{"id":"7fc898b6-1d78-4449-8cb7-8544da49f76b","arxiv_id":"2411.16899","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An expanded GNILC algorithm using a DCT along the frequency axis, a bias-corrected AIC for foreground degrees of freedom, and an RPCA-embedded blind variant recovers simulated 21-cm power spectra from SKA-MID and BINGO with roughly 10-20 percent bias in selected multipole ranges.","lead":"New single-dish 21-cm experiments need to strip enormous radio foregrounds to see the hydrogen signal. This paper upgrades the GNILC cleaning method with a frequency-domain transform, a bias-corrected model-selection rule, and a blind variant, recovering simulated SKA-MID and BINGO power spectra with about 10-20 percent power loss.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline recovery claim depends on foreground-rank selection via the modified AIC (Eq. 48), but Np is only heuristic and this AIC is never validated against the original AIC or known ranks.","rationale":"The reader's weakest assumption points at the same step I consider most load-bearing. The modified AIC in Eq. (48) is the device that determines how many foreground modes are removed; in this paper it is a novel analytical contribution, and the SKA-MID and BINGO results are produced with it. Yet Np is never defined quantitatively, and the modified criterion is never benchmarked against the original AIC or against a known foreground rank. This makes the central 10-20% power-loss claim conditional on a calibration that is not demonstrated. The concern is not fatal: the underlying ILC mathematics is standard, the bias derivation in Appendix B is internally consistent, and the 1000-realization check in Fig. 4 is a genuine stability test. But that stability test does not cover the AIC selection itself. Secondary issues, such as the abstract's ≤20% wording versus the 21.3% entry in Table 2 and the post-hoc multipole ranges, also support a conditional verdict, but the AIC/Np uncertainty is the one step whose failure would change all headline numbers. Since the reader already reached CONDITIONAL on essentially this ground, I do not propose a different verdict.","tokens_in":28476,"tokens_out":6665,"duration_ms":66314,"concrete_test":"Recompute the SKA-MID and BINGO cases on identical inputs with three AIC choices: (i) the original Olivari AIC (Eqs. 36-37); (ii) the modified Eq. (48) with the claimed Np; and (iii) Eq. (48) with Np scaled by 0.5 and 2.0. For each, record the selected m and the resulting Table-2 power-loss entries. On a noiseless power-law-only foreground, also compare the selected m with the rank of the true foreground covariance matrix. If m changes with Np or Table-2 entries shift by more than about 5 percentage points, the reported power-loss claim is not robust to the model-selection step.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To produce a cleaned map, eGNILC must first choose the foreground degrees of freedom m by minimizing the modified AIC in Eq. (48). This m fixes the dimensionality of the ILC constraint W S = S (Eqs. 26, 27, and 35), so every recovered power-spectrum entry in Table 2 inherits it. Eq. (48), however, depends on Np, the effective independent-pixel count, which is specified only through the heuristic \"effective theta_FWHM\" prescription in Sec. 3.1.1. The paper gives no formula for Np, shows no sensitivity test for Np, and never compares Eq. (48) with the original Olivari et al. (2016) AIC (Eqs. 36-37) on the same simulations. If Np is mis-estimated, the penalty term nch[1/(1-2m/Np)+ln(1-2m/Np)] changes, m shifts, and with it the recovered auto/cross spectra and the claimed 10-20% power-loss numbers. The authors' own caveat that the modified AIC should not be combined with RPCA means the SKA-MID and BINGO headline results rely on this AIC step, not on the blind variant. Figures 6 and 7 and Table 2 show degradation with more complex foregrounds, but no error bars or m-selection audit are provided, so the central recovery claim is conditional on this model-selection step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an extension of the GNILC foreground-removal method for 21-cm intensity mapping, called eGNILC. The extensions are (i) a Discrete Cosine Transform along the frequency axis before needlet-space ILC, (ii) a modified AIC criterion for selecting the number of foreground degrees of freedom, derived from an analytic calculation of the ILC bias in Appendix B, and (iii) an RPCA-based blind estimate of the 21-cm covariance that replaces the simulated-signal prior. The method is demonstrated on simulated SKA-MID-like and BINGO-like noiseless skies with a frequency-dependent Airy beam, with reported power losses of roughly 10-20% over the quoted multipole ranges. The paper claims that the DCT step reduces power loss at low multipoles and that the eGNILC bias depends on the averaging-domain size and foreground dof but not on the 21-cm signal itself.","tokens_in":28704,"tokens_out":5546,"duration_ms":53253,"significance":"If validated, the analytic bias result in Appendix B is a useful contribution: it gives a closed-form correction to the ILC bias that depends on m and Np rather than on the assumed signal, and it motivates a modified AIC. The