{"id":"363d8897-dec2-4aa3-af2f-126c5d32bc66","arxiv_id":"2412.02246","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An FFT-based estimator recovers the angular bispectrum from gridded radio-interferometric visibilities across all triangle shapes, validated on simulated MWA data at 154.25 MHz.","lead":"This paper presents an FFT-based estimator that measures the angular bispectrum, a three-point statistical summary of the radio sky, directly from gridded interferometric visibility data. It validates the method on simulated Murchison Widefield Array observations, recovering a known non-Gaussian signal within about 10 to 15 percent across many triangle shapes and sizes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 12 omits the Q^{-3} factor required by Eq. 8, so the estimator as written is not normalized to recover the angular bispectrum in brightness-temperature units.","rationale":"The paper's central claim is that the binned estimator of Eq. 12 recovers the angular bispectrum. The most load-bearing condition for that claim is that the estimator's normalization correctly converts the three-visibility product into brightness-temperature units. Eq. 8 gives the three-visibility correlation as A Q^3 B, with A = \\pi\\theta_0^2/3. Eq. 12 uses only 1/A, omitting Q^{-3}. This is not a subtle modeling assumption but a direct dimensional inconsistency in the central formula: if a reader implements the equations as printed, the estimator will not recover the ABS. The simulation results in Figs. 2-4 presumably used the correct normalization in code, but the text as written cannot reproduce them without an undeclared Q=1 convention or a hidden Q^{-3} factor. This concern is more fundamental than the Gaussian-beam and slow-variation assumptions flagged by the reader: even if the beam is perfectly Gaussian and B is slowly varying, Eq. 12 as printed is still off by Q^3. The fix is straightforward but must be stated for the paper to be reproducible. The gridding damping factor (~0.89) and low-\\ell beam convolution are additional systematics, but they are secondary to the missing normalization. The reader's weakest_assumption focused on the beam model and slow-variation; the Q^{-3} omission is a distinct and more direct issue, hence partial agreement.","tokens_in":15826,"tokens_out":20350,"duration_ms":207080,"concrete_test":"Implement Eq. 12 exactly as printed on the paper's simulated visibilities, which are generated via Eq. 4 with the stated Q = 2k_B/\\lambda^2, and compare the output to the analytic B_Ana of Eq. 14. If the output is Q^3 times B_Ana (i.e., many orders of magnitude too large), the published estimator is missing a Q^{-3} factor. Alternatively, re-derive the expectation of Eq. 12 by substituting Eq. 8 into the sums; the derivation will show \\langle\\hat{B}\\rangle = Q^3 B(\\ell_1,\\ell_2,\\ell_3), proving the missing factor.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central estimator is defined in Eq. 12 as \\hat{B} = (1/A) \\sum_\\theta D_1 D_2 D_3 / \\sum_\\theta I_1 I_2 I_3, with A = \\pi\\theta_0^2/3. However, the visibility-bispectrum relation in Eq. 8 contains an explicit factor Q^3: \\langle V(U_1)V(U_2)V(U_3+\\Delta U)\\rangle = (\\pi\\theta_0^2/3) Q^3 B(\\ell_1,\\ell_2,\\ell_3). Since D and I are linear in the gridded visibilities V_g, the ratio in Eq. 12 estimates the Q^3-weighted visibility product. To recover B in K^3, the estimator must divide by A Q^3, not A alone. The paper never states that Q is set to unity in the simulations, nor that visibilities are pre-divided by Q. For the MWA at 154.25 MHz, Q = 2k_B/\\lambda^2 is far from unity (hundreds of Jy sr^{-1} K^{-1}), so Q^3 is many orders of magnitude large. An independent implementation of the printed equations would therefore not reproduce the claimed 10-15% recovery shown in Figs. 2-4. This normalization error is independent of the Gaussian-beam and slow-variation assumptions and directly affects the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an FFT-based estimator for the binned angular bispectrum (ABS) from gridded radio-interferometric visibility data. The estimator follows Shaw et al. (2021) and uses the three-visibility correlation relation derived for a Gaussian primary beam to normalize the binned bispectrum. The estimator is validated using 500 noiseless simulated realizations of an MWA observation at 154.25 MHz, with a non-Gaussian sky model whose ABS is known analytically. The authors report recovery of the input ABS to 10–15% accuracy over ℓ1 ∈ [46, 1320] and a range of triangle shapes (μ, t).","tokens_in":16105,"tokens_out":6248,"duration_ms":61771,"significance":"If the normalization issue identified below is corrected, the estimator would provide a computationally efficient method (scaling as N_t^2 log N_t) to measure the angular bispectrum from interferometric observations, covering all closed triangle configurations. The paper's validation strategy is solid: the analytical bispectrum is externally specified, and the comparison includes the discrete sampling correction [B_Ana]_d. The derivation in Appendix A is clean, and the paper gives a clear statement of the slow-variation and small-baseline caveats. However, the printed estimator is not self-consistent in its units, which directly affects the central claim.","major_comments":[{"comment":"The estimator omits the factor Q^{-3} required by the visibility-bispectrum relation of Eq. (8) (derived in Eq. A12). Eq. (8) states ⟨V(U1)V(U2)V(U3+ΔU)⟩ = (πθ0^2/3) Q^3 exp(...) B(ℓ1,ℓ2,ℓ3). Since D and I in Eq. (12) are constructed from the gridded visibilities V_g without any division by Q, the ratio in Eq. (12) estimates Q^3 times the weighted average of the three-visibility product. Dividing by A = πθ0^2/3 alone therefore yields Q^3 B, not B. With Q = 2k_B/λ^2 ≈ 700 Jy sr^{-1} K^{-1} at 154 MHz, Q^3 is many orders of magnitude away from unity. The manuscript nowhere states that Q is set to unity, nor that the visibilities are pre-divided by Q. An independent implementation of the printed equations would not reproduce the 10–15% recovery shown in Figs. 2–4. The estimator should be explicitly normalized by A Q^3, or the visibility products should be defined with V_g/Q, or the text must state and justify a convention with Q = 1.","section":"§2, Eq. (12)"}],"minor_comments":[{"comment":"The caption defines ΔB = |B−[B_Ana]_d|/B, whereas the text defines ΔB = |B−[B_Ana]_d|/[B_Ana]_d; please reconcile these definitions, presumably using [B_Ana]_d in the denominator.","section":"§4, Fig. 4 caption"},{"comment":"The paper states that the factor exp(−π^2 θ0^2 ΔU^2/3) has value 0.89 for a typical (ΔU)^2 = (ΔU_g)^2/2, but this correction is not applied to the estimator. A brief comment on why the residual 0.89 factor does not introduce a systematic bias in the binned estimates would help the reader understand the validation.","section":"§3"},{"comment":"The definitions of D and I use a continuous θ variable, but the sums are over a discrete grid; it would be clearer to state explicitly that θ denotes the coordinate in the Fourier-plane image and that the FFT is used to evaluate these sums.","section":"§2, Eqs. (10)–(11)"},{"comment":"Shaw et al. (2021) is cited as arXiv:2107.14564 with no journal reference; please confirm the publication status and update the reference if a peer-reviewed version exists.","section":"References"},{"comment":"The text contains the typo 'for for ℓ≥80'; this should be corrected.","section":"§4"}],"recommendation":"major_revision","confidential_remarks":"The missing Q^{-3} factor is the main technical issue; it appears to be a units oversight rather than a fundamental flaw in the method, but the manuscript must be corrected before publication. The validation is limited to noiseless simulations with no foregrounds or systematics, which is acknowledged and appropriate for a first methods paper, though the abstract's 'faithfully recover' wording should be tempered to reflect these restrictions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper does a real service: it takes the FFT-based binned bispectrum technique from Shaw et al. and works out the visibility-space relation for the angular bispectrum (Eqs. 7 and 8). The derivation in Appendix A is clean, and the validation against a known non-Gaussian input model is thorough—all triangle shapes, good coverage across ℓ, deviations mostly within 1-2σ and 10-15%. For a methods paper, that is solid evidence the estimator works on noiseless gridded visibilities.\n\nBut the written estimator has a normalization error that needs to be fixed before this can be used as a reference. Eq. 8 gives ⟨VVV⟩ = (πθ0²/3) Q³ B. The estimator in Eq. 12 divides by A = πθ0²/3, but not by Q³. Since the visibilities in Eq. 4 are in Jy, the ratio in Eq. 12 has units Jy³, and as written it returns Q³ times the true B in K³. The paper never says Q=1 in the simulations nor that the visibilities are pre-divided by Q. An independent implementation of the printed equations will not reproduce the claimed 10-15% recovery. I suspect the code includes the 1/Q³ factor and the equation is just missing it, but for a methods paper this is load-bearing and must be stated explicitly.\n\nOther soft spots are lighter. The validation is noiseless and foreground-free, which is disclosed, but it means the 10-15% claim is about an ideal case. The Gaussian beam and slow-variation approximation is standard but its range of validity is not quantified; they drop the ℓ≈20 bin but do not say where the bias becomes significant. No code or data are shipped, so reproducibility rests on the printed equations—which makes the Q factor more serious.\n\nIf the normalization is clarified, this is a useful methods paper for 21-cm cosmology and radio interferometry. It deserves peer review; the referee should ask for a corrected Eq. 12 and a demonstration that the code matches the equation (for example, a unit test with Q set to a known value). I would not cite it until that is sorted.