{"id":"7c37c856-7da0-4b2e-b10a-a9a127a9330f","arxiv_id":"2412.10580","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A U-Net trained on a Wiener-filter loss reconstructs polarized CMB E and B modes from masked noisy maps, making power spectrum estimation fast and less biased at low multipoles.","lead":"Cosmologists need clean maps of the cosmic microwave background's polarization, especially the faint B-mode signal that could reveal primordial gravitational waves. This paper trains a neural network to reproduce the optimal Wiener filter for Q/U polarization maps with realistic noise and sky masks, and uses it to estimate E- and B-mode power spectra faster, with smaller errors on large scales than the standard pseudo-Cℓ method.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The residual-is-pure-B assumption behind the J_B loss (Sec. 2.1) is untested; if residual E-mode leaks into the final B input, the low-ℓ B-band powers—the headline target—can be biased, and the mean over 100 maps does not fully rule this out.","rationale":"I read the paper as making a two-part claim: (1) DeepWiener approximates the Wiener filter for polarization maps, and (2) the resulting filtered maps support a simulation-based optimal quadratic estimator whose E and B power spectra are unbiased with errors matching the inverse Fisher matrix. The first part is well supported: the J_Q,U loss (eq. 2.6) is the exact Wiener-filter chi-squared, the network is validated against an independent PCG solver, and the architecture is a direct generalization of previously published temperature-map work. The second part is credible for the E-mode and for the J_Q,U branch, where the same loss is used throughout and the 100-map means are unbiased.\n\nThe soft spot is the J_B branch. The paper's own Sec. 2.1 admits that J_B is an approximation that assumes the B-mode map is free of ambiguous modes. Everything downstream—the B-mode cross-correlation, the power-spectrum estimation, and the comparison with pseudo-Cℓ—depends on that assumption. The iterative E-subtraction is designed to reduce leakage, but the paper provides no diagnostic that directly measures residual E-to-B leakage in the final input map. The cross-correlation with PCG is a valuable sanity check, but it does not distinguish a clean B-mode estimate from one that tracks a mixture of B and leaked E.\n\nI agree with the reader's weakest_assumption. The concern is not that the method is internally inconsistent; it is that a key premise is unverified. The proposed E-only null test would settle it directly. Until such a test is performed, the conditional verdict is appropriate. I do not recommend rejection because the J_Q,U branch and the overall methodology have independent support, and no gross bias appears in the reported means. I also note the abstract's blanket claim that errors are smaller than pseudo-Cℓ is stronger than the body, where purified pseudo-Cℓ errors are smaller beyond ℓ ≈ 750; this is a presentation concern, secondary to the leakage issue.","tokens_in":24082,"tokens_out":4903,"duration_ms":52409,"concrete_test":"Run the full pipeline, including the iterative E-subtraction and the J_B-trained model, on 100 signal-only realizations with E-mode power equal to the paper's perturbed 'true' spectrum and B = 0, using the same masks and noise realizations. If the residual-is-pure-B assumption holds, the mean estimated C_BB should be consistent with zero after noise-bias subtraction. Any significant positive low-ℓ signal is direct evidence of residual E-to-B leakage biasing the headline B-mode band. Repeating with E power equal to the fiducial spectrum separates the leakage term proportional to (C_EE^true − C_EE^fid) from a constant leakage floor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The B-mode branch of the pipeline rests on the claim that after iterations (eqs. 2.7–2.11) the input to the J_B-trained network is dominated by B-modes, so the scalar loss J_B (eq. 2.12) is a valid Wiener-filter objective. The paper explicitly states that J_B 'is not derived from (2.6) but rather is an approximation where the B-mode map is not contaminated by ambiguous modes', and that sigma*^2 is itself estimated from simulations of B_obs − B_sky. This is a self-identified limitation and the single most load-bearing assumption in the central claim.\n\nIf residual E-mode power survives the iterative subtraction, the B_obs^(3) map contains E-to-B leakage that is linear in the E-mode signal. The simulation-based quadratic estimator subtracts a bias term computed from fiducial-spectrum simulations (eq. 2.23). That subtraction cancels leakage proportional to the fiducial E power, but not leakage proportional to (C_EE^true − C_EE^fid). Since the paper validates against a 'true' spectrum that is deliberately different from the fiducial one (Fig. 13), any such leakage biases exactly the low-ℓ B bins where the paper claims 85–99% error reduction over pseudo-Cℓ.