{"id":"954ea22d-8370-42c5-8cda-cc3ca4ff93e6","arxiv_id":"2501.03921","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"LSBI approximates simulated likelihoods as linear Gaussians and iteratively refines posteriors, reaching accuracy comparable to neural SBI on CMB power spectrum emulations.","lead":"This paper introduces Linear Simulation-based Inference (LSBI), a simple, equation-based alternative to neural-network methods for estimating cosmological parameters from forward simulations. It shows that, on toy models and emulated CMB power spectrum data, LSBI matches standard nested sampling after about four or five rounds of ten thousand simulations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The CMB test's own likelihood (Eq. 23) violates the homoscedastic Gaussian assumption of Eq. (1), and the paper provides no coverage check; if the resulting LSBI posteriors are overconfident, the claim that they agree with the underlying parameters is unsupported.","rationale":"The reader's weakest_assumption identified the linear-Gaussian surrogate as the fragile part of the argument. My concern sharpens that: the paper's own CMB test uses a likelihood whose noise is heteroscedastic and parameter-dependent (Eq. 23), which violates Eq. (1) even after sequential narrowing. This is more than a generic 'linearity' worry; it is a concrete misspecification in the flagship application. The paper's Figure 5 gives a hint of overconfidence (LSBI DKL above the nested-sampling band), but no calibration statistics are provided. I considered whether the more glaring issue is the lack of any quantitative comparison to neural SBI, which the reader also flagged. However, that is an overclaim about competitiveness rather than a threat to the internal correctness of the method; the load-bearing condition for 'posteriors that agree' is calibration, and that is untested. The proposed coverage test would settle whether the concern lands: if LSBI credible intervals have near-nominal coverage on repeated simulated datasets, the central claim is credible; if not, the method is overconfident and the claim should be downgraded. Since the reader already recommended a CONDITIONAL verdict pending additional validation, my concern does not change that recommendation: it strengthens the condition that calibration be demonstrated. I therefore keep the verdict as CONDITIONAL (unchanged).","tokens_in":11601,"tokens_out":7070,"duration_ms":72937,"concrete_test":"Generate M=100 observed CMB data vectors by sampling the true likelihood (Eq. 23) at a fixed θ* with independent noise realizations. For each dataset, run Sequential LSBI exactly as in the paper (k=10^4, five rounds, N=100) and record the resulting posterior. Compute the empirical coverage of the LSBI 68% and 95% credible intervals for θ*, and compare with nested sampling coverage on the same datasets. If the true parameters fall inside the credible regions less often than the nominal confidence (e.g., below 50% for the 68% interval), the posterior is overconfident and the central claim of agreement fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that sequential LSBI generates posteriors that agree with the underlying parameters, but the paper never quantifies calibration. The CMB example uses a scaled chi-squared likelihood (Eq. 23) whose variance depends on the model parameters through C_ell(θ) + N_ell. This directly contradicts the homoscedastic Gaussian assumption of Eq. (1). Sequential narrowing can make the linear-in-mean approximation plausible, but it does not remove the parameter dependence of the noise covariance; the estimated C in Eq. (15) is a single pooled covariance, not a function of θ. The paper's own Figure 5 shows the LSBI KL divergence slightly exceeding the nested-sampling band, which is a warning sign of overconfidence. Yet no coverage or empirical calibration test is reported, and the comparison with nested sampling is based on one dataset and visual inspection. Without a demonstration that the 68% and 95% credible regions contain the true parameters at the advertised frequency, the strongest claim does not follow from the evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops Linear Simulation-based Inference (LSBI), which approximates the likelihood as a homoscedastic Gaussian with a mean that is linear in the parameters. The hyperparameters m, M, and C are inferred from simulator outputs via closed-form posterior distributions, stated for broad uniform priors in Eqs. (13)-(15) and derived for conjugate