{"id":"4f6fdf59-3c17-42c8-8f4d-0e037ece40cc","arxiv_id":"2505.05274","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"The authors derive a marginalized Gaussian-process likelihood for wideband pulsar timing that models dispersion measure noise with arbitrary variability models, generalizing the existing piecewise-constant approach.","lead":"Pulsar timing arrays hunt for gravitational waves by measuring radio pulse delays, but gas in the galaxy delays pulses in a way that drifts over time. This paper derives a general method to model that drifting gas delay directly in 'wideband' pulsar data, making it easier to separate this noise from gravitational wave signals.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified. The central Gaussian marginalization in Eq. (17) is correct, and the linearized timing model is the standard, well-justified approximation it is claimed to be.","rationale":"The paper's central claim is that Eq. (17) provides a valid marginalized wideband likelihood for arbitrary DM variation models, including Gaussian processes. This is a straightforward application of Gaussian linear-model marginalization, and the derivation is sound given the linearized timing model. The reader's identified weakest assumption, the linearization in Eq. (10), is indeed the most fragile part, but it is universally used in pulsar timing and is well-justified for the high-quality, well-constrained timing solutions typical of wideband data. The paper's simulation exercises the full pipeline and recovers injected parameters, supporting the claim. I considered whether the use of improper priors for timing parameters undermines the derivation: as in standard practice, this is handled by taking a broad-Gaussian limit and dropping a constant factor, which does not affect the hyperparameter posteriors or the comparisons made here. The computational-complexity argument is also correct. No concrete error or hidden assumption was found, so the reader's ACCEPT verdict stands.","tokens_in":14073,"tokens_out":28116,"duration_ms":261641,"concrete_test":"Re-analyze the simulated dataset with a full non-linear timing and noise analysis (e.g., Vela.jl SPNTA) and compare the posterior distributions of EFAC, DMEFAC, and the spin/DM noise spectral parameters against the marginalized linearized likelihood; agreement within sampling error would confirm that the linearization is adequate for typical wideband data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I find no load-bearing concern. The central result, Eq. (17), is the standard analytic Gaussian marginalization of a linear model with a Gaussian prior; the matrix determinant lemma ensures the normalization is correct. The extension to wideband data by stacking TOA and DM residuals in a single vector y and including DM-related columns in the design matrix M is sound. The weakest assumption is the linearized timing model of Eq. (10), which is the usual first-order expansion near a best-fit timing solution; this is standard and appropriate for pulsar timing. The treatment of improper priors for timing parameters is conventional: the marginalized likelihood is understood as the limit of a broad Gaussian prior, and any resulting constant factor does not affect the noise-parameter posteriors that are the focus of the paper. The simulation provides a reasonable sanity check, including the comparison with DMX that highlights a known limitation of DMX. I find no circularity, internal inconsistency, or unsupported key step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper derives an analytically marginalized likelihood for wideband pulsar timing with arbitrary dispersion-measure (DM) variability models, including Gaussian-process (DMGP) models. Starting from a linearized wideband timing model y = ybar + M a and a Gaussian prior a ~ N(0, Phi), the authors marginalize over the parameter deviations to obtain Lambda = exp(-1/2 y^T C^{-1} y) / sqrt(det(2 pi C)) with C = N + M Phi M^T (Eqs. 17-18). The paper then derives a factorized ENTERPRISE-friendly form of this likelihood (Appendix B), compares it with the Alam et al. (2021) wideband likelihood (Appendix A), and derives a maximum-likelihood estimator for PINT (Appendix C). The method is demonstrated on a simulated 500-epoch dataset with injected spin and DM power-law noise; the posteriors recover the injected values, and a comparison with the DMX model shows a known bias in DMEFAC. Code and data are provided as supplementary material.","tokens_in":14235,"tokens_out":20321,"duration_ms":196640,"significance":"The central result, Eq. (17), is a clean and standard Gaussian marginalization. If it stands, it provides a general, computationally practical way to include DMGP, DMX, spline, and solar-wind DM models in wideband