\n\nThe 100-map mean spectra are encouraging but not dispositive: a leakage coefficient that is small, or a true-minus-fiducial difference that is modest, can produce a bias below the plotted error bars. Cross-correlation with the PCG solution (Fig. 10) is a consistency check, not a proof of leakage-free input; a filter can track the PCG B-mode while still propagating leaked E power. No E-mode-only null test is reported, so the unbiasedness of the headline B-mode measurement is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DeepWiener, a U-Net trained to approximate the Wiener filter for CMB polarization maps with inhomogeneous noise and a sky mask. Because a single network trained on Q and U maps with the quadratic loss J_Q,U reconstructs the E-mode well but the B-mode poorly, the authors propose an iterative scheme: subtract successive E-mode estimates from the observed maps and eventually train on the residual B-mode map using a scalar loss J_B. The filtered maps are then fed into a simulation-based optimal quadratic estimator to obtain E- and B-mode band powers, with the noise bias and Fisher matrix computed from fiducial-spectrum simulations through the same network. The method is validated against the conjugate-gradient Wiener filter solution, the mean of 100 single-map power-spectrum estimates is found to be unbiased, the per-map errors are shown to match the inverse Fisher matrix, and the errors are compared with pseudo-C_ell estimates from NaMaster.","tokens_in":24302,"tokens_out":8514,"duration_ms":84816,"significance":"If the central claim holds, the paper provides a practical, low-cost route to near-optimal polarized CMB map filtering and power-spectrum estimation: after training, the network applies the Wiener filter roughly an order of magnitude faster than the PCG method, making the simulation-based quadratic estimator feasible (Table 1). Strengths of the manuscript are that the training loss is exactly the Wiener-filter chi-squared, so the network's target is well defined; the public code and detailed appendices support reproducibility; the comparison with PCG is a genuine external convergence check; and the internal consistency between measured errors and the inverse Fisher matrix (Fig. 14) is a useful validation of the pipeline. The main risk is the B-mode branch: the J_B loss is explicitly an approximation resting on the assumption that the iterative subtraction leaves a residual dominated by B-modes, and this assumption is not independently quantified.","major_comments":[{"comment":"The scalar loss J_B in eq. (2.12) is introduced as an approximation in which the residual maps Q^(3)_obs and U^(3)_obs contain only B-mode signal. This is the load-bearing assumption for the B-mode branch: if residual E-mode power survives the iterative subtraction, the low-ℓ B input is contaminated by E-to-B leakage that is linear in the E signal. Because the noise-bias term in eq. (2.23) is computed from fiducial-spectrum simulations, that subtraction cannot cancel leakage proportional to (C_EE^true − C_EE^fid). The 100-map mean in Fig. 13 and the cross-correlation with the PCG solution in Fig. 10 are supportive but not dispositive; a small leakage coefficient or a modest true-minus-fiducial difference can hide the bias. Please add a quantitative test: for example, measure the cross-correlation between B^(3)_obs (or the final B_NN) and the true E field as a function of ℓ, and repeat the power-spectrum validation with a strongly displaced true E spectrum, or with the E amplitude artificially boosted, to show that the low-ℓ B bias remains consistent with zero.","section":"§2.1 (eqs. 2.7–2.12); §5.1 (Fig. 10)"},{"comment":"The estimator is unbiased only if the filter applied to the data is the same as the filter used to construct b_ℓ and F. The network is an approximation to the Wiener filter, and the paper validates the estimator for a single true spectrum that is close to the fiducial one. This leaves open a residual bias proportional to (C_true − C_fid) that would not be visible in an internal consistency check. I recommend validating with at least a second true spectrum, ideally with a larger displacement, and reporting the per-bin bias, or alternatively deriving the linear response of the estimator to C_true − C_fid and showing that it is subdominant to the statistical errors.","section":"§5.2 (eqs. 2.23–2.26; Fig. 13)"},{"comment":"The error bars in Fig. 14 match the inverse Fisher matrix, but that Fisher matrix is computed with the same approximate network filter; this demonstrates self-consistency of the pipeline rather than optimality of the filter. The claim that the method outperforms pseudo-Cℓ by a large factor would be strengthened by comparing against a quadratic estimator built from the exact PCG Wiener filter on a subset of maps, or against the analytic Fisher bound for the exact filter. Without such a comparison, the 99% and 95% error reductions quoted for the first bin are relative to the apodized purified pseudo-Cℓ pipeline and may overstate the contribution of the filter itself as opposed to the choice of