{"id":"1e4dbed9-46ad-4569-9c2f-d9635a756ab7","arxiv_id":"2505.21220","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A Wavelet Flow generative model jointly produces CMB lensing convergence and cosmic infrared background maps whose power spectra and Minkowski functionals match the training simulation within a few percent.","lead":"This paper uses a machine learning method called Wavelet Flow to generate simulated maps of two cosmic signals: the lensing of the cosmic microwave background and the cosmic infrared background. If the method holds up, it offers a fast way to produce realistic foreground models for next-generation CMB experiments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validation cutouts may overlap training cutouts: 65,000 random 2° patches from one full-sky map cover roughly 7 times the sky, so the few-percent 'held-out' accuracy could partly reflect leakage rather than generalization.","rationale":"The reader's weakest assumption was that in-distribution fidelity transfers to real data, a limitation the paper explicitly acknowledges. While that is a valid external-validity concern, I find a more immediate internal-validity issue: the paper does not establish that the validation cutouts are truly held out. The numbers of cutouts and their sizes make random training-validation overlap almost certain unless the split was designed to prevent it. If validation patches overlap training patches, the reported agreement could be inflated by memorization or by sharing the same large-scale structures, which would undermine the central claim even in-distribution. This is a concrete, checkable concern because the code is public; it does not require reinterpreting the science, only inspecting the data pipeline. I therefore recommend keeping the CONDITIONAL verdict, but the condition should be verification of a spatially disjoint validation split rather than only the fixed-parameter generalization issue. I disagree with the reader's selection of the single weakest assumption because the leakage issue is more directly load-bearing for the paper's stated claim of 'statistically recovers the input to high accuracy.'","tokens_in":16230,"tokens_out":4928,"duration_ms":61554,"concrete_test":"Inspect the cutout-generation and data-split code in the linked GitHub repository to determine how the 65,000 cutouts were positioned and partitioned. Compute the angular separation between each validation cutout centre and the nearest training cutout centre, and report the fraction of validation cutouts whose separation is smaller than the cutout size (about 2.13°) or that share any pixel with a training cutout. If overlap exists, retrain (or at minimum re-evaluate) using a strictly disjoint split, for example by tiling the full-sky map and holding out entire non-adjacent tiles. If the power-spectrum and Minkowski-functional biases remain within the reported few-percent bands on the disjoint split, the leakage concern is resolved; if they degrade, the headline accuracy estimate must be revised downward.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim in the abstract and Section 5 is that a trained Wavelet Flow 'statistically recovers the input to high accuracy' on held-out maps. Section 4.1 states that the training data are 256×256, 0.5-arcmin cutouts projected from the full-sky Agora maps, with 52,000 training and 13,000 validation samples. The paper does not state that the validation cutouts are spatially disjoint from the training cutouts. A 256-pixel cutout at 0.5 arcmin is about 2.13° on a side; 65,000 such cutouts have roughly 7 times the area of the full sky, so if cutout centres were drawn randomly, substantial overlap is unavoidable. Since the power spectra and Minkowski functionals in Figures 3 and 4 are averaged over 13,000 validation samples, even a modest overlap fraction can bias the comparison toward agreement. This is the most load-bearing threat to the central claim because the claim is specifically about statistical recovery of held-out inputs; a leaky split would make the reported few-percent errors an artifact of the sampling procedure rather than evidence about the learned joint distribution. This concern is distinct from the fixed-parameter limitation that the paper itself concedes in Section 6, which is about transfer to real data; the leakage concern is about whether the in-distribution validation numbers are trustworthy at all.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains a Wavelet Flow (Glow-based normalizing flows applied to Haar wavelet coefficients) to jointly model 256×256, 0.5-arcmin cutouts of CMB lensing convergence κ and cosmic infrared background (CIB) maps drawn from the Agora simulations. A key methodological contribution is a 'Hybrid Component Correlated' (HCC) prior: white-noise priors at coarse wavelet scales and a component-correlated Gaussian prior, built from training-set auto- and cross-power spectra, at the finest wavelet scale. The authors report that generated samples reproduce validation power spectra to within a few percent across scales and Minkowski functionals to within 1--2.5%, and they argue this is the first flow-based model for correlated multi-component CMB secondaries at field level. The paper also releases code and trained models.","tokens_in":16566,"tokens_out":3460,"duration_ms":44316,"significance":"If the reported accuracy is robust, the paper demonstrates a useful and practical capability: fast, field-level generation of correlated κ and CIB maps for simulation-based inference, covariance estimation, and mock-sky pipelines. The scale-dependent prior design is an interesting extension of Wavelet Flow and is presented with