{"id":"5e7e83d4-4eb9-40d3-aa05-2689c70762df","arxiv_id":"2412.11894","paper_version":4,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A ridge-regularized Gaussian form of the PTA likelihood lets pulsars be analyzed individually first and then combined globally for a gravitational-wave background search, reproducing standard ENTERPRISE results on EPTA DR2new data.","lead":"The paper presents a regularized, Fourier-space version of the pulsar timing array likelihood, so each pulsar is reduced to a Gaussian summary that a later full-array gravitational wave search can reuse. It allows white noise to be marginalized out per pulsar and is designed to make combining radio and gamma-ray pulsar data easier.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Gaussian summary in Eqs. (18)-(19) has no error diagnostic; the failed PCA experiment in Appendix C shows moment-matched compression can silently narrow posteriors on EPTA data, so 'retains all information' is not established.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the Gaussian summary approximation of Eq. (18)-(19). My stress test agrees, and I found no additional concern more central. The fixed-theta version of the derivation is standard Gaussian algebra and is correct; the real-data validation is genuine and the authors' inclusion of the failed PCA attempt is honest. However, the central methodological claim that the two-step procedure retains all information when white-noise and deterministic parameters are marginalized rests entirely on the unquantified accuracy of the single-Gaussian summary. The paper supplies no posterior-predictive check, no measure of Gaussianity, no sensitivity run, and no injection test for Eq. (18)-(19). The Appendix C result demonstrates on the paper's own data that a moment-based compression can silently narrow posteriors, which raises the prior probability that the same failure can occur in the main method. A concrete rerun of Step 2 with the exact mixture summary would settle whether the approximation error is negligible for the EPTA GWB results or large enough to shift the headline posteriors. Because the reader already marked the verdict CONDITIONAL and flagged exactly this assumption, I do not move the verdict; I would only strengthen the requested condition: the paper should provide the mixture-vs-Gaussian comparison before claiming full information retention.","tokens_in":22980,"tokens_out":4098,"duration_ms":42249,"concrete_test":"Run Step 1 exactly as in the paper on EPTA DR2new (at minimum for J1738+0333, ideally all 25 pulsars), saving the theta chain. For each saved theta_i, draw Fourier-coefficient samples from N(ahat0(theta_i),Sigma0(theta_i)) to build the exact mixture marginal p(a|delta_t,rho0). Replace the single Gaussian N(ahat0,Sigma0) in Eqs. (21)-(22) with this exact mixture summary and re-run the Step-2 GWB search. If the resulting posteriors on log10 A_GWB and gamma_GWB differ from those obtained with Eq. (21) by more than the Monte Carlo sampling error, the Gaussian summary is biased and the 'retains all information' claim must be qualified; if they agree within error, the approximation is validated for this dataset.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing approximation is the replacement, in Eqs. (18)-(19), of the true marginal p(a|delta_t,rho0) = integral dtheta N(a|ahat0(theta),Sigma0(theta)) p(theta|delta_t,rho0) by a single Gaussian with matched first two moments. This is exact only if the conditional means depend linearly on theta and the conditional covariances are theta-independent; the paper's heuristic that theta are 'not strongly correlated' with a (Sec. II D) is not the correct criterion, since a Gaussian mixture can be strongly non-Gaussian even when those correlations vanish. No diagnostic, error estimate, or injection test is provided for Eq. (19). This matters because Eq. (21) multiplies the summary by the reweighting ratio p(a|rho)/p(a|rho0), so any summary error propagates directly into the Step-2 GWB posterior. The concern is not hypothetical: Appendix C applies a related moment/PCA compression on the same EPTA pulsar J1738+0333 and finds the resulting RN posteriors are visibly narrower than the full time-domain result (Fig. 4), a silent loss of information. The Fig. 2 comparison is visual only, and both compared analyses use the same Gaussian-summary route for Step 1, so it cannot validate the approximation. The analytic-equivalence claim in the Conclusions is valid only when theta are held fixed; the marginalized case, which the abstract describes as retaining 'all information', is not covered by that proof.