single-pulsar noise analysis and pulsar-timing-array analyses, removing the current DMX-only limitation in ENTERPRISE. The simulation validates the implementation, and the numerical equivalence check in Appendix B supports the factorized form. The paper also makes concrete contributions toward ENTERPRISE and PINT integration. The derivation is self-contained and builds on well-established PTA methods, and the released simulated dataset and notebook aid reproducibility. Overall, this is a useful methods paper for the pulsar-timing community, though a few presentation and consistency issues should be fixed.","major_comments":[],"minor_comments":[{"comment":"The covariance matrix after marginalizing over the correlated-noise amplitudes alpha with prior covariance Psi should be G = N + U Psi U^T, not G = N + U^T Psi^{-1} U as printed. The printed expression is dimensionally inconsistent (N is 2N_toa x 2N_toa while U^T Psi^{-1} U is n x n) and would propagate incorrectly into Eqs. (C3)-(C6). Please correct this.","section":"Appendix C, Eq. (C2)"},{"comment":"The text in Section 4 states that all parameter estimates are consistent with the injected values within 2-sigma uncertainties, while the caption of Figure 1 says within 3-sigma uncertainties. Please make these statements consistent and verify which is correct.","section":"Section 4 and Figure 1"},{"comment":"The Fourier/Taylor expansion for the rotational phase is written as a sum starting at j = 1, which omits the linear spin-frequency term F0 (tau_i - tau_0). Since Table 1 lists F0 as the spin frequency, the sum should start at j = 0 (or the indexing should be adjusted) to include the zeroth-order term.","section":"Section 2, Eq. (2)"},{"comment":"The Introduction says the linearization and analytic marginalization are described in Section 2, but these are actually presented in Section 3. Please correct the cross-reference.","section":"Section 1"},{"comment":"The discussion of improper priors states that timing parameters have infinite diagonal elements in Phi. In the limit of infinite variance, C in Eq. (18) is not a finite matrix, and Eq. (17) is understood only as a limiting expression. Please state the limiting argument explicitly, for example by referring to a sequence of broad proper priors and noting that divergent constants cancel in the posterior for the noise parameters.","section":"Section 3, after Eq. (18)"},{"comment":"The repeated analysis with a smaller number of harmonic bins is described only qualitatively ('EFAC and DMEFAC being overestimated'). Reporting the numerical values or a figure would make this robustness check reproducible and quantitative.","section":"Section 4"},{"comment":"The quantity B G^T \\bar L^{-1} delta is the posterior mean (or MAP under the prior B) of the dispersion parameters beta given the DM measurements, not the maximum-likelihood estimate. Please adjust the wording for accuracy.","section":"Appendix B, Eq. (B14a)"}],"recommendation":"minor_revision","confidential_remarks":"The main result is sound and well-suited to MNRAS. The issues are local and fixable; no reanalysis is required. The Appendix C covariance typo should be corrected before publication to avoid confusion for implementers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a clean methods paper that does what it says. The main result is Eq. (17): a single Gaussian likelihood for the combined TOA and DM residuals, with covariance C = N + M Φ Mᵀ. The design matrix M can encode any DM variation model, including a Gaussian process, and the shared DM parameters couple the two residual blocks automatically. As far as I can tell, this isn't in the literature yet: Alam et al. (2021) only give a factorized DMX likelihood, and Vela.jl's DMGP is in the full non-linear SPNTA setting. The paper is careful to distinguish those cases and shows in Appendix A that Alam's separated form emerges when DMX parameters are left unmarginalized.\n\nThe derivation is standard Bayesian linear algebra, correctly using Woodbury and determinant identities, and the computational cost discussion is sensible. The simulation is a reasonable sanity check: recovered parameters within 2σ, whitened residuals, and a comparison showing DMX underestimates DMEFAC—a useful warning for the community. They also ship the simulated dataset and a Jupyter notebook, so the result is reproducible.\n\nSoft spots, in proportion, are minor. It's a single simulated dataset, so no systematic study of cadence or noise levels; the linearized timing model is the usual first-order expansion, well-justified for timing data; and the improper-prior limiting argument is deferred to references, which is standard practice. Appendix B's factorized form is a bit terse but numerically verified against Eq. (17). No circularity, no internal inconsistency. The citation pattern looks appropriate; the key prior work is cited.