concrete comparisons (Figures 5--7). The release of code and trained models is a strength that will facilitate reproducibility and follow-up work. The central limitation is that validation is performed on cutouts from the same simulation used for training, so the in-distribution fidelity claims need to be scrutinized before they can support the paper's conclusions.","major_comments":[{"comment":"The validation protocol does not establish that validation cutouts are spatially disjoint from training cutouts. The paper states (Section 4.1) that 52,000 training and 13,000 validation 256×256 cutouts are projected from the full-sky Agora maps, but it never states that the validation patch centers are chosen to avoid the training patch centers. A 256-pixel cutout at 0.5 arcmin corresponds to about 2.13° on a side, so 65,000 cutouts cover roughly seven times the full sky. If centers were drawn randomly, almost every validation patch overlaps one or more training patches, and the reported few-percent power-spectrum and Minkowski-functionals agreement in Figures 3 and 4 could be substantially inflated by this leakage. This is load-bearing because the abstract and Section 5 claim statistical recovery of held-out inputs. The authors should specify the sampling scheme for training and validation patches, demonstrate that validation patches are spatially separated from all training patches, or re-run the validation on an independent Agora realization (or otherwise non-overlapping sky regions).","section":"§5, Figure 4"},{"comment":"The Minkowski functional comparisons are presented without any error bars or uncertainty intervals. The text states that model outputs remain 'within 1% for all κ functionals' and 'within 2.5% for all CIB functionals,' but with no statistical uncertainty on either the validation estimates or the model-sample estimates, these numbers cannot be assessed for significance. The paper should add error bars (e.g., standard errors over the 13,000 validation and generated samples) or a quantitative compatibility measure such as a χ² statistic, and should state the number of thresholds/bins used and how the percentile range was chosen.","section":"§3.4 and §5"},{"comment":"The finest-scale CC prior is constructed directly from the training-set power and cross-spectra (Eqs. 3.10--3.14), and the text notes that this makes the wavelet coefficients 'exhibit the same power spectrum as the prior or latent distribution.' Consequently, the agreement of the generated small-scale power spectra with the validation data is partly enforced by construction rather than learned by the flow. This does not invalidate the method, but it means the power-spectrum comparisons at the finest scale are not an independent test of the learned model. The paper should state this explicitly and quantify what the flow adds at that scale, for example by comparing against samples drawn directly from the CC prior and by quoting the all-WN-prior results from Figure 6 in the main text.","section":"§3.4, §5"}],"minor_comments":[{"comment":"The heading 'W avelet Flow' contains an unintended space; it should read 'Wavelet Flow.'","section":"§3.3 heading"},{"comment":"The heading 'T raining Data and preprocessing' contains a stray capital 'T'; it should read 'Training Data and Preprocessing.'","section":"§4.1 heading"},{"comment":"There are typos in the prior-selection appendix: 'combintations' should be 'combinations' and 'leve1' should be 'level 1' in the Figure 5 caption and surrounding text.","section":"Appendix A"},{"comment":"The statement that 'the κ power spectrum bias remains within 1% for most multipoles' should specify the multipole range and the binning used in Figure 3, and the text should clarify whether the quoted percentages refer to the mean fractional difference or the maximum across bins.","section":"§5"},{"comment":"The description of the κ map preprocessing says Gaussian noise is added 'after the Nyquist frequency'; since a map has no power beyond Nyquist, this should be clarified as adding noise to the highest-frequency wavelet coefficients or to the map before the DWT, and the exact amplitude of the added noise should be stated.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the potential leakage between training and validation cutouts, which directly affects the central accuracy claim. If the authors can demonstrate a spatially disjoint validation split or re-validate on an independent simulation, the paper's core result is likely acceptable; otherwise the headline few-percent numbers would not be trustworthy. The Minkowski-functionals error bars and the circularity caveat for the finest-scale CC prior are secondary but should be addressed in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe paper is a solid proof of concept for applying Wavelet Flow to jointly simulate correlated κ and CIB fields. As far as I can tell, this is the first flow-based model trained on two CMB secondary components at field level, and the HCC prior – a correlated prior at the finest wavelet scale with white noise elsewhere – is a modular choice that gives a real improvement over white-noise-only models. The power-spectrum and Minkowski-functional comparisons on 13,000 validation maps are carefully presented, with concrete training details and public code. That is real work, and the in-distribution numbers are plausible.