{"id":"0b423fda-5665-481d-ba5b-ad5a3f6ba7fc","arxiv_id":"2504.15149","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A simulation-based inference pipeline recovers unbiased Omega_m and S8 posteriors from noise-free mock void lensing signals, though realistic shape noise currently erases most of the constraining power.","lead":"Astronomers trained a neural network to estimate the universe's matter density and clustering amplitude from the weak lensing signal of cosmic voids, using mock galaxies instead of an analytical model. The method works on noise-free simulations, but adding realistic galaxy shape noise makes its constraining power drop almost to the prior.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline constraints come from a noise-free data vector; the paper's own Appendix A shows that adding realistic shape noise collapses the posteriors toward the priors, so the claim of observational applicability is not supported.","rationale":"The paper is a careful methods demonstration: the SBI training, the cosmology-based train/validation split, the TARP coverage test, and the 200-realization fiducial test are executed properly, and the recovery of input parameters on self-consistent mocks is credible. The weakness is not internal inconsistency; it is the gap between the noise-free, simulation-only data vector and an observable void-lensing measurement. The authors themselves flag this in Sec. 6 ('All of the results in this work are made in a self-consistent way') and in Appendix A, where including realistic shape noise reduces the posteriors to nearly the priors. This is the single most load-bearing assumption because the tight contours in Fig. 8 and the unbiased recovery in Figs. 5-10 depend on a measured quantity that, under the paper's own noise model, carries little information. A full forward-model test with noise would settle whether the Appendix A result is an artifact of the simplified covariance injection or the true signal-to-noise of the statistic. I therefore agree with the reader's CONDITIONAL verdict: the framework is promising and the SBI machinery works, but the claim that it can constrain cosmology with observational void lensing data is not yet supported.","tokens_in":15430,"tokens_out":9163,"duration_ms":92480,"concrete_test":"Run a full forward-model validation with noise: for each of the 100 validation cosmologies, generate a lightcone with the same FastPM settings, ray-trace it as in Appendix A, add shape noise at n_eff = 20 arcmin^-2 and sigma_e = 0.288, measure the stacked void-shear ESD using Eq. A.1, and train a new SBI estimator on these noisy data vectors. Compute the marginalized 68% widths and TARP coverage for Omega_m and S8. If the widths remain comparable to the priors, the Appendix A collapse is confirmed and the main claims must be restricted to noise-free mocks; if the widths tighten substantially, the appendix's simplified noise injection was the cause of the collapse and the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The main results are self-consistent in the SBI sense: an NPE trained on noise-free mocks recovers parameters on held-out noise-free mocks (Figs. 5-10, TARP). This is necessary but not sufficient for the paper's advertised conclusion that void lensing can be applied to observational data. The missing condition is that the 15-bin ESD summary statistic (Sec. 3.3) retains constraining power once realistic measurement noise is included. Appendix A is the only test of this condition, and it finds the opposite: with Stage-III or Stage-IV shape noise, the posteriors for Omega_m, S8, and sigma are 'almost similar to the priors' (Fig. A.1). Taking this result at face value, the tight contours in Sec. 5 are properties of a noiseless simulation quantity, not of an observable. Thus the load-bearing assumption is not the neural estimator itself but the assumption that the noise-free ESD remains informative after measurement noise; the paper's own evidence says it does not. Even if the appendix's covariance-injection procedure is approximate, it is the only noise-aware test offered, and without it no evidence exists that the chosen data vector carries usable information at realistic noise levels.