priors in Appendix A. A sequential version iterates, using the previous posterior as the sampling distribution for the next round of simulations. The method is tested on a 50-dimensional quadratic toy model with Gaussian and non-Gaussian noise, and on an emulated CMB temperature power spectrum. The authors report convergence in four to five rounds of O(10^4) simulations, competitive with neural SBI, and conclude that LSBI yields posteriors that agree with the underlying parameters while remaining explainable.","tokens_in":11803,"tokens_out":5991,"duration_ms":57582,"significance":"If the claims hold, LSBI would be a useful interpretable alternative to neural density estimation for problems whose likelihood is locally linear-Gaussian. The paper has several strengths: the conjugate-prior derivation in Appendix A appears correct, the method is simple and transparent, and the authors provide code. However, the main-text uniform-prior result is underived, and the CMB validation relies on visual comparison with nested sampling for a single dataset without coverage checks, so the central claim of agreement with underlying parameters is not yet established. The non-Gaussian toy results also show a clear discrepancy for Laplacian noise, which the conclusion does not fully acknowledge.","major_comments":[{"comment":"The main-text posterior distributions for m, M, and C are stated as following from 'broad uniform priors', but no derivation is given and the prior is not specified exactly. A uniform prior on the covariance C is not a well-defined objective prior, and the inverse-Wishart form in Eq. (15) with ν = k - d - n - 2 is not the usual standard normal-inverse-Wishart regression result. Since Eqs. (13)-(15) are the basis for every result in the paper, please provide the derivation or a precise statement of the improper or reference prior used, and explain how the degrees of freedom arise.","section":"II.B, Eqs. (13)-(15)"},{"comment":"The CMB test uses the scaled chi-squared likelihood (2ℓ+1)/(Cℓ+Nℓ) times a chi-squared variate, whose variance depends on the model parameters through Cℓ(θ)+Nℓ. This violates the homoscedastic assumption of Eq. (1), and the pooled covariance C estimated in Eq. (15) is not a function of θ. The paper compares LSBI posteriors with nested sampling by eye and by KL divergence on one dataset; Figure 5 shows KL above the nested-sampling band, which is consistent with overconfidence. No coverage or calibration test is reported. Please add a simulation-based coverage check (e.g., credible-interval coverage over many observed datasets) or otherwise quantify the calibration of the CMB posteriors. Without this, the claim that the posteriors agree with the underlying parameters is unsupported.","section":"III.B, Eq. (23) and Eq. (1)"},{"comment":"For the asymmetric Laplacian noise case, the text states that the KL divergence does not converge to within the error bars of the nested-sampling estimate, and the conclusion nevertheless claims that LSBI generates posteriors that agree with the underlying parameters. This is an inconsistency between the presented evidence and the abstract/conclusion. Please either soften the claim to reflect the non-Gaussian results or investigate the source of the discrepancy.","section":"III.A, Figs. 3-4"},{"comment":"The CMB runs use N = 100 Monte Carlo samples of the hyperparameters, while Section II.C recommends N of order 10^3 and notes that overconfidence grows as N decreases. The reported KL excess in Figure 5 may therefore reflect an undersampled integral rather than method performance. Please repeat the CMB analysis with larger N, or provide a convergence check in N, and report the resulting KL and posterior calibration.","section":"III.B, Eq. (19) and Fig. 5"}],"minor_comments":[{"comment":"The prior mean for n_s is written as 0.964 × 10^-1, which equals 0.0964; if the intended value is n_s ≈ 0.964, this is a typo that should be corrected for reproducibility.","section":"Appendix C, Eq. (C1)"},{"comment":"The caption says the posteriors are computed from k = 10^6, 500, 2500, and 10000 samples, but it is ambiguous which k corresponds to which panel and round; please specify the values explicitly.","section":"Figure 3 caption"},{"comment":"The notation uses θ and D for both the full vectors and their sample means; using overlines or a different symbol for the means would improve clarity.","section":"Eqs. (8)-(11)"},{"comment":"The Bayes-factor definition appears both in Eq. (20) and in Eq. (B1); please cite the