{"id":"902296c5-a0e2-42fe-928a-b5ec9d0ad9c3","arxiv_id":"2505.23413","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A fast 2-D Gaussian process method applied to Jodrell Bank pulsar data finds significant profile variations in 21 pulsars and new spin-down correlations in 7.","lead":"By flattening pulse profiles into a single periodic time axis and applying a fast Gaussian process approximation, the authors track small pulse-shape changes in 21 of 26 pulsars. They report spin-down-correlated shape changes in seven pulsars for the first time, and show one pulsar, B0740-28, switching between correlated and inverted states.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 2-D GP is actually a 1-D kernel on a flattened coordinate, entangling phase and time lags; this misspecification plus post-hoc eigenprofile selection undermines the significance of the new detections.","rationale":"The reader's weakest assumption—that the GP reconstruction, including its formal covariance, is faithful—is correct and important, but the paper contains a more specific and deeper problem: the flattened-coordinate kernel is not a true 2-D product kernel. The mathematics in §2.2 shows that the covariance between two profile points depends on the sum of phase and time differences rather than on their product, because the time kernel is evaluated at δφ + δT. This is a form of model misspecification that can produce artificial phase-dependent time smoothing, exactly matching the admitted excess noise at profile edges. Because the paper uses the formal GP covariance to set significance masks and to generate the 1000 realisations that yield the r uncertainties, the claimed significances inherit this misspecification. In addition, the post-hoc selection of eigenprofiles for B1826-17 (fourth) and B0611+22/B0919+06/J2043+2740 (sum of first two) without multiple-comparison control means the reported correlations are not robust. The paper does have real strengths: the software and data are released, 11 previously known correlations are reproduced, and the 1-D vs 2-D comparison in Figures 28–29 gives some qualitative reassurance. However, these do not settle whether the new detections are spurious, because the known correlations are strong and might survive the kernel artefact while weak new ones do not. A direct separable-kernel comparison, or at least a permutation test on the reported correlations, would resolve whether the central claim holds. Until then, the results are plausible but not established; the appropriate verdict is conditional acceptance pending this check.","tokens_in":25725,"tokens_out":10844,"duration_ms":105689,"concrete_test":"Implement a true separable 2-D GP with the same Celerite approximation: use coordinates (φ, T) with kernel K((φ,T),(φ',T')) = Θ(φ-φ')·τ(T-T'), where Θ is the 1-periodic Gaussian and τ is the Matérn-3/2 in days; this can be written as a Kronecker product of two Celerite kernels and evaluated with standard 2-D GP methods. Run this model on the seven pulsars claimed as new detections (B0105+65, B0611+22, B0626+24, B1740-03, B1826-17, B1917+00, B2148+63) and compare the GP difference maps, eigenprofiles, and Pearson r values with the published flattened-coordinate results. If the correlation coefficients or the apparent significance change by more than the quoted 1σ errors, the flattened kernel is biasing the central claim.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim—21 pulsars with significant profile variations and 7 new spin-down correlations—depends on the GP reconstruction being a faithful model of intrinsic profile variation. That assumption fails in a specific way: the kernel in §2.2 is not a genuine 2-D product kernel. The coordinate flattening x = φ/Δφ_onpulse + T/days means the kernel k(δx) = Θ(δx)τ(δx) is a function of the single lag δx = δφ + δT (with Δφ_onpulse normalized to 1 and δT an integer). Because Θ is 1-periodic in x, the phase part becomes Θ(δφ), but the time part becomes τ(δφ + δT), so the covariance is Θ(δφ)·τ(δφ + δT). This entangles phase and time: for two points at the same phase, the time correlation is τ(δT); for points separated by phase δφ, the effective time lag is increased by δφ, so different phase bins are smoothed with different time length scales. This is not the outer product of a phase-only and a time-only kernel, despite the §2.2 claim. The artifact appears as excess noise at the profile edges—exactly what §4 admits—and the formal GP covariance used for significance and for the 1000-realisation error on r (§2.4) does not account for this model misspecification. Consequently, the significance of the 21 detections and the 7 new correlations (especially B0626+24 r=0.13±0.14 and B1740-03 r=0.14±0.10, plus the post-hoc eigenprofile choices