{"id":"5a302a5d-e525-43c7-9fa1-023d72f41137","arxiv_id":"2506.09170","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"HyFoReS reduces simulated gain-induced foreground leakage in 21-cm power spectra by up to 1000 times, reaching the thermal noise level only for bandpass errors below about 1e-4.","lead":"This paper tests HyFoReS, a foreground-removal algorithm, on simulated Canadian Hydrogen Intensity Mapping Experiment data with injected antenna gain errors. It finds the method suppresses gain-induced foreground bias by up to three orders of magnitude, and below thermal noise for small bandpass errors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Point sources are turned off because the KL filter cannot model them; with real point-source foregrounds, the core condition KvF<<vHI in Eq. (5) fails, so the demonstrated below-noise suppression is unproven for the real sky.","rationale":"The paper is a well-structured simulation study with an explicit formalism and controlled numerical tests. The algebraic construction of the window matrix, the treatment of the complex filter nonlinearity in Eq. (17), and the identification of noise in the gain estimates in Fig. 6 are genuinely supportive. The strongest claim is appropriately qualified: below thermal noise for bandpass RMS at or below 10^-4, three-order suppression for antenna gains, with second-order perturbations and noise in gain estimation identified as limiting factors. The most load-bearing unaddressed assumption is the foreground model. Equation (5) requires KvF << vHI. The KL filter's covariance contains only Galactic synchrotron, so point sources are removed from the simulation (Sec. III A). Since real foregrounds include point sources that the KL filter does not model, those sources will contribute a residual K vF that is not proportional to the gain errors, contaminating the gain estimator in Eq. (10) and escaping the subtraction in Eq. (15). The authors are transparent about this limitation in Sec. V C, and the synchrotron foreground is indeed the dominant large-scale component, so the results are not vacuous. However, the central numerical claims are established only in the regime where the core assumption is guaranteed. A rerun with point sources enabled, ideally with and without a point-source term in the KL covariance, would settle whether the demonstrated margin survives a realistic sky. Since the paper already discusses this limitation and no code or data are provided to test it independently, the conditional verdict stands without change.","tokens_in":23056,"tokens_out":3951,"duration_ms":44939,"concrete_test":"Re-run the Fig. 2 bandpass setup at 10^-4 RMS with point sources enabled in the foreground map, using the existing CHIME point-source generator, while leaving the KL covariance unchanged. Compare the post-HyFoReS power-spectrum bias |P - Pfid|/sigma with the point-source-free run at each k bin. If the bias exceeds approximately 1 outside the foreground wedge, or if the suppression factor drops below three orders of magnitude, the claim is conditional on the point-source-free foreground model. As a control, also run with a KL covariance that includes point-source modes; if suppression is restored, the bottleneck is the foreground model rather than HyFoReS itself.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition is stated in Eq. (5): after KL filtering, the estimated signal is dominated by systematics-coupled foregrounds, requiring KvF << vHI so that the residual is approximately KGvF. The paper's simulations turn off point sources because 'the KL foreground filter does not include a covariance of point sources' (Sec. III A). The real radio sky contains point sources, which are spectrally smooth and would survive the KL filter as unmodeled signal-like modes. In that regime, KvF can be comparable to or larger than vHI, and the gain estimator in Eq. (10) receives an extra term K vF that is not proportional to the gain errors. The subtraction in Eq. (15) then cannot separate systematics-induced leakage from intrinsic point-source leakage. The authors explicitly flag this limitation in Sec. V C, but the consequence is that the central numerical claims—'below thermal noise' for 10^-4 bandpass errors and the three-order-of-magnitude suppression factor—are demonstrated only for foregrounds that are smoother than the real sky. This is not a peripheral realism issue; it is the regime in which the core approximation is guaranteed to hold.