{"id":"447be252-db1c-4cb1-b8e5-6be0f3e7ffe4","arxiv_id":"1909.02653","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"Optimal anisotropic CMB filtering plus a second-stage filter on the reconstructed lensing convergence map reduces lensing power spectrum variance by factor 2-5 relative to isotropic filtering, with up to 30% additional improvement on large scales.","lead":"Gravity bends the cosmic microwave background, and this paper shows how to measure that bending more precisely when telescope maps have uneven noise. Upcoming ground-based experiments should get lensing power spectra with substantially smaller errors by optimally filtering the maps and then filtering the reconstructed lensing map.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Factor-2-5 variance gain conflates optimal filtering with extra sky area; matched-area test needed.","rationale":"The paper is a solid methods contribution with 500 Monte Carlo simulations and a validated patch model; no internal inconsistency is apparent. My concern is interpretive: the central quantitative claim in the abstract is presented as a comparison of filtering methods, while the baselines differ in effective sky area by a factor near 3. Because variance scales approximately as 1/f_sky, area alone can account for much of the reported factor. The matched-area control proposed here is straightforward and can settle the attribution. This does not invalidate the pipeline or the up-to-30% kappa-filtering result, which is measured at fixed pipeline and supported by simulations; it changes how the factor-2-5 headline should be read. The reader's weakest-assumption concern about the patch approximation is real but is mitigated by Monte Carlo corrections and the simulation agreement shown in Fig. 9; the area confound is more directly tied to the headline number. I therefore recommend conditional acceptance pending clarification or a displayed matched-area comparison.","tokens_in":21142,"tokens_out":22506,"duration_ms":268316,"concrete_test":"Recompute the SO MV variance comparison using the optimal anisotropic pipeline applied to maps multiplied by the same apodized mask used for the isotropic-masked baseline in Sec. II A 2, so both pipelines cover identical sky area. If the variance ratio to the isotropic baseline drops from roughly 2-5 to near unity, the reported factor is an area effect and the abstract's filtering-specific claim is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The headline factor-2-5 variance gain is attributed to optimal anisotropic filtering (Abstract; Sec. V), but the comparison does not match sky coverage. The anisotropic runs use the full hit-count area (fsky about 0.13); the isotropic-masked baseline uses an apodized mask cutting roughly 46% of that area (effective fsky about 0.04), and the weighted baseline has effective fsky about 0.01 (footnote after Eq. 3.9; Sec. II A 2). Lensing variance scales roughly inversely with f_sky for cosmic-variance-dominated modes, and similarly for reconstruction-noise-dominated modes, so an area ratio near 3 predicts a substantial part of the reported factor 2-5 even if the filter itself gives no gain. The authors disclose this and test several masks, but the abstract's wording implies the gain is from filtering. The paper's useful method does not depend on this attribution, but the headline quantitative claim is not isolated, so the factor should be re-verified on a matched-area control or reworded as a full-pipeline (area plus filter) gain.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops and tests an optimized quadratic-estimator pipeline for CMB lensing reconstruction under anisotropic noise. The first filtering stage applies inverse-signal-plus-noise (optimal anisotropic) filtering to the input CMB maps via conjugate-gradient inversion; the second stage approximately normalizes the reconstructed convergence map with a local effective response and applies a Wiener-like filter using a pixel-dependent effective reconstruction noise. Bias subtraction uses paired mean-field estimates and 500 Monte Carlo simulations for the N0 and N1 contractions, and the analytic 'patch' model is used to compute the response normalization and to predict variance improvements. On flat-sky SO- and S4-like simulations, the paper reports a factor of 2-5 variance reduction relative to isotropic masked/weighted baselines and an additional reduction of up to about 30% from the κ-filtering step on large scales.","tokens_in":21289,"tokens_out":9910,"duration_ms":118078,"significance":"If the reported gains are robust, the paper is practically important for upcoming ground-based lensing analyses. Its strengths are substantial: the simulation pipeline is unusually complete (500 lensed simulations, independent mean-field pairs, Monte Carlo N0 and N1 debiasing, and documented convergence-tolerance tests in Appendix A); the analytic patch normalization is explicitly compared with Monte Carlo results rather than tuned; and the κ-filtering improvement is predicted analytically and then verified in independent simulations. The main caveat is that the headline factor-2-5 comparison is not matched in sky coverage, so the paper's central quantitative claim mixes the effect of optimal filtering with the effect of using a larger effective sky area. This is fixable either with a matched-area control or