{"id":"455c65dc-e095-400e-ae51-370df708239c","arxiv_id":"2505.04603","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"ABI estimates the distance between posterior distributions via conditional quantile regression and uses adaptive rejection sampling with generative proposals to produce likelihood-free posterior approximations that converge to the true posterior.","lead":"A new method called Adaptive Bayesian Inference (ABI) performs likelihood-free inference by comparing posterior distributions directly, using a new distance called MSW and a neural network that predicts posterior quantiles. It avoids summary statistics and data-space discrepancies, and the paper claims better accuracy than Wasserstein ABC, summary-based ABC, and neural likelihood-free methods on four benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.5's proof for p>1 relies on a false convexity inequality: MSW_p is not convex in its first argument, so the oracle posterior convergence claim is unproven for p>1.","rationale":"The reader's verdict is CONDITIONAL, citing the unquantified quantile-network estimator as the weakest assumption while also noting that the main convergence theorem is incomplete for p>1. My independent check isolates a concrete mathematical flaw in the proof of Theorem 3.5: the Jensen-type inequality at line (E.16) is valid for p=1 but false for p>1, as shown by a simple two-point example. This is load-bearing because Theorem 3.5 is the paper's theoretical guarantee that the oracle ABI posterior converges to the true posterior, and the theorem is stated for all p≥1. The flaw does not necessarily invalidate the method: the p=1 case survives, the empirical comparisons may still be informative, and the theorem might be repaired. It does, however, mean the strongest_claim as stated ('the oracle ABI posterior converges... as the tolerance vanishes') is not currently established for p>1. The reader's separate concern about the gap between the oracle MSW and the network-estimated MSW remains valid and is not addressed by my finding. I do not see a reason to move beyond CONDITIONAL: the empirical contribution is substantial, the p=1 oracle result appears sound, and the missing steps are identifiable rather than fatal signs of invalidity.","tokens_in":38398,"tokens_out":5119,"duration_ms":54414,"concrete_test":"Check whether inequality (E.16) holds for a simple d=1, p=2 instance: let μ_1=δ_0, μ_2=δ_2, ν=δ_0, and α=1/2. Since MSW_2=W_2 on R, compute MSW_2(αμ_1+(1−α)μ_2,ν)=√2 and αMSW_2(μ_1,ν)+(1−α)MSW_2(μ_2,ν)=1. If this is accepted as a counterexample, re-derive Theorem 3.5 for p>1 without invoking (E.16), or remove the p>1 claim and state the theorem for p=1 only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical guarantee of the paper, Theorem 3.5, is stated for p∈[1,∞), but its proof uses inequality (E.16): MSW_p(π_ABI^(ε), π*) ≤ ∫_{A_ε} f_X(x) MSW_p(π_{Θ|x}, π*) dx / ∫_{A_ε} f_X(x) dx. This is a convexity/Jensen inequality for MSW_p in its first argument. It holds for p=1, because W_1 and hence MSW_1 are convex in the first argument via the dual representation. It fails for p>1. For example, on R with d=1, take μ_1=δ_0, μ_2=δ_2, ν=δ_0, λ=1/2. Then 0.5·W_2(δ_0,ν)+0.5·W_2(δ_2,ν)=1, but W_2(0.5δ_0+0.5δ_2,ν)=√2>1. Since W_2 = MSW_2 for d=1, the inequality (E.16) is false. The same non-convexity affects MSW_p for every p>1. Consequently, the displayed argument does not establish Theorem 3.5 for p>1, and the manuscript offers no alternative proof for this case. This is not about the practical quantile-network estimator; it is a gap in the oracle statement itself. The theorem may be repairable by restricting the claim to p=1, or by proving convergence through Theorem 3.6 and a separate argument controlling MSW_p(π_{Θ|X∈A_ε}, π*) without the Jensen step, but as written the guarantee is unsupported for p>1.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Adaptive Bayesian Inference (ABI), a likelihood-free inference framework that replaces data-space ABC discrepancies with a distance between posterior distributions, called the Marginally-augmented Sliced Wasserstein (MSW) distance. MSW combines coordinate marginals with random one-dimensional projections and admits a quantile representation, which the authors estimate by deep conditional quantile regression. The algorithm performs adaptive rejection sampling with decreasing tolerance thresholds and updates proposals via generative density estimation. The theoretical sections establish metricity, an IPM representation for p=1, topological equivalence with the Wasserstein distance, a convergence rate for the empirical trimmed MSW distance, and convergence of the oracle ABI posterior to the true posterior as the tolerance vanishes. Experiments