{"id":"9346e45b-1f0e-4d50-90d0-10e530ed1472","arxiv_id":"2505.10141","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Vela.jl implements the first public, non-linear Bayesian wideband pulsar timing and noise analysis pipeline, demonstrated on NANOGrav wideband data for PSR J1923+2515.","lead":"This paper describes Vela.jl, a Julia/Python package that performs Bayesian pulsar timing and noise analysis using wideband observations, where each measurement yields one arrival time and one dispersion measure. A demonstration on a NANOGrav pulsar shows the pipeline recovers DM noise and places a limit on achromatic red noise.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central 'first implementation' claim is supported only by a self-run demo; no cross-check against existing wideband-capable software or synthetic data is provided, leaving implementation correctness undemonstrated.","rationale":"The stress-test pass focused on the central claim in Sections I and V: Vela.jl is the first public implementation of the full non-linear wideband timing and noise model. For this claim to be meaningful, the implementation must be correct. The paper's demonstration on J1923+2515 is a real-data run without any independent verification. The reader's weakest-assumption about the diagonal TOA-DM covariance (Section II) is a genuine concern and should be flagged, but it concerns the validity of the demonstration's scientific conclusions rather than the software's claimed functionality. Even if the covariance were included, the software's output would still need validation. The absence of a cross-check against existing wideband-capable software (ENTERPRISE for the linearized limit, or TEMPONEST for non-linear narrowband) or a synthetic data injection leaves the central functionality claim ungrounded. This is a standard expectation for software papers: authors typically show agreement with an established code or recover known injected signals. Without such evidence, a reader cannot distinguish a correct implementation from a plausible-looking but buggy one. The public code and data make this check straightforward, so the concern is not that the claim is wrong, but that it is not yet demonstrated. This does not change the reader's conditional verdict; it reinforces it. We recommend keeping the conditional acceptance, with the condition being the inclusion of an external validation step before or after publication.","tokens_in":15036,"tokens_out":8372,"duration_ms":80716,"concrete_test":"Re-analyze the J1923+2515 dataset with ENTERPRISE using the same timing model linearized about the best-fit parameters, the same white-noise parameters, and the same DM-noise GP harmonics, and compare the posterior distributions for the DM-noise amplitude and spectral index, the EFACs, and the EQUADs against Vela.jl's results. If the marginalized posteriors differ by more than the quadrature sum of their statistical uncertainties (plus the expected linearization error), this indicates an implementation error in Vela.jl. A complementary test is to generate a synthetic wideband dataset with a known injected DM-noise signal and verify that Vela.jl recovers the injected parameters within the 68% credible interval.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that Vela.jl is the first public package for full non-linear wideband Bayesian timing and noise analysis. For this to hold, the implementation must actually compute the wideband likelihood correctly. The only evidence offered is the Section IV demonstration on PSR J1923+2515, which shows posterior distributions but no ground truth. Because the data are real, there is no known answer to compare against; a bug in the Gaussian-process marginalization or the wideband residual calculation could produce the same qualitative output. The paper acknowledges the diagonal-N approximation for TOA-DM covariances, and that is a limitation, but it is not the most load-bearing issue: even with diagonal N, the correctness of the likelihood evaluation is unverified. Without a numerical cross-check against an independent implementation (e.g., ENTERPRISE's linearized wideband mode) or a synthetic-data injection test, the functionality claim is plausible but unproven. The code and data are public, so such a check is feasible, but it is not included in the paper.