{"id":"f3ea8b79-5c07-45b7-a442-1dda35785eb6","arxiv_id":"2607.12609","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"PSRDISP is a Gaussian-process framework that fits dispersion-measure and solar-wind noise directly to epoch-wise dispersion measures, recovering injected signals in simulated pulsar data.","lead":"PSRDISP models the random and regular delays that interstellar gas and the solar wind imprint on pulsar radio signals by analyzing epoch-by-epoch dispersion measures directly, rather than pulse arrival times. It gives pulsar timing arrays a new, independent way to check their noise models, which is important for searches for gravitational waves.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (13) inverts the DM-to-delay scaling: a DM amplitude converts to time via D/ν_ref^2, not ν_ref^2/D. If used literally, the DMN covariance is mis-scaled by ~2.2e5 in variance, so the reported recovery cannot be reproduced from the text as written.","rationale":"The reader's weakest assumption was the diagonal covariance of epoch-wise DMs and the absence of scattering bias. Those are real limitations, but the paper states them explicitly and conditions the method's applicability on them, so they do not undermine the central simulation-based claim within the stated scope. The more load-bearing issue is internal consistency: the mathematical description of the amplitude transformation in Eq. (13) contradicts the dispersion relation in Eq. (4). If a reader implements the method from the equations, the DMN covariance normalization is wrong by orders of magnitude, and the claimed comparability with conventional ToA-domain noise amplitudes fails. The reported recovery figures suggest the actual implementation likely uses the correct inverse ratio, but the text as written cannot be reproduced as-is. This is a correctable but essential flaw in the presentation of the core method, and it justifies keeping the paper conditional pending correction and a reproducibility check. The paper otherwise has genuine strengths: a standard GP framework, explicit simulations in both narrowband and wideband paradigms, and honest discussion of the independence and scattering limitations. The concern here is not about the validity of the underlying idea but about the published derivation of its key conversion step.","tokens_in":26364,"tokens_out":15433,"duration_ms":160572,"concrete_test":"Re-derive Eq. (13) from Eq. (4) with units: δt = (D/ν_ref^2)·δDM, so the equivalent time-domain amplitude is A_DM·(D/ν_ref^2), not A_DM·(ν_ref^2/D). Then rerun the narrowband injection-recovery using the transformation exactly as printed in Eq. (13) and check whether the recovered log10 A_DMN shifts by ~2.67 dex away from the injected −13.5. If it shifts, Eq. (13) as written is wrong and must be corrected to A_DM → A_DM·(D/ν_ref^2). If it does not shift, the code already uses the correct ratio and the text should be amended to remove the dimensional inconsistency.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism of PSRDISP is the translation between DM-domain Fourier amplitudes and the conventional time-delay PSD amplitudes used in ToA-based SPNA. Eq. (4) gives Δt = D·DM/ν^2, so a DM fluctuation of amplitude δDM produces a time delay at reference frequency ν_ref of amplitude δt = (D/ν_ref^2)·δDM. Eq. (13) instead multiplies the DM-domain amplitude by ν_ref^2/D, which is dimensionally incorrect: it has units of (pc cm^-3)/s rather than s/(pc cm^-3), and the variance normalization entering Φ (Eq. 21) would be off by (ν_ref^2/D)^2 ≈ 2.23×10^5 in variance, or ≈2.67 dex in log10 A. The posterior in Fig. 3 recovers log10 A_DMN ≈ −13.52 against an injected −13.5, which is consistent only if the code does not use Eq. (13) as written. This is not a mere wording issue: the paper's claim that PSRDISP gives an independent check on conventional SPNA depends on the fitted DMN amplitude being correctly comparable to the ToA-domain amplitude. As published, the transformation preventing that comparison is internally inconsistent with Eq. (4), making the method's description irreproducible and its claimed comparability unsupported.