compare ABI with WABC, ABC-SS, SNLE, SNPE, and WGAN-GP on four benchmark models.","tokens_in":38911,"tokens_out":9161,"duration_ms":86797,"significance":"The posterior-space matching idea is timely and potentially influential: it avoids hand-crafted summary statistics, remains meaningful for non-identifiable parameters and dependent observations, and the reduction of MSW to conditional quantile regression is a clean and practical algorithmic insight. The martingale-based proof of ABC posterior continuity (Theorem 3.6) is a genuine technical contribution. The empirical study is broad and the results are encouraging. However, the central theoretical guarantee is currently established only for p=1 in the oracle setting, and the implemented quantile-network estimator is not covered by the stated convergence theorems; these gaps must be repaired before the paper's main claims are fully supported.","major_comments":[{"comment":"Theorem 3.5 is stated for all p in [1,∞), but the proof hinges on inequality (E.16), which is a Jensen-type convexity inequality for MSW_p in its first argument. This inequality is valid for p=1, since W_1 and hence MSW_1 admit an IPM dual representation (Theorem 3.1), but it fails for p>1. For example, on R with d=1, take mu1=delta_0, mu2=delta_2, nu=delta_0, and lambda=1/2; then 0.5*W_2(delta_0,nu)+0.5*W_2(delta_2,nu)=1, while W_2(0.5*delta_0+0.5*delta_2,nu)=sqrt(2)>1. Since W_2=MSW_2 for d=1, Eq. (E.16) is false for p>1, and no alternative proof for p>1 is supplied. This is a load-bearing gap in the oracle convergence guarantee. Please either restrict Theorem 3.5 to p=1, or prove convergence for p>1 through a different argument, for example by combining Theorem 3.6 with a separate bound on MSW_p(pi_{Theta|X in A_epsilon}, pi*) that does not use the Jensen step.","section":"Section 3.2 / Appendix E.3.1, Eq. (E.16)"},{"comment":"The theoretical guarantees in Theorems 3.4 and 3.5 concern the oracle MSW distance, while the deployed algorithm (Algorithm 2, Eq. (2.4)) uses the estimated distance \\hat{MSW}_{p,delta,K,H} obtained from a neural quantile regression network trained on samples from the current proposal. The manuscript provides no bound on the error between \\hat{MSW} and the true MSW, and no statement that the estimated acceptance region approximates the oracle region A_t. Consequently, Theorem 2.2 (the sample complexity of ARS) and Theorem 3.5 do not apply to the implemented procedure, and the sequence of partial posteriors could be driven by a systematically biased metric. Please either provide an estimation-error bound under the stated network and training assumptions, or clearly state in Sections 2 and 5 that the convergence guarantees apply only to the oracle MSW and that the neural estimator is an approximation validated only empirically.","section":"Sections 2.1.2 and 2.2 vs. Theorems 3.4 and 3.5"},{"comment":"The advertised parametric rate O(m^{-1/(2p)}) in Remark 3.2 is a rate for the oracle empirical MSW distance between i.i.d. samples from fixed measures, not for the estimator used in Algorithm 3. The practical estimator additionally depends on the number of projections K, the number of quantile levels H, the network architecture, the Huber threshold kappa, and the optimization error, none of which appear in the bound. Since Algorithm 2 makes acceptance decisions using this estimated distance, the claim that ABI achieves a parametric convergence rate is not supported for the implemented method. The manuscript should either extend the analysis to the full estimator or explicitly restrict the rate claim to the oracle quantity.","section":"Section 3.1.2, Theorem 3.4 and Remark 3.2"}],"minor_comments":[{"comment":"The symbol delta is used both for the trimming parameter and for the confidence level in the assumptions and in the bound of Theorem 3.4, which makes formulas such as sqrt(log(16d/delta)) ambiguous. Rename the confidence level, for example to eta, and use it consistently with the theorem statement's \\bar{delta}.","section":"Theorem 3.4 and Appendix E.2.5"},{"comment":"In the paragraph on determining the sequence of tolerance levels, the displayed chain is ϵ0(alpha)>ϵ1(alpha)>...>ϵT(alpha), while Algorithm 2 and the surrounding text define the sequence as ϵ1>...