DCT improvement at low multipoles is clearly visible in Fig. 3, and the RPCA-embedded variant is shown to work in the noiseless demonstrations of Fig. 2. The paper also includes a useful stability check against cosmic variance in Fig. 4. However, the headline SKA-MID and BINGO claims rest on a model-selection step whose key input, Np, is only heuristic, and the validation is partially self-referential because the signal prior is generated with the same CORA code used for the simulated input maps. No public code is mentioned, so the derivations are not machine-checked by the reader, but the algebraic steps in Appendix B are internally consistent.","major_comments":[{"comment":"The modified AIC is the only criterion used to set the foreground degrees of freedom m in the SKA-MID and BINGO runs, so every recovered spectrum in Figs. 6-7 and Table 2 inherits this m via Eqs. (32)-(35). However, Eq. (48) depends on Np, the effective number of independent pixels, which is introduced only through the qualitative \"effective theta_FWHM\" prescription in Sec. 3.1.1. No formula or algorithm for Np is given, no sensitivity test to Np is reported, and the modified AIC is never compared with the original AIC of Eqs. (36)-(37) on the same simulations. A mis-estimated Np shifts m and changes all of the quoted power-loss numbers, and the paper's own caveat that the modified AIC should not be combined with RPCA (Sec. 3.2, Sec. 6) means the headline results rely on this AIC step. Please add an explicit Np prescription, an m-selection audit, and a comparison of Eq. (48) with the original AIC.","section":"Sec. 3.1.1 and Sec. 3.1.2, Eq. (48)"},{"comment":"The abstract states that SKA-MID exhibits ≲20% power loss, but Table 2 lists 21.3% for the 1069.7 MHz auto-spectrum with B=2.5 and the \"Extrapol+Unresol\" foregrounds over 20<ell<250. This is an internal contradiction in the central quantitative claim. In addition, Table 2's header gives the BINGO multipole range as 30<ell<300, while the abstract and Sec. 6 state [20,300]; these numbers must be reconciled, and the abstract should report either the cell-wise values or the maximum rather than a bound that excludes one of the listed cells.","section":"Table 2 and Abstract"},{"comment":"The validation of the headline SKA-MID and BINGO results is partially self-referential. The mixing matrix is built from the simulated 21-cm covariance prior generated with the same CORA code that produced the input maps (Sec. 2.1), the tests omit instrumental noise, and no independent signal simulator is used. Because the RPCA variant is explicitly not used with the modified AIC in these runs, the claimed applicability to real data rests on a prior that is derived from the same simulation machinery as the input. Please quantify the sensitivity of the recovered spectra to the assumed Hi power-spectrum amplitude and shape, or test with an independent signal generator, before drawing conclusions about real BINGO and SKA-MID data.","section":"Sec. 3.1.1, Secs. 5.1-5.2"}],"minor_comments":[{"comment":"The text in Sec. 5 says the beam is not deconvolved, but Sec. 6 states that the Airy-disk beam was applied and then \"deconvolved after foreground removal\"; this contradiction should be resolved because it changes the interpretation of the high-ell residuals.","section":"Sec. 5 vs Sec. 6"},{"comment":"The DCT definition sums over n=1 to N-1, which appears to omit the n=0 term, and the scaling factor f combines constants in an unconventional way; please check this against the standard type-II DCT convention used by the scipy implementation named in the text.","section":"Eq. (1)"},{"comment":"The phrase \"imaginary mixing matrix\" is confusing, since the matrix S is real; a term such as \"effective mixing matrix\" or \"frequency-basis mixing matrix\" would be clearer.","section":"Sec. 3.3"},{"comment":"The BINGO rows for the 1257.5 x 1252.5 MHz cross-spectrum are empty even though Fig. 7 shows this cross-power spectrum; please either report the values or state explicitly why they are omitted.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of the journal and the core analytic derivation appears sound. My major comments are all addressable within the manuscript's scope: an explicit Np prescription plus a comparison of the modified AIC with the original AIC, a sensitivity test of the recovery to the signal prior, and a corrected statement of the headline power-loss numbers. If these are provided, the paper could be suitable for publication. No concerns about citation behavior or novelty disclosure beyond the points already listed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on 21-cm intensity mapping foreground cleaning. The genuinely new pieces are the DCT-before-needlets step, the bias-corrected AIC in Eq. (48), and the RPCA-embedded mixing-matrix estimation. On reading, the ILC math is standard, the bias derivation in Appendix B is internally consistent, and the demonstration plots do show the DCT step helping at low multipoles. Credit where it is due: the paper tests on SKA-MID-like and BINGO-like mocks with varying Airy beams and masks, reports auto and cross spectra, and explicitly warns that the modified AIC should not be combined with RPCA. That is honest and keeps the headline experiments tied to the tested pipeline.