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents PSRDISP, a Bayesian Gaussian-process framework for modeling dispersive noise processes—DM noise, solar-wind noise, and deterministic DM/SW trends—directly from epoch-wise dispersion-measure estimates rather than from times of arrival. The method uses a reduced-rank Fourier-domain GP with a power-law PSD, analytically marginalizes over Fourier amplitudes and deterministic DM polynomial parameters, and samples the remaining hyperparameters with emcee. The authors validate the method on simulated narrowband and wideband datasets, reporting good recovery of injected DMN and SWN amplitudes and spectral indices, and reconstructing time-domain realizations. They argue that working in the DM domain makes the method largely immune to achromatic red noise and provides an independent check on ToA-based single-pulsar noise analyses.","tokens_in":26798,"tokens_out":15087,"duration_ms":163770,"significance":"The approach is timely for PTA-era noise characterization and, if correct, would offer a complementary data domain for estimating chromatic noise. The paper's strengths are the use of the standard van Haasteren–Levin marginalization, the explicit likelihood derivation, the inclusion of both narrowband and wideband simulated validation, and the candid statement of key limitations (independence of DM estimates, scattering biases). The claimed advantage over ToA-based analyses, however, is not quantitatively demonstrated, and one equation in the central derivation is dimensionally inconsistent. These issues must be addressed before the comparability and reproducibility claims can be accepted.","major_comments":[{"comment":"The transformation in Eq. (13) is dimensionally inconsistent with Eq. (4). Since Δt = D·DM/ν², converting a DM-domain Fourier amplitude (units pc cm⁻³) to the conventional time-delay amplitude requires multiplication by D/ν_ref² ≈ 2.12×10⁻³ s/(pc cm⁻³), not by ν_ref²/D ≈ 472 (pc cm⁻³)/s. As written, Eq. (13) would change the variance entering Φ (Eq. 21) by (ν_ref²/D)² ≈ 2.23×10⁵, i.e., ≈2.67 dex in log10 A. The recovery in Fig. 3 (log10 A_DMN = −13.52 vs injected −13.5) cannot be reproduced from the text as written. This is load-bearing for the paper's central comparability claim. Please correct the factor and clarify whether Eqs. (21)–(22) define the PSD for DM-domain or delay-domain amplitudes.","section":"§2, Eqs. (12)–(13) and (21)–(22)"},{"comment":"The headline advantage over ToA-based SPNA—reduced achromatic-red-noise contamination—is asserted but not demonstrated. In the wideband simulation, ARN is injected into the ToAs only; the DM data used by PSRDISP are unaffected by ARN by construction. The paper does not report a comparison with ENTERPRISE on the same simulated ToAs, with and without ARN, to quantify leakage into DMN/SWN recovery. Without such a benchmark, the 'almost unaffected' claim and the method's value as an independent check remain plausible but unquantified. Please add this comparison or soften the claim.","section":"§4 and Appendix A"}],"minor_comments":[{"comment":"The posterior corner plots omit the minus sign on log10 A_DMN and log10 A_SWN; e.g., 'log10 Adm = 13.52' should read '−13.52', consistent with Table 1 and the text.","section":"Figs. 3 and A.3"},{"comment":"The units of DM are written as 'pc/cm³' in several places; the standard notation is pc cm⁻³. Also, check the units of D in Eq. (4) against the factor in Eq. (13).","section":"§3 and Fig. 2"},{"comment":"The text says γ_DMN is recovered at ~2σ, but the quoted value 3.31±0.20 versus the injected 3.0 corresponds to ~1.6σ. Please reconcile the statement with the quoted uncertainty.","section":"§3"},{"comment":"The sentence 'we used DMEFAC=EFAC=1.2' and the later statement that DISPEFAC is close to the injected DMEFAC should be clarified: DISPEFAC is defined in Eq. (15) for DM-domain white noise, while DMEFAC/EFAC are ToA-domain parameters. Explain the expected relationship.","section":"Appendix A"},{"comment":"The Data Availability section says the simulated datasets are shared as supplementary material, but no statement is made about availability of the PSRDISP analysis code. Please add a software availability statement.","section":"Software/Data Availability"}],"recommendation":"major_revision","confidential_remarks":"The key blocker is the Eq. (13) inconsistency. I recommend asking the authors to verify their implementation with a simple numerical check: does the code multiply the DM-domain Fourier amplitudes by D/ν_ref² or by ν_ref²/D? If the former, the reported recovery is internally consistent and the paper is likely sound after a correction; if the latter, the recovery cannot be reproduced and the paper needs substantial rework. I also recommend requesting a direct ENTERPRISE comparison on identical simulated data before acceptance, since the paper's main advantage over existing methods is otherwise not demonstrated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper repackages the standard Fourier-GP noise machinery into a standalone DM-domain fit. That is a reasonable idea and it may give PTA analysts a useful cross-check on DMN/SWN parameters. But the key selling point—that fitting epoch-wise DMs avoids achromatic red-noise contamination—is never benchmarked against the conventional ToA-based analysis, and the scaling transformation in Eq. (13) is written upside-down relative to Eq. (4). As written, the paper is not fully reproducible.