by reframing the claim as a full-pipeline comparison.","major_comments":[{"comment":"The factor-2-5 variance reduction is presented in the Abstract and Conclusions as a gain from optimal CMB map filtering, but the comparison is not isolated in sky coverage. The optimal anisotropic runs use all non-zero hit pixels (fsky≈0.13), while the isotropic-masked baseline has effective area fraction fA,L=0.13 of the flat-sky map (fsky≈0.04) and the weighted baseline has fA,L=0.03 (fsky≈0.01), as stated in the footnote after Eq. (3.9). Since lensing power-spectrum variance scales approximately inversely with effective sky fraction for both cosmic-variance- and noise-dominated modes, a substantial fraction of the reported factor can be attributable to area alone, especially for the weighted baseline. I request either a matched-area comparison (for example, an isotropic filter applied over the same scanned region, or an anisotropic-filter run restricted to the masked baseline region) or a rewording of the Abstract and Sec. V so that the factor-2-5 claim is explicitly described as a full-pipeline (area plus filter) gain rather than a filter-only improvement.","section":"Sec. IV, Fig. 9; Sec. V; footnote after Eq. (3.9)"}],"minor_comments":[{"comment":"The phrase \"pre-trained neutral network\" should read \"pre-trained neural network.\"","section":"Sec. V"},{"comment":"Please define the notation C^{κκ}_fid explicitly as a diagonal matrix in harmonic space; as written, the symbol could be confused with the lensing power spectrum C^{κκ}_L.","section":"Eq. (2.8)"},{"comment":"The sentence \"mask all pixels with (temperature map) noise over ~9.6 µK arcmin\" should read \"noise above ~9.6 µK arcmin\" for clarity.","section":"Sec. II A 2"},{"comment":"The effective area fractions fA,L are reported in one place as fractions of the flat-sky map and in another as full-sky fractions (fsky≈0.04 and 0.01). Please state the convention at each occurrence, since this is directly relevant to interpreting the variance improvements.","section":"Fig. 5 and footnote after Eq. (3.9)"},{"comment":"Please state more explicitly in the caption or text whether the plotted quantity is C^{φφ}_{L,fid} - <Ĉ^{φφ}_L>, and note that this is an additive correction applied after the fA,L normalization; the current caption is easy to misread as a bias subtraction akin to N0/N1.","section":"Fig. 6"},{"comment":"The variance expression neglects N1 variance, which the paper notes can matter on small scales. Since Fig. 7 quotes the theoretical variance improvement over the full L-range, it would be helpful to indicate in the figure which scales are most affected by this approximation rather than only in the text.","section":"Sec. III, Eq. (3.13)"}],"recommendation":"major_revision","confidential_remarks":"I do not see a correctness problem in the estimator construction or in the Monte Carlo validation; the major-revision request is about the framing and quantification of the headline factor-2-5 claim. If the authors add a matched-area comparison, or alternatively reword the claim as a full-pipeline gain and remove the filter-only attribution, the paper would be acceptable. The patch approximation and κ-filter method are sound enough that no re-derivation is needed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the second-stage Wiener filtering of the reconstructed convergence map using a locally defined effective reconstruction noise, plus the analytic patch model for the response. The 30% variance improvement from that step is a same-area comparison, and it holds up: the analytic patch prediction matches Monte Carlo across L ~ 100–1500, with only small additive corrections at low L. That is a credible and useful method for SO/S4-style analyses.\n\nThe factor-2-5 variance gain quoted in the abstract is a different beast. It compares optimal anisotropic filtering over the full hit-count area (fsky ~ 0.13) against isotropic filtering over apodized masks with effective fsky ~ 0.04 or weighted maps with effective fsky ~ 0.01. Lensing variance scales roughly inversely with area, so an area ratio of ~3 already predicts a large chunk of the reported gain. The authors do disclose this in Sec IV — they say explicitly \"over a reduced sky area\" — but the abstract does not. That is a real soft spot, and the headline number should be re-verified on a matched-area control or reworded as a full-pipeline gain.\n\nThe simulation pipeline is otherwise solid: 500 lensing simulations, independent mean-field pairs, Monte Carlo N0 and N1 debiasing, and convergence tolerance tests in Appendix A. There is no circularity: the analytic patch model predicts variance improvements that are then checked against simulations, not fitted. The effective-noise band (40 ≤ L ≤ 90) used in the kappa filter is a modeling choice, but the paper shows it is close to using the full N0,L and does not tune it to the final variances.\n\nLimitations are stated honestly: flat sky, uncorrelated noise, single frequency, no foregrounds. The patch approximation is quasi-local and breaks down at low L, where they add a small MC correction. Minor concerns, not load-bearing.