{"id":"ecbb283c-3b72-4a6e-83c7-cb4b5ae807b3","arxiv_id":"2412.00968","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Applying a hybrid reconstruction, then cross-correlating the squared potential with the density, yields a simulated f_NL forecast up to three times tighter than the unreconstructed bispectrum.","lead":"This paper proposes reconstructing the initial density field with a perturbation theory plus neural network hybrid, then cross-correlating the squared potential with density to constrain primordial non-Gaussianity. On simulations, the forecast for f_NL improves by up to a factor of three, before galaxy bias and other survey complications are included.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fNL=0-trained CNN does not provably preserve the PNG signal: Section 6.3.2 and Table 2 show a residual ~9% multiplicative attenuation in reconstructed fNL, so the factor-of-three claim rests on network systematics that are not yet calibrated.","rationale":"I read the paper as a proof of concept: the cross-power estimator is known to be near-optimal for linear fields, and the reconstruction improves the signal-to-noise in simulations. The load-bearing assumption for the central claim is that the CNN trained on fNL=0 simulations preserves the fNL signal in reconstructed fields. The paper itself demonstrates that this is only approximately true: Table 2 shows a residual multiplicative attenuation of roughly 9% after removing the additive fNL=0 shift, with the reconstructed-field bias larger than the pre-reconstruction bias at z=0. This is exactly the reader's weakest assumption, and it is the right place to focus because every quantitative headline (factor 1.5-3 improvement, sigma(fNL)=17.4 at kmax=0.2) is computed with this network. The proposed test with intermediate fNL values would settle whether the distortion is a simple calibratable transfer function. If it is, the method stands as advertised modulo the disclosed idealizations; if not, the central claim is overstated. I therefore agree with the reader's conditional verdict and see no reason to change it. The paper's honesty about the bias and its plan for calibration are strengths, but they do not remove the need for the calibration to be demonstrated before the factor-of-three claim is used for survey projections.","tokens_in":30348,"tokens_out":11667,"duration_ms":117502,"concrete_test":"Run the fNL=0-trained CNN+HE18 pipeline on phase-matched Quijote-PNG (or equivalent) simulations with fNL = -100, -50, -10, +10, +50, +100, fit each with the Section 6.2 template procedure at z=1 and kmax=0.2, and regress recovered fNL (after subtracting the fNL=0 offset) against injected fNL. If the best-fit slope differs from unity by more than the ~1% statistical error, or if the relation shows curvature, the residual multiplicative shift in Table 2 is not a fixed calibration factor and the factor-of-three forecast must be revised. As a secondary cross-check, recompute the Section 5.4 Fisher error with a two-parameter (fNL, b2) model to see whether the single-parameter improvement survives marginalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central forecast in Section 5.4 assumes the hybrid reconstruction trained only on fNL=0 simulations (Section 5.1) transfers to fNL != 0 fields without destroying the primordial signal. This is not a small technicality: the CNN is a nonlinear map optimized to reproduce fNL=0 initial conditions, and there is no guarantee it commutes with the fNL perturbation in Eq. 2.4. The paper's own template fits in Section 6.3.2 provide the strongest evidence against the assumption. After subtracting the fNL=0 offset, Table 2 shows that CNN+HE18 at z=1, kmax=0.2 recovers fNL=90.9 and -90.8 for injected +/-100, i.e. a ~9% multiplicative attenuation; the corresponding pre-reconstruction nonlinear field recovers +/-98.4. At z=0 the fNL=0 offset is -11.7 for CNN+HE18 versus +4.9 for the nonlinear field, so the reconstruction injects a larger additive bias than the field it was meant to clean. Because the Section 5.4 Fisher forecast measures its derivative from fNL=+/-100 simulations passed through the same network, the quoted sigma already includes part of this attenuation; the unresolved question is whether the distortion is a simple, calibratable transfer function or depends on fNL amplitude, cosmology, and survey selection. If it is not calibratable, the method cannot deliver unbiased fNL constraints, and the advertised factor-of-three improvement applies only to a biased summary statistic. The optimality proof in Section 3.3 is also for linear Gaussian fields, so it does not by itself certify the reconstructed