{"id":"b76a9b25-e662-4857-9d7c-0d6757364f7f","arxiv_id":"2412.07068","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A semi-analytic covariance matrix for ACT DR6 power spectra, combining inhomogeneous survey-depth weights and a simulation-based eigenbasis correction, reaches sub-percent agreement with Monte Carlo simulations.","lead":"ACT's latest data release needs accurate error bars on its cosmic microwave background power spectra. This paper builds a semi-analytic covariance matrix that matches expensive simulations to better than one percent after a new simulation-based correction, and could be reused by future CMB experiments.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The sub-percent validation is internal to the tiled Monte Carlo model; the known large-scale polarized noise excess and the Appendix E ~30% large-scale polarization discrepancies leave the Monte-Carlo-as-truth premise incompletely tested.","rationale":"The reader's weakest assumption is the Monte Carlo covariance as ground truth, and I agree that this is the load-bearing premise. My concern is more specific than 'simulations might be wrong': the paper itself provides two pieces of evidence that the tiled Monte Carlo is not an unbiased proxy for ACT DR6 at large-scale polarization - the measured noise excess in Fig. 12 and the roughly 30% unfiltered discrepancies in Appendix E. Because the d^2 validation draws from the same tiled model, it cannot detect this bias. The directional-wavelet comparison does not fully address the concern, since both noise models share the filter-induced excess. The proposed renormalization test directly removes the known excess and would show whether the quoted sub-percent agreement survives anchoring to data. If it does, the conditional verdict is sufficient; if it does not, the suitability claim needs to be restricted to scales where the Monte Carlo is unbiased. No internal inconsistency was found in the derivation of Eq. 13 or in the correction procedure, and the independent-simulation split (1,000 for correction, 600 for validation) is a commendable safeguard. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":35626,"tokens_out":8452,"duration_ms":97700,"concrete_test":"Renormalize the tiled noise simulations so that, after the Fourier-space filter, the ensemble-mean noise pseudospectrum matches the measured dr6.02 data noise pseudospectrum at all retained multipoles (or, equivalently, replace the tiled noise draws with Gaussian realizations built from the measured data noise power and inverse-variance weights). Re-run the Sec. 4.3 correction using this data-anchored Monte Carlo covariance, then evaluate the d^2 distribution on 600 independent simulations from the same renormalized ensemble. If any retained EE diagonal changes by more than about 1-2% or the d^2 mean moves away from 1763 by more than about 0.3%, the Monte-Carlo-as-truth premise fails for the affected scales. As a cross-check, repeat the Appendix E no-holes comparison with the Fourier-space filter and the actual DR6 scale cuts included.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.2 concedes that the tiled simulations have excess large-scale polarized noise after the Fourier-space filter (Fig. 12), inflating a few low-ell EE error bars by 1-2%. Appendix E (Figs. 14-15) reports roughly 30% Monte Carlo versus analytic discrepancies for large-scale polarization in the unfiltered no-holes test, attributed to a breakdown of the approximate noise model (Eq. 10) as well as NKA. The headline d^2 validation (Sec. 5.1, Fig. 9) uses 600 simulations drawn from the same tiled model that supplies the Monte Carlo covariance used for the Sec. 4.3 correction. It therefore validates the corrected matrix against the model, not against ACT DR6 data. The claim that the semi-analytic matrix is suitable for the DR6 likelihood requires the tiled Monte Carlo covariance to be an unbiased proxy for the true data covariance at the sub-percent level; the paper quantifies only the diagonal effect on three bins and leaves open the possibility that the corrected covariance inherits a model-level large-scale polarization bias. The directional-wavelet comparison in Sec. 5.2 is a useful robustness test, but it compares two models sharing the same filter-induced excess and does not anchor to the data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a semi-analytic pipeline for the ACT DR6 power-spectrum covariance matrix. The authors generalize the MASTER pseudo-C_ell covariance to include inhomogeneous survey depth through separate effective signal