\n\nWho is this for? CMB lensing methodologists and anyone planning lensing analyses for SO or S4. The kappa-filtering idea is the actual contribution and deserves to be in the literature. The factor-2-5 claim needs a matched-area test or a less ambiguous description. I would send it to a serious referee, and I would expect the referee to ask for that clarification before acceptance.","headline":"Solid methods paper: the new kappa-filtering step is real and well tested, but the headline factor-2-5 variance gain is a full-pipeline comparison that mostly reflects extra sky area, not the filter alone.","tokens_in":21872,"tokens_out":2654,"would_cite":true,"duration_ms":28624,"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":"For CMB lensing reconstruction from maps with anisotropic noise, optimally filtering the input CMB maps and then Wiener-filtering the reconstructed convergence map cuts lensing power-spectrum variance by a factor of 2–5, with a further…","keywords":["CMB lensing reconstruction","anisotropic noise","inverse-variance filtering","quadratic estimator","lensing power spectrum","Wiener filter","patch approximation","conjugate gradient"],"falsifier":"Run the identical pipeline on a simulated scan whose hit-count map varies sharply, for example narrow stripes or many small disconnected patches, and compare the analytic patch normalization $f_{A,L}$ with the full Monte Carlo normalization: a deviation larger than a few percent in the multipole range $100\\lesssim L\\lesssim 1500$, or the disappearance of the factor 2–5 variance gain relative to isotropic filtering, would show that the quasi-locality assumption is not satisfied for that noise pattern.","tokens_in":20877,"feed_emoji":"🌌","tokens_out":11060,"duration_ms":112241,"temperature":0.7,"pith_summary":"This paper is about how to extract the cosmic-microwave-background lensing signal from maps with uneven, cut-sky noise. The authors claim that quadratic-estimator lensing reconstruction should use two filtering steps: an inverse-variance filter on the input CMB maps, solved by conjugate gradients, and then a Wiener-like filter applied to the reconstructed convergence map. Compared with the simple isotropic filters used in earlier analyses, the first step lowers the reconstructed lensing spectrum variance by a factor of 2–5 for Simons Observatory- and CMB Stage-4-like noise, and the second step reduces it further by up to about 30% on large scales. An analytic patch model for the anisotropic response keeps the pipeline unbiased at the percent level and fast enough to run on hundreds of simulations.","feed_headline":"Two-stage filtering cuts CMB lensing variance by 2-5x","feed_subtitle":"Patchy-noise CMB maps yield sharper lensing spectra when both the input map and the convergence map are filtered.","key_machinery":"The load-bearing machinery is the combination of the inverse-variance CMB map filter and the independent-patch approximation. The filter is the solution of $(\\mathbf{b}C_{\\mathrm{fid}}\\mathbf{b}^{\\top}+\\mathbf{N})^{-1}X$, computed iteratively with a conjugate-gradient solver, which automatically down-weights masked and high-noise pixels. The patch approximation splits the sky into regions of nearly constant noise, computes the local quadratic-estimator response $R^p_L$ and reconstruction noise $N^p_{0,L}$ analytically in each patch, and assembles them into the normalization factor $f_{A,L}$ and the effective white noise map $N^{\\kappa}_{0,\\mathrm{eff}}(x)$. This patch model motivates and calibrates the second-stage Wiener filter $C^{\\kappa\\kappa}_{\\mathrm{fid}}(C^{\\kappa\\kappa}_{\\mathrm{fid}}+N^{\\kappa}_{0,\\mathrm{eff}})^{-1}$, which converts an inverse-noise-weighted map that is optimal on small scales into one that also correctly weights large-scale signal-dominated modes.","core_discovery":"The central discovery is that a near-optimal, still-quadratic lensing pipeline for anisotropic noise is a two-stage filter. Stage one applies the optimal inverse-signal-plus-noise filter to the observed CMB maps, using a preconditioned conjugate-gradient inversion rather than an isotropic harmonic-space filter. Stage two takes the resulting quadratic-estimator convergence map, locally normalizes it through an analytic patch response, and applies the Wiener filter $C^{\\kappa\\kappa}_{\\mathrm{fid}}(C^{\\kappa\\kappa}_{\\mathrm{fid}}+N^{\\kappa}_{0,\\mathrm{eff}})^{-1}$ using a slowly varying effective reconstruction-noise map. On SO- and S4-like simulations the paper finds a factor 2–5 variance reduction from stage one relative to isotropic masked or weighted filtering, and up to roughly 30% further variance reduction from stage two on large scales, with the analytic response model matching the Monte Carlo normalization to within a few percent.","pith_inferences":["The analytic variance expressions could be used before observing to rank candidate scan strategies: compute the expected variance reduction from the hit-count map alone and choose the scan whose noise pattern gives the largest gain.","Because the second stage improves signal-dominated large-scale modes, the filtered convergence maps should also be more effective for CMB delensing, an application the paper mentions but does not quantify.","The reported gains are lower bounds relative to a fully optimal iterative maximum-likelihood estimator, which the paper notes could do still better in the high-signal-to-noise regime; the practical value of the present pipeline is that it stays quadratic and fast."],"forward_implications":["For SO- and S4-like noise and the scan