pipeline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes estimating local primordial non-Gaussianity by cross-correlating the squared reconstructed primordial potential with the reconstructed density field, ⟨Φ²δ⟩, using a hybrid reconstruction that combines Hada–Eisenstein perturbation-theory reconstruction with a convolutional neural network trained on f_NL = 0 simulations. Section 3 presents a Fisher calculation on linear Gaussian fields aiming to show that this compressed estimator carries the same f_NL information as the full matter bispectrum. Sections 5–6 apply the estimator to Quijote-PNG simulations, reporting that reconstruction reduces the gravity-induced bispectrum and yields a factor of 1.5 (k_max = 0.1 h/Mpc) to 3 (k_max = 0.2 h/Mpc) improvement in a single-parameter, unmarginalized f_NL forecast at z = 1. The paper also develops a perturbation-theory template fit for the reconstructed fields and documents additive and multiplicative biases in the recovered f_NL, which are left for future calibration.","tokens_in":30671,"tokens_out":8138,"duration_ms":77505,"significance":"If the method can be shown to deliver unbiased f_NL constraints, it would provide a computationally inexpensive bispectrum-like statistic and would demonstrate that reconstruction extends the usable k-range for PNG searches. The paper has clear strengths: the estimator derivation and covariance are given in detail, the analysis is carried out on the public Quijote-PNG simulations, the perturbation-theory template fits are carefully documented, and the residual biases are discussed explicitly rather than hidden. However, the central transfer assumption — that a CNN trained on f_NL = 0 simulations preserves the f_NL signal when applied to f_NL ≠ 0 fields — is only partially supported by the paper's own Table 2, which shows roughly 9–10% multiplicative attenuation and a z = 0 additive offset of −11.7. The significance of the factor-of-three forecast is therefore conditional on successful calibration of these distortions.","major_comments":[{"comment":"The central transfer assumption of the method is not yet established. The CNN reconstruction is trained exclusively on f_NL = 0 simulations (Section 5.1) but is then applied to f_NL = ±100 fields. Table 2 shows that after subtracting the f_NL = 0 offset, CNN+HE18 at z = 1 with k_max = 0.2 h/Mpc recovers f_NL = 90.9 and −90.8 for injected +100 and −100, a ~9% multiplicative attenuation; at z = 0 the f_NL = 0 offset is −11.7 and the recovered values are 92.4/−92.0 (k_max = 0.1) and 89.4/−88.9 (k_max = 0.15), while the nonlinear field recovers ±97.7/±98.4. Because the Fisher forecast in Section 5.4 defines its derivative from the f_NL = ±100 reconstructed cross-power, the quoted sigma already includes this attenuation; however, the additive bias at f_NL = 0 and the f_NL dependence of the distortion are not calibrated. As the paper states, calibration with a range of f_NL values is left to future work (Section 6.3.2). Until such a test is performed, the factor-of-three improvement is demonstrated only for a biased summary statistic. I request either a calibration test with intermediate f_NL values (e.g., f_NL = ±10, ±30) or a re-framing of the headline claim as an uncalibrated proof of concept.","section":"Section 5.1 / Section 6.3.2, Table 2"},{"comment":"The abstract states that the cross-power estimator 'has the same information content as the full matter bispectrum,' but the optimality demonstration in Section 3.3 is restricted to a linear Gaussian field with a known potential and a single-parameter Fisher calculation. It does not establish that the estimator applied to the reconstructed (or nonlinear) fields contains the same information as the bispectrum of those fields; nor does it account for the parameter degeneracies discussed in Section 6.3. The comparison with Coulton et al. in Section 5.4 is for the pre-reconstruction nonlinear field and is not a direct equivalence test. I recommend rewording the abstract and conclusion to say that the estimator is near-optimal for the ideal linear field and is competitive with, but not proven equivalent to, bispectrum analyses of the reconstructed fields.","section":"Section 3.3 and Abstract"},{"comment":"The headline improvement factors of 1.5 and 3 are based on a single-parameter, unmarginalized Fisher forecast. The paper acknowledges this, and Appendix B shows that including b_2 inflates σ(f_NL) by ~30% at k_max = 0.1 h/Mpc; Figure 8 shows a strong f_NL–b_2 degeneracy after reconstruction. The abstract's 'up to a factor of three improvement' is thus not a forecast for a marginalized f_NL constraint. Since the factor of three is the paper's central