for B1826-17 and B0611+22) is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents psrcelery, an efficient approximation to a 2-D Gaussian process for pulsar profile stacks, implemented by flattening the phase/time grid into a periodic 1-D coordinate and using Celerite. The GP reconstruction is followed by PCA to extract eigenprofiles and their time series, which are then correlated with independently derived spin-down-rate time series. Applied to 26 Jodrell Bank pulsars, the authors report significant profile variations in 21 pulsars, confirm spin-down-correlated shape changes in 11 previously known cases, and claim 7 new correlations (PSRs B0105+65, B0611+22, B0626+24, B1740-03, B1826-17, B1917+00, B2148+63). The paper also presents a detailed case study of PSR B0740-28, where the shape-spin-down correlation appears to switch sign between quasi-stable states. The manuscript includes open-source code and a data release.","tokens_in":26105,"tokens_out":6790,"duration_ms":73046,"significance":"If the statistical claims survive scrutiny, this is a valuable contribution: it offers a computationally tractable way to model phase- and time-correlated profile variability, and it would roughly double the number of pulsars with evidence for spin-down-correlated shape changes, strengthening the magnetospheric state-switching interpretation. The 11 confirmations of previously reported correlations provide important external validation, and the direct comparisons with the 1-D GP approach (Figs. 28-29) illustrate genuine gains in capturing phase-coherent features. The open-source package and data availability are strengths. However, the headline numbers rest on the GP covariance model being a faithful description of the intrinsic variability and on post-hoc PCA selections, and these foundations need stronger validation before the new detections can be considered secure.","major_comments":[{"comment":"The kernel in Eq. (3) is not a genuine 2-D product kernel once the flattening of Eq. (1) is used. With x = φ/δφ + T/days, the lag is δx = δφ + δT (with δT an integer), so the covariance is Θ(δφ)·τ(δφ + δT). The time kernel is therefore evaluated at a lag that includes the phase separation: points in different phase bins are smoothed with different effective time length scales, and the covariance is not the outer product of a phase-only kernel and a time-only kernel as claimed in §2.2. The admitted excess noise at the leading and trailing profile edges (§4) is the kind of artifact this phase-time entanglement could produce. Because the 3-σ masks in the figures and the 1000-realisation errors on r (§2.4) use the formal covariance of this misspecified model, the significance of the 21 detections and of the new correlations is not established by the current analysis. I would like to see either a proper 2-D product kernel or a simulation study (injecting known profile variations and recovering them) demonstrating that the flattened-kernel approximation does not bias the detections or the reported r uncertainties.","section":"§2.1-2.2, Eq. (1) and Eq. (3)"},{"comment":"The paper never defines what makes a detection of profile-shape variation 'significant'. The reconstructed maps are shown with 3-σ masks based on the formal GP uncertainty, but §4 cautions that 'this formal error is likely underestimated'. Without a quantitative detection criterion, such as a likelihood-ratio test against a constant-profile model or a cross-validated prediction score, the statement that 21 of 26 pulsars show significant variations is not falsifiable from the text. Please add a defined threshold and report which pulsars pass it, ideally using a null-hypothesis test that does not rely solely on the GP's own covariance.","section":"§4, Table 1"},{"comment":"The Pearson r uncertainties are computed from realisations of the maximum-likelihood GP model, so they inherit any model misspecification and do not include systematic errors from template alignment, scaling, backend changes, or interstellar scintillation. Several of the newly claimed correlations are individually consistent with zero at the quoted level: B0626+24 has r=0.13±0.14, B1740-03 has r=0.14±0.10, and B2148+63 has r=0.37±0.18 and is described in §4.1.11 as 'barely significant' with 'firm conclusions hard to draw'. In addition, B1917+00 has r=0.73±0.19, but §4.2.9 notes that this is 'surprisingly strong relative to the impression of the plot' and is partly driven by two early observations, with r≈0.4 after excluding them. At minimum these three or four sources should be reclassified as candidates, or the uncertainties should be propagated through an explicit model of systematic errors.","section":"§2.4 and Table 1"},{"comment":"The eigenprofile used for the correlation is selected