},{"comment":"The benchmark of FPW versus astropy Lomb-Scargle may not be apples-to-apples because FPW is implemented in C++ while the comparison implementation is Python-based; the text should state explicitly what code path is being timed.","section":"Section 4"},{"comment":"The definition of X is implicit; state that X is the threshold corresponding to a tail probability of 2e-9 for chi-square with the relevant DOF.","section":"Appendix B, Eq. (B22)"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claims depend on a statistical calibration that the authors explicitly defer to Paper II. I would recommend requiring the authors to either include a proper null-distribution test and code-statistic consistency check in this paper, or clearly downgrade the quantitative claims to preliminary. The code mismatch in Appendix D is a serious error that will confuse readers; the published version should correct it in both the equation and the code. The editor may also want to verify that the GitHub repository matches the printed code."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely useful methods paper with a real problem. The GP derivation in Appendix A is clean, and the statistic is a sensible weighted phase-binning chi-square. The paper is honest about its lineage – it explicitly states equivalence to Scargle's Bayesian blocks and kinship with epoch folding and the 1924 Whittaker-Robinson method. That is not a flaw; it is good citational hygiene. The multi-M computation trick in Section 3.4 is practically valuable, and the code plus timing benchmark against astropy Lomb-Scargle is a real contribution.\n\nBut the load-bearing quantitative claims – the chi-square(M-1) null distribution, the 75% sinusoid sensitivity, the 3x eclipse sensitivity, the 2e-9 false-alarm thresholds – rest on a statistic that the printed code does not compute. In Appendix D, deltaChi2 sets Minv = 1/(2*VtCinvV[i]) and accumulates ytCinvV[i]*Minv*ytCinvV[i], with no 1/alpha^2 term. That is exactly half of the alpha->infinity limit of Eq. 1. So either the code is wrong or Eq. 1 is not what is being run; in either case the significance calibration as printed does not apply to the code as printed. The authors also say Paper II will evaluate the statistics; there is no Monte Carlo null test in this paper. A retained 1/alpha^2 term changes each bin term's distribution, and subtracting a mean from heteroscedastic data need not reduce the degrees of freedom by exactly one. These are not cosmetic.\n\nThe real-data examples (the 6.94-minute white dwarf and the eclipsing double white dwarf) are suggestive but not a substitute for calibration. The '1.5 billion objects' deployment claim in the abstract is not supported by results in this paper. The paper should either fix the code listing, or state that the code implements a different variant and re-derive the null distribution for that variant.\n\nWho is this for? Practitioners running period searches on survey-scale data, and methods people working on fast periodograms. It deserves a serious referee – the underlying idea is sound and the code is useful – but the referee should demand a corrected code listing and a Monte Carlo null validation (or a clear Paper II pointer) before the quantitative claims are accepted.\n\nMy recommendation: send to peer review, but with major revision expected.","headline":"A useful, honestly-written phase-binning periodicity method whose printed code disagrees with its own Eq. 1, leaving the chi-square calibration and sensitivity ratios unsupported; deserves a serious referee but needs a corrected code listing and Monte Carlo null test before the numbers can be trusted.","tokens_in":18044,"tokens_out":2258,"would_cite":false,"duration_ms":22417,"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":"The paper claims FPW, a Gaussian-process-derived phase-binning statistic, detects arbitrary periodic waveforms nearly as well as Lomb-Scargle on sinusoids, several times better on narrow eclipses, at a cost that scales as frequencies…","keywords":["time series analysis","period search","periodic variable stars","phase binning","Lomb-Scargle periodogram","eclipsing binaries","Zwicky Transient Facility","Gaussian process"],"falsifier":"Generate many Gaussian-noise light