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a simulation-based inference (SBI) framework for cosmological parameter estimation from void lensing. The forward model starts from Omega_m and S8, runs FastPM simulations, identifies halos with rockstar, populates galaxies with Halo Abundance Matching including a scatter parameter sigma, finds voids with DIVE, and measures a 15-bin void lensing excess surface density profile. One thousand cosmologies with ten HAM scatter values each yield 10,000 mock data vectors; the authors train an ensemble of Masked Autoregressive Flows and Mixture Density Network neural posterior estimators using the ltu-ili package. Validation is carried out on held-out cosmologies, on 200 independent realizations of a fiducial cosmology, and with the TARP coverage metric; the noise-free analyses recover unbiased Omega_m and S8 with approximately calibrated uncertainties. Appendix A adds Stage-III and Stage-IV shape noise and finds that the posteriors become almost indistinguishable from the priors.","tokens_in":15670,"tokens_out":7381,"duration_ms":72092,"significance":"If viewed strictly as a mock-data methods paper, the contribution is solid and useful: the internal validation is well designed, with held-out cosmologies, an explicit nuisance parameter for the galaxy-halo connection, 200 fiducial realizations, and TARP coverage checks. The use of public simulation and inference codes aids reproducibility. The critical limitation is that the headline constraints come from a noise-free data vector: the only noise-aware test in the paper, Appendix A, shows that realistic shape noise destroys the constraining power of the chosen 15-bin ESD summary statistic. The paper therefore establishes that SBI can learn a posterior from noise-free mocks of the same forward model, but it does not establish the observational applicability promised in the abstract. The self-consistency of the validation is acknowledged in Sec. 6, and that acknowledgment is itself a strength; however, the abstract and conclusions should be aligned with this limitation.","major_comments":[{"comment":"The abstract and Sec. 5 claim that SBI can effectively learn posterior distributions for Omega_m and S8 from void lensing, and the abstract extends this to \"the potential to apply void lensing analysis to observational data.\" Appendix A (Fig. A.1) directly undercuts this extension: when Stage-III or Stage-IV shape noise is injected through the only noise-aware procedure in the paper, the posteriors for Omega_m, S8, and sigma are \"almost similar to the priors.\" Taking this test at face value, the tight contours in Figs. 8-10 are properties of a noiseless simulation summary, not of an observable, and no evidence is provided that the chosen 15-bin ESD data vector remains informative at realistic noise levels. Please either revise the abstract and conclusions to scope the claims explicitly to noise-free mock data, or add a demonstration that a modified data vector (for example, with tomographic redshift bins, optimized void-radius selection, or learned compression) retains constraining power under realistic shape noise.","section":"Abstract, Sec. 5, Appendix A"},{"comment":"The ESD summary statistic is computed from projected matter profiles around voids, with no source redshift distribution or shape noise in the main pipeline, and with the same forward model used for both training and validation. Consequently, any error in the modeling steps (FastPM with B=2 and 40 steps, rockstar halos, HAM galaxies, DIVE voids, and the projection approximation to lensing) is invisible to the validation. The authors acknowledge this in Sec. 6, which is commendable, but the Appendix A statement that \"our noise-free void lensing model is accurate enough\" is not justified by self-consistency: a posterior width much smaller than the shape-noise level says nothing about the bias introduced by the approximate forward model. A calibration test against a higher-fidelity N-body or ray-tracing pipeline, or at minimum a quantitative estimate of the bias from each approximation, is needed before the framework can be presented as observationally ready.","section":"Sec. 3.3 and Sec. 6"},{"comment":"The validation statistics in Figs. 5 and 6 are described as covering 1,000 validation sets, but these contain only 100 independent cosmologies; each cosmology contributes ten HAM realizations that share the same FastPM density field and halo catalog. The effective number of independent samples for the cosmological-parameter unbiasedness test is therefore much smaller than 1,000, and the comparison with a standard normal distribution in Fig. 6 should account for this correlation. Please state the effective sample size, or compute a block-bootstrap estimate that resamples whole cosmologies, before claiming that the 1D distributions are consistent with a standard normal.","section":"Sec. 4 and Sec. 5.2"}],"minor_comments":[{"comment":"The Gaussian likelihood expression is missing the minus sign and the factor of 1/2; it should read p(d|theta) proportional to exp[-(1/2)(d - f(theta))^T C^{-1}(d - f(theta))].","section":"Eq. (2)"},{"comment":"The first sentence of the caption is duplicated verbatim.","section":"Fig. 7 