>ϵT. Please align the indexing.","section":"Section 2.2.2"},{"comment":"Algorithm 2 line 3 uses the notation pi_X and pi_{x*} before the shorthand pi_x is defined in Section 2.1.2 after Eq. (2.4). Define this notation immediately after Definition 2.2 or when the algorithm is first presented.","section":"Algorithm 2 and Section 2.1.2"},{"comment":"The constant in the proof is written as Cd,lambda, but the symbol Cd,lambda is also used for the constant in Proposition E.1; the proof defines it as ~Cd(1-lambda)^{-1/(d+1)} after previously using Cd,p,lambda. Please use distinct names to avoid confusion.","section":"Appendix E.2.3, Proposition E.2"},{"comment":"There is a typo in the citation: 'Thoerem 3.4, Kallenberg and Kallenberg 1997' should be 'Theorem 3.4, Kallenberg and Kallenberg (1997)'.","section":"Appendix E.3.1"},{"comment":"The caption states that the best results are in bold, but for parameters theta2 and theta5 the best values belong to SNLE and WGAN, respectively, and are not bolded. Please either adjust the bolding or clarify the statement in the text.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript does not include a reproducibility statement or code availability. Given the strong dependence of the practical algorithm on the quantile network and the POTNet generative model, releasing code and seeds would substantially increase confidence in the empirical claims. Also, the introduction's claim of being 'the first non-Monte-Carlo-based sequential ABC' is strong; the revision should either support it with a careful literature comparison or soften it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the paper with interest. The core idea is genuinely attractive: replace data-space distances in ABC with a distance between posterior distributions, estimate that distance via quantile regression, and adapt proposals using a generative model. If it works, it sidesteps both the summary statistic problem and the curse of dimensionality in ABC. The empirical sections are extensive and, on the face of it, the method beats WABC, ABC-SS, SNLE, SNPE, and WGAN on four benchmark models. The trimmed MSW distance with marginal augmentation is a sensible construction, and the IPM characterization for p=1 plus the parametric rate on the empirical distance are real contributions.\n\nThe main problem is Theorem 3.5. It is stated for p in [1,∞), but the proof uses a Jensen-type inequality (Eq. E.16) that is false for p>1. The stress-test example is correct: in one dimension, MSW_2 = W_2, and W_2 is not convex in its first argument. So the displayed argument does not prove the theorem for p>1. That matters because the paper does not restrict itself to p=1 and the practical procedure may use p=2. I do not see an alternative proof in the manuscript. The theorem may be repairable, but as written the oracle convergence claim is unsupported for p>1.\n\nA separate but related gap: the theory is for the oracle MSW distance, while the algorithm uses a quantile network trained on the current proposal. There is no bound on the error between the learned and true MSW, so the theory does not cover the actual algorithm. This is a common gap in this literature, but it deserves to be flagged. The empirical claims also lack error bars and code. I would not hold that against the method itself, but it does weaken the evidence.\n\nOn balance, this is the kind of paper I would want refereed. The central idea is new and worth developing, the p=1 case may well be sufficient for many uses, and the experiments suggest real practical gains. Send it out. The referee should press on the p>1 gap and ask either for a full proof or a clean restriction to p=1, plus a handle on the estimation error.","headline":"A novel and empirically strong ABC method whose headline convergence theorem is unproven for p>1 due to a false convexity inequality; deserves peer review but needs a fix or a caveat.","tokens_in":39330,"tokens_out":2587,"would_cite":true,"duration_ms":26566,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62G08","62G05"],"pacs":[],"model":"deepseek-v4-flash","headline":"ABI claims that likelihood-free inference can be made efficient and exact in the limit by comparing posterior distributions with a marginally-augmented sliced Wasserstein distance estimated by conditional quantile regression.","keywords":["approximate Bayesian computation","likelihood-free inference","sliced Wasserstein distance","conditional quantile regression","adaptive rejection sampling","generative density estimation","posterior matching","simulator-based inference"],"falsifier":"For a model with a tractable exact posterior, compute the oracle MSW distance between posteriors for simulated datasets close to $x^*$ and compare it to the quantile-network estimate; if the network systematically misorders candidate datasets near $x^*$, ABI's acceptance region is mis-specified and its output will diverge from the true posterior even as the tolerance shrinks.","tokens_in":38226,"feed_emoji":"🎯","tokens_out":7105,"duration_ms":59827,"temperature":0.7,"pith_summary":"Adaptive Bayesian Inference (ABI) is a new likelihood-free inference method that replaces the usual comparison of simulated and observed data with a