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes the implementation of wideband pulsar timing and noise analysis in the Vela.jl package. It introduces a WidebandTOA data type, extends the TimingModel with wideband support, and provides two likelihood kernels: a WhiteNoiseKernel for Eq. (6) and a WoodburyKernel for the analytically marginalized likelihood in Eq. (12). The paper demonstrates the package on the NANOGrav 12.5-year wideband dataset for PSR J1923+2515, reporting timing parameters, DM noise detection, and an upper limit on achromatic red noise. The central claim, made in Sections I and V and in the abstract, is that Vela.jl is the first software package to provide full non-linear wideband Bayesian timing and noise analysis.","tokens_in":15278,"tokens_out":4279,"duration_ms":42781,"significance":"If the implementation is correct, this fills a genuine gap in the pulsar timing software ecosystem: ENTERPRISE offers wideband noise analysis only with a linearized timing model, and TEMPONEST handles non-linear timing models but only for narrowband data. The software is publicly available with a Python interface, and the demonstration runs on real, publicly released data. The analytic marginalization of Gaussian-process noise amplitudes follows published derivations and, as implemented, should improve sampling efficiency. The paper also contributes a command-line interface and a performance benchmark. The main risk to the significance claim is the absence of a numerical cross-check demonstrating that the implementation actually computes the wideband likelihood and posterior correctly.","major_comments":[{"comment":"The central claim that Vela.jl is the first public implementation of full non-linear wideband Bayesian timing and noise analysis is not supported by any independent numerical validation. The demonstration on PSR J1923+2515 uses real data without known ground truth, so a bug in the Gaussian-process marginalization or the wideband residual calculation could produce the same qualitative output. The authors should add a synthetic-data injection test (simulate wideband TOA/DM datasets with known parameters and noise, then verify recovery) or a numerical comparison against an independent implementation such as ENTERPRISE's linearized wideband mode on the same dataset. The code is public, so such a test is straightforward and would substantially strengthen the paper's load-bearing claim.","section":"Section IV and Section V"},{"comment":"The diagonal approximation of the TOA-DM covariance matrix N is explicitly acknowledged in Section II, but the DM noise detection presented in Section IV relies on this approximation. If the zero-covariance frequency is inaccurate for the L-wide or 430 MHz data, or if pulse jitter introduces frequency-correlated noise, the recovered DM noise parameters could be biased. The paper notes that public datasets do not provide covariance estimates, but it should either justify the approximation for PSR J1923+2515 (e.g., using the Hessian-based estimate from PulsePortraiture, if available) or perform a sensitivity analysis by adding simulated nonzero covariances. Without this, the robustness of the demonstrated DM noise detection is unclear.","section":"Section II and Section IV"},{"comment":"The analysis uses 'cheat priors' centered on frequentist PINT estimates with a width of 100 times the frequentist uncertainty for several parameters. The paper states that increasing the prior widths has 'no discernible impact' on the posterior, but no quantitative evidence is provided. Because these priors are informative and derived from the same dataset, the posterior distributions could be artificially narrowed. The authors should show a comparison of the posterior distributions obtained with different prior widths, or at least report the numerical differences in the parameter estimates, to support the robustness claim.","section":"Section IV and Table I"}],"minor_comments":[{"comment":"The notation '2ntoa' is ambiguous; it should be typeset as 2 n_toa or with a defined symbol to avoid confusion between a product and a subscript.","section":"Section II, Eq. (6)"},{"comment":"The benchmark is computed for a simulated narrowband dataset, but the text claims that wideband execution times are similar. Either provide a wideband benchmark or explicitly qualify that this is an order-of-magnitude expectation not measured for wideband data.","section":"Figure 4 caption"},{"comment":"The actual MCMC settings for the analysis (number of walkers, total iterations, burn-in, thinning) are not stated; the script in Figure 3 is illustrative. The authors should report these settings or point to a configuration file in the repository, to enable reproducibility of the presented posterior distributions.","section":"Section IV"},{"comment":"The prior for TNREDGAM (spectral index) is listed as Uniform[0, 7] and the posterior is 'consistent with its prior', but this is noted only in a table footnote. It would be clearer to state this directly in the text or to show the posterior-prior comparison in Figure 5.