caption"},{"comment":"The stated units of Delta Sigma, \"h^2 M_sun/pc^2/Mpc\", look like a typo; excess surface density is more conventionally quoted in h M_sun/pc^2 or M_sun/pc^2, so please check the units.","section":"Sec. 5.1 and Fig. 4"},{"comment":"Thiele et al. 2024a and 2024b are listed with identical journal, volume, and page (ApJ 969, 89); if these are intended to be distinct papers, the second entry needs a different bibliographic record.","section":"References"},{"comment":"The description of the blue line as \"a normal distribution whose mean equals the average of predictions\" is incomplete; please specify whether the width is the mean predicted standard deviation or the standard deviation of the predictions themselves.","section":"Sec. 5.3 and Fig. 10"},{"comment":"The description of the 9:1 split is clear, but it would help to state explicitly that 900 cosmologies (with all ten HAM realizations each) are used for training and 100 cosmologies for validation.","section":"Sec. 4"}],"recommendation":"major_revision","confidential_remarks":"For the editor: this is a technically competent mock-data methods paper, and the internal validation is genuinely careful. The main problem is that the title and abstract promise more than the results deliver, because Appendix A shows that the chosen summary statistic loses its constraining power under realistic shape noise. I recommend requiring the authors to either add a noise-robust data vector or substantially qualify the observational claims. The duplicate Thiele references and the typo in Eq. (2) should also be corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this is the first SBI application to void lensing, and the forward-modeling assembly is genuinely new. The authors connect FastPM, rockstar halo finding, HAM galaxy assignment, DIVE voids, and pyfcfc correlation functions into an end-to-end simulator that outputs the excess surface density around voids, then train a neural posterior estimator on 10,000 mock catalogs. That is real work, and the internal validation is careful: held-out cosmologies recover input parameters, TARP coverage sits on the diagonal, and 200 fiducial realizations show no systematic bias. If you need a template for SBI-plus-void-lensing validation, this is a useful reference.\n\nThe soft spot is not the SBI machinery—it is the observability of the summary statistic. All headline constraints come from a noise-free data vector. The paper's own Appendix A injects shape noise at Stage-III and Stage-IV levels and finds the posteriors for Omega_m, S8, and the HAM scatter sigma become essentially the priors (Fig. A.1). The authors frame this as a data-processing issue to be optimized later, but taken at face value it says the 15-bin ESD profile, as constructed, carries little usable information at realistic noise. That matters because the abstract and Section 5 advertise the method as a route to observational void lensing constraints. The paper also lacks a baseline comparison: no standard likelihood or alternative compression is tested, so we do not know whether SBI is buying anything beyond what a Gaussian likelihood on the same noise-free vector would give. Forward-model fidelity is explicitly deferred to future work; that is fine for a methods paper, but it means the validation demonstrates self-consistency, not accuracy.\n\nThe internal validation is self-consistent in the SBI sense, not circular—no parameters are fit to validation data. But the shared-generator limitation is real, and the authors acknowledge it in Section 6. I would not call this a fatal flaw; it is a standard limitation of simulation-based inference. The missing piece is any evidence that the chosen data vector survives realistic noise. Appendix A currently argues the opposite.\n\nWho is this for? Someone building an SBI pipeline for void statistics or testing summary statistics for void lensing will get value from the pipeline design and the validation protocol. The cosmological constraints themselves should not be cited as evidence about the universe. I would give it a serious referee: the methods contribution is worth publishing, but the authors should be pushed to either improve the data vector/compression so that shape noise does not erase the signal, or reframe the paper as a noise-free proof-of-concept with the observational claim sharply downgraded. My recommendation: send to peer review, ask for major revision, and treat the appendix as a central limitation rather than a side note.","headline":"First SBI void-lensing pipeline with careful noise-free validation, but its own appendix shows shape noise collapses the signal to priors, so the observational claim is not yet