post hoc: PSR B1826−17 uses the fourth eigenprofile, and PSRs B0611+22, B0919+06 and J2043+2740 use the sum of the first two eigenprofiles. These choices are made after inspecting the correlation with ν-dot, so the reported r values are maxima over a set of candidate shape parameters. The 1000-realisation error in §2.4 does not account for this selection. Please report the correlations for all eigenprofiles, apply a multiple-comparison correction, or define the shape parameter independently of the ν-dot comparison.","section":"Table 1 footnote and §4.1.2, §4.1.8, §4.2.3"}],"minor_comments":[{"comment":"'Principle component analysis' should be 'principal component analysis'; the same typo appears elsewhere in the text.","section":"§2.4"},{"comment":"'Eignenvalues' should be 'eigenvalues'.","section":"§4.1.1"},{"comment":"'Busrts' should be 'bursts'.","section":"§4.2.3"},{"comment":"There are pulsar-name inconsistencies: Figure 4 is captioned 'PSR B1740−28' where the text and Table 1 refer to B0740−28, and §4.2.7 refers to 'PSR B1824+14' where the heading and Table 1 give B1842+14.","section":"Figure 4 caption and §4.2.7"},{"comment":"The Fourier coefficient θ_n = √(2πσ²) exp(−2πσ² n²) appears to lack a factor of π in the exponent for the standard Fourier transform of a Gaussian under ω0=2π; please verify the coefficient convention or state the normalization used.","section":"Eq. (10)"},{"comment":"The caption says the colour map fades to white where the uncertainty 'drops below 3-σ'; this seems likely to be 'exceeds 3-σ' and should be clarified.","section":"Figure 3 caption"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope and the authors are clearly expert in the observational side. The main risk is that the headline claims outpace the statistical validation: the kernel misspecification, the undefined detection threshold, and the post-hoc eigenprofile selection all affect the central numbers. I recommend major revision rather than rejection because the confirmations and the method are sound enough to be salvageable, but I would ask for simulation-based calibration of the GP significances and a conservative reclassification of the marginal new correlations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful methods paper with a genuinely new computational trick, and the 11 confirmations give the approach real credibility. The seven new spin-down correlations are plausible but not all solid; the paper needs a firmer significance criterion before those go into the literature as detections.\n\nWhat's new: the flattened periodic Celerite construction in Sections 2.2–2.3 is a clever way to approximate a 2-D GP at linear cost, and the psrcelery package is released with data. The method demonstrably recovers the known correlations from L10/S22/N22/B24, which is strong external validation. The B0740-28 analysis, with the apparent inversion of the profile-shape/nu-dot relation between quasi-stable states, is the most interesting physical result in the paper.\n\nWhere I'd push back: the kernel is not actually a 2-D outer product in the original (phase, time) space; it's a 1-D kernel on a flattened coordinate, with the time kernel evaluated at delta-phi+delta-T. The paper's wording in Section 2.2 oversells it as an outer product. That's a modeling choice, and the method still works, but the covariance is not what you'd get from a separable 2-D GP, and the formal errors likely miss this misspecification. More importantly, the paper openly says the GP formal error is 'likely underestimated' yet uses that formal error to set the transparency masks and to compute errors on r via 1000 realizations. So the quoted uncertainties on the Pearson coefficients are probably too small. Add the post-hoc selection of eigenprofiles (fourth eigenprofile for B1826-17, sums of first two for B0611+22, B0919+06, J2043+2740) with no multiple-comparison correction, and the weakest new detections — B0626+24 at r=0.13±0.14, B1740-03 at r=0.14±0.10 — aren't significant on their own. B1917+00 drops from r=0.73 to ~0.4 when you exclude two early points.