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The paper's novelty claim is strong and the software appears to be well-engineered, but the lack of a numerical cross-check is the key issue. The self-citation is heavy, though the code is public and the results are externally checkable. I would encourage the editor to request a validation section (synthetic data or comparison with ENTERPRISE) before acceptance, as this is standard for software papers claiming new functionality."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real engineering contribution—the first public non-linear Bayesian wideband pulsar timing implementation—and the central claim is probably true, but the paper does not yet prove the implementation is correct. I would send it to review, but I would condition acceptance on an independent validation.\n\nWhat is genuine: the package implements the wideband likelihood with Gaussian-process marginalization following published derivations, and it ships public code and data. The J1923+2515 demo is a new analysis and produces sensible posteriors: DM noise is detected, achromatic red noise is not, and the red noise upper limit is consistent with the common-signal amplitude reported by the PTAs. The author is also honest about limitations, explicitly flagging the diagonal covariance approximation, the lack of TOA-DM covariance estimates in public datasets, and the use of cheat priors.\n\nSoft spots, in proportion: the missing cross-check is the one that matters most. The demo has no ground truth, so a bug in the Gaussian-process marginalization or the wideband residual calculation could produce posteriors that look qualitatively similar. A synthetic injection test or a numerical comparison with ENTERPRISE's linearized wideband mode on the same dataset would nail the correctness claim. That is a concrete, feasible addition, and its absence is the main reason I stop short of saying the claim is fully proven. Second, the diagonal-N approximation is a real limitation; the paper flags it and says a future publication will investigate it. For a low-S/N pulsar like J1923+2515, the covariance could be non-negligible, so the DM noise detection could be somewhat mis-specified. This is a caveat, not a fatal flaw, because the author acknowledges it and the central capability claim does not depend on the covariance being zero. Third, the cheat priors anchored to PINT are slightly unaesthetic, but the width check and the explicit discussion make this a minor issue. Self-citation is not a problem here; the cited formalism is directly relevant and the code is externally checkable.\n\nWho this is for: pulsar timing practitioners and PTA method developers. It is a useful tool and the paper is a fair presentation of it. I would bring it to a reading group, and I would cite it if I were doing wideband timing analyses. The right outcome is serious peer review with a request for an independent numerical validation before the strongest claims are taken as established.","headline":"A genuinely useful software contribution with a plausible first-implementation claim, but the paper stops short of proving the implementation is correct; it deserves peer review with a request for independent validation.","tokens_in":15779,"tokens_out":1712,"would_cite":true,"duration_ms":19829,"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 presents Vela.jl as the first public software package to perform full non-linear Bayesian pulsar timing and noise analysis on wideband datasets, where each observation yields a time-of-arrival and dispersion-measure pair.","keywords":["pulsar timing","wideband timing","Bayesian inference","dispersion measure","dispersion-measure noise","Gaussian process","timing noise analysis","Vela.jl"],"falsifier":"A direct estimate of the TOA–DM noise correlation in a wideband observation, computed without applying the zero-covariance-frequency adjustment, would be the decisive check: if that correlation is large relative to the quoted uncertainties and re-analyzing PSR J1923+2515 with it shifts the reported DM-noise amplitude beyond its posterior uncertainty, the paper's diagonal-covariance assumption is falsified.","tokens_in":14778,"feed_emoji":"📡","tokens_out":14830,"duration_ms":130234,"temperature":0.7,"pith_summary":"This paper presents Vela.jl as the first public software package to run full non-linear Bayesian pulsar timing and noise analysis on wideband datasets, in which each observation yields a time of arrival and a dispersion measure measured jointly from the full frequency band rather than from separate sub-bands. The paper's motivation is that wideband receivers are now standard in pulsar timing, while the available