direct comparison of the posterior distributions those data would induce. The paper claims this posterior-space comparison is more robust than data-space discrepancies when sample sizes are small, observations are dependent, or parameters are non-identifiable. It proves that, using the oracle distance, the ABI posterior converges weakly to the true posterior as the tolerance threshold shrinks to zero, and that a trimmed version of its new distance can be estimated at the parametric rate $O(m^{-1/2})$ when $p=1$. The practical algorithm estimates the distance by conditional quantile regression and refines proposals through adaptive rejection sampling with generative density estimation. A sympathetic reader would care because this recasts a notoriously inefficient rejection problem as a supervised learning problem with sequential refinement.","feed_headline":"Posteriors, not data, are the key to fast likelihood-free inference","feed_subtitle":"New method compares posteriors via quantile regression and proves convergence to exact posterior as tolerance vanishes","key_machinery":"The load-bearing object is the Marginally-augmented Sliced Wasserstein distance, defined as a convex combination of the average trimmed $p$-Wasserstein distance between coordinate marginals and the sliced $p$-Wasserstein distance averaged over random sphere projections. Its quantile representation rewrites univariate Wasserstein distances as $L^p$ differences of quantile functions, so computing MSW between posteriors becomes a sequence of one-dimensional conditional quantile regression tasks solved by a single shared ReLU network. The other component is adaptive rejection sampling: at each iteration a budget-constrained rejection step samples from the joint proposal, parameters are pruned by the estimated MSW threshold, and a generative model fitted to the accepted draws becomes the next proposal. This mechanism converts posterior matching into a supervised distributional regression problem and permits sequential refinement without explicit prior density evaluation.","core_discovery":"The central claim is that approximate Bayesian inference can be done by matching full posterior distributions rather than data summaries. ABI defines the approximate posterior as the law of $\\theta$ conditional on the event that the simulated data's induced posterior $\\pi(\\theta \\mid X)$ lies within tolerance $\\epsilon$ of the observed posterior $\\pi(\\theta \\mid x^*)$ under the Marginally-augmented Sliced Wasserstein (MSW) distance. Theorem 3.5 shows that as $\\epsilon \\downarrow 0$ this oracle ABI posterior converges weakly in $P_p(\\Omega)$ to the true posterior $\\pi(\\theta \\mid x^*)$. The MSW distance is a metric that metrizes weak convergence on $P_p(\\mathbb{R}^d)$, is an integral probability metric when $p=1$, and its trimmed empirical version converges to the population value at rate $O(m^{-1/(2p)})$, recovering the parametric rate at $p=1$. These guarantees are established for the oracle distance; the implemented algorithm replaces it with a quantile-regression network estimate.","pith_inferences":["If the quantile-network estimate is accurate, the same posterior-matching kernel could be plugged into sequential Monte Carlo or population Monte Carlo schemes; the paper suggests this direction but does not develop it.","The key open question the theory leaves is whether the MSW estimate from the network is close enough to the oracle; a direct numerical comparison of estimated vs oracle MSW near $x^*$ would test this and could guide network capacity and tolerance schedules.","One could test whether the marginal augmentation term or the sliced term drives the empirical gains by ablating $\\lambda$; the paper reports performance for fixed choices but does not isolate these contributions.","The posterior-space view suggests a general recipe: any distributional metric with a quantile or dual representation could replace data-space discrepancies; comparing MSW with an energy-distance analogue would clarify how much of the gain is specific to slicing versus posterior matching itself."],"forward_implications":["When the data are fixed and the tolerance vanishes, the oracle ABI posterior converges weakly to the exact posterior instead of merely to a summary-based approximation.","The trimmed MSW distance can be estimated from finite samples at the parametric $O(m^{-1/2})$ rate for $p=1$, whereas ordinary Wasserstein ABC suffers the slow $O(n^{-1/s})$ rate for data dimension $s\\ge 3$.","Because samples are compared in posterior space, ABI remains meaningful for small observed sample sizes, dependent observations, and non-identifiable parameters where data-space IPM comparisons degrade.","The approximate rejection sampling