\n\nWhat is genuinely new: applying the Lentati/van-Haasteren GP likelihood directly to epoch-wise DM time series, adding a DISPEFAC white-noise parameter, and showing it works for both narrowband and wideband simulated data. The authors are also honest about their assumptions: they explicitly flag the independence of DM measurements, the scattering-bias caveat, and they warn against circular use of the posteriors as priors elsewhere. The simulations look carefully constructed, and the recovery is internally consistent, which is what you would hope for from a same-model self-consistency test.\n\nThe main soft spot is Eq. (13). Eq. (4) gives Δt = (D/ν^2)·δDM, so converting a DM amplitude to a time delay means multiplying by D/ν^2. The paper multiplies by ν^2/D instead. That is dimensionally wrong, and if taken literally it would shift the DMN amplitude by roughly 2.7 dex in log10 A. The fact that the posteriors recover the injected values suggests the code actually uses the correct factor, but the text doesn't say this, the code isn't released, and the claim that the fitted amplitude is comparable to conventional SPNA amplitudes rests on this transformation. That needs to be fixed and verified, not just asserted.\n\nThe other weakness is that the headline advantage—reduced achromatic red-noise contamination—is never quantified. A direct injection/recovery comparison with ENTERPRISE on the same dataset would settle it. The independence assumption is named, which is fair, but it means the method may not apply to DMs from global DMX/DMMODEL fits, and scattering biases are a real concern for wideband DMs.\n\nWho is this for? PTA noise analysts who want a complementary, DM-only diagnostic. It deserves a serious referee, but not as-is. I would send it back asking for a corrected scaling relation, a code release or at least a worked numerical example, and a head-to-head benchmark against ToA-based SPNA. Without those, the paper remains a promising tool in need of validation rather than a demonstrated result.","headline":"A sensible DM-domain GP noise fit with a real payoff if corrected, but Eq. (13) inverts the DM-to-delay scaling and the claimed advantage over ToA-based SPNA is not actually demonstrated.","tokens_in":27257,"tokens_out":3205,"would_cite":false,"duration_ms":37836,"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":"This paper argues that dispersive noise in pulsar timing can be modeled directly from epoch-wise dispersion-measure time series, with no time-of-arrival delays used anywhere, and that the recovered noise parameters match injections in simul","keywords":["pulsar timing","dispersion measure","Gaussian process","single-pulsar noise","DM noise","solar wind noise","wideband timing","narrowband timing"],"falsifier":"Take DM time series produced by a global fit (for example a piecewise-constant or spline solution) that is known to create inter-epoch correlations, run PSRDISP on them, and check whether the recovered noise amplitude and spectral index shift outside the nominal posterior uncertainties. Alternatively, inject a scattering delay proportional to ν^-4 into simulated sub-banded arrival times before extracting DMs, and test whether the dispersive-noise parameters become biased.","tokens_in":26300,"feed_emoji":"📡","tokens_out":4587,"duration_ms":48864,"temperature":0.7,"pith_summary":"This paper tries to establish that dispersive single-pulsar noise—the stochastic and deterministic variations in the dispersion measure (DM) caused by the interstellar medium and the solar wind—can be characterized accurately using epoch-wise DM estimates alone, without ever forming time-of-arrival (ToA) delays. The method models the DM series as a constant plus polynomial deterministic terms, Fourier-domain Gaussian processes for DM noise and solar wind noise, and a white-noise term; a Gaussian likelihood with a diagonal covariance leads to analytically marginalised posteriors for the noise hyper-parameters. In simulated narrowband and wideband datasets with injected noise, the recovered amplitudes and spectral indices agree with injections at about the one-sigma level, and the reconstructed time-domain realisations track the injected processes. If the paper is right, it gives an independent, achromatic-red-noise-insensitive channel for validating conventional ToA-based noise analyses, which matters for pulsar timing array sensitivity.","feed_headline":"Model pulsar noise straight from dispersion measures","feed_subtitle":"A Fourier-domain Gaussian process on epoch-wise DMs recovers injected DM and solar-wind