\n\nThe soft spots are real but not fatal. The main one is that Eq. (48) depends on Np, the effective independent pixel count, which is never concretely defined. The paper gestures at an effective theta_FWHM and says the method is insensitive to the domain shape, but it shows no sensitivity test for Np and never compares the modified AIC with the original AIC from Olivari et al. (2016) on the same simulations. Since m sets the ILC constraint, any error in Np moves the recovered spectra, and the force of the 10-20% power-loss claims depends on this step not being tuned. Second, the abstract's \"less than 20%\" claim is contradicted by the 21.3% entry in Table 2 (SKA-MID 1069.7 MHz, Extrapol+Unresolved, B=2.5). That is a small inconsistency, but it undercuts the headline and should be fixed. Third, the recovered spectra are from single realizations; there are no error bars on Table 2, so the differences between B=2.5 and B=1.7 are hard to judge. Lack of code is a practical hindrance for a methods paper, and using CORA-generated signal covariances to clean CORA-generated signal maps makes the validation partly self-referential. That last point is common in this literature, but it should be stated.\n\nBottom line: this is a competent, incremental methods paper. The central idea is not flawed, but the headline claims are conditional on the unvalidated AIC/Np step. A serious referee would want Np specified or tested, a direct AIC comparison, error bars on the power-loss numbers, and ideally code. The likely readers are people building single-dish 21-cm pipelines for BINGO, MeerKAT, or SKA-MID.\n\nRecommendation: yes, send it to peer review. It is the kind of paper that improves with referee input and should be in the literature, but not with the current over-broad abstract.","headline":"A credible incremental extension of GNILC for 21-cm foreground cleaning, but the headline power-loss numbers rest on an unvalidated model-selection step.","tokens_in":29410,"tokens_out":1860,"would_cite":false,"duration_ms":17224,"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":"eGNILC recovers the 21-cm power spectrum in simulated SKA-MID and BINGO skies with 10-20 percent power loss.","keywords":["21-cm intensity mapping","foreground removal","GNILC","needlet internal linear combination","discrete cosine transform","Akaike information criterion","robust principal component analysis","SKA-MID"],"falsifier":"Run eGNILC on the same SKA-MID mock twice, once with the heuristic \\(N_p\\) multiplied by 0.5 and once by 2; if the recovered power spectrum moves by more than the quoted 10-20 percent, the result depends on an unvalidated \\(N_p\\) and the AIC step is the fragile link.","tokens_in":28182,"feed_emoji":"📡","tokens_out":6227,"duration_ms":56013,"temperature":0.7,"pith_summary":"This paper expands the Generalized Needlet Internal Linear Combination method into eGNILC for 21-cm intensity mapping, where foregrounds outshine the signal by roughly four orders of magnitude. The key moves are a discrete cosine transform along the frequency axis, a bias-corrected Akaike criterion for choosing the foreground degrees of freedom, and an optional robust principal component analysis step that makes the method blind. In noiseless SKA-MID-like and BINGO-like simulations with a varying Airy beam, eGNILC recovers the 21-cm auto- and cross-power spectra with about 10-20 percent power loss over multipoles \\(\\ell \\in [20,250]\\) and \\(\\ell \\in [20,300]\\), respectively. The paper also derives an eGNILC bias that depends on the averaging domain size and foreground degrees of freedom rather than on the 21-cm signal itself, and shows that the frequency transform reduces power loss specifically at low multipoles. A reader should care because successful foreground removal is the main obstacle to measuring the low-redshift 21-cm power spectrum with single-dish telescopes.","feed_headline":"eGNILC recovers 21-cm spectra with 10-20 percent power loss","feed_subtitle":"DCT basis plus bias-corrected AIC cleans SKA-MID-like and BINGO-like skies over multipoles 20-250 and 20-300.","key_machinery":"The central machinery is the needlet-space internal linear combination with mixing matrix \\(S = $R_s^{{1/2}}$U_s\\), where \\(R_s\\) is the 21-cm signal covariance and \\(U_s\\) is the signal subspace obtained from the whitened data covariance. The discrete cosine transform of the frequency axis makes the foreground contribution low-rank; the modified AIC, \\(\\mathrm{AIC}(m, N_p) = 2m + n_{ch}\\left[1/(1-2m/N_p) + \\ln(1 - 2m/N_p)\\right] + \\sum_{i=1}^{n_{ch}-m}[\\mu_i - \\ln\\mu_i - 1]\\), selects the foreground degrees of freedom \\(m\\); and robust principal component analysis splits the data covariance into a low-rank foreground part plus a sparse signal part to supply a blind estimate of the mixing matrix.","core_discovery":"eGNILC performs the internal linear combination in needlet space but first transforms the frequency axis with a discrete cosine transform, so that spectrally smooth foregrounds occupy only a few low-order modes. The mixing matrix is built from the 21-cm signal subspace, and the number of foreground modes \\(m\\) is selected