\n\nBottom line: the 11 confirmations show the method is sound in aggregate; the seven new claims need a proper significance test that accounts for eigenprofile selection and the underestimated GP variance. The B0740-28 state-switching result is worth reading carefully, and the psrcelery tool is a real contribution. I'd send this to review, but the referee should ask for a quantitative detection threshold, a multiple-comparison correction, and a simple null test (e.g., shuffle nu-dot or use only the first eigenprofile) before accepting the new correlations as established.","headline":"A clever and well-validated GP method for pulsar profile variability, but the seven new spin-down correlations are not all on solid statistical footing as presented.","tokens_in":26690,"tokens_out":5130,"would_cite":true,"duration_ms":51066,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper builds a fast 2-D Gaussian-process model of pulsar pulse profiles over time and phase, detects significant shape variations in 21 of 26 pulsars, and reports spin-down-correlated changes in 7 pulsars for the first time.","keywords":["pulsars","pulse profile variations","Gaussian process regression","spin-down rate","principal component analysis","magnetospheric state switching","radio pulsars","Celerite"],"falsifier":"Run the pipeline on simulated pulse stacks with synthetic profile variations of the claimed amplitudes (1-2 percent of peak flux) and timescales injected into the real noise, and count how often the model reports significant variations and spin-down correlations; a false-positive rate above the nominal threshold, or Monte Carlo Pearson-$r$ errors wider than the reported $\\pm 0.1$ to $\\pm 0.2$ values, would show the formal covariance is too optimistic and weaken the new detections.","tokens_in":25532,"feed_emoji":"📡","tokens_out":11277,"duration_ms":104911,"temperature":0.7,"pith_summary":"Pulsar radio profiles are usually stable, but a small population shows time-correlated shape changes, a handful of which track the star's spin-down rate. This paper builds a fast two-dimensional Gaussian-process model that treats the pulse profile as a map over time and rotational phase, improving on earlier one-dimensional approaches that ignored phase correlations. Applied to 26 pulsars with long monitoring datasets, the model detects significant profile variations in 21, confirms spin-down-correlated changes in 11 previously known pulsars, and reports evidence for 7 new spin-down-correlated cases. For one pulsar, B0740-28, the correlation between profile shape and spin-down switches between quasi-stable states, and the profile shape associated with greater spin-down sometimes inverts, which complicates simple magnetospheric explanations.","feed_headline":"7 pulsars show spin-down-linked profile changes for the first time","feed_subtitle":"A fast 2-D Gaussian-process model detects 21 variable pulsars, 18 with shapes tied to spin-down rate.","key_machinery":"The central object is the 2-D Gaussian-process kernel $k(\\delta x)=\\Theta(\\delta x)\\,\\tau(\\delta x)$, the outer product of a periodic phase kernel $\\Theta$ (approximated as a Fourier series of a Gaussian) and a time kernel $\\tau$ (exponential or Matérn-3/2), rendered as a Celerite sum of complex-exponential terms, with Celerite being a fast GP solver whose cost grows linearly with data size. By flattening the pulse stack into a periodic 1-D series, the kernel borrows signal across both phase and observation epoch, recovering few-percent component-intensity variations from noise-dominated individual profiles. Principal-component analysis of the GP reconstruction then produces eigenprofiles and their time series, and the Pearson correlation of these with $\\dot{\\nu}$ is the paper's main statistic.","core_discovery":"The central claim is that pulse-profile variability can be separated from noise by a two-dimensional Gaussian process whose kernel is the product of a periodic phase kernel and a time kernel, implemented with the fast Celerite solver so the computation remains feasible. Analysing 26 pulsars with this model, the paper finds significant time-correlated profile shape variations in 21 of them. A principal-component analysis of the reconstructed profiles shows that the first principal component time series correlates with spin-down rate $\\dot{\\nu}$ in 18 pulsars, including 7 first-time detections (B0105+65, B0611+22, B0626+24, B1740-03, B1826-17, B1917+00, and B2148+63). The remaining 5 pulsars show no significant variations. In B0740-28 the correlation is not fixed: sliding-window PCA shows the eigenprofile itself changes with time, at some epochs inverting the shape associated with higher spin-down.","pith_inferences":["Beyond this paper, the separable time-phase kernel construction could be applied to other quasi-periodic signals where correlations are smooth in both dimensions, such as frequency-resolved pulsar spectra or mode-switching analyses at other wavelengths.","The seven new $\\dot{\\nu}$ correlations include Pearson $r$ values as low as 0.13 with errors around 0.1-0.2, and the paper itself warns that formal GP errors are likely underestimated; an independent reanalysis with more conservative noise modelling would test whether the weakest new claims survive.","If B0740-28's sign inversions are physical, then pulsar emission models that derive magnetospheric geometry from a one-to-one relation between profile shape and spin-down