appendix at first use or remove the duplication.","section":"Eq. (20) and Appendix B"},{"comment":"The sentence 'The estimated error nested sampling is also shown as a gray band' is missing a word and should read 'The estimated error of the nested-sampling result is also shown as a gray band.'","section":"Figure 5 caption"},{"comment":"The number of hyperparameter samples N used for the toy-model results is not stated in the text; please provide the values used so the results are reproducible.","section":"II.B, Eq. (19)"}],"recommendation":"major_revision","confidential_remarks":"I agree with the conditional assessment. The method is promising and the conjugate-prior derivation is a useful contribution, but the underived uniform-prior result and the absence of calibration for the CMB case are load-bearing gaps. A major revision with a proper derivation and coverage tests would make the central claim credible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clean but modest methods paper. It takes a known statistical object—Bayesian linear regression with unknown covariance—and packages it as a simulation-based inference algorithm with sequential rounds. That framing is genuinely useful as a transparent baseline against neural SBI, and the derivations in the main text and appendix are correct as far as I checked. But the paper overclaims in two places: the 'competitive with state-of-the-art' claim is not quantified, and the CMB test uses a likelihood (Eq. 23) that violates the paper's own homoscedastic Gaussian assumption, with no coverage check to show the resulting posteriors are calibrated.\n\nWhat's actually new: very little mathematically. Eqs. 13-15 are standard posterior results for multivariate linear regression; the sequential scheme is the same round-by-round idea as SNL. SELFI and MOPED already exploit linearity in SBI. The value is in the presentation: a non-neural, interpretable SBI recipe with released code. That's worth having if done carefully.\n\nThe paper has real strengths. The conjugate-prior appendix is worked out properly. The authors are honest about failure modes: LSBI returns the prior for Cauchy noise, and they flag that the DKL is slightly overestimated in the CMB case. They also discuss the effect of finite N. This is honest work, not a marketing exercise.\n\nWhere it's soft: first, the main text says 'broad uniform priors' for m, M, C. That's not a well-defined statement for a covariance matrix, and the posterior in Eq. 15 must come from a specific improper or inverse-Wishart prior. The claim needs to be corrected or clarified. Second, the CMB likelihood in Eq. 23 is chi-squared with variance depending on C_ell(theta)+N_ell; this is parameter-dependent noise, so the homoscedastic assumption of Eq. 1 is violated. Sequential narrowing makes the linear-in-mean approximation better, but it does not make the covariance constant. Since the paper does not run any coverage or empirical calibration test, the 'posteriors agree with underlying parameters' claim is not backed up. The KL excess in Fig. 5 is consistent with overconfidence, so this is not a nitpick. Third, the neural comparison is not quantitative; no wall-clock times or accuracy metrics for neural methods are given.\n\nWho is this for: someone building an SBI pipeline who wants a simple, explainable first pass, or a teaching tool for SBI. It is not a paper that changes how cosmology does inference. It deserves a serious referee, but the referee should ask for prior specification, a proper calibration study, and a toned-down competitiveness claim. If those are fixed, it could be a solid methods note.","headline":"An honest, modest SBI baseline built on standard Bayesian linear regression; the calibration claim in the CMB section is not yet supported.","tokens_in":12338,"tokens_out":2516,"would_cite":false,"duration_ms":24959,"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":"This paper argues that fast, explainable simulation-based inference can be achieved by replacing neural density estimators with a sequentially updated linear-Gaussian likelihood, reaching nested-sampling-quality CMB posteriors in about…","keywords":["simulation-based inference","linear likelihood approximation","sequential inference","Bayesian parameter estimation","CMB power spectrum","neural density estimation","explainable inference","CosmoPower"],"falsifier":"Run sequential LSBI on a realistic CMB likelihood that includes non-Gaussian foreground and beam systematics and compare to nested sampling