\n\nThis is an enabling tool for pulsar timing array analyses, not a discovery. It deserves a serious referee. I'd send it.","headline":"Correct, clean derivation of a wideband DMGP likelihood; the DMX comparison is a nice practical bonus.","tokens_in":14774,"tokens_out":3933,"would_cite":true,"duration_ms":38491,"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":"Wideband pulsar timing can now absorb Gaussian-process dispersion-measure noise through an analytically marginalized likelihood.","keywords":["pulsar timing","dispersion measure noise","wideband timing","Gaussian process","Bayesian noise analysis","pulsar timing arrays","analytic marginalization","interstellar medium"],"falsifier":"Run the closed-form likelihood on a simulated wideband dataset in which the timing-model parameters are deliberately displaced far from their best-fit values, and compare its posterior with the full nonlinear posterior sampled directly; if the two disagree by more than the statistical uncertainty, the linearization premise fails. A similar comparison with a non-Gaussian prior on the parameter deviations would quantify how much the Gaussian-prior assumption costs.","tokens_in":13868,"feed_emoji":"📡","tokens_out":8881,"duration_ms":84079,"temperature":0.7,"pith_summary":"This paper derives a closed-form likelihood for wideband pulsar timing, the paradigm in which each observation yields both a time of arrival and a dispersion measure (DM). It shows that when the timing model is linearized and the model parameters are given Gaussian priors, those parameters can be integrated out analytically, leaving a Gaussian likelihood whose covariance is the white-noise covariance plus a design-matrix projection of the prior covariance. Because the design matrix and prior covariance can encode any DM variation model, the result brings Gaussian-process models of stochastic DM variation into single-pulsar and pulsar-timing-array analyses. This matters because existing wideband analyses relied on piecewise-constant DMX bins or sampled many per-epoch parameters, whereas a Gaussian process captures long-timescale correlated DM variations with a spectral prior and keeps the computation analytically tractable.","feed_headline":"Wideband timing gains closed-form Gaussian-process DM noise","feed_subtitle":"The new likelihood handles any DM variation model, replacing piecewise-constant bins with Gaussian processes.","key_machinery":"The central object is the analytically marginalized Gaussian likelihood identity: for a linear model $y = \\bar{y} + M a$ with a zero-mean Gaussian prior $a \\sim N(0, \\Phi)$, the nuisance parameter $a$ can be integrated out exactly, leaving a Gaussian likelihood in $y$ with covariance $C = N + M \\Phi M^{T}$. This identity carries the argument because it converts the choice of a DM variability model into a choice of design-matrix columns $M$ and prior covariance $\\Phi$, so no per-epoch DM parameters need to be sampled.","core_discovery":"Under the standard linearized wideband timing model $y = \\bar{y} + M a$, with a zero-mean Gaussian prior on the parameter deviation vector $a$ with covariance $\\Phi$, the paper derives the analytically marginalized likelihood $\\Lambda \\propto \\exp(-\\tfrac{1}{2} y^{T} C^{-1} y) / \\sqrt{\\det(2\\pi C)}$, where $C = N + M \\Phi M^{T}$. This expression has the same Gaussian form as the narrowband likelihood, so any DM variation model that can be written as design-matrix columns with Gaussian coefficients, including Fourier-basis Gaussian processes, DMX bins, constrained splines, and solar-wind models, can be included and marginalized analytically. The paper shows that a previously published wideband likelihood appears as the special case in which DMX and DMJUMP parameters are left unmarginalized, and that marginalizing those parameters brings it to the new form. A simulated dataset with injected spin noise and DM noise recovers the input parameters within $2\\sigma$ uncertainties, and a comparison analysis with the DMX model finds that DMEFAC is underestimated because piecewise-constant DMX bins absorb short-timescale DM measurement noise.","pith_inferences":["The same $M \\Phi M^{T}$ marginalization could be applied to other per-observation chromatic delays, such as scattering-variation delays, by adding design-matrix columns; the paper notes scattering is not yet well understood in wideband timing, but the mathematics does not depend on the physical origin of the delay.","The factorized form of the likelihood in Appendix B suggests a modular wideband pulsar-timing-array analysis in which DM residuals are modeled first and TOA residuals are then conditioned on the