estimator.","section":"§5.3 (Figs. 14 and 17)"}],"minor_comments":[{"comment":"The second expression should read ⟨|s^B_s|^2⟩ = S^B_fid; as printed it repeats the E-mode expression.","section":"§5.2 (eqs. 5.2–5.3)"},{"comment":"The assumption d^Q_n = d^U_n, i.e., identical noise realizations for Q and U, is physically restrictive and should be justified, or the analysis repeated with independent Q and U noise, since real polarization noise is not identical in the two Stokes parameters.","section":"§5.2 (after eq. 5.5)"},{"comment":"The error bars are not defined; please state whether they are the standard deviation of single-map estimates or the standard error of the 100-map mean.","section":"Figs. 13 and 16"},{"comment":"The architecture section gives the general U-Net structure but not the layer count, filter sizes, stride, or activation functions; since the code is public this is not blocking, but a short table or explicit reference to the repository's configuration would improve reproducibility.","section":"§3"},{"comment":"The stopping criterion for the iterative E-subtraction is the B-mode cross-correlation with PCG, but the actual correlation values at each iteration are not reported; adding them would make the convergence of the iteration transparent.","section":"§5.1 and Appendix C"},{"comment":"The notation sY^B_l and the differential operator D^B_s are not defined; the spin-weighted spherical harmonics and the purification operator should be introduced explicitly for readers not familiar with NaMaster.","section":"§5.3 (eq. 5.9)"}],"recommendation":"major_revision","confidential_remarks":"The self-identified limitation in §2.1 is the key risk: the J_B loss is only valid if the residual after iterative subtraction is dominated by B-modes, and this is not demonstrated quantitatively. The requested leakage diagnostic and a robustness test against a larger true-minus-fiducial E difference should be feasible within the scope of this work. The manuscript is otherwise well within the scope of JCAP, builds on the authors' prior temperature-map work with sufficient novelty, and the central methodological framework is sound. I see no grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid methods paper. The genuinely new pieces are the iterative E-mode subtraction scheme (Secs. 2.1, 5.1, App. C) and the scalar J_B loss for B-mode Wiener filtering once the residual is assumed to be mostly B, applied to polarization maps with inhomogeneous noise and a simulation-based quadratic estimator built on the filtered maps. The core derivation is standard: the J_Q,U loss is the Wiener filter chi-squared, the estimator follows Seljak and Horowitz et al., and the PCG comparison is an honest convergence check. The mean B power spectra over 100 maps are unbiased and the measured errors track the inverse Fisher matrix. The ~10x speedup over PCG is relevant if you need thousands of simulations for bias and Fisher estimation.\n\nThe soft spots are real but not fatal. First, the abstract states that their errors are smaller than pseudo-Cℓ, but the body (Fig. 17) shows purified pseudo-Cℓ is actually smaller beyond ℓ≈750; the abstract overclaims. Second, the headline first-bin reductions partly come from the apodization imposed on the pseudo-Cℓ baseline. Third, and most important, the J_B loss is an approximation, as the authors themselves say in Sec. 2.1. The input to that network is assumed to be free of E-to-B leakage after a few rounds of subtraction, but there is no E-only null test to show that. The quadratic estimator's bias subtraction cancels leakage proportional to the fiducial E power, so any leakage proportional to (C_EE_true - C_EE_fid) would bias exactly the low-ℓ B bins where they claim 85-99% error reduction. The 100-map mean does not fully rule this out. Fourth, the iteration counts are post hoc and the cross-correlation curves in Figs. 9-10 have no error bars.\n\nNone of this breaks the central claim: the J_Q,U pipeline is sound, and the J_B variant is a reasonable approximation that works on the tests shown. But before adopting it for survey science, I would want an E-only null test and a quantification of the leakage residual into B as a function of C_EE_true - C_EE_fid.\n\nThe paper is for people building CMB polarization pipelines for S4/LiteBIRD/QUBIC who need fast approximate Wiener filtering and will check systematics themselves. It deserves a serious referee: the method is well-motivated, the code is public on GitHub, and the limitations are at least partly acknowledged. I would send it to review, asking the authors to address the leakage question and moderate the abstract.","headline":"Solid methods paper for CMB polarization with a genuinely useful iterative E-subtraction trick, but the B-mode branch rests on an untested leakage assumption and the abstract oversells the pseudo-Cℓ comparison.","tokens_in":25083,"tokens_out":4794,"would_cite":true,"duration_ms":40375,"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":"A U-Net trained on the Wiener-filter loss