tools either approximate the timing model as linear for such data or handle non-linear models only in the narrowband case. The implementation pairs the joint likelihood of timing and dispersion-measure residuals with Gaussian-process models for dispersion-measure noise and slow 'red' timing noise, and analytically marginalizes many timing parameters so a generic sampler can handle the rest. A demonstration on the 12.5-year dataset of PSR J1923+2515 detects dispersion-measure noise and bounds the amplitude of achromatic red noise, showing the workflow on real data. The gap the paper fills is that wideband Bayesian inference without linearization now has a public implementation.","feed_headline":"One package now runs full Bayesian wideband pulsar timing","feed_subtitle":"Treats time and dispersion measure jointly, so wideband data get full non-linear Bayesian analysis.","key_machinery":"The central object is the joint wideband likelihood built from a $2n_{\\rm toa}$-dimensional vector $y$ that stacks the timing residual and dispersion-measure residual of each observation against a covariance matrix $N$; time-correlated processes enter as reduced-rank Gaussian processes, so the marginalized covariance becomes $C = N + U\\Phi U^T$. The package realizes this with two kernels, the WhiteNoiseKernel for a diagonal $N$ and the WoodburyKernel for the reduced-rank case, with the Woodbury identity supplying the matrix inverse. Analytic marginalization over the Gaussian-process amplitudes and over nearly linear timing parameters is what keeps the remaining sampling problem low-dimensional, and the WidebandTOA type bundles the TOA, DM, uncertainties, and observing metadata so narrowband and wideband data share one workflow.","core_discovery":"The central claim is that the full non-linear wideband timing model—including the joint likelihood of timing and dispersion-measure (DM) residuals, white-noise parameters, and Fourier-domain Gaussian processes for spin and DM noise—can be implemented in a public Bayesian inference package with support for arbitrary deterministic and stochastic dispersion-measure models, and that Vela.jl is the first such implementation. The paper demonstrates that the likelihood can be computed with either a diagonal noise covariance or a reduced-rank covariance using the Woodbury identity, and that Gaussian-process amplitudes and approximately linear timing parameters can be marginalized analytically. Applying the workflow to the 12.5-year dataset of PSR J1923+2515, the paper reports a detection of dispersion-measure noise with spectral index $2.7 \\pm 0.8$, a 95.45% upper limit of $-13.7$ on the log-amplitude of achromatic red noise (the slow, frequency-independent part of the timing residuals), a proper-motion measurement, and a parallax consistent with zero at $3\\sigma$.","pith_inferences":["The paper leaves the TOA–DM noise correlation unmodeled because public data do not constrain it; the same Woodbury machinery would accept such a covariance if future data releases measured it, so re-running demonstrations like J1923+2515 with measured covariances is a natural next test.","The Gaussian-process DM model could also be applied to multi-band wideband observations to separate solar-wind dispersion from interstellar dispersion, a separation this paper does not attempt.","If jitter-induced TOA–DM correlation is substantial, part of the reported DM-noise signal could be a noise-model artifact; comparing the same data against a correlated-noise alternative would settle how much of the detection is real."],"forward_implications":["Wideband pulsar-timing-array datasets can be analyzed with a full non-linear timing model in a Bayesian framework, rather than only with a linearized approximation.","Dispersion-measure variability can be modeled as a Gaussian process instead of only as piecewise-constant DMX steps, enabling searches for smooth DM-variation signals.","Analytic marginalization over noise amplitudes and nearly linear timing parameters lowers the dimensionality so generic ensemble and nested samplers can be used.","The same Python interface handles narrowband and wideband data, so existing Vela.jl workflows extend to wideband without new analysis patterns.","The command-line interface makes standard wideband Bayesian analysis possible without writing sampler code."],"supporting_citations":[{"why":"Supplies the elementary wideband timing method and the zero-covariance-frequency adjustment that public wideband measurements rely on; it is the paradigm's foundation.","marker":"[18]"},{"why":"Formulates wideband timing by treating the integrated pulse profile as a single two-dimensional entity, motivating the paradigm Vela.jl