result, this caveat should appear in the abstract, and ideally the single-parameter numbers should be accompanied by an estimate of the marginalized degradation for the same k_max and smoothing choices.","section":"Section 5.4, Table 1 and Appendix B"}],"minor_comments":[{"comment":"Several entries in Table 1 are left blank (e.g., IC z = 0 with cosine filter at k_max = 0.2 h/Mpc); please add a note explaining why these forecasts are omitted.","section":"Section 5.4, Table 1"},{"comment":"The low-k divergence is regulated by replacing P_Φ with a form proportional to (k + 10⁻⁴)^{n_s−4}; please state explicitly that this is a numerical regulator and show the sensitivity of Figure 1 to the chosen value of 10⁻⁴.","section":"Section 3.3, Eq. (3.14)"},{"comment":"The phrase 'at low k, wearers these differences amplify' contains a typo; it should read 'at low k, whereas these differences amplify'.","section":"Section 5.3"},{"comment":"The parenthetical adjusted f_NL means are quoted without uncertainties; please report the scatter of the adjusted values or explicitly state that they are mean-only values.","section":"Section 6.3.2, Table 2"},{"comment":"The pairs of vertical lines marking the cosine filter boundaries are not described; please explain in the caption which k_max values correspond to the filter edges.","section":"Figure 1 caption"},{"comment":"The sentence 'our measurement error σ(f_NL) ∼ 50' should read 'our forecast error σ(f_NL) ∼ 50', since the paper does not perform a measurement from data.","section":"Section 7.3"}],"recommendation":"major_revision","confidential_remarks":"This is a solid proof-of-concept with an acknowledged but unresolved systematics issue. The main risk is over-interpretation of the factor-of-three claim; a calibration test with intermediate f_NL values and a marginalized forecast would make the paper substantially stronger. The scope is appropriate for JCAP, and I would not reject on the current evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this one. It's a clean proof-of-concept for using the <Phi^2 delta> cross-power on reconstructed fields to get at fNL. The combination of the Schmittfull estimator with hybrid PT+CNN reconstruction is new, and the paper does a solid job showing the cross-power carries the same information as the full bispectrum for a linear field. The field-level perturbative model for the reconstructed field is a nice touch and fits the simulations well, which makes the method more than a black-box CNN exercise.\n\nThe headline factor-of-three improvement at z=1 with kmax=0.2 is real within the stated setup: unmarginalized, real-space matter field, with the derivative measured from fNL=+/-100 simulations run through the same reconstruction. So the forecast already includes some of the network's response. The soft spot is the one the paper itself flags in Section 6.3.2: the CNN is trained only on fNL=0 simulations, and Table 2 shows it attenuates the injected signal by about 9% (adjusted means of ~91 instead of 100). That is a systematics issue, not a fatal one, but it needs a calibration strategy. The residual multiplicative shift could depend on fNL amplitude, cosmology, and survey selection, and the paper doesn't yet show that it is a simple transfer function. The additive bias at z=0 (-11.7) is also larger than the nonlinear-field bias, so the reconstruction is not clearly cleaning that channel.\n\nThe other caveats are known: forecasts are unmarginalized, ignore galaxy bias, shot noise, RSD, and survey geometry. The paper says so plainly, and the comparison with Shirasaki et al. and Flöss & Meerburg is fair. The optimality proof in Section 3.3 is for linear Gaussian fields, so it motivates but doesn't certify the reconstructed pipeline.\n\nWorth sending to peer review. A serious referee will ask for a multi-fNL calibration test, a marginalized forecast including b2, and maybe a first pass at biased tracers. But the idea is promising and the analysis is honest. I'd bring it to our reading group and cite it when discussing reconstruction-based PNG approaches.","headline":"Solid proof-of-concept for using the <Phi^2 delta> cross-power on reconstructed fields to improve fNL constraints; the advertised factor-of-three holds in the idealized setup, but the CNN's distortion of the PNG signal remains an uncalibrated systematic.","tokens_in":31268,"tokens_out":3126,"would_cite":true,"duration_ms":30074,"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":"The paper claims that the cross-power between the squared primordial potential and the reconstructed density field carries the same $f_{\\rm NL}$ information as the full matter bispectrum and, after