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper extends the HyFoReS systematics-foreground-removal formalism to gain-type errors in simulated CHIME data. After reviewing the general algorithm, in which a linear foreground filter produces a signal estimate dominated by systematics-coupled foregrounds and a foreground estimate is cross-correlated with it to estimate the perturbation parameters, the authors treat bandpass errors and, more elaborately, complex antenna-dependent gains on stacked visibilities. The complex case requires separate real and imaginary derivative matrices, a block window matrix, and a restriction to co-polarization modes. The numerical tests inject gain errors with RMS amplitudes from 1e-5 to 1e-2 into full- and half-pathfinder telescope simulations and measure the two-dimensional HI power-spectrum bias before and after HyFoReS. The paper reports below-thermal-noise foreground suppression for bandpass errors at 1e-5 and 1e-4, and roughly three orders-of-magnitude suppression for 1e-3 and 1e-2 bandpass or antenna errors. It explicitly lists second-order gain terms, noise in the gain estimator, the exclusion of point sources from the KL foreground covariance, and the sparse baseline coverage of the half-pathfinder as limitations.","tokens_in":23222,"tokens_out":13016,"duration_ms":137512,"significance":"The paper is a solid, transparent simulation study. Its main strengths are that it tests the estimator on injected gains rather than fitting the power spectrum, it explicitly verifies the noise-gain term in Eq. (49) against simulation, it handles the non-commutativity of complex conjugation and the KL filter by splitting gains into real and imaginary parts, and it reports quantitative suppression factors with the limitations stated in the text. The claimed thresholds (e.g., below-thermal-noise performance for 1e-4 bandpass errors) are concrete and testable. If the approach survives inclusion of realistic point-source foregrounds, it would be a useful extension of a method already demonstrated on CHIME beam errors. The point-source exclusion is the principal reason the results do not yet establish the abstract's broad claim for real 21-cm telescopes.","major_comments":[{"comment":"Point sources are removed from the simulations because the KL foreground filter does not include a point-source covariance, but this directly controls the regime in which the central approximation holds. Eq. (5) states that after the linear filter the estimated signal is dominated by the systematics-coupled foreground, KvF << vHI; the gain estimator in Eq. (10) and the subtraction in Eq. (15) are derived under that condition. On the real radio sky, point sources are spectrally smooth and are not represented in the KL covariance used here, so the filter will not remove them as it removes the diffuse synchrotron component. An unmodeled point-source term K vF would enter the estimated signal and the cross-correlations in Eqs. (7) and (10), biasing the gain estimates and leaving point-source leakage that the subtraction in Eq. (15) cannot distinguish from the HI signal. The abstract and Sec. VI extend the conclusions to '21-cm telescopes' and 'real CHIME telescope data' without this caveat. The manuscript should either (i) include a point-source term in the KL covariance and rerun the central tests, (ii) measure the degradation when point sources are added without modifying the filter, or (iii) explicitly restrict the headline claims to diffuse Galactic synchrotron foregrounds and state that point-source leakage is not yet quantified.","section":"Sec. III A; Eq. (5); Sec. V C"},{"comment":"The antenna-dependent-gain test on the half-pathfinder telescope has an unperturbed power spectrum that already shows roughly an order-of-magnitude foreground bias at low k_perp (Fig. 4, top-left panel), attributed to sparse baseline sampling. As a result, the residual low-k_perp bias in the cleaned panels for the 1e-3 and 1e-2 cases may be the telescope's intrinsic alias floor rather than gain-induced leakage that HyFoReS failed to remove. The quoted three-order suppression is a before/after statement within this configuration, and it does not by itself show that HyFoReS brings the spectrum to the level that would be achieved by a perfect-gain version of the same telescope. The paper should quantify the residual relative to the unperturbed spectrum (for example, (P_cleaned - P_unperturbed)/sigma) and report how much of the residual at low k_perp would persist with zero gain errors.","section":"Sec. IV B; Fig. 4; Sec. V A"}],"minor_comments":[{"comment":"The half-pathfinder integration time listed as 18,250 days (50 years) is surprising; please clarify whether this is an effective integration after redundant-baseline stacking or a typo, because the noise-gain floor in Sec. IV B and the comparison in Sec. V A depend on the