noise without ToA delays.","key_machinery":"The central object is the epoch-wise DM time series and its decomposition into deterministic and stochastic parts. The load-bearing mechanism is a Fourier-domain Gaussian process on DMs: a Fourier basis matrix φ and amplitude covariance Φ build the effective noise covariance Ξ = ζ + φΦφ^T, where ζ is the diagonal white-noise covariance; a power-law spectral density with dispersive scaling ν^-2 controls the stochastic amplitudes. This lets the likelihood be analytically marginalised over the noise amplitudes, leaving only hyper-parameters and deterministic coefficients to be sampled.","core_discovery":"The central claim is that dispersive processes leave a direct imprint on the DM time series itself, so the noise covariance can be constructed from DM fluctuations (δDM) rather than from delays. The total DM is decomposed as DM0 plus polynomial DM and solar-wind terms, Fourier-basis stochastic DM noise and solar-wind noise, and white measurement noise. The likelihood is multivariate Gaussian with a diagonal white-noise covariance, and the Fourier amplitudes are integrated out analytically via a reduced-rank covariance. Because no ToA delays enter the model at any point, the treatment is almost unaffected by achromatic red noise and applies equally to narrowband-derived or wideband-measured D","pith_inferences":["If it holds on real data, PSRDISP could flag occasions where ToA-based analyses misattribute chromatic noise because of spectral leakage, since the two methods should agree when models are correct.","A straightforward extension is to model the solar wind as a spatially correlated process across an array of pulsars; the present framework already has the Fourier-basis structure to accommodate this.","A testable use is to run both this DM-only method and a joint ToA+DM wideband fit on the same dataset; agreement would corroborate both, and disagreement would localise the source of bias.","The main limitation the authors leave open—scattering-induced biases in DM estimates—suggests the method could be extended by jointly fitting a scattering term with a known frequency scaling, providing a path to separate scattering from true dispersive noise."],"forward_implications":["Provides an independent, ToA-free cross-check on conventional single-pulsar noise analyses.","Characterisation of dispersive noise becomes largely insensitive to achromatic red noise and white ToA noise.","The same machinery applies to both narrowband and wideband DM estimates, as long as epoch-wise DMs are independent.","The recovered DM-noise and solar-wind-noise parameters can be used to tune or validate the noise budget of a pulsar timing array.","In simulations, recovered noise amplitudes and spectral indices are consistent with injections, and time-domain realisations track the injected noise process."],"fun_headline_variants":["No ToA delays: pulsar noise from epoch-wise DMs","Pulsar noise from DM fluctuations, not ToA delays","Agnostic pulsar noise analysis using epoch-wise DM fluctuations","PSRDISP: Gaussian process on DMs models pulsar noise","Dispersion measures alone capture pulsar noise, no ToA needed"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method assumes epoch-wise dispersion measures are statistically independent with a diagonal covariance matrix, and that the DMs are not biased by scatter broadening; if either fails, the recovered noise parameters are no longer reliable.","fun_headline_variants_meta":{"raw":{"variants":["No ToA delays: pulsar noise from epoch-wise DMs","Pulsar noise from DM fluctuations, not ToA delays","Agnostic pulsar noise analysis using epoch-wise DM fluctuations","PSRDISP: Gaussian process on DMs models pulsar noise","Dispersion measures alone capture pulsar noise, no ToA needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001368,"raw_usage":{"total_tokens":5392,"prompt_tokens":759,"completion_tokens":4633,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":4543}},"tokens_in":503,"tokens_out":4633,"duration_ms":35762,"temperature":1.0,"reasoning_tokens":4543,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T01:54:38.648110+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take DM time series produced by a global fit (for example a piecewise-constant or spline solution) that is known to create inter-epoch correlations, run PSRDISP on them, and check whether the recovered noise amplitude and spectral index shift outside the nominal posterior uncertainties. Alternatively, inject a scattering delay proportional to ν^-4 into simulated sub-banded arrival times before extracting DMs, and test whether the dispersive-noise parameters become biased.","supporting_citations":[],"review_version":3}