after five rounds of $10^4$ simulations; if the KL divergence does not approach the nested-sampling value, the paper's central claim that LSBI is competitive on realistic cosmological data is falsified.","tokens_in":11342,"feed_emoji":"🌌","tokens_out":6892,"duration_ms":66963,"temperature":0.7,"pith_summary":"The paper introduces Linear Simulation-based Inference (LSBI), a likelihood-free parameter estimation method that approximates the simulator's output as a Gaussian whose mean is linear in the parameters. It derives closed-form posterior distributions for the unknown linear-likelihood hyper-parameters directly from simulator samples, then iterates the procedure around the observed data to handle nonlinear models. On toy models and an emulated CMB temperature power spectrum, four or five rounds of roughly ten thousand simulations per round produce posteriors that agree with nested sampling while remaining fully explainable. The broader point is that fast simulation-based inference does not require a neural network, which addresses the lack of intellectual oversight that black-box approaches can create.","feed_headline":"Linear SBI matches neural-net quality using ~10^4 simulations","feed_subtitle":"A no-neural-network likelihood approximation converges in four to five rounds on CMB power-spectrum data.","key_machinery":"The load-bearing object is the linear simulation-based likelihood, $D|\\theta \\sim \\mathcal{N}(m + M\\theta, C)$, with hyper-parameters $m$, $M$, and $C$ treated as unknown. The paper derives their posterior distributions directly from simulator samples (Eqs. 13\\textendash 15), avoiding neural density estimation; the sequential version then uses the previous posterior as the sampling distribution for the next round's simulations, which is what makes the linear approximation hold for nonlinear models. This machinery converts simulation-based inference into a sequence of closed-form Gaussian, matrix-normal, and inverse-Wishart updates.","core_discovery":"Under the linear-Gaussian approximation $D|\\theta \\sim \\mathcal{N}(m + M\\theta, C)$, LSBI obtains closed-form posteriors for the unknown likelihood hyper-parameters from $k$ simulated pairs $\\{(\\theta^{(i)}, D^{(i)})\\}$: $m$ is Gaussian, $M$ is matrix-normal, and $C$ is inverse-Wishart (Eqs. 13\\textendash 15). The parameter posterior is then a Gaussian mixture obtained by sampling these hyper-parameter posteriors. Sequentially re-running LSBI with simulation parameters drawn from the previous round's posterior narrows the region where the linear approximation is valid, and on the CMB power spectrum this converges after four or five rounds of $10^4$ simulations to posteriors consistent with nested sampling. The paper claims this makes fast, credible simulation-based inference achievable without neural networks and with full explainability.","pith_inferences":["The linear-Gaussian framework suggests a natural active-learning strategy: using the current posterior covariance to choose the next round's simulation points could further reduce the $10^4$ simulation budget.","The data compression proposed in Appendix B, if implemented, would make LSBI attractive for very high-dimensional datasets such as full-sky CMB maps or 3D galaxy surveys, where inverting $d \\times d$ matrices becomes the bottleneck.","The evidence estimate in Eq. 20, combined with the proposed linear compression, points toward an explainable route to Bayesian model comparison that neural simulation-based inference methods typically struggle to provide.","A likely failure mode not stressed in the paper is that the Gaussian mixture posterior may become overconfident when the number of hyper-parameter samples $N$ is small; the paper notes this dependence, and testable diagnostics for ensemble spread could be developed."],"forward_implications":["Sequential LSBI reaches the same posterior as nested sampling with a total simulation budget of roughly four to five times $10^4$ calls on the CMB temperature power spectrum, comparable to neural density estimators.","The method tolerates deviations from Gaussian noise: uniform and Student-$t$ noise cases converge after about four rounds, while asymmetric Laplacian noise shows slower but still useful convergence.","Because the posterior is a Gaussian mixture over analytically sampled hyper-parameters, the computation can be parallelized and does not require training a neural network.","The same framework yields an estimate of the Bayesian evidence, so model comparison can in