\n\nRecommendation: send to review, with the normalization error as the main technical point to resolve.","headline":"Missing Q^3 factor in the printed estimator makes the central equation wrong, but the underlying method and validation are otherwise solid.","tokens_in":16693,"tokens_out":7264,"would_cite":false,"duration_ms":64839,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A fast FFT-based estimator recovers the binned angular bispectrum from gridded radio visibilities, matching simulated MWA sky signals to about 10–15% across multipoles $46 \\le \\ell \\le 1320$.","keywords":["angular bispectrum","radio interferometry","visibility correlation","Murchison Widefield Array","FFT estimator","non-Gaussianity","21-cm cosmology","binned bispectrum"],"falsifier":"Run the estimator on simulated visibilities generated with the true MWA sinc-squared beam instead of the Gaussian approximation, using a sky whose input bispectrum is known; if the recovered values deviate from the prediction by more than the 10–15% claimed in the valid multipole range, the normalization is mis-calibrated. A sharper check is to compute the estimator for baselines $U \\le 1/\\theta_0$, where the paper says its Eq. (8) should not hold, and to compare the recovered $B(\\ell_1,\\mu,t)$ with an independent prediction; the paper currently drops this regime without quantifying where the bias becomes unacceptable.","tokens_in":15631,"feed_emoji":"📡","tokens_out":11494,"duration_ms":90488,"temperature":0.7,"pith_summary":"The paper aims to establish that the angular bispectrum (ABS)—the lowest-order statistic of a non-Gaussian random field—can be measured directly from gridded interferometric visibility data using a fast FFT-based estimator. The authors derive the relation between the three-visibility correlation and the ABS under a Gaussian primary-beam approximation, and they use it to build a binned estimator that covers all possible triangle shapes and sizes. They validate the estimator against simulated Murchison Widefield Array (MWA) observations at 154.25 MHz, using 500 independent realizations of a sky signal whose ABS has a known analytical form, and find recovery within about 10–15% across the multipole range $46 \\le \\ell \\le 1320$. If correct, this makes three-point statistics of the Epoch of Reionization 21-cm signal computationally tractable, going beyond the power spectrum to capture phase correlations that the power spectrum misses.","feed_headline":"FFT bispectrum estimator recovers radio sky to 10–15%","feed_subtitle":"A fast route to the lowest-order statistic sensitive to non-Gaussianity in the cosmic radio signal.","key_machinery":"The central object is the binned angular bispectrum estimator of Eq. (12): $\\hat{B}(\\ell_1,\\ell_2,\\ell_3) = \\frac{1}{A} \\frac{\\sum_{\\theta} D(\\ell_1,\\theta) D(\\ell_2,\\theta) D(\\ell_3,\\theta)}{\\sum_{\\theta} I(\\ell_1,\\theta) I(\\ell_2,\\theta) I(\\ell_3,\\theta)}$, where $D(\\ell_m,\\theta)$ is the inverse Fourier transform of the gridded visibilities restricted to an annular ring $a_m$ in the $(u,v)$ plane, and $I(\\ell_m,\\theta)$ is the corresponding transform of the gridding weights. The ratio of products acts as a weighted average over all closed triangles whose three vertices fall in the chosen rings; the parameters $\\mu$ and $t$ specify the triangle shape, and $\\ell_1$ its size. The normalization $A=\\pi\\theta_0^2/3$ comes from evaluating the three-visibility correlation under the assumptions that the primary beam is Gaussian with width $\\theta_0=0.6\\theta_F$ and that the true bispectrum is nearly constant over the beam's aperture width in the $(u,v)$ plane. The use of FFTs makes the computation scale as $\\sim N_t^2 \\log N_t^2$, which is what makes a full-shape bispectrum measurement practical.","core_discovery":"The central claim is that the binned angular bispectrum estimator of Eq. (12), with normalization $A=\\pi\\theta_0^2/3$ inherited from the three-visibility correlation of Eq. (8), recovers the true angular bispectrum $B(\\ell_1,\\mu,t)$ from gridded visibilities. The paper demonstrates this by constructing a non-Gaussian sky field whose ABS is known analytically to first order in the non-Gaussianity parameter $f_{\\rm NG}$, simulating the corresponding MWA visibilities, and comparing the estimator output with the analytical prediction averaged over the discrete modes in each bin. The estimated values agree with the prediction within statistical fluctuations for the majority of the analyzed bins spanning the full $(\\mu,t)$ shape parameter space, with typical fractional deviations of 10–15%. The estimator's computational cost scales as $\\sim N_t^2 \\log N_t^2$ rather than the direct-correlation cost, which is what makes a survey of all triangle shapes feasible. The paper also states that the estimator remains unbiased when the system noise is included, because the noise contributions to different visibilities are uncorrelated.","pith_inferences":["The ratio form of Eq. (12) may immunize the estimator against gridding artifacts and baseline-density variations, since the denominator tracks the effective weight of each triangle; a clean