\n\nThe soft spots are about how strongly the evidence supports the headline accuracy. First, the validation set may not be spatially independent of the training set. They use 65,000 random 2° cutouts from one full-sky map; that is about seven times the sky area, so overlap is unavoidable if cutouts are drawn randomly. The paper does not say the validation cutouts are spatially disjoint from training. If they are not, the few-percent power-spectrum agreement partly reflects leakage rather than generalization, and the effective sample size is smaller than 13,000 independent maps. The authors should fix this by splitting by sky region or validating on a separate simulation.\n\nSecond, the finest-scale CC prior is built directly from the training-set power and cross-spectra. The paper states that this ensures the wavelet coefficients have the same power spectrum as the prior, so the small-scale two-point agreement is partly by construction. That does not invalidate the model – it still has to learn the non-Gaussian and cross-correlation structure – but it lowers the significance of that particular agreement.\n\nThe Minkowski-functional comparison would be stronger with error bars, and the paper already concedes the fixed-cosmology limitation. None of these are fatal; the leakage issue is the one that needs a response, and it is easy to address with a cleaner split.\n\nWho is this for? Anyone building mock-sky pipelines or simulation-based inference for CMB secondary analyses. It is a useful baseline even with the caveats, and the code matters. I would send it to peer review, with the expectation that the train/validation overlap is fixed or explicitly discussed. I would not cite the few-percent accuracy as a general statement until that is sorted out.","headline":"A useful proof of concept for joint κ–CIB generation with flows, but the validation split likely leaks spatial information and the finest-scale prior makes small-scale power agreement partly by construction.","tokens_in":17081,"tokens_out":3742,"would_cite":true,"duration_ms":36792,"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":"A Wavelet Flow jointly trained on lensing convergence and cosmic infrared background maps generates new samples whose power spectra match the simulation within a few percent, with Minkowski functionals matching to 2.5%.","keywords":["cosmic microwave background","CMB lensing convergence","cosmic infrared background","normalizing flows","wavelet flow","field-level simulation","Minkowski functionals","foreground simulation"],"falsifier":"Evaluate the trained flow on an independent $\\kappa$-CIB dataset\\u2014for example, maps from a different simulation pipeline or reconstructed lensing convergence and Planck CIB data\\u2014and compute the same auto-spectra, cross-spectrum, and Minkowski functionals; if the bias against this external dataset exceeds a few percent, the claimed field-level accuracy does not generalize beyond the training simulation.","tokens_in":16028,"feed_emoji":"🌌","tokens_out":8459,"duration_ms":83809,"temperature":0.7,"pith_summary":"The paper asks whether one generative model can reproduce the joint, field-level statistics of two correlated extragalactic CMB signals: the lensing convergence field $\\kappa$ and the cosmic infrared background (CIB). It claims yes: a Wavelet Flow trained on cutouts from the Agora simulation generates new $\\kappa$ and CIB maps whose average auto- and cross-power spectra match held-out validation maps to within about 2\\u20132.5% at every scale, and whose Minkowski functionals match to within 2.5%. This matters because current and upcoming CMB experiments need many realistic realizations of non-Gaussian foregrounds for simulation-based inference, covariance estimation, and pipeline testing, and no analytic form for these joint distributions exists.","feed_headline":"Flow model mimics CMB foregrounds within a few percent","feed_subtitle":"Jointly generates correlated lensing and infrared maps matching simulated power spectra and shape statistics.","key_machinery":"The carrying mechanism is the Wavelet Flow: a Haar discrete wavelet transform decomposes each $256\\times256$ map into seven resolution levels, and a separate Glow-style normalizing flow models the conditional distribution of the detail coefficients at each level using affine coupling layers. Sampling starts from the coarsest $2\\times2$ block and rebuilds the image through inverse wavelet transforms, so large-scale correlations and small-scale features are handled at separate resolutions. The hybrid component-correlated prior assigns a white-noise prior at coarse scales and, at the finest scale only, a prior whose covariance is the measured auto- and cross-power spectra of the wavelet coefficients; this is what lets the model capture small-scale non-Gaussianity while avoiding the near-100% large-scale failures seen when correlated priors are used everywhere.","core_discovery":"The central claim is that a Wavelet Flow, jointly trained on $\\kappa$ and CIB maps, statistically recovers the input fields themselves: generated samples reproduce the mean power spectra of the Agora validation maps to within a few percent across all scales, including the $\\kappa$-CIB cross-spectrum, and their Minkowski functionals are accurate at the 2.5% level, demonstrating that the model has learned non-Gaussian structure beyond two-point statistics. The paper presents this as the first flow-based model for correlated multi-component CMB secondaries at field level. A hybrid prior\\u2014white noise at coarse wavelet scales and a component-correlated prior, built from the measured per-scale auto- and cross-spectra, at the finest scale\\u2014improves small-scale power recovery by up to 3 percentage points and reduces the maximum Minkowski-functional bias from roughly 10% to 2.5%.","pith_inferences":["If the few-percent accuracy transfers to real sky data\\u2014an untested step, since the model is trained at fixed cosmological and