principle be performed without additional simulator calls.","A minimum of $k_{\\min} = n + 2d + 2$ simulations is required for well-defined distributions, and the recommended number of hyper-parameter samples $N$ is of order $10^3$ to avoid overconfidence.","Simulation-free inference is achievable without the use of neural networks, preserving explainability and intellectual oversight.","When applied to the Cosmic Microwave Background power spectrum, sequential LSBI produces posteriors that agree with the underlying parameters after four to five rounds of $O(10^4)$ simulations."],"supporting_citations":[{"why":"Supplies the simulation-based inference paradigm and the motivation for likelihood-free parameter estimation.","marker":"[1]"},{"why":"Define the DELFI neural density estimation approach that LSBI is compared against for simulation budget.","marker":"[5, 6]"},{"why":"Describes TMNRE, a state-of-the-art neural simulation-based inference method whose convergence LSBI aims to match.","marker":"[7]"},{"why":"Documents overconfident posterior approximations in neural simulation-based inference, motivating the explainability goal.","marker":"[19]"},{"why":"Introduces a black-box linear analysis for cosmological surveys, the closest prior linear simulation-based work.","marker":"[20]"},{"why":"Introduces the MOPED linear data compression used in the proposed compression scheme.","marker":"[21]"},{"why":"Provides the matrix-normal and inverse-Wishart distributional machinery used in the analytic posteriors.","marker":"[22]"},{"why":"Provide the CosmoPower and CosmoPowerJAX emulators used to generate CMB power-spectrum simulations.","marker":"[23, 24]"},{"why":"Supplies nested sampling via dynesty, the reference posterior used for comparison.","marker":"[25]"}],"fun_headline_variants":["Linear SBI matches neural nets in just five rounds","Explainable SBI: 10^4 sims per round, 4-5 rounds total","No neural nets: LSBI rivals SOTA on CMB data","Fast, transparent SBI: LSBI needs only 10^4 sims","Sequential linear SBI reaches neural-net grade posteriors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that, within each round, the simulator's output is well described by a Gaussian whose mean is linear in the parameters; if the true noise has undefined variance (as with Cauchy noise) or the model is strongly nonlinear over the sampled region, the approximation breaks down.","fun_headline_variants_meta":{"raw":{"variants":["Linear SBI matches neural nets in just five rounds","Explainable SBI: 10^4 sims per round, 4-5 rounds total","No neural nets: LSBI rivals SOTA on CMB data","Fast, transparent SBI: LSBI needs only 10^4 sims","Sequential linear SBI reaches neural-net grade posteriors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000254,"raw_usage":{"total_tokens":1522,"prompt_tokens":852,"completion_tokens":670,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":574}},"tokens_in":468,"tokens_out":670,"duration_ms":6506,"temperature":1.0,"reasoning_tokens":574,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:43:56.236211+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run sequential LSBI on a realistic CMB likelihood that includes non-Gaussian foreground and beam systematics and compare to nested sampling after five rounds of $10^4$ simulations; if the KL divergence does not approach the nested-sampling value, the paper's central claim that LSBI is competitive on realistic cosmological data is falsified.","supporting_citations":[{"cited_title":"It can be noted by eye that the posterior coincides well with the result of nested sampling after four to five rounds of LSBI","cited_arxiv_id":null,"evidence_quote":"Describes TMNRE, a state-of-the-art neural simulation-based inference method whose convergence LSBI aims to match."},{"cited_title":"Castelvecchi, Can we open the black box of ai?, Nature News 538, 20 (2016)","cited_arxiv_id":null,"evidence_quote":"Documents overconfident posterior approximations in neural simulation-based inference, motivating the explainability goal."},{"cited_title":"Leclercq, W","cited_arxiv_id":null,"evidence_quote":"Introduces the MOPED linear data compression used in the proposed compression scheme."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the matrix-normal and inverse-Wishart distributional machinery used in the analytic posteriors."},{"cited_title":"Spurio Mancini, D","cited_arxiv_id":null,"evidence_quote":"Supplies nested sampling via dynesty, the reference posterior used for comparison."}],"review_version":1}