backed.","tokens_in":16231,"tokens_out":3681,"would_cite":true,"duration_ms":29831,"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":"Simulation-based inference recovers unbiased cosmological parameters from void lensing, even though no analytical model of void lensing exists.","keywords":["cosmic voids","void lensing","simulation-based inference","weak gravitational lensing","neural posterior estimation","galaxy-halo connection","cosmological parameters","FastPM simulations"],"falsifier":"Train the same estimator on the paper's mocks, then run the pipeline on high-fidelity N-body mocks with known input cosmology; if the posterior means are offset by more than the reported $1\\sigma$ uncertainties, or if TARP coverage fails on the new data, the claim of unbiased parameter recovery from void lensing is refuted.","tokens_in":15231,"feed_emoji":"🕳️","tokens_out":9630,"duration_ms":78651,"temperature":0.7,"pith_summary":"This paper establishes that void lensing, the gravitational shear signal around cosmic voids, can be used for cosmological inference without any analytical model of the effect. The authors replace the missing likelihood with a neural posterior estimator trained on forward-modeled mock observations that include cosmology and a galaxy-halo connection. On held-out mocks and on multiple realizations of a fiducial cosmology, the estimator recovers $\\Omega_m$ and $S_8$ with mean values consistent with the input truth and error bars that pass a coverage test. The payoff is that void lensing, which has resisted analytic modeling, can still be exploited as a cosmological probe. The paper's own results are noise-free, and its appendix shows that realistic shape noise weakens the constraints substantially.","feed_headline":"Void lensing constrains cosmology with no analytical model","feed_subtitle":"A neural estimator trained on mock void signals recovers Omega_m and S8 with calibrated uncertainties.","key_machinery":"The load-bearing object is the end-to-end forward model feeding a neural posterior estimator. FastPM particle-mesh simulations with $1024^3$ particles in a $1\\,h^{-1}\\mathrm{Gpc}$ box, force resolution $B=2$, and 40 time steps generate dark matter fields; rockstar identifies halos; halo abundance matching with a scatter parameter $\\sigma$ populates galaxies; DIVE constructs void catalogs; and the void-matter cross-correlation is converted to the excess surface density $\\Delta\\Sigma$ using the integral relation between tangential shear and convergence. The resulting 15-dimensional $\\Delta\\Sigma(R_p/R_v)$ data vectors are paired with their input parameters ($\\Omega_m$, $S_8$, $\\sigma$) to train an ensemble of three masked autoregressive flow and three mixture density network estimators, and the posteriors are validated with the TARP coverage metric.","core_discovery":"The paper claims that a simulation-based inference pipeline can recover unbiased posterior distributions for $\\Omega_m$ and $S_8$ from void lensing excess surface density data vectors, despite the absence of an analytical void lensing model. It trains an ensemble of masked autoregressive flows and mixture density networks as a neural posterior estimator on 10,000 mocks built from 1,000 cosmologies simulated with FastPM, with ten halo abundance matching scatter values per cosmology. Validations against held-out mocks show predicted means scattering around the truth consistently with the estimated $1\\sigma$ uncertainties, TARP coverage follows the diagonal, and the average of 200 fiducial-cosmology realizations centers on the truth. The paper also finds a strong $\\Omega_m$-$S_8$ anti-correlation ($\\rho=-0.85$) while the galaxy-halo connection parameter is only weakly correlated with either cosmological parameter ($|\\rho|<0.25$), and shows in an appendix that including Stage-III or Stage-IV shape noise degrades the constraints substantially.","pith_inferences":["Because the validation is self-consistent, the next test is to run the trained estimator on mocks from a higher-fidelity N-body code; if the posterior means shift, the fast gravity approximation is the first component to blame.","The strong $\\Omega_m$-$S_8$ anti-correlation suggests that combining this void-lensing probe with a clustering-based simulation-based inference pipeline in the same forward model could break the degeneracy without any new analytic theory.","The shape-noise result implies a concrete extension: adding tomographic source bins or a compression network to the forward model should restore constraining power, and this is directly testable in the same self-consistent framework.","If real-data application is the goal, the forward model must add photo-z scatter, shear calibration, and source redshift distributions; the current validation cannot