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-step, Fourier-domain reformulation of the pulsar timing array likelihood. Step 1 runs per-pulsar inference over white-noise and deterministic-signal parameters while holding the red-noise/DM/GWB hyperparameters fixed at reference values rho0, and compresses the result into a Gaussian summary N(a|ahat0,Sigma0) for the Fourier coefficients. Step 2 multiplies these per-pulsar summaries by the reweighting ratio p(a|rho)/p(a|rho0) and by the hyperparameter prior, yielding a full-array posterior for the GWB and red-noise parameters; the Fourier coefficients are then marginalized analytically in Eqs. (22)-(23). The authors show that when the white-noise/deterministic parameters theta are held fixed, the construction is algebraically identical to the standard time-domain marginalized likelihood, and they demonstrate the method on the EPTA DR2new dataset: a single-pulsar noise analysis for J1738+0333 (Fig. 1) and a 25-pulsar GWB search (Fig. 2). The main approximation is introduced in Eqs. (18)-(19), where the theta-marginalized conditional posterior of the Fourier coefficients is replaced by a single Gaussian with matched first two moments.","tokens_in":23291,"tokens_out":6481,"duration_ms":61753,"significance":"If the two-step construction is truly information-preserving, this is a useful contribution to PTA methodology. The Gaussian algebra in Eqs. (12)-(23) is clean and standard; the reweighting formula is derived inside the paper rather than assumed; and the comparison against an independent codebase (ENTERPRISE) on real EPTA DR2new data is a concrete implementation check. The authors also make code and tutorials available. The proposed split into per-pulsar summaries and an array-level search is attractive for combining radio and gamma-ray datasets and for reusing expensive per-pulsar noise analyses. However, the 'retains all information' claim currently rests on the unsupported Gaussian-mixture approximation in Eq. (19). Appendix C shows that a closely related moment/PCA compression can silently narrow posteriors on the same class of data, and no diagnostic or injection test is given for Eq. (19). The practical significance of the method therefore depends on closing this gap.","major_comments":[{"comment":"The final step of Eq. (18)-(19) replaces the true marginal p(a|delta_t,rho0)=integral dtheta N(a|ahat0(theta),Sigma0(theta)) p(theta|delta_t,rho0), which is generally a Gaussian mixture, with the single Gaussian N(a|ahat0,Sigma0) obtained by matching first and second moments. This replacement is exact only under restrictive conditions (e.g., ahat0(theta) linear in theta and Sigma0(theta) independent of theta); the stated heuristic that the white-noise parameters are 'not strongly correlated' with the Fourier coefficients is not the correct criterion, since a symmetric two-component mixture can have zero correlation and still be strongly non-Gaussian. The approximation error propagates directly into Eq. (21) through the reweighting ratio p(a|rho)/p(a|rho0), so it biases the Step-2 GWB posterior whenever the summary is inaccurate. The paper provides no error estimate, no diagnostic, and no injection test for this step. This is load-bearing because the abstract's claim that the method 'retains all information' goes beyond the Conclusions' equivalence statement, which is explicitly restricted to the case where theta is held fixed.","section":"Sec. II D, Eqs. (18)-(19)"},{"comment":"The admitted failure of the PCA-based moment compression in Appendix C is directly relevant to the soundness of Eq. (19). On the same EPTA pulsar J1738+0333, the moment-summary approach of Eqs. (C10)-(C11) produces RN posteriors that are visibly narrower than the full time-domain result (Fig. 4). The authors attribute this to poor frequency coverage and non-Gaussianity, but the same pulsar is used in Fig. 1 to illustrate Eq. (19). This does not by itself invalidate Eq. (19), but it removes the default assumption that Gaussian moment summaries are benign in this setting. The manuscript needs either a diagnostic that distinguishes the regime where Eq. (19) is safe from the regime where the Appendix C compression fails, or a restriction of the information-preservation claim to cases satisfying that diagnostic.","section":"Appendix C, Fig. 4"},{"comment":"The validation in Fig. 2 is visual only and its scope is narrower than the abstract suggests. The blue reference curve is a standard time-domain analysis with the white-noise parameters fixed to single-pulsar maximum-likelihood values, while the orange curve uses the Fourier-domain summary with those parameters marginalized. Agreement between the two therefore demonstrates