rate need revision; a natural next check is whether polarisation position-angle sweeps change alongside the inversions.","A calibration experiment that injects synthetic profile variations of known amplitude and timescale into the same noise would quantify how faint a shape change the 2-D GP can recover, and would put the 21-of-26 detection rate on a firmer statistical footing."],"forward_implications":["The number of pulsars known to have spin-down-correlated profile changes rises from roughly 11 to 18, making magnetospheric state switching appear to be a common rather than exceptional phenomenon.","The 2-D GP plus PCA pipeline can automate profile-variation searches, replacing hand-picked shape parameters and enabling surveys over the full catalogue of hundreds more pulsars.","The recurring 'M'-shaped eigenprofile, with shoulders growing relative to the centre, points toward a change in the relative strength of core and conal radio emission as the underlying physical pattern.","B0740-28 shows that no fixed shape-versus-spin-down mapping holds for all pulsars: the sign of the coupling can invert on timescales of years, so inferences about torque changes from profile data must allow for state changes.","For pulsars whose profile fluctuations are faster than the observing cadence, the measured spin-down rate may correlate with the time-averaged fraction spent in each state rather than with an instantaneous spin-down value, as suggested by B0950+08 and B1822-09."],"supporting_citations":[{"why":"Identified the first six pulsars whose profile shapes track spin-down rate, establishing the phenomenon this paper extends.","marker":"L10"},{"why":"Applied the previous 1-D GP method to the same 17-pulsar sample and found one additional spin-down-correlated case; supplies the comparison baseline and much of the dataset.","marker":"S22"},{"why":"Introduced GP regression for separating underlying pulse-shape variations from noise-dominated profile stacks.","marker":"Karastergiou et al. (2011)"},{"why":"Refined the 1-D GP algorithm and found time-correlated profile variations in five southern pulsars, a direct predecessor of this method.","marker":"Brook et al. (2016)"},{"why":"Provides the Celerite fast GP solver (linear scaling, complex-exponential kernels) that makes the 2-D model computationally practical.","marker":"Foreman-Mackey et al. (2017)"},{"why":"Showed GP regression can model profiles along the phase dimension, motivating the product kernel used here.","marker":"Johnston & Karastergiou (2019)"},{"why":"Supplies the simultaneous timing and GP spin-noise model from which the $\\dot{\\nu}$ time series used in the correlations are derived.","marker":"Keith & Niţu (2023)"},{"why":"Independent long-term profile-change detections for B1842+14 and B1914+09, used as comparison cases in this paper.","marker":"B24"}],"fun_headline_variants":["7 pulsars show first-ever spin-down-linked profile changes","Spin-down ties to pulsar shape changes in 7 new pulsars","2-D GP model finds 21 pulsars vary, 18 spin-down tied","Pulsar shape-spin link flips in B0740-28"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The Gaussian-process reconstruction, including its formal covariance, faithfully represents the intrinsic profile variation; the paper itself cautions (Section 4) that the formal error is likely underestimated and that excess noise is visible on the leading and trailing edges, so unmodelled systematics or underestimated GP variance would inflate the significance of the 21 detections and the uncertainties on the new correlations.","fun_headline_variants_meta":{"raw":{"variants":["7 pulsars show first-ever spin-down-linked profile changes","Spin-down ties to pulsar shape changes in 7 new pulsars","2-D GP model finds 21 pulsars vary, 18 spin-down tied","Pulsar shape-spin link flips in B0740-28"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000632,"raw_usage":{"total_tokens":2957,"prompt_tokens":1021,"completion_tokens":1936,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":1857}},"tokens_in":637,"tokens_out":1936,"duration_ms":15526,"temperature":1.0,"reasoning_tokens":1857,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:46:29.219320+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on simulated pulse stacks with synthetic profile variations of the claimed amplitudes (1-2 percent of peak flux) and timescales injected into the real noise, and count how often the model reports significant variations and spin-down correlations; a false-positive rate above the nominal threshold, or Monte Carlo Pearson-$r$ errors wider than the reported $\\pm 0.1$ to $\\pm 0.2$ values, would show the formal covariance is too optimistic and weaken the new detections.","supporting_citations":[],"review_version":1}