actual integration time.","section":"Table I; Sec. IV B"},{"comment":"The color scale is saturated and uses the same range for all panels; because the key claim is 'bias below the one-sigma error', the unperturbed and cleaned panels would be easier to evaluate with a symmetric scale and visible colorbar limits.","section":"Figs. 2 and 4"},{"comment":"Equation (16) summarizes the cleaned signal as vHI + G^2 vF, but the noise-gain term identified in Eq. (47) is not shown; adding it to the display, or explicitly saying it is omitted for clarity, would make the later discussion in Sec. IV B more direct.","section":"Eq. (16); Sec. IV B"},{"comment":"The sentence describing Eq. (44) says the pseudoinverse is taken first and then rows and columns for cross-polarizations are zeroed, but the equation notation could also be read as a pseudoinverse of the restricted matrix; since these operations do not commute in general, please specify the exact order.","section":"Eq. (44)"},{"comment":"The abstract states that bandpass perturbations are reduced below thermal noise when the RMS is on the order of 1e-4 or lower, but the body of the paper qualifies this as 'over most k bins' with the wedge region excepted; the abstract should carry the same qualification.","section":"Abstract; Sec. IV A"}],"recommendation":"major_revision","confidential_remarks":"The point-source limitation is the main barrier and is a genuine correctness-risk concern rather than a style issue: it defines the regime in which the core approximation in Eq. (5) is guaranteed to hold. The other concerns (unperturbed alias floor, integration time, presentation) are manageable. I see no evidence of circularity in the headline results, and the authors' explicit caveats are a credit. If the authors can add point-source tests or honestly restrict the claims to diffuse synchrotron foregrounds, a revised version could be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: a competent, honest extension of HyFoReS to gain-type errors in polarized visibilities. The real/imaginary decomposition for complex antenna gains is a clean piece of formalism, and the tests are controlled and informative. The main caveat is not hidden: the paper turns off point sources because the KL filter doesn't model them, so the \"below thermal noise\" claim only holds for smooth Galactic foregrounds. That is exactly the regime where Eq. (5) is guaranteed to hold. With real point sources, unmodeled modes can survive KL filtering and break the KvF << vHI condition. The authors flag this in Sec. V C, and it's a fair limitation, but it means the headline numbers are not yet established for the real sky.\n\nWhat's genuinely new: previous HyFoReS work was beam perturbations in map space and a toy simulation. Here it's complex antenna gains in realistic polarized visibilities, with a parametrization that handles the complex-nonlinear KL filter. The paper also diagnoses the failure mode for small antenna gains: spurious foreground-noise correlation in the gain estimates (Fig. 6 is a nice check). That's real work.\n\nSoft spots: (1) point-source foregrounds off; (2) no code or data; (3) the antenna-gain case never reaches the thermal noise floor, so the main \"success\" for that case is a three-order-of-magnitude bias reduction, not below-noise cleanliness. All three are explicitly acknowledged. The computational cost claim (O(10^3) CPU hours) is plausible but there's no runtime benchmark.\n\nCitation pattern looks normal. The self-citations are to the papers that actually introduced HyFoReS, so they're warranted.\n\nThis paper deserves a serious referee. It's not ready as-is because the point-source gap affects the central quantitative claims, but the formalism and diagnostics are solid. A referee should ask for point-source tests (or a KL filter that models them) and for code/data release.","headline":"Useful, honest extension of HyFoReS to gain errors in polarized visibilities; the headline below-noise claim is only proven for smooth foregrounds because point sources are switched off.","tokens_in":23804,"tokens_out":2174,"would_cite":true,"duration_ms":21937,"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":"HyFoReS, a two-step foreground filter, estimates and subtracts antenna gain errors from simulated 21-cm data, reducing power-spectrum foreground bias by up to three orders of magnitude and below the noise floor for small bandpass errors.","keywords":["21-cm intensity mapping","foreground subtraction","HyFoReS","antenna gain calibration","CHIME","power spectrum bias","Karhunen-Loève filter","polarized visibilities"],"falsifier":"Repeat the identical bandpass-error simulation with point