DM fit, which could simplify gravitational-wave searches.","A testable prediction follows from the simulation: on real wideband datasets with short-timescale DM fluctuations, DMX analyses will systematically underestimate DMEFAC relative to a Gaussian-process DM analysis, which can be checked with existing wideband pulsar data."],"forward_implications":["Single-pulsar noise analysis of wideband data can now include Gaussian-process DM noise alongside the piecewise-constant DMX model, so long-timescale correlated DM variations are represented by a spectral prior instead of many per-epoch bins.","Wideband analyses no longer depend on infinitely wide improper priors on DMX parameters, making them usable when low-frequency or high-bandwidth observations leave individual DMs poorly constrained.","Pulsar timing array searches with wideband data can analytically marginalize DM variations, including Gaussian-process DM models, while looking for cross-pulsar correlated signals such as the nanohertz gravitational-wave background.","The likelihood evaluation cost is $O(N_{\\mathrm{data}} p^{2})$, and for equivalent observations the narrowband likelihood is slower by a factor of $N_{\\mathrm{subband}}/2$, so wideband Gaussian-process DM analyses are computationally cheaper as well as more general.","The same formalism covers piecewise-constant DMX, constrained spline, and solar-wind DM models as special cases, giving one unified likelihood for all common DM variability models."],"supporting_citations":[{"why":"supplies the analytic marginalization steps that turn the joint posterior over parameters into the marginalized likelihood of Eq. (17).","marker":"van Haasteren et al. (2009)"},{"why":"establishes the zero-mean Gaussian prior on timing-model parameter deviations with covariance $\\Phi$, the conjugate prior used here.","marker":"van Haasteren & Levin (2013)"},{"why":"provides the spectral Gaussian-process priors for spin and DM noise amplitudes that the new likelihood makes available for wideband data.","marker":"Lentati et al. (2013)"},{"why":"gives the existing wideband likelihood that this paper generalizes; the comparison shows the new expression reduces to it when DMX and DMJUMP parameters are left unmarginalized.","marker":"Alam et al. (2021)"},{"why":"introduces the wideband TOA-DM measurement paradigm and the portrait method, the data model this work starts from.","marker":"Pennucci et al. (2014)"},{"why":"develops the wideband timing technique, including the simultaneous TOA-DM estimation used in the paper's likelihood.","marker":"Liu et al. (2014)"},{"why":"defines the power-law spectral covariance for Gaussian-process DM noise and spin noise that anchors the prior covariance $\\Phi$.","marker":"van Haasteren & Vallisneri (2014a)"},{"why":"defines the DMX piecewise-constant DM model that is the current standard and a special case of the new framework.","marker":"Arzoumanian et al. (2015)"}],"fun_headline_variants":["Closed-form Gaussian-process DM noise for wideband pulsar timing","Wideband timing folds any DM model into closed-form likelihood","Analytic Gaussian-process DM noise for wideband timing","GP DM noise in wideband timing now has closed-form likelihood"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The derivation assumes the wideband residuals are exactly linear in the timing and noise parameter deviations, with a Gaussian prior on those deviations; when the timing solution is far from optimal or the priors are strongly non-Gaussian, the closed-form likelihood is only an approximation.","fun_headline_variants_meta":{"raw":{"variants":["Closed-form Gaussian-process DM noise for wideband pulsar timing","Wideband timing folds any DM model into closed-form likelihood","Analytic Gaussian-process DM noise for wideband timing","GP DM noise in wideband timing now has closed-form likelihood"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000615,"raw_usage":{"total_tokens":2830,"prompt_tokens":890,"completion_tokens":1940,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":1871}},"tokens_in":506,"tokens_out":1940,"duration_ms":15128,"temperature":1.0,"reasoning_tokens":1871,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:08:44.213796+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the closed-form likelihood on a simulated wideband dataset in which the timing-model parameters are deliberately displaced far from their best-fit values, and compare its posterior with the full nonlinear posterior sampled directly; if the two disagree by more than the statistical uncertainty, the linearization premise fails. A similar comparison with a non-Gaussian prior on the parameter deviations would quantify how much the Gaussian-prior assumption costs.","supporting_citations":[],"review_version":1}