hybrid reconstruction, improves…","keywords":["primordial non-Gaussianity","local-type PNG","fNL","density field reconstruction","convolutional neural network","cross-power spectrum","bispectrum","large-scale structure"],"falsifier":"Run the same pipeline on simulations with $f_{\\rm NL}=\\pm 50$ and $\\pm 200$ (not used in training or calibration) and check whether, after subtracting the $f_{\\rm NL}=0$ additive shift, the recovered $f_{\\rm NL}$ is linear in the input with unit slope within the quoted errors; a residual multiplicative offset or a seed-dependent shift at these intermediate amplitudes would falsify the claim that the factor-of-three forecast translates into an unbiased $f_{\\rm NL}$ measurement.","tokens_in":30059,"feed_emoji":"🌌","tokens_out":7827,"duration_ms":66224,"temperature":0.7,"pith_summary":"The paper proposes constraining local-type primordial non-Gaussianity, parametrized by $f_{\\rm NL}$, by first removing late-time gravitational evolution from the observed density field and then cross-correlating the squared primordial potential, $\\Phi^2$, with the reconstructed density, $\\delta$. It argues that this cross-power estimator, $\\langle\\Phi^2\\delta\\rangle$, contains the same $f_{\\rm NL}$ information as the full matter bispectrum while needing a data vector that scales only with the number of $k$-bins. Applied to hybrid reconstruction on matched simulations, the estimator improves the single-parameter $f_{\\rm NL}$ forecast by a factor of 1.5 at $k_{\\max}=0.1\\ h/{\\rm Mpc}$ and a factor of 3 at $k_{\\max}=0.2\\ h/{\\rm Mpc}$ at $z=1$ over the unreconstructed field. The paper presents this as a proof of concept that does not yet include galaxy bias, redshift-space distortions, or survey realism, and it openly leaves calibration of residual biases to future work.","feed_headline":"Reconstruction triples probe power for primordial non-Gaussianity","feed_subtitle":"A cheap cross-power estimator on reconstructed fields matches the full bispectrum's f_NL sensitivity.","key_machinery":"The central object is the cross-power estimator $P_{\\Phi^2\\delta}(k)$, defined by $\\langle\\Phi^2(\\mathbf{k})\\delta(\\mathbf{k}')\\rangle=(2\\pi)^3\\delta_D(\\mathbf{k}+\\mathbf{k}')P_{\\Phi^2\\delta}(k)$. Its power comes from the identity that squaring the potential in configuration space is a convolution in Fourier space, which collapses the $f_{\\rm NL}$ bispectrum template into a three-point cross-correlation that is near-optimal for the squeezed limit. The second component is hybrid reconstruction: an iterative 2LPT-based algorithm preprocessing the density field, followed by a convolutional neural network trained on $f_{\\rm NL}=0$ simulations, which removes the growth, shift, and tidal quadratic terms. A cosine smoothing filter band-limits the $\\Phi^2$ field so that the estimator is not dominated by small-scale modes where reconstruction degrades.","core_discovery":"On the paper's own terms, the central discovery is that the near-optimal bispectrum estimator for local $f_{\\rm NL}$, derived from the product-separable template of the primordial bispectrum, reduces to a cross-power spectrum $P_{\\Phi^2\\delta}(k)$ between the squared primordial potential and the linear density field. When the same $k$-modes are available, this compressed statistic returns the same Fisher error on $f_{\\rm NL}$ as the full matter bispectrum. The new step is to feed this estimator with density fields that have been reconstructed to remove gravitational nonlinearities: the hybrid reconstruction reduces the second-order gravitational terms nearly to zero and extends the usable $k$-range, dropping the single-parameter forecast $\\sigma(f_{\\rm NL})$ from about 54.5 (pre-reconstruction) to 17.4 at $z=1$ with $k_{\\max}=0.2\\ h/{\\rm Mpc}$, a factor of three. The paper also documents an additive bias at $f_{\\rm NL}=0$ and a residual multiplicative shift at $f_{\\rm NL}=\\pm 100$, and it explicitly identifies calibration of these biases as necessary follow-up work.","pith_inferences":["Beyond the paper's explicit claims, if the residual multiplicative bias proves calibratable with a small set of $f_{\\rm NL}$-scaling simulations, the method could be adapted to galaxy samples with lower number densities, where reconstruction fidelity is lower but the compressed estimator's low dimensionality makes the covariance tractable.","The factor-of-three gain assumes matter fields at number density $\\sim 10^{-1}\\,(h/{\\rm Mpc})^3$; at DESI-like galaxy densities the paper itself notes the recoverable scale drops to roughly $k\\approx 0.13\\ h/{\\rm Mpc}$, so an honest end-to-end galaxy forecast would likely show a smaller