reproduces the exact Wiener filter for masked, noisy CMB polarization maps, and the resulting E and B power spectra are unbiased, with errors that beat pseudo-Cℓ at the scales where primordial…","keywords":["cosmic microwave background polarization","Wiener filter","convolutional neural network","U-Net","E/B-mode decomposition","E-to-B leakage","power spectrum estimation","optimal quadratic estimator"],"falsifier":"Train the full pipeline on signal-only simulations containing E-modes and no input B-modes; any recovered low-ℓ B band power above zero would reveal residual E-to-B leakage. A cleaner quantitative test: compute the cross-spectrum between the final residual B map and the true input E map; a nonzero low-ℓ correlation would mean the pure-B assumption is violated and the claimed unbiasedness is not guaranteed.","tokens_in":1431,"feed_emoji":"🌌","tokens_out":4329,"duration_ms":90663,"temperature":0.7,"pith_summary":"The paper sets out to show that a convolutional U-Net, trained with the same quadratic loss that defines the optimal Wiener filter, can stand in for the expensive exact filter on CMB polarization maps. On masked skies with realistic inhomogeneous noise, the network's E- and B-mode reconstructions closely track the conjugate-gradient Wiener solution after a few iterations of E-mode subtraction. Using those filtered maps, a simulation-based optimal quadratic estimator returns E and B power spectra that are unbiased over 100 maps, with single-map errors matching the inverse Fisher matrix and, in the low-multipole bin where primordial B-modes are sought, smaller than pseudo-Cℓ errors by 99% for the simpler mask and 95% for the complex mask. The practical payoff of the claim is near-optimal filtering and spectrum estimation at about one tenth of the conjugate-gradient compute cost.","feed_headline":"Neural filter cuts CMB B-mode error by 99% at low multipoles","feed_subtitle":"A U-Net plus iterative E-mode subtraction delivers unbiased E/B spectra with errors near the Fisher bound.","key_machinery":"DeepWiener is a U-Net autoencoder whose input has a linear channel for the Q and U (or B) maps plus non-linear channels for the noise variance map and the mask, with the non-linear outputs multiplying the linear channel so that the filter remains linear in the data. The training objective is the Wiener-filter loss J_{Q,U}, equation (2.6): pixel-space noise-weighted residuals plus Fourier-space signal-covariance terms in the E/B basis. Because B is far weaker than E, the method iterates: it subtracts the network's E estimate, retrains on the residual Q and U maps, and repeats; once the residual is assumed B-dominated, it switches to the scalar loss J_B, equation (2.12), whose noise variance is estimated from simulations of B_obs - B_sky. The filtered maps then enter a simulation-based optimal quadratic estimator: a Fisher matrix and noise-bias vector are computed by applying the trained network to hundreds of fiducial simulations, and the band powers are recovered by inverting the Fisher relation. The machinery replaces the matrix inversion in ($S^{{-1}}$+R^T $N^{{-1}}$ R)^{-1} R^T $N^{{-1}}$ d with a trained feed-forward pass.","core_discovery":"The central claim is that the exact Wiener-filter solution for masked, inhomogeneously noisy polarization maps can be reproduced by DeepWiener, a U-Net trained with the Wiener-filter quadratic loss J_{Q,U}, and that the E-to-B leakage caused by the mask can be suppressed by iteratively subtracting the network's E-mode estimate from the Q and U maps. After three to five such rounds the residual maps are treated as B-dominated, and the network can instead be trained with a scalar loss J_B applied directly to the observed B-mode map. From these filtered maps the paper constructs a simulation-based optimal quadratic estimator; over 100 test maps the resulting E and B band powers are centered on the true spectrum, their errors match the inverse-Fisher diagonal, and at the low-ℓ bin relevant to primordial gravitational waves the error is 99% (Mask1) or 95% (Mask2) smaller than the purified pseudo-Cℓ error. The paper further states that this makes near-optimal polarization filtering and spectrum estimation available at roughly one tenth of the conjugate-gradient computation cost.","pith_inferences":["The flat-sky, single-patch framework leaves open whether the same network and loss transfer to curved-sky, all-sky maps; if transferable, this style of filtering could serve as a fast map-level preprocessing step in end-to-end likelihood analyses.","A direct comparison against an unpurified and unapodized pure-B estimator, or against the exact quadratic estimator driven by the conjugate-gradient Wiener filter, would separate the gain due to the neural filter from the loss due to mask apodization in the pseudo-Cℓ baseline.","The iterative E-subtraction acts as a decontamination step that could in principle be applied inside existing pipelines before any spectrum estimator, not exclusively before the optimal quadratic