by a modified Akaike Information Criterion that includes the eGNILC bias \\(1 - 2m/N_p\\), where \\(N_p\\) is the effective number of independent pixels in the covariance-estimation domain. In simulations, the recovered 21-cm auto- and cross-power spectra match the input to within roughly 20 percent for SKA-MID and 10 percent for BINGO over the quoted multipole ranges when no instrumental noise is added; the frequency-dependent Airy beam causes serious errors only at large multipoles. The paper additionally shows that when adjacent frequency channels are highly correlated, the algorithm must be applied to decimated subsets of channels, and that the eGNILC bias is negligible for simple power-law foregrounds outside the Galactic plane when no beam is present.","pith_inferences":["The paper leaves thermal and \\(1/f\\) noise untreated, so a natural extension is to add instrumental noise to the SKA-MID and BINGO mocks and test whether the RPCA sparse component absorbs the noise as the text suggests.","Because high adjacent-channel correlation forced the SKA-MID analysis to split 500 channels into 10 subsets, eGNILC's performance at very fine frequency resolution will depend on how that grouping is chosen; a formal criterion for choosing the number of subsets would settle this dependence.","The modified AIC is never compared with the original AIC on the same mocks; an explicit comparison would show whether the \\(N_p\\) term actually changes the selected foreground degrees of freedom and the recovered spectrum."],"forward_implications":["For simple power-law foregrounds with no beam, eGNILC leaves the 21-cm power spectrum essentially unbiased outside the Galactic plane.","With a realistic frequency-dependent Airy beam and no thermal noise, eGNILC recovers SKA-MID power spectra within about 10-20 percent over multipoles \\(\\ell \\in [20,250]\\) and BINGO spectra within about 10 percent over \\(\\ell \\in [20,300]\\).","Applying the discrete cosine transform along frequency reduces the power loss at low multipoles, where plain GNILC lacks accuracy.","When adjacent frequency channels are highly correlated, eGNILC must be run on decimated uncorrelated subsets; BINGO-like channel spacing avoids this need.","The usable multipole range is limited at low \\(\\ell\\) by the number of independent samples available for covariance estimation and at high \\(\\ell\\) by the frequency-varying beam."],"supporting_citations":[{"why":"Introduces the needlet internal linear combination (GNILC) method that this paper expands.","marker":"Remazeilles et al. (2011)"},{"why":"First applies GNILC to 21-cm intensity mapping and supplies the original AIC criterion that the modified version refines.","marker":"Olivari et al. (2016)"},{"why":"Derives the ILC bias due to limited covariance samples, which the eGNILC bias calculation adapts to 21-cm data.","marker":"Delabrouille et al. (2009)"},{"why":"Provides the simulated 21-cm signal and unresolved synchrotron and point-source foreground models used in the mocks.","marker":"Shaw et al. (2014)"},{"why":"Supplies the robust principal component analysis low-rank/sparse split used to make the mixing-matrix estimate blind.","marker":"Zuo et al. (2018)"},{"why":"Defines the SKA-MID telescope configuration used for the single-dish simulation forecast.","marker":"Square Kilometre Array Cosmology Science Working Group et al. (2020)"},{"why":"Defines the BINGO telescope design and specifications used for the second forecast.","marker":"Battye et al. (2013)"}],"fun_headline_variants":["eGNILC uses DCT to save 21-cm signal from foregrounds","eGNILC keeps 80-90% of 21-cm power","Blind eGNILC cleans 21-cm skies with no foreground model","eGNILC: DCT + bias-corrected AIC for 21-cm foreground removal","Airy beam only hurts eGNILC at large multipoles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The recovery claim rests on the AIC correctly identifying the foreground degrees of freedom \\(m\\), which requires knowing the effective number of independent pixels \\(N_p\\); the paper sets \\(N_p\\) through a heuristic beam-based estimate and never validates it against the original AIC.","fun_headline_variants_meta":{"raw":{"variants":["eGNILC uses DCT to save 21-cm signal from foregrounds","eGNILC keeps 80-90% of 21-cm power","Blind eGNILC cleans 21-cm skies with no foreground model","eGNILC: DCT + bias-corrected AIC for 21-cm foreground removal","Airy beam only hurts eGNILC at large multipoles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000669,"raw_usage":{"total_tokens":3150,"prompt_tokens":1143,"completion_tokens":2007,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":759,"completion_tokens_details":{"reasoning_tokens":1898}},"tokens_in":759,"tokens_out":2007,"duration_ms":14653,"temperature":1.0,"reasoning_tokens":1898,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:46:24.367170+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run eGNILC on the same SKA-MID mock twice, once with the heuristic \\(N_p\\) multiplied by 0.5 and once by 2; if the recovered power spectrum moves by more than the quoted 10-20 percent, the result depends on an unvalidated \\(N_p\\) and the AIC step is the fragile link.","supporting_citations":[],"review_version":1}