and noise masks (Eq. 13, Appendix B.1), use the improved narrow-kernel approximation, and model the Fourier-space filter with isotropic transfer functions t^alpha with exponents alpha_2pt and alpha_4pt fitted to dedicated simulations (Section 4.1.2, Appendix B.2). The analytic matrix is then corrected with a new shrinkage procedure: the Monte Carlo covariance is rotated into the analytic eigenbasis, its block-diagonal elements are smoothed with a Gaussian process, and the result is rotated back (Eqs. 17-19). Validation uses 1,600 tiled-model simulations: the uncorrected inhomogeneous matrix gives a simulation chi-squared mean of 1812.8 +/- 1.5 against an expected 1763 (about 2.8% high, corresponding to roughly 1.4% error-bar underestimation), while the homogeneous prescription is about 15.7% discrepant; after correction based on 1,000 simulations and evaluation on the independent 600, both corrected matrices have chi-squared means of 1764.7 and 1767.3, respectively, consistent with 1763. The paper also compares against a directional-wavelet noise model, tests the effect of point-source holes, and shows that their correction outperforms a correlation-matrix-preserving alternative.","tokens_in":35929,"tokens_out":6259,"duration_ms":72221,"significance":"If the claimed accuracy holds, this is a timely and useful methodological contribution. The explicit derivation of the polarized pseudo-spectrum covariance in Appendix B.1, the use of an independent 600-simulation evaluation set, and the systematic comparison of two correction schemes are all strengths. The paper also gives a practical framework that should transfer to Simons Observatory and other high-resolution CMB experiments. However, the validation is carried out almost entirely against the authors' own tiled Monte Carlo model, and the manuscript itself documents large-scale polarized-noise deficiencies in that model. The significance of the paper therefore depends on how much weight one places on the Monte-Carlo-as-truth assumption, and the suitability claim for the ACT DR6 likelihood is currently stronger than the evidence provided.","major_comments":[{"comment":"The correction and the headline validation share the same tiled Monte Carlo model as reference. Section 5.1 correctly uses independent simulation draws for the d^2 test, but that test cannot detect any bias that is common to the model and the corrected matrix. Section 5.2 and Figure 12 show that the tiled simulations have excess large-scale polarized noise after the Fourier filter, inflating some low-ell EE error bars by 1-2%, and Section 4.3 constructs the corrected matrix from exactly this Monte Carlo covariance. Because the paper claims the matrix is well-suited for the DR6 likelihood, it should either propagate the known simulation bias through the full covariance, including off-diagonal blocks rather than only the quoted three bins, or explicitly restrict the suitability claim to scales and spectra for which the simulation model is validated against the data.","section":"Sections 5.1-5.2 and 4.3, Figure 12"},{"comment":"The unfiltered, no-holes control test reports roughly 30% Monte Carlo versus analytic discrepancies for large-scale polarization and attributes them to a breakdown of the approximate noise model of Equation 10 as well as to the NKA. This is a load-bearing limitation because the baseline comparison in Figure 8 cannot distinguish a failure of the analytic ansatz from the known simulation noise excess; both are partly absorbed by the Fourier-filter transfer-function fit described in Appendix B.2. The paper should provide a quantitative demonstration that the Fourier filter and the DR6 scale cuts suppress this polarization discrepancy in the baseline configuration, or it should soften the conclusion that the semi-analytic matrix is well-suited for the DR6 likelihood.","section":"Appendix E, Figures 14-15"},{"comment":"The 'uncorrected analytic' covariance is not fully analytic: the exponents alpha_2pt and alpha_4pt are fitted to 500 mock simulations, and the text itself notes that the fit can absorb NKA-induced errors. The reported better-than-3% agreement therefore conflates the analytic ansatz with the fitted transfer-function model. Please report the statistical uncertainty on the fitted exponents and the sensitivity of the quoted 3% and 1.4% figures to those values, so that the genuinely analytic part of the prescription can be separately assessed.","section":"Section 4.1.2, Appendix B.2"}],"minor_comments":[{"comment":"The polarization four-point exponent is quoted as alpha_4pt = 0.504 in the Figure 4 caption but as 0.469 in the Appendix B.2 fit shown in Figure 11; please clarify whether these are different polarization cases or different fit setups, and specify which value enters the baseline covariance.","section":"Figure 4 versus Figure 11"},{"comment":"The footnote states that minor revisions were made to the dr6.02 maps, masks, scale cuts, fiducial signal spectra, and simulations after the paper was completed, but that these should not affect the conclusions; please provide a versioned record of the exact inputs used for the reported numbers, since the unverifiable nature of this statement makes the numerical results difficult to reproduce.","section":"Footnote 1"},{"comment":"The statement 'This procedure has no free parameters' is confusing immediately after two Gaussian-process hyperparameters are optimized; consider rephrasing to 'no hand-tuned parameters' or explicitly stating which quantities are held fixed.","section":"Section 4.3, text after Eq. (19)"},{"comment":"In the second robustness test, the phrases 'TT, TE, and EE spectra' and 'TB, EB, and BB spectra' should be clarified as power-spectrum data vectors or covariance blocks rather than individual spectra, since each of these labels denotes multiple array pairs.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":"This is a solid methods paper with a careful derivation and a mostly well-designed validation scheme. The main question for the editor is whether the suitability claim for ACT DR6 can stand without a more direct test of the Monte Carlo reference model at large-scale polarization, where the authors themselves document 30%-level discrepancies in a control test. I see this as fixable either by additional propagation/robustness checks or by explicitly narrowing the claim, and I would not recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does what it says: builds a semi-analytic covariance for ACT DR6 that is accurate at the percent level relative to the Monte Carlo ensemble, and the validation is done the right way — 1,000 simulations for the correction, an independent 600 for the chi-squared test. That split is a genuine strength.\n\nWhat's new: first application of the improved narrow-kernel approximation to a CMB power-spectrum covariance; an inhomogeneous effective-mask treatment that extends the Planck 2016 structure to ACT's non-white, stripy noise; a Fourier-filter transfer-function model with fitted two- and four-point exponents; and a shrinkage method based on the eigenbasis rotation that outperforms the common correlation-matrix-preserving correction. The analytic derivation in Eq. 13 and the appendices is clean, and the paper is transparent about where the approximations sit.\n\nSoft spots: the reference frame is the Monte Carlo covariance, and the simulations have known imperfections — excess large-scale polarized noise after the Fourier filter (Fig. 12), inflating a few EE bins by 1-2%, and Appendix E shows ~30% level analytic-versus-MC discrepancies for large-scale polarization in the unfiltered no-holes test. That means the sub-percent agreement is with the tiled simulation model, not with the data itself. The authors acknowledge this and quantify the diagonal effect on the DR6 scale cuts; it is a caveat, not a fatal flaw. The bigger issue for reproducibility is that no code or data products are released, so outside the collaboration the central numbers cannot be checked. The fitted alpha exponents and the GP smoothing are arguably ad-hoc but they are tested against the independent simulations, and the comparison to the homogeneous prescription shows the method is not fragile.\n\nVerdict: this is a serious methods paper. The central claim holds up as a claim about the simulations; the jump to 'suitable for the DR6 likelihood' inherits the simulation-fidelity assumption. CMB analysts, especially those working on ground-based data and future SO analyses, will get real value from it. It deserves a serious peer review. I would recommend acceptance with the request that the authors state the simulation-fidelity caveat prominently and make the code and the corrected matrix available if the collaboration permits.","headline":"Solid, careful covariance pipeline for ACT DR6; the sub-percent validation is real but conditional on the simulation model, and the paper is honest about that.","tokens_in":36532,"tokens_out":3189,"would_cite":true,"duration_ms":34461,"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 claims that a semi-analytic covariance matrix, upgraded to include inhomogeneous survey