pattern tested, replacing isotropic filtered maps with optimally inverse-variance filtered maps lowers the reconstructed lensing spectrum variance by a factor of 2–5 over most of the analysed multipole range.","Adding the approximate convergence ($\\kappa$) filtering stage reduces large-scale lensing power variance by up to about 30%, in line with the analytic patch prediction, at modest extra numerical cost.","The analytic patch response model leaves only a small percent-level Monte Carlo correction, so the optimized quadratic-estimator pipeline stays fast enough to run over many simulated skies.","Polarization-only reconstructions benefit most from the second filtering stage because their reconstruction noise is more anisotropic; temperature reconstructions, being more signal-dominated, benefit less.","The reconstructed convergence maps after the second filter are more correlated with the true lensing field on large scales than maps from optimal CMB filtering alone, which matters for applications that use the maps rather than just their spectra."],"supporting_citations":[{"why":"Supplies the quadratic-estimator derivation and the inverse-variance filtered form of the lensing estimator that the whole pipeline builds on.","marker":"[12]"},{"why":"Establishes the Planck lensing analysis baseline, including mean-field subtraction, Monte Carlo debiasing, and the demonstration that anisotropic filtering helps polarization reconstruction.","marker":"[2]"},{"why":"First demonstrated multigrid-preconditioned conjugate-gradient inverse-variance filtering for lensing, the numerical method used for the optimal CMB map filter.","marker":"[18]"},{"why":"Provides the quadratic-estimator weights and the response normalization $R_L$ that the analytic patch model uses for each locally isotropic patch.","marker":"[26]"},{"why":"Supplies the Monte Carlo $N_0$ and $N_1$ debiasing estimators that turn the raw lensing spectrum into an unbiased estimate.","marker":"[6]"},{"why":"Provides the scanning-strategy hit-count map used to simulate realistic anisotropic noise in the comparisons.","marker":"[17]"},{"why":"Provides the Simons Observatory experimental specifications (sensitivity, beam, observation time) used in the SO-like simulations.","marker":"[3]"},{"why":"Provides the CMB-S4 experimental specifications used in the S4-like simulations.","marker":"[4]"}],"fun_headline_variants":["Two-stage filter cuts CMB lensing variance 2-5x","Second-stage filter boosts lensing variance 30%","Optimal two-stage filter sharpens CMB lensing spectra","Anisotropic noise tamed by two-step lensing filter","Patchy CMB noise beaten by dual filtering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The analytic response model and the second filter rest on the assumption that lensing convergence estimators are quasi-local, so that the sky can be treated as a patchwork of independent regions of nearly constant noise; if the noise varies on scales comparable to the estimator's real-space support, the analytic normalization and the reported gains would have to be revalidated by simulation.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage filter cuts CMB lensing variance 2-5x","Second-stage filter boosts lensing variance 30%","Optimal two-stage filter sharpens CMB lensing spectra","Anisotropic noise tamed by two-step lensing filter","Patchy CMB noise beaten by dual filtering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000991,"raw_usage":{"total_tokens":4227,"prompt_tokens":1000,"completion_tokens":3227,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":3148}},"tokens_in":616,"tokens_out":3227,"duration_ms":25709,"temperature":1.0,"reasoning_tokens":3148,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:43:37.376703+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical pipeline on a simulated scan whose hit-count map varies sharply, for example narrow stripes or many small disconnected patches, and compare the analytic patch normalization $f_{A,L}$ with the full Monte Carlo normalization: a deviation larger than a few percent in the multipole range $100\\lesssim L\\lesssim 1500$, or the disappearance of the factor 2–5 variance gain relative to isotropic filtering, would show that the quasi-locality assumption is not satisfied for that noise pattern.","supporting_citations":[{"cited_title":"(2.6) is trivial to invert in harmonic space if the noise (and beam) is taken as isotropic, as both the theory and noise covariance matrices are then diagonal in harmonic space","cited_arxiv_id":null,"evidence_quote":"Establishes the Planck lensing analysis baseline, including mean-field subtraction, Monte Carlo debiasing, and the demonstration that anisotropic filtering helps polarization reconstruction."},{"cited_title":"CMB lensing reconstruction using cut sky polarization maps and pure-$B$ modes","cited_arxiv_id":"1403.3911","evidence_quote":"Provides the quadratic-estimator weights and the response normalization $R_L$ that the analytic patch model uses for each locally isotropic patch."},{"cited_title":"This corresponds to using the filter ( bCfidb⊤ +N )−1 WX , where W is diagonal in pixel space and N is taken to be isotropic","cited_arxiv_id":null,"evidence_quote":"Provides the Simons Observatory experimental specifications (sensitivity, beam, observation time) used in the SO-like simulations."}],"review_version":1}