the fixed-theta identity of Eqs. (20)-(23) together with the practical insignificance of the WN marginalization for this dataset, but it does not validate the information-preservation property of Eq. (19). A quantitative comparison (e.g., posterior overlap or a Bayes-factor check between the two formulations) and, more importantly, an injection-recovery test with a known GWB signal would be needed to support the 'retains all information' wording in the abstract.","section":"Sec. III B, Fig. 2"}],"minor_comments":[{"comment":"Equation (18) appears to contain a typo: the line after the proportionality sign repeats the integral twice ('integral dtheta N(...)p(theta) x integral dtheta N(...)p(theta)'), whereas the intended expression is a single integral. Please correct.","section":"Eq. (18)"},{"comment":"The notation for the fixed-theta conditional quantities in Eqs. (16)-(17) is not kept distinct from the marginalized quantities in Eq. (19); the primes/superscripts are introduced and then dropped, and Table I lists only ahat0 and Sigma0. Please introduce consistent symbols such as ahat0(theta) and Sigma0(theta) for the fixed-theta quantities and use them throughout.","section":"Eqs. (16)-(17) and Table I"},{"comment":"The text says the regularized-likelihood posteriors are 'equivalent' to a full single-pulsar noise analysis while also noting that they are slightly wider because the white-noise parameters are marginalized. Please rephrase to distinguish 'equivalent in sampling target' from 'numerically identical'.","section":"Sec. III A"},{"comment":"The sentence 'the distribution at the denominator has to be wider than the condition number at the numerator' presumably means that the covariance matrix of the denominator distribution must be larger in the relevant sense; the reference to 'condition number' is confusing and should be rewritten.","section":"Appendix A 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the central fixed-theta derivation is sound. My recommendation is driven by the unsupported Gaussian-summary approximation: it is the single load-bearing point for the information-preservation claim, and the paper's own Appendix C provides a concrete warning that moment compression can silently narrow posteriors in this setting. I would be satisfied by a revision that either supplies a diagnostic/injection test for Eq. (19) on the EPTA data or scales back the abstract's information-retention claim to the fixed-theta case and clearly labels the marginalized case as an approximation to be validated per application."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper gives PTA analysts a modular two-step workflow: per-pulsar Fourier summaries as Gaussian means and covariances, then an array-level reweighted Gaussian likelihood for the GWB search. The central derivation, Eqs. (21)-(23), is clean and correct, and the validation on the 25-pulsar EPTA DR2new dataset against ENTERPRISE is genuine evidence that the pipeline works. The authors also deserve credit for reporting the failed PCA variant and for stating plainly that there is no speedup over existing codes.\n\nThe genuinely new piece is the reweighting trick: regularize with a ridge prior at rho0, summarize each pulsar's posterior as a Gaussian, then multiply by the prior ratio p(a|rho)/p(a|rho0) to undo the regularization. That specific derivation is not in the cited literature, and the two-step separation is a real contribution even though all ingredients are standard.\n\nThe load-bearing approximation is Eq. (18)-(19): after marginalizing over white noise and deterministic parameters, the per-pulsar posterior of the Fourier coefficients is replaced by a single Gaussian matched to the conditional means and covariances. There is no error estimate or diagnostic for this step. The heuristic that theta are \"not strongly correlated\" with a is not the right criterion; a mixture can be strongly non-Gaussian even when those correlations vanish. So the abstract's \"retains all information\" is not established for the marginalized case. The analytic equivalence claim is proved only when theta is held fixed, which the conclusions do state, but the abstract overstates it. The Fig. 2 match is real but visual. I do not agree with the stress-test claim that both compared analyses use the same Gaussian-summary route: the blue curve is the standard time-domain ENTERPRISE likelihood with fixed white noise, so the comparison does provide some validation. It just does not bound the approximation error on noisier or more complex data. The Appendix C PCA failure is the honest warning: the same kind of compression can silently narrow posteriors.