depth and Fourier-filter transfer functions, matches 1,600 Monte Carlo simulations to better than 3% for ACT DR6 power spectra, and that a…","keywords":["cosmic microwave background","ACT DR6","power spectrum covariance","MASTER formalism","Monte Carlo simulations","matrix shrinkage","semi-analytic covariance","inhomogeneous noise"],"falsifier":"Build a covariance from simulations whose large-scale polarized noise is forced to match the data at ℓ below about 500 and rerun the chi-squared test; if the corrected matrix no longer lands within about 0.3% of expectation, the claimed accuracy was absorbing a simulation artifact rather than modeling the data.","tokens_in":35410,"feed_emoji":"📡","tokens_out":4627,"duration_ms":47935,"temperature":0.7,"pith_summary":"The paper tries to establish that a semi-analytic covariance matrix, built from the MASTER pseudo-Cℓ formalism but upgraded to include inhomogeneous survey depth and Fourier-filter transfer functions, can describe ACT DR6 power-spectrum error bars accurately enough for cosmological inference. The uncorrected inhomogeneous prescription agrees with 1,600 Monte Carlo simulations to about 3%, whereas the previous homogeneous prescription differs by about 16%; after a new simulation-based correction both reach sub-percent agreement. If the claim is right, the DR6 likelihood can use a covariance nearly as accurate as Monte Carlo but far cheaper to compute, and future ground-based CMB experiments with atmospheric noise inherit the same recipe.","feed_headline":"Covariance recipe cuts CMB error mismatch to under 1 percent","feed_subtitle":"Adding survey-depth weighting and a simulation-based fix brings ACT DR6 error bars in line with Monte Carlo truth.","key_machinery":"The load-bearing object is the rotated Monte Carlo covariance matrix Σ_R = $Σ_A^{{-1/2}}$ Σ_M $Σ_A^{{-T/2}}$; if the analytic and Monte Carlo eigenbases nearly coincide, its diagonal gives the smooth ratio of the two eigenspectra, which is fit block-by-block with Gaussian processes and rotated back to produce the corrected matrix. The analytic part rests on the MASTER pseudospectrum covariance using the improved narrow-kernel approximation (INKA), which replaces each steep spectrum by a normalized pseudospectrum, together with separate effective masks for signal and noise and simulated transfer functions for the Fourier filter.","core_discovery":"The paper's central claim is that the covariance bias introduced by the MASTER approximations can be reduced to the percent level for ACT DR6 by three moves: using different effective masks for signal and noise to encode survey depth, modeling the Fourier-space filter with two-point and four-point isotropic transfer functions whose exponents are fit from dedicated simulations, and correcting the remaining bias with a shrinkage step that assumes the analytic and Monte Carlo covariances share an eigenbasis and that their eigenspectrum ratio is smooth. The validation is a chi-squared test on simulated data vectors: the simulation mean chi-squared improves from 1812.8 for the uncorrected inhomogeneous prescription and 2039.0 for the homogeneous prescription toward the expected 1763, and after correction becomes 1764.7 and 1767.3, within about 0.3% of nominal. The paper concludes that the corrected semi-analytic matrix, built from either prescription, is well-suited for use in the ACT DR6 likelihood.","pith_inferences":["If the eigenbasis assumption degrades for surveys with sharper mask holes or stronger striping, the shrinkage could require more than one Gaussian-process length scale; a natural test is to inspect the off-diagonal structure of Σ_R for such a survey.","The admitted excess large-scale polarized noise in the simulations means the current validation is probably pessimistic for large-scale EE bins, so forcing simulated noise to match the data at ℓ below about 500 could shift those error bars down by 1 to 2 percent.","The same effective-mask technology that separates signal and noise depth could improve analytic covariances for galaxy clustering surveys with strong depth gradients, not just CMB temperature and polarization.","Because the correction discards bin-wise off-diagonal elements beyond the within-block diagonals, the final covariance implicitly assumes the true cross-bin correlation structure is analytic; this could be tested with a simulation ensemble large enough to resolve |b − b′| ≥ 2 correlations directly."],"forward_implications":["The corrected inhomogeneous matrix can replace the homogeneous default in the ACT DR6 likelihood with