curves with the same irregular sampling and per-point error bars as ZTF, run the published FPW code over the same frequency grid, and compare the empirical tail of the statistic with the claimed $\\chi^2_{M-1}$ distribution; a factor-of-two offset in the statistic would show up directly as a gross mismatch in false-alarm rates.","tokens_in":16829,"feed_emoji":"🔭","tokens_out":12590,"duration_ms":105080,"temperature":0.7,"pith_summary":"This paper introduces Fast Periodicity Weighting (FPW), a phase-binning statistic for finding periodic signals in unevenly sampled time series without assuming the shape of the phase-folded waveform. The authors argue that a single FPW periodogram can serve the full mixture of waveforms found in Zwicky Transient Facility data, from nearly sinusoidal rotators to eclipsing white-dwarf binaries. They report relative sensitivities of about 75% of Lomb-Scargle for pure sinusoids and about three times Lomb-Scargle for narrow eclipses, and they show that predictable sub-harmonic peaks can reveal periods shorter than the nominal search cutoff. Because the per-light-curve cost is the number of trial frequencies times the number of observations with a small constant, they argue the method is practical for billion-source surveys.","feed_headline":"FPW period search: 75% of Lomb-Scargle on sinusoids, 3x on eclipses","feed_subtitle":"It stays fast enough to search a billion light curves while catching eclipses Lomb-Scargle misses.","key_machinery":"The central object is the phase-binning statistic of Eq. 1, $$S_{\\rm FPW} = \\sum_{m=1}^M \\frac{\\left(\\sum_{j\\in m} x_{m,j}/\\sigma_{x_{m,j}}^2\\right)^2}{\\sum_{j\\in m} 1/\\sigma_{x_{m,j}}^2 + 1/\\$alpha^{2}$},$$ a sum over $M$ phase bins of the squared inverse-variance-weighted bin mean, regularized by a Gaussian prior of width $\\alpha$ on the signal in each bin. It is derived from a Gaussian Process model in which the periodic signal is one free amplitude per phase bin, and a standard low-rank matrix update reduces the marginal-likelihood difference to these simple per-bin sums. Because the numerator and denominator are accumulated at the finest binning, every divisor $M$ of $M_{\\max}$ can be evaluated from the same sums, so $M=2,4,5,10,20$ come at almost the cost of one bin count. This object carries the argument: all sensitivity and speed claims follow from the form of these accumulated sums.","core_discovery":"The central claim is that the statistic $S_{\\rm FPW}$ defined in Eq. 1 is a valid detection statistic for periodic signals of arbitrary phase-folded shape. For each trial frequency, observations are binned by phase; in each bin, the inverse-variance-weighted sum of mean-subtracted fluxes is squared and normalized by the total inverse-variance, plus a regularization term $1/\\alpha^2$ from a Gaussian prior on the phase-bin amplitudes. The paper derives this quantity as a chi-square difference between a phase-binned periodic model and a constant model, compares its significance with Lomb-Scargle at a fixed false-alarm level, and demonstrates on ZTF light curves that it detects both quasi-sinusoidal rotating white dwarfs and an eclipsing double white dwarf. The paper treats $S_{\\rm FPW}$ as $\\chi^2$ with $M-1$ degrees of freedom under the null, which yields the reported 75% and 3x sensitivity ratios.","pith_inferences":["Extension: the same per-bin weighted sums that feed FPW also feed the BLS statistic, so a combined FPW-plus-BLS periodogram could be evaluated with almost no additional arithmetic, giving both a waveform-agnostic scan and a tuned eclipse detector in one pass.","Extension: because FPW produces sub-harmonic peaks by design, a blind peak-ranking stage must be harmonic-aware; otherwise the true period may be mistaken for one of its sub-harmonics, especially at large $M$.","Extension: if the $\\chi^2_{M-1}$ null calibration is confirmed after correcting the factor-of-two issue in the published code, the waveform-agnostic property makes FPW a plausible building block for multi-band period searches, though the paper defers that step.","Extension: the paper's connection to Bayesian blocks suggests adaptive bin widths chosen per candidate period could improve sensitivity to extremely narrow eclipses beyond the fixed $M=20$ case."],"forward_implications":["A single run with $M_{\\max}=20$ gives periodograms for $M=2,4,5,10,20$ at nearly the cost of one bin count, so a survey can search for sinusoids, intermediate waveforms, and eclipses simultaneously.","Sub-harmonic peaks are expected in FPW periodograms, so a detection at a sub-harmonic can reveal a true period shorter than the shortest period explicitly searched, extending effective frequency coverage by a factor of $M/2$ for even $M$.","For eclipse duty cycles up to about 1/20, FPW's detection distance is about 30% larger than Lomb-Scargle's, implying more eclipsing systems recoverable from the same data.","For pure sinusoids, FPW's significance is about 75% of Lomb-Scargle's, so the maximum detection distance is about 7.5% smaller, a modest cost for waveform generality.","With $\\mathcal{O}(N_{\\rm freqs} N_{\\rm data})$ complexity and a small constant, FPW is faster than a fast Lomb-Scargle implementation for light curves with up to roughly 500 points, making billion-source blind searches feasible."],"supporting_citations":[{"why":"Describes ZTF, the survey whose billion-source, thousand-epoch light curves motivate the need for a fast waveform-agnostic period search.","marker":"Bellm et al. 2019"},{"why":"Provides the Lomb-Scargle periodogram and false-alarm treatment used as the baseline for FPW's relative sensitivity claims.","marker":"VanderPlas 2018"},{"why":"Establishes the chi-square distribution with two degrees of freedom for Lomb-Scargle that anchors the significance comparison.","marker":"Scargle 1982"},{"why":"Gives the fast FFT-based Lomb-Scargle implementation whose complexity and constant factor define the speed benchmark FPW is compared against.","marker":"Press & Rybicki 1989"},{"why":"Defines BLS, the standard box-shaped eclipse search that FPW's eclipse sensitivity is contrasted with.","marker":"Kovacs et al. 2002"},{"why":"Supplies the Bayesian-block log-likelihood that the paper shows is equivalent to the per-bin FPW contribution.","marker":"Scargle et al. 2013"},{"why":"Provides the 6.94-minute rotating white dwarf whose quasi-sinusoidal light curve demonstrates FPW's sub-harmonic behavior.","marker":"Caiazzo et al. 2021"},{"why":"Provides the eclipsing double white dwarf whose 20.3-minute peak FPW detects more cleanly than Lomb-Scargle in the truncated dataset.","marker":"Burdge et al. 2020"}],"fun_headline_variants":["FPW period search: 3x better on eclipses, still catches sinusoids","New period detection algorithm handles arbitrary waveforms fast","FPW: period search that doesn't miss eclipses","Fast periodicity weighting: catches waveforms Lomb-Scargle misses","Period detection for any shape, fast enough for a billion light curves"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything hinges on the assumption that $S_{\\rm FPW}$ follows a $\\chi^2$ distribution with $M-1$ degrees of freedom when the light curve is pure noise, and the published code computes half of Eq. 1, so that calibration has to be verified before the 75% and 3x sensitivity statements can be relied on.","fun_headline_variants_meta":{"raw":{"variants":["FPW period search: 3x better on eclipses, still catches sinusoids","New period detection algorithm handles arbitrary waveforms fast","FPW: period search that doesn't miss eclipses","Fast periodicity weighting: catches waveforms Lomb-Scargle misses","Period detection for any shape, fast enough for a billion light curves"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000185,"raw_usage":{"total_tokens":1331,"prompt_tokens":966,"completion_tokens":365,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":277}},"tokens_in":582,"tokens_out":365,"duration_ms":3766,"temperature":1.0,"reasoning_tokens":277,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T19:41:14.147244+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate many Gaussian-noise light curves with the same irregular sampling and per-point error bars as ZTF, run the published FPW code over the same frequency grid, and compare the empirical tail of the statistic with the claimed $\\chi^2_{M-1}$ distribution; a factor-of-two offset in the statistic would show up directly as a gross mismatch in false-alarm rates.","supporting_citations":[],"review_version":1}