astrophysical parameters\\u2014field-level simulation-based inference for CMB analyses becomes practical without expensive hydrodynamical simulations.","The finding that correlated priors hurt at coarse scales but help at the finest scale suggests a general recipe for other non-Gaussian cosmological fields: use physically motivated priors only where the flow struggles, which could be tested directly on tSZ or 21-cm maps.","A practical bottleneck is map size: the current $2^\\circ\\times2^\\circ$ cutouts fall far short of the thousands of square degrees covered by current experiments, so tiling or outpainting would be needed for full-sky applications.","Because the paper validates only sample statistics, the flow's exact log-likelihood remains untested on real $\\kappa$-CIB patches; scoring held-out sky patches by log-likelihood would be a sharper test than summary-statistic agreement."],"forward_implications":["Generated $\\kappa$-CIB pairs can serve as the foreground-simulation step in simulation-based inference, producing foreground-added lensed CMB maps for realistic mock skies.","The learned joint distribution can act as a prior in posterior-based delensing and foreground-cleaning pipelines, where no current method includes a non-Gaussian foreground model.","At few-percent power-spectrum accuracy, the model contributes a small fraction of the roughly 10%-per-bin lensing measurement uncertainty and can differentiate foreground biases to lensing spectra of order 5\\u201310%.","The scale-dependent prior recipe\\u2014independent optimization of each wavelet level\\u2014extends to additional foreground components such as tSZ, kSZ, and radio sources, which the paper identifies as the next step.","Because the flow gives tractable log-likelihoods, the same trained model is usable for density estimation and inference, not only for generating samples."],"supporting_citations":[{"why":"Supplies the Agora training and validation data: the $\\kappa$ and CIB map pairs whose statistics define the target distribution.","marker":"[23]"},{"why":"Introduces the Wavelet Flow architecture that decomposes images with the discrete wavelet transform and trains separate flows per scale.","marker":"[36]"},{"why":"Prior multiscale normalizing-flow treatment of cosmological fields; source of the small post-Nyquist noise addition used in preprocessing.","marker":"[37]"},{"why":"Defines the Glow affine-coupling blocks and invertible 1x1 convolutions that form each scale's flow.","marker":"[30]"},{"why":"Earlier normalizing-flow model for $\\kappa$ alone; motivates field-level modeling and modified priors in cosmology.","marker":"[31]"},{"why":"Provides the CIB log transform and the GAN-based microwave-sky model that serves as the comparison baseline.","marker":"[38]"},{"why":"Defines the orthonormal Haar wavelets used for all wavelet decompositions in the paper.","marker":"[60]"},{"why":"The wavelet-based normalizing-flow implementation that the authors' code builds on.","marker":"[63]"}],"fun_headline_variants":["Wavelet flow hits few-percent accuracy on CMB foregrounds","Flow model jointly generates lensing and CIB maps","Wavelet flow matches CMB foreground power spectra","Non-Gaussian CMB foregrounds simulated by wavelet flow","Flow model recreates CMB foregrounds within 2.5%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The few-percent accuracy is measured on held-out cutouts from the same Agora simulation used for training, so the central claim assumes that this simulation's $\\kappa$-CIB statistics are close enough to the real sky that in-distribution fidelity transfers to actual CMB analyses.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet flow hits few-percent accuracy on CMB foregrounds","Flow model jointly generates lensing and CIB maps","Wavelet flow matches CMB foreground power spectra","Non-Gaussian CMB foregrounds simulated by wavelet flow","Flow model recreates CMB foregrounds within 2.5%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1375,"prompt_tokens":965,"completion_tokens":410,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":327}},"tokens_in":581,"tokens_out":410,"duration_ms":4702,"temperature":1.0,"reasoning_tokens":327,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:32:45.511468+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the trained flow on an independent $\\kappa$-CIB dataset\\u2014for example, maps from a different simulation pipeline or reconstructed lensing convergence and Planck CIB data\\u2014and compute the same auto-spectra, cross-spectrum, and Minkowski functionals; if the bias against this external dataset exceeds a few percent, the claimed field-level accuracy does not generalize beyond the training simulation.","supporting_citations":[{"cited_title":"Wavelet Flow: Fast Training of High Resolution Normalizing Flows","cited_arxiv_id":"2010.13821","evidence_quote":"Introduces the Wavelet Flow architecture that decomposes images with the discrete wavelet transform and trains separate flows per scale."},{"cited_title":"Dai and U","cited_arxiv_id":null,"evidence_quote":"Prior multiscale normalizing-flow treatment of cosmological fields; source of the small post-Nyquist noise addition used in preprocessing."},{"cited_title":"Haar,Zur theorie der orthogonalen funktionensysteme","cited_arxiv_id":null,"evidence_quote":"Defines the orthonormal Haar wavelets used for all wavelet decompositions in the paper."},{"cited_title":"Efficient Out-of-Distribution Detection of Melanoma with Wavelet-based Normalizing Flows","cited_arxiv_id":"2208.04639","evidence_quote":"The wavelet-based normalizing-flow implementation that the authors' code builds on."}],"review_version":1}