sub-percent agreement to simulations.","Because the homogeneous matrix also reaches sub-percent after correction, the new shrinkage method can rescue even roughly 16%-biased analytic inputs.","The eigenbasis/smooth-ratio shrinkage outperforms the common correlation-preserving reweighting, which leaves 1.7% to 4.9% biases in the chi-squared test.","Future CMB experiments with atmospheric noise can use the same pipeline, and the inhomogeneous prescription may need a smaller simulation ensemble than the homogeneous one.","The Fourier-filter transfer-function treatment raises analytic covariances by about 5% at ℓ = 500 relative to the DR4 fixed-exponent relation, changing error bars on that scale."],"supporting_citations":[{"why":"Supplies the improved narrow-kernel approximation (INKA) that the inhomogeneous matrix uses to handle steep signal and noise spectra.","marker":"Nicola et al. (2021)"},{"why":"Establishes the inhomogeneous-survey-depth structure with separate signal and noise effective masks that this paper adapts to ACT.","marker":"Planck Collaboration et al. (2016)"},{"why":"Provides the tiled and directional wavelet noise models used for the Monte Carlo simulations and the earlier finding that homogeneous prescriptions underestimate ACT covariance by up to about 20%.","marker":"Atkins et al. (2023)"},{"why":"Sets the fiducial cosmology and foreground power spectra and the DR4-style Fourier-filter correction that the homogeneous matrix inherits and this paper refines.","marker":"Choi et al. (2020)"},{"why":"Gives the MASTER pseudospectrum estimator and covariance framework that the analytic matrix generalizes.","marker":"Efstathiou (2004)"},{"why":"Defines the spin-dependent coupling matrices (00, 0+, ++, --) used in the multi-field covariance expression.","marker":"Brown et al. (2005)"},{"why":"Proposes the correlation-preserving reweighting that the paper tests and finds inferior to its eigenbasis-based shrinkage.","marker":"Hamimeche & Lewis (2009)"},{"why":"Documents roughly 10% underestimation of Planck covariances when survey depth is ignored, motivating the depth-weighted prescription.","marker":"Li et al. (2023)"}],"fun_headline_variants":["ACT DR6 covariance: Monte Carlo match at 0.3%","Covariance fix shrinks ACT DR6 error bars to 0.3%","Simulation-tuned covariance for ACT DR6: sub-1% errors","Semi-analytic matrix plus shrinkage matches simulations to 0.3%","ACT DR6 covariance: simulation-corrected to sub-1% accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The Monte Carlo covariance built from the simulated noise models is an unbiased stand-in for the true ACT DR6 covariance, so any bias in the simulations is inherited by the corrected matrix.","fun_headline_variants_meta":{"raw":{"variants":["ACT DR6 covariance: Monte Carlo match at 0.3%","Covariance fix shrinks ACT DR6 error bars to 0.3%","Simulation-tuned covariance for ACT DR6: sub-1% errors","Semi-analytic matrix plus shrinkage matches simulations to 0.3%","ACT DR6 covariance: simulation-corrected to sub-1% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001682,"raw_usage":{"total_tokens":6690,"prompt_tokens":991,"completion_tokens":5699,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":5597}},"tokens_in":607,"tokens_out":5699,"duration_ms":42946,"temperature":1.0,"reasoning_tokens":5597,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:10:32.325015+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a covariance from simulations whose large-scale polarized noise is forced to match the data at ℓ below about 500 and rerun the chi-squared test; if the corrected matrix no longer lands within about 0.3% of expectation, the claimed accuracy was absorbing a simulation artifact rather than modeling the data.","supporting_citations":[{"cited_title":"(2023) for this model","cited_arxiv_id":null,"evidence_quote":"Provides the tiled and directional wavelet noise models used for the Monte Carlo simulations and the earlier finding that homogeneous prescriptions underestimate ACT covariance by up to about 20%."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Sets the fiducial cosmology and foreground power spectra and the DR4-style Fourier-filter correction that the homogeneous matrix inherits and this paper refines."},{"cited_title":"A field that follows the model of Equation 9 has zero mean and is Gaussian, so its statistics are fully specified by its covariance","cited_arxiv_id":null,"evidence_quote":"Gives the MASTER pseudospectrum estimator and covariance framework that the analytic matrix generalizes."}],"review_version":1}