implements.","marker":"[19]"},{"why":"Provides the principal-component-analysis template method used in practice to derive wideband TOA–DM pairs.","marker":"[20]"},{"why":"Supplies the linearized wideband Bayesian framework that is the main point of comparison for the paper's non-linear implementation.","marker":"[28, 29]"},{"why":"The non-linear Bayesian timing approach for narrowband data that the wideband implementation extends.","marker":"[30]"},{"why":"The 12.5-year wideband dataset of PSR J1923+2515 used for the demonstration.","marker":"[31]"},{"why":"Develops the Gaussian-process representation of dispersion-measure noise for wideband datasets and the analytic marginalization scheme Vela.jl adopts.","marker":"[34]"},{"why":"The companion Vela.jl overview defining the package's types, priors, and workflow that this work extends to wideband data.","marker":"[35]"},{"why":"Provides the low-rank Fourier-basis Gaussian-process construction and harmonic selection for the spin and DM noise covariance.","marker":"[41]"},{"why":"Supplies the partial analytic marginalization over timing parameters used to reduce the number of sampled dimensions.","marker":"[42]"}],"fun_headline_variants":["First Bayesian wideband pulsar timing software","Vela.jl brings Bayesian analysis to wideband pulsar timing","Full non-linear Bayesian timing for wideband pulsar data","Wideband pulsar timing with joint TOA-DM analysis","First to implement full wideband Bayesian pulsar timing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the noise in each wideband observation's TOA and DM is uncorrelated, so the covariance matrix $N$ is diagonal; public wideband datasets provide no estimate of the TOA–DM covariance, and if that covariance is not negligible for jitter-dominated data, every reported posterior is mis-specified.","fun_headline_variants_meta":{"raw":{"variants":["First Bayesian wideband pulsar timing software","Vela.jl brings Bayesian analysis to wideband pulsar timing","Full non-linear Bayesian timing for wideband pulsar data","Wideband pulsar timing with joint TOA-DM analysis","First to implement full wideband Bayesian pulsar timing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000575,"raw_usage":{"total_tokens":2664,"prompt_tokens":848,"completion_tokens":1816,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":1735}},"tokens_in":464,"tokens_out":1816,"duration_ms":12898,"temperature":1.0,"reasoning_tokens":1735,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:15:32.689916+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct estimate of the TOA–DM noise correlation in a wideband observation, computed without applying the zero-covariance-frequency adjustment, would be the decisive check: if that correlation is large relative to the quoted uncertainties and re-analyzing PSR J1923+2515 with it shifts the reported DM-noise amplitude beyond its posterior uncertainty, the paper's diagonal-covariance assumption is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the elementary wideband timing method and the zero-covariance-frequency adjustment that public wideband measurements rely on; it is the paradigm's foundation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Formulates wideband timing by treating the integrated pulse profile as a single two-dimensional entity, motivating the paradigm Vela.jl implements."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the principal-component-analysis template method used in practice to derive wideband TOA–DM pairs."},{"cited_title":"Young and M","cited_arxiv_id":null,"evidence_quote":"The non-linear Bayesian timing approach for narrowband data that the wideband implementation extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The 12.5-year wideband dataset of PSR J1923+2515 used for the demonstration."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Develops the Gaussian-process representation of dispersion-measure noise for wideband datasets and the analytic marginalization scheme Vela.jl adopts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The companion Vela.jl overview defining the package's types, priors, and workflow that this work extends to wideband data."},{"cited_title":"Susobhanan, Bayesian pulsar timing and noise analy- sis with vela.jl: An overview, The Astrophysical Journal 980, 165 (2025)","cited_arxiv_id":null,"evidence_quote":"Provides the low-rank Fourier-basis Gaussian-process construction and harmonic selection for the spin and DM noise covariance."},{"cited_title":"Damour and N","cited_arxiv_id":null,"evidence_quote":"Supplies the partial analytic marginalization over timing parameters used to reduce the number of sampled dimensions."}],"review_version":1}