detect errors from their omission."],"forward_implications":["Void lensing can yield cosmological parameter constraints without an analytic likelihood, so the analysis is no longer blocked by the absence of a void lensing model.","The galaxy-halo connection can be absorbed into the forward model as a nuisance parameter; because $\\sigma$ is only weakly degenerate with $\\Omega_m$ and $S_8$, marginalizing over it does not strongly degrade the cosmological constraints.","The same simulation-based inference recipe transfers to other summary statistics whose likelihoods are intractable but whose forward models are runnable.","Posterior evaluation is fast after training, since neural posterior estimation avoids MCMC sampling.","The noise-free constraints are optimistic: Appendix A shows that Stage-III and Stage-IV shape noise broadens the posteriors substantially, so extracting useful information from real data will require optimized void size bins, tomography, or data compression."],"supporting_citations":[{"why":"Supplies the FastPM particle-mesh simulation code that generates all training and validation dark matter fields.","marker":"Feng et al. 2016"},{"why":"Provides the rockstar algorithm used to identify halos in the FastPM density fields.","marker":"Behroozi et al. 2012"},{"why":"Supplies the halo abundance matching approach used to populate galaxies in halos.","marker":"Rodríguez-Torres et al. 2016"},{"why":"Provides the specific halo abundance matching implementations and scatter model used in the forward modeling pipeline.","marker":"Yu et al. 2022, 2023"},{"why":"Provides the DIVE void finder that constructs void catalogs from the galaxy catalogs.","marker":"Zhao et al. 2016"},{"why":"Supplies the FCFC/pyfcfc correlation-function code used to measure void-matter cross-correlations, from which the void lensing signal is derived.","marker":"Zhao 2023"},{"why":"Motivates sampling halo abundance matching parameters jointly with cosmology in the simulation-based inference training set, shaping the 10,000-mock design.","marker":"Hahn et al. 2023b"},{"why":"Defines the TARP coverage test used to validate that the posterior uncertainties are neither over- nor under-confident.","marker":"Lemos et al. 2023a"},{"why":"Provides the training code used to build the neural density estimators (masked autoregressive flows and mixture density networks).","marker":"Ho et al. 2024"}],"fun_headline_variants":["SBI learns void lensing cosmology with no analytic model","No analytical model? SBI extracts void lensing cosmology","Void lensing without an analytic model? SBI constrains cosmology","Simulation-based inference recovers void lensing cosmology, no analytic model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"All validation uses mocks produced by the same forward model (FastPM with force resolution $B=2$, 40 steps, rockstar halos, HAM galaxies, DIVE voids, and a projected-profile lensing signal that ignores shape noise), so if that model is a poor description of the real Universe, the reported unbiased posteriors will not transfer to observations.","fun_headline_variants_meta":{"raw":{"variants":["SBI learns void lensing cosmology with no analytic model","No analytical model? SBI extracts void lensing cosmology","Void lensing without an analytic model? SBI constrains cosmology","Simulation-based inference recovers void lensing cosmology, no analytic model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001745,"raw_usage":{"total_tokens":6859,"prompt_tokens":876,"completion_tokens":5983,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":5910}},"tokens_in":492,"tokens_out":5983,"duration_ms":36684,"temperature":1.0,"reasoning_tokens":5910,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:32:05.393917+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same estimator on the paper's mocks, then run the pipeline on high-fidelity N-body mocks with known input cosmology; if the posterior means are offset by more than the reported $1\\sigma$ uncertainties, or if TARP coverage fails on the new data, the claim of unbiased parameter recovery from void lensing is refuted.","supporting_citations":[{"cited_title":"2016, MNRAS, 463, 2273 Gil-Marín, H., Verde, L., Noreña, J., et al","cited_arxiv_id":null,"evidence_quote":"Supplies the FastPM particle-mesh simulation code that generates all training and validation dark matter fields."},{"cited_title":"S., Wechsler, R","cited_arxiv_id":null,"evidence_quote":"Provides the rockstar algorithm used to identify halos in the FastPM density fields."},{"cited_title":"2016, MNRAS, 459, 2670","cited_arxiv_id":null,"evidence_quote":"Provides the DIVE void finder that constructs void catalogs from the galaxy catalogs."}],"review_version":1}