\n\nThis paper is for PTA methodologists, especially anyone wanting to combine radio and gamma-ray data or modularize noise and GWB inference. It deserves serious peer review. A referee should ask for diagnostics or injection tests on the summary approximation, and for a qualified abstract. I would accept it with revision, not desk reject.","headline":"A useful two-step PTA workflow with correct Gaussian algebra and real-data validation; the main caveat is the unquantified Gaussian-mixture summary, whose limits the paper itself demonstrates in Appendix C.","tokens_in":23862,"tokens_out":2151,"would_cite":true,"duration_ms":20944,"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":"The paper shows that a pulsar timing array likelihood can be regularized into per-pulsar Gaussian Fourier summaries, so the array-level gravitational-wave background search becomes an analytic reweighting of those summaries.","keywords":["pulsar timing arrays","gravitational wave background","Fourier domain likelihood","likelihood regularization","Bayesian inference","Hellings-Downs correlation","gamma-ray pulsars","two-step analysis"],"falsifier":"Simulate a PTA in which a strong deterministic signal overlaps the lowest Fourier frequencies, run both the two-step method and the full time-domain likelihood, and compare the GWB amplitude posteriors; if they differ by more than sampling noise, the Gaussian summary has discarded information the true posterior retains.","tokens_in":22690,"feed_emoji":"🔭","tokens_out":10951,"duration_ms":91720,"temperature":0.7,"pith_summary":"This paper tries to make pulsar timing array searches modular: instead of fitting one giant joint model over every pulsar's noise and the gravitational-wave background (GWB), it proposes a two-step pipeline. In Step 1 each pulsar is analyzed alone, and all its timing information is compressed into a Gaussian distribution over Fourier coefficients, with a ridge-regression prior used to make the likelihood normalizable. In Step 2 the full array combines those per-pulsar summaries and reweights them analytically to search for the common, spatially correlated signal. On a 25-pulsar dataset the two-step results match the standard time-domain posteriors for both single-pulsar noise and the GWB amplitude and spectral index. If the method is correct, array-level searches become reusable artifacts that no longer need to re-run each pulsar's detailed noise model.","feed_headline":"Compress each pulsar into a Gaussian, then search the array","feed_subtitle":"Each pulsar becomes one Gaussian; the array-level search reweights it to match the standard 25-pulsar result.","key_machinery":"The load-bearing object is the regularized per-pulsar likelihood: a ridge-regression prior $p(a|\\rho_0)$ with fixed reference hyperparameters is added after marginalizing over timing-model parameters, turning the per-pulsar likelihood into an exact Gaussian $\\mathcal{N}(a|\\hat{a}_0,\\Sigma_0)$ in the Fourier coefficients. Step 2 divides out that same Gaussian through the analytic ratio $p(a|\\rho)/p(a|\\rho_0)$, so the dummy prior cancels and the true prior, including inter-pulsar correlations, takes over. Assembling $\\hat{a}_0$ and $\\Sigma_0$ from MCMC samples of the noise parameters uses the covariance update formula of Eq. (19), and the flat-tail power-law reference spectrum keeps the covariance matrices well conditioned.","core_discovery":"The central result is Eq. (21): the full-array posterior factorizes as the product of per-pulsar Gaussian summaries $\\mathcal{N}(a_k|\\hat{a}_{0,k},\\Sigma_{0,k})$, a reweighting ratio $p(a|\\rho)/p(a|\\rho_0)$ that removes the regularization prior, and the array-level prior $p(a|\\rho)$ that carries the Hellings-Downs correlation, the quadrupolar angular correlation expected from a gravitational-wave background. This factorization is an analytical identity when the white-noise and deterministic-signal parameters are held fixed, and an approximation when they are marginalized over. The paper's concrete demonstration is that this Fourier-space formulation recovers the same single-pulsar noise and GWB posteriors as the standard time-domain likelihood on the 25-pulsar dataset used for validation.","pith_inferences":["The per-pulsar Gaussian summaries could be archived as a standard data product, letting future array-level searches test new common-signal models without redoing single-pulsar noise inference.","A practical safety diagnostic would compare the Gaussian summary's covariance with the sample covariance of Fourier coefficients drawn in the full model; where they diverge, Step 1 should be rerun rather than trusted.","The reference spectrum $\\rho_0$ is a free knob, and its choice affects conditioning and summary fidelity, so the particular flat-tail choice used here should be validated on each pulsar rather than assumed universal.","A direct testable extension is to pass gamma-ray photon-to-photon Fourier samples through Eq. (21) and compare a combined radio+gamma-ray GWB posterior with the radio-only one."],"forward_implications":["A GWB search can be split into per-pulsar noise analyses and a global Fourier-domain analysis that never touches the raw arrival times.","White-noise and deterministic signals, such as exponential DM dips, can be marginalized over in the per-pulsar step instead of being fixed to best-fit values, which is statistically more complete.","The Fourier summaries stay reusable: changing the array-level model only changes the reweighting factor and the prior, not the per-pulsar processing.","Radio and gamma-ray pulsar data, which currently use different likelihoods, could feed the same Fourier-space summary representation.","On the 25-pulsar dataset, the two-step search reproduces the standard GWB amplitude and spectral-index posteriors."],"supporting_citations":[{"why":"Supplies the time-domain likelihood and its analytic marginalization over timing and Fourier parameters, which the new formulation must reproduce.","marker":"[38,41]"},{"why":"Provides the standard time-domain analysis implementation used as the baseline for the posterior comparisons.","marker":"[36]"},{"why":"Gives the 25-pulsar dataset used in the single-pulsar and GWB search demonstrations.","marker":"[49]"},{"why":"Defines the Hellings-Downs correlation that the array-level prior uses to distinguish a gravitational-wave background from noise.","marker":"[7]"},{"why":"Supplies the per-pulsar noise models (including chromatic noise and DM dips) that Step 1 follows for the dataset.","marker":"[50]"},{"why":"Introduces the photon-to-photon gamma-ray likelihood that the Fourier-space formulation is designed to admit.","marker":"[30]"}],"fun_headline_variants":["Compress pulsars to Gaussians, then search for GWB","Fourier-space factorization simplifies PTA data combination","Per-pulsar Gaussian summaries enable unified PTA analysis","One Gaussian per pulsar, then combine for GWB search","Regularized likelihood path to Fourier-space GWB search"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim stands or falls on treating the per-pulsar posterior of the Fourier coefficients, after white-noise and deterministic parameters are marginalized, as a single Gaussian; if any of those parameters is strongly correlated with the Fourier coefficients, the summary is lossy and the reweighting is biased.","fun_headline_variants_meta":{"raw":{"variants":["Compress pulsars to Gaussians, then search for GWB","Fourier-space factorization simplifies PTA data combination","Per-pulsar Gaussian summaries enable unified PTA analysis","One Gaussian per pulsar, then combine for GWB search","Regularized likelihood path to Fourier-space GWB search"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000727,"raw_usage":{"total_tokens":3231,"prompt_tokens":894,"completion_tokens":2337,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":2255}},"tokens_in":510,"tokens_out":2337,"duration_ms":16504,"temperature":1.0,"reasoning_tokens":2255,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:31:11.757875+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a PTA in which a strong deterministic signal overlaps the lowest Fourier frequencies, run both the two-step method and the full time-domain likelihood, and compare the GWB amplitude posteriors; if they differ by more than sampling noise, the Gaussian summary has discarded information the true posterior retains.","supporting_citations":[{"cited_title":"regularization","cited_arxiv_id":null,"evidence_quote":"Provides the standard time-domain analysis implementation used as the baseline for the posterior comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the 25-pulsar dataset used in the single-pulsar and GWB search demonstrations."},{"cited_title":"In reality, we are actually interested in pða; ρjδtÞ","cited_arxiv_id":null,"evidence_quote":"Defines the Hellings-Downs correlation that the array-level prior uses to distinguish a gravitational-wave background from noise."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the per-pulsar noise models (including chromatic noise and DM dips) that Step 1 follows for the dataset."},{"cited_title":"Kerr, Astrophys","cited_arxiv_id":null,"evidence_quote":"Introduces the photon-to-photon gamma-ray likelihood that the Fourier-space formulation is designed to admit."}],"review_version":1}