sources included in the sky and in the KL filter's foreground covariance: if the cleaned power spectrum retains foreground bias well above the noise floor at $10^{-4}$ gains, the premise that filtered residuals are dominated by gain-coupled foregrounds is false. As a separate check, verify that post-cleaning bias scales with the square of the injected gain amplitude; a linear scaling would mean the gain estimator itself is biased rather than limited by second-order terms.","tokens_in":82,"feed_emoji":"📡","tokens_out":8048,"duration_ms":150964,"temperature":0.7,"pith_summary":"21-cm intensity mapping tries to map cosmic hydrogen, but radio foregrounds swamp the signal by orders of magnitude, and telescope gain errors imprint spectral structure on the foregrounds so that standard linear filters no longer remove them. This paper tests HyFoReS, an algorithm that first applies a linear foreground filter and then cross-correlates the filtered signal with a foreground-dominated data estimate to isolate the gain-coupled residual emission and subtract it. In CHIME-style simulations, HyFoReS brings foreground bias in the power spectrum below thermal noise for bandpass gain errors at the $10^{-5}$ and $10^{-4}$ level, and reduces the bias by roughly three orders of magnitude for larger bandpass errors and for complex antenna-dependent gains. The result matters because gain calibration errors at the percent level are realistic, and the same machinery works in visibility space or map space for any parametrizable systematic.","feed_headline":"Filter scrubs antenna gain errors from 21-cm data","feed_subtitle":"In CHIME-like simulations the gain-induced foreground bias falls by up to 1,000x, easing calibration demands.","key_machinery":"The central object is the window matrix $W$, which maps the true gain perturbations into the gains that the cross-correlation step actually measures; HyFoReS pseudoinverts $W$ to recover the gain modes that survive the KL filter, and uses those unwindowed gains to subtract foreground residuals. For complex antenna gains, each gain is split into real and imaginary parts with derivative matrices $\\Gamma$ and $\\Delta = i\\Gamma$, the window matrix is built as a block matrix coupling the two parts, and only the XX and YY co-polarization visibilities enter the fit because cross-polarization foregrounds are too weak to estimate reliably. The argument runs on the foreground-signal hierarchy (foregrounds dominate by roughly $10^5$) and on treating gains as small, so that first-order perturbation theory suffices and the cleaned signal carries only second-order foreground residuals.","core_discovery":"The paper's claim is that HyFoReS can estimate antenna gain errors directly from the data, without a sky model, and subtract the foregrounds those errors leak into the 21-cm signal. After the KL foreground filter has removed the intrinsic smooth foregrounds, the residual in the filtered data is dominated by foregrounds multiplied by gain perturbations; HyFoReS cross-correlates that residual with the unfiltered visibilities, a good foreground proxy because foregrounds outshine the signal by a factor of about $10^5$, to estimate the gains, corrects the estimates through the pseudoinverse of a window matrix that accounts for modes the KL filter removed, and subtracts the first-order foreground residual. For frequency-only bandpass errors this leaves the power spectrum unbiased at the noise level for RMS perturbations of $10^{-5}$ and $10^{-4}$, while $10^{-3}$ and $10^{-2}$ level errors are suppressed by about three orders of magnitude. For complex antenna-dependent gains, the real and imaginary parts are estimated separately and co-polarization visibilities alone are used; the bias is again cut by three orders of magnitude, with the residual set by noise in the gain estimates and by second-order ($G^2 v_F$) foreground terms.","pith_inferences":["The simulations omit point sources because the KL filter's foreground covariance does not model them, so the tested sky is smoother than the real one; on real data, point-source residuals could violate the premise that only gain-coupled foregrounds survive the linear filter, and the authors' ongoing point-source model work will determine how much this degrades performance.","The noise floor in gain estimation scales with the dimensionality of the gain parameter space, so reducing that space, for example by exploiting smooth frequency dependence of per-feed gains or longer time-stationarity, could extend HyFoReS' sub-noise performance to the antenna-dependent case.","An iterative or Wiener-filtered variant of HyFoReS, which the paper suggests, could be tested directly