test would compare its variance with an unnormalized triple-product estimator on identical simulated data.","The paper's own observation of large deviations near the squeezed limit suggests a testable extension: use finer annular rings toward $\\ell_3 \\to 0$ to see whether the deviations persist, or whether they come from the bispectrum changing rapidly within the bin.","Extending the method to instruments with non-Gaussian beams would require recomputing the normalization $A$ numerically from the measured aperture pattern; the current $A=\\pi\\theta_0^2/3$ is tied to the Gaussian assumption.","Real observations will contain foregrounds that are correlated across baselines, so the noise-unbiasedness argument does not cover them; the severity of foreground-induced bias in the bispectrum is a question the current validation does not address."],"forward_implications":["The estimator makes a full-shape survey of the angular bispectrum computationally feasible for arrays like the MWA, covering all triangle shapes rather than restricting to equilateral or isosceles configurations.","It provides a direct probe of three-point correlations of the diffuse radio sky, including foregrounds and Galactic synchrotron emission, extending the established angular power spectrum analyses to higher-order statistics.","Because the estimator is unbiased when system noise is uncorrelated across visibilities, it is ready to be applied to real data once foregrounds and systematics are handled.","The extension to multi-frequency observations would turn the angular bispectrum into a three-dimensional redshifted 21-cm bispectrum, the statistic most directly tied to the topology of reionization."],"supporting_citations":[{"why":"Establishes that the three-visibility correlation directly probes the bispectrum, the theoretical basis for Eq. (8).","marker":"Bharadwaj & Pandey 2005"},{"why":"Introduces the Gaussian beam approximation and the two-visibility correlation, which the three-visibility derivation in Appendix A extends.","marker":"Choudhuri et al. 2014"},{"why":"Supplies the $(\\ell_1,\\mu,t)$ parametrization of triangle shapes used to bin the estimator.","marker":"Bharadwaj et al. 2020"},{"why":"Provides the FFT-based binned bispectrum estimator structure that Eq. (12) adapts to visibility data.","marker":"Shaw et al. 2021"},{"why":"Source of the FFT technique that reduces the computation to $\\sim N_t^2 \\log N_t^2$.","marker":"Sefusatti et al. 2006"},{"why":"Supplies the drift-scan baseline distribution used to define the gridded visibilities in the validation.","marker":"Patwa et al. 2021"},{"why":"Further FFT-based bispectrum formalism that supports the fast computation strategy.","marker":"Scoccimarro 2015"}],"fun_headline_variants":["FFT bispectrum estimator: 10–15% accuracy across all triangle shapes","FFT-speed bispectrum estimator: 10-15% accuracy for all triangles","Fast radio bispectrum estimator validated to 10-15%","Bispectrum from visibilities: FFT estimator validated with MWA","All triangle shapes, FFT speed: new bispectrum estimator"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The estimator's normalization and its relation to the sky bispectrum assume that the primary beam is well described by a Gaussian of width $\\theta_0 = 0.6\\theta_F$ and that the true bispectrum varies slowly over the beam's aperture width in the $(u,v)$ plane; if either fails at the baselines being used, every estimated value carries a systematic bias, and the paper itself excludes the lowest multipole bin ($\\ell \\approx 20$) because the relation is not expected to hold there.","fun_headline_variants_meta":{"raw":{"variants":["FFT bispectrum estimator: 10–15% accuracy across all triangle shapes","FFT-speed bispectrum estimator: 10-15% accuracy for all triangles","Fast radio bispectrum estimator validated to 10-15%","Bispectrum from visibilities: FFT estimator validated with MWA","All triangle shapes, FFT speed: new bispectrum estimator"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001322,"raw_usage":{"total_tokens":5379,"prompt_tokens":937,"completion_tokens":4442,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":4341}},"tokens_in":553,"tokens_out":4442,"duration_ms":32192,"temperature":1.0,"reasoning_tokens":4341,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:41:18.634022+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the estimator on simulated visibilities generated with the true MWA sinc-squared beam instead of the Gaussian approximation, using a sky whose input bispectrum is known; if the recovered values deviate from the prediction by more than the 10–15% claimed in the valid multipole range, the normalization is mis-calibrated. A sharper check is to compute the estimator for baselines $U \\le 1/\\theta_0$, where the paper says its Eq. (8) should not hold, and to compare the recovered $B(\\ell_1,\\mu,t)$ with an independent prediction; the paper currently drops this regime without quantifying where the bias becomes unacceptable.","supporting_citations":[],"review_version":1}