gain.","A natural testable extension is to apply the same reconstruction-plus-cross-power pipeline to the equilateral and orthogonal templates on the corresponding Quijote-PNG simulations, checking whether the factor-of-three gain is preserved or exceeded.","The additive $f_{\\rm NL}=0$ bias, which changes with redshift and $k_{\\max}$, suggests the reconstruction network may be over-Gaussianizing the field; comparing reconstruction residuals between $f_{\\rm NL}=0$ and $f_{\\rm NL}\\neq 0$ simulations could reveal a simple correction."],"forward_implications":["If the estimator is as informative as the full bispectrum, $f_{\\rm NL}$ constraints can be obtained with a data vector of length $O(k_{\\max})$ instead of $O(k_{\\max}^3)$, making PNG analyses far cheaper computationally.","At $z=1$ with $k_{\\max}=0.2\\ h/{\\rm Mpc}$, reconstruction yields a factor-of-three improvement in unmarginalized single-parameter $\\sigma(f_{\\rm NL})$ over the unreconstructed matter field, and a factor of 1.5 at $k_{\\max}=0.1\\ h/{\\rm Mpc}$.","Because hybrid reconstruction removes the second-order gravitational terms, a simple perturbative model of the reconstructed field describes the measured cross-power, so the method does not depend on a simulation-based likelihood.","Reconstruction extends the range of scales usable for $f_{\\rm NL}$, shifting strategy away from relying only on the lowest $k$ modes that dominate the scale-dependent bias.","The same product-separable cross-power construction applies to equilateral and orthogonal PNG shapes, and reconstruction may help even more for the equilateral shape, whose signal is most contaminated by gravitational nonlinearities."],"supporting_citations":[{"why":"Supplies the near-optimal product-separable bispectrum estimator from which the $\\langle\\Phi^2\\delta\\rangle$ cross-power is derived.","marker":"[47]"},{"why":"Develops the hybrid CNN+HE18 reconstruction method whose performance enables the factor-of-three gain.","marker":"[59]"},{"why":"Provides the iterative 2LPT reconstruction algorithm used as the perturbation-theory preprocessing step.","marker":"[65]"},{"why":"Provides the Quijote and Quijote-PNG N-body simulations with matched Gaussian seeds used for training and $f_{\\rm NL}$ forecasts.","marker":"[54–56]"},{"why":"Demonstrates for weighted skew-spectra that a cross-power-type statistic carries the same $f_{\\rm NL}$ information as the bispectrum, which this paper verifies for $\\langle\\Phi^2\\delta\\rangle$.","marker":"[50]"},{"why":"Provides the Quijote-PNG matter-field bispectrum forecast used as the baseline comparison for the cross-power error.","marker":"[41]"}],"fun_headline_variants":["Initial field reconstruction triples f_NL sensitivity","Reconstructed density fields tighten non-Gaussianity bounds","Cross-power estimator on reconstructed fields matches bispectrum","Factor-of-three gain in f_NL constraints via reconstruction","Reconstructing initial conditions sharpens non-Gaussianity probes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the neural-network reconstruction, trained only on $f_{\\rm NL}=0$ simulations, preserves the primordial non-Gaussian signal when applied to $f_{\\rm NL}\\neq 0$ fields, which the paper's own template fits show is only approximately true (a residual multiplicative shift of roughly $\\pm 10$ remains after subtracting the additive bias).","fun_headline_variants_meta":{"raw":{"variants":["Initial field reconstruction triples f_NL sensitivity","Reconstructed density fields tighten non-Gaussianity bounds","Cross-power estimator on reconstructed fields matches bispectrum","Factor-of-three gain in f_NL constraints via reconstruction","Reconstructing initial conditions sharpens non-Gaussianity probes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000475,"raw_usage":{"total_tokens":2359,"prompt_tokens":946,"completion_tokens":1413,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":1335}},"tokens_in":562,"tokens_out":1413,"duration_ms":10662,"temperature":1.0,"reasoning_tokens":1335,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:48:18.197067+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same pipeline on simulations with $f_{\\rm NL}=\\pm 50$ and $\\pm 200$ (not used in training or calibration) and check whether, after subtracting the $f_{\\rm NL}=0$ additive shift, the recovered $f_{\\rm NL}$ is linear in the input with unit slope within the quoted errors; a residual multiplicative offset or a seed-dependent shift at these intermediate amplitudes would falsify the claim that the factor-of-three forecast translates into an unbiased $f_{\\rm NL}$ measurement.","supporting_citations":[],"review_version":1}