estimator.","Because the bias-subtraction step is built from fiducial-spectrum simulations, the same pipeline could be extended to estimate E-B cross-spectra or to propagate the full band-power covariance, including off-diagonal terms, into parameter forecasts."],"forward_implications":["A trained DeepWiener model can be applied to any number of maps with the same mask and noise properties, making noise-bias and Fisher-matrix simulations at the scale of 2000 maps practical rather than prohibitive.","Because the estimator remains unbiased when the true spectrum differs from the fiducial, the pipeline can measure an unknown B-mode signal rather than only validating on simulations.","The first-bin B-mode error reduction of 99% (Mask1) and 95% (Mask2) over purified pseudo-Cℓ, if correct, translates directly into stronger constraining power on the tensor-to-scalar ratio in masked survey patches.","More complex masks require more iterations (5 versus 4 under J_{Q,U}; 4 versus 3 under J_B), so the method's cost scales with mask complexity while staying about an order of magnitude below the conjugate-gradient method.","Per-map prediction cost is independent of the number of maps, so the approach becomes more favorable as survey volume grows."],"supporting_citations":[{"why":"Supplies the WienerNet architecture and the training loss that DeepWiener adapts to polarization maps.","marker":"[35]"},{"why":"Establishes the prior demonstration of neural-network Wiener filtering for temperature maps, including the inhomogeneous-noise channel that this work generalizes.","marker":"[28]"},{"why":"Provides the efficient optimal-reconstruction and band-power formalism that the simulation-based quadratic estimator is built on.","marker":"[30]"},{"why":"Introduces the likelihood-based optimal quadratic estimator that the paper implements after Wiener filtering.","marker":"[29]"},{"why":"Provides the public pseudo-Cℓ library (NaMaster) used as the comparison baseline for power-spectrum errors.","marker":"[36]"},{"why":"Supplies the conjugate-gradient algorithm that the paper uses as the exact Wiener-filter benchmark and whose runtime is compared in Table 1.","marker":"[21]"},{"why":"Motivates the iterative E-mode subtraction by framing pure E and B maps as achievable through successive Wiener filtering.","marker":"[34]"},{"why":"Defines the pure pseudo-Cℓ formalism used to mitigate E-to-B leakage in the comparison method.","marker":"[32]"}],"fun_headline_variants":["Neural Wiener filter slashes CMB B-mode error by 99%","DeepWiener nails CMB polarization spectra to Fisher limit","AI Wiener filter achieves 99% B-mode error cut at low multipoles","Neural net reproduces exact Wiener filter for CMB polarization","U-Net solves CMB polarization Wiener filter at 10x speed"],"cache_read_input_tokens":26880,"weakest_assumption_plain":"The method assumes that after a few rounds of subtracting the network's E-mode estimate, the remaining Q and U maps are essentially pure B-mode, so the scalar loss used for the final B-mode training does not secretly fit leftover E-mode signal.","fun_headline_variants_meta":{"raw":{"variants":["Neural Wiener filter slashes CMB B-mode error by 99%","DeepWiener nails CMB polarization spectra to Fisher limit","AI Wiener filter achieves 99% B-mode error cut at low multipoles","Neural net reproduces exact Wiener filter for CMB polarization","U-Net solves CMB polarization Wiener filter at 10x speed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0009,"raw_usage":{"total_tokens":3890,"prompt_tokens":977,"completion_tokens":2913,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":2830}},"tokens_in":593,"tokens_out":2913,"duration_ms":18808,"temperature":1.0,"reasoning_tokens":2830,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:52:18.206829+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the full pipeline on signal-only simulations containing E-modes and no input B-modes; any recovered low-ℓ B band power above zero would reveal residual E-to-B leakage. A cleaner quantitative test: compute the cross-spectrum between the final residual B map and the true input E map; a nonzero low-ℓ correlation would mean the pure-B assumption is violated and the claimed unbiasedness is not guaranteed.","supporting_citations":[{"cited_title":"Efficient Optimal Reconstruction of Linear Fields and Band-powers from Cosmological Data","cited_arxiv_id":"1810.00503","evidence_quote":"Provides the efficient optimal-reconstruction and band-power formalism that the simulation-based quadratic estimator is built on."},{"cited_title":"Press, S.A","cited_arxiv_id":null,"evidence_quote":"Supplies the conjugate-gradient algorithm that the paper uses as the exact Wiener-filter benchmark and whose runtime is compared in Table 1."},{"cited_title":"Pure E and B polarization maps via Wiener filtering","cited_arxiv_id":"1610.03345","evidence_quote":"Motivates the iterative E-mode subtraction by framing pure E and B maps as achievable through successive Wiener filtering."}],"review_version":1}