in these simulations: if residuals currently scale as $G^2 v_F$, a second pass should reduce them further unless noise in the first-pass gain estimates dominates."],"forward_implications":["Bandpass gain errors at RMS $10^{-4}$ or below no longer bias the 21-cm power spectrum above thermal noise after HyFoReS, so calibration tolerances at those frequencies can be relaxed.","At the percent-level gain errors typical of calibrated interferometers, foreground bias is still reduced by about three orders of magnitude, improving the prospects for detecting the 21-cm auto-power spectrum.","Because the method requires only a linear foreground filter and a foreground-dominated data estimate, it carries over to any parametrizable, time-independent systematic and to map-space analyses, not just the gain cases demonstrated here.","The residual after cleaning is dominated by second-order gain-foreground coupling and by noise in the gain estimates, so further gains would require iterative cleaning or noise-filtered gain estimators."],"supporting_citations":[{"why":"Original derivation of HyFoReS; supplies the cross-correlation estimator, window matrix, and pseudoinverse formalism the paper extends.","marker":"[32]"},{"why":"Demonstrates HyFoReS on real CHIME maps for beam perturbations; establishes the map-space precedent this paper extends to gains in visibility space.","marker":"[33]"},{"why":"Defines the KL foreground filter and m-mode formalism used as the linear filter inside HyFoReS throughout the simulations.","marker":"[28]"},{"why":"Provides the CHIME simulation pipeline that generates sky maps, visibilities, and power spectra used for the tests.","marker":"[27]"},{"why":"Describes the CHIME pathfinder telescope that the simulated full and half pathfinders are modeled on.","marker":"[36]"},{"why":"Reports the real CHIME gain calibration level (sub-percent) and beam systematics that motivate the perturbation amplitudes and the auto-spectrum goal.","marker":"[15]"}],"fun_headline_variants":["HyFoReS cuts gain-induced foreground bias by 1000x","Antenna gain errors scrubbed from 21-cm data","New filter tames antenna gain errors in CHIME sims","HyFoReS reduces foreground bias to noise level","Gain error mitigation for 21-cm intensity mapping"],"cache_read_input_tokens":25984,"weakest_assumption_plain":"The method assumes the linear foreground filter has already removed the smooth galactic emission almost completely, a condition the simulations secure only by deleting point sources from the sky model because the filter cannot handle them.","fun_headline_variants_meta":{"raw":{"variants":["HyFoReS cuts gain-induced foreground bias by 1000x","Antenna gain errors scrubbed from 21-cm data","New filter tames antenna gain errors in CHIME sims","HyFoReS reduces foreground bias to noise level","Gain error mitigation for 21-cm intensity mapping"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1344,"prompt_tokens":1015,"completion_tokens":329,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":244}},"tokens_in":631,"tokens_out":329,"duration_ms":3994,"temperature":1.0,"reasoning_tokens":244,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:55:10.887428+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the identical bandpass-error simulation with point sources included in the sky and in the KL filter's foreground covariance: if the cleaned power spectrum retains foreground bias well above the noise floor at $10^{-4}$ gains, the premise that filtered residuals are dominated by gain-coupled foregrounds is false. As a separate check, verify that post-cleaning bias scales with the square of the injected gain amplitude; a linear scaling would mean the gain estimator itself is biased rather than limited by second-order terms.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Original derivation of HyFoReS; supplies the cross-correlation estimator, window matrix, and pseudoinverse formalism the paper extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the KL foreground filter and m-mode formalism used as the linear filter inside HyFoReS throughout the simulations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CHIME simulation pipeline that generates sky maps, visibilities, and power spectra used for the tests."},{"cited_title":"Bandura, G","cited_arxiv_id":null,"evidence_quote":"Describes the CHIME pathfinder telescope that the simulated full and half pathfinders are modeled on."},{"cited_title":"Amiri, K","cited_arxiv_id":null,"evidence_quote":"Reports the real CHIME gain calibration level (sub-percent) and beam systematics that motivate the perturbation amplitudes and the auto-spectrum goal."}],"review_version":1}