error decays exponentially in the simulation budget $R$ under a local-positivity condition, so the adaptive scheme's bias can be controlled explicitly.","The sequential generative update allows inference when the prior is intractable, requiring only a simulator for the prior rather than its density."],"supporting_citations":[{"why":"Supplies the standard ABC posterior-convergence framework that ABI's martingale proof of Theorem 3.6 contrasts with.","marker":"Barber et al., 2015"},{"why":"Defines the Wasserstein ABC baseline and motivates ABI's move away from data-space IPMs due to dependent-data and small-sample issues.","marker":"Bernton et al., 2019"},{"why":"Provides the ABC-SS summary-statistic baseline whose first-order accuracy ABI aims to improve upon.","marker":"Jiang et al., 2017"},{"why":"Supplies the SNLE baseline and the sequential neural likelihood framework against which ABI is compared.","marker":"Papamakarios et al., 2019"},{"why":"Supplies the SNPE baseline against which ABI's sequential posterior approximation is compared.","marker":"Greenberg et al., 2019"},{"why":"Gives minimax rates for ReLU-network quantile regression that justify ABI's conditional quantile estimation.","marker":"Padilla et al., 2022"},{"why":"Provides the trimmed sliced Wasserstein convergence bound used in Theorem 3.4 for the sliced component of MSW.","marker":"Manole et al., 2022"},{"why":"Establishes the slow $O(n^{-1/s})$ empirical Wasserstein rate that motivates avoiding data-space Wasserstein comparisons.","marker":"Talagrand, 1994"},{"why":"Shows sliced Wasserstein metrizes weak convergence, which ABI uses for MSW's topological equivalence.","marker":"Nadjahi et al., 2019"},{"why":"Supplies the Wasserstein metric theory, including metricity and dual representation, on which MSW properties rest.","marker":"Villani et al., 2009"}],"fun_headline_variants":["Posterior-space matching beats data-space ABC","ABI: Bayesian inference without likelihoods, via posterior matching","Match posteriors, not summaries: new ABC variant","Quantile regression makes likelihood-free inference converge faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approximate algorithm's accuracy depends on the quantile regression network estimating the true MSW distance accurately, especially near observed $x^*$; the paper proves guarantees for the oracle MSW only, not for the network-estimated version.","fun_headline_variants_meta":{"raw":{"variants":["Posterior-space matching beats data-space ABC","ABI: Bayesian inference without likelihoods, via posterior matching","Match posteriors, not summaries: new ABC variant","Quantile regression makes likelihood-free inference converge faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000244,"raw_usage":{"total_tokens":1547,"prompt_tokens":974,"completion_tokens":573,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":520}},"tokens_in":590,"tokens_out":573,"duration_ms":6223,"temperature":1.0,"reasoning_tokens":520,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:25:25.473507+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a model with a tractable exact posterior, compute the oracle MSW distance between posteriors for simulated datasets close to $x^*$ and compare it to the quantile-network estimate; if the network systematically misorders candidate datasets near $x^*$, ABI's acceptance region is mis-specified and its output will diverge from the true posterior even as the tolerance shrinks.","supporting_citations":[{"cited_title":"The rate of convergence for approximate bayesian computation","cited_arxiv_id":null,"evidence_quote":"Supplies the standard ABC posterior-convergence framework that ABI's martingale proof of Theorem 3.6 contrasts with."},{"cited_title":"Approximate bayesian computation with the wasserstein distance","cited_arxiv_id":null,"evidence_quote":"Defines the Wasserstein ABC baseline and motivates ABI's move away from data-space IPMs due to dependent-data and small-sample issues."},{"cited_title":"Learning summary statistic for approximate bayesian computation via deep neural network","cited_arxiv_id":null,"evidence_quote":"Provides the ABC-SS summary-statistic baseline whose first-order accuracy ABI aims to improve upon."},{"cited_title":"Minimax confidence intervals for the sliced wasserstein distance","cited_arxiv_id":null,"evidence_quote":"Provides the trimmed sliced Wasserstein convergence bound used in Theorem 3.4 for the sliced component of MSW."},{"cited_title":"The transportation cost from the uniform measure to the empirical measure in dimension 3","cited_arxiv_id":null,"evidence_quote":"Establishes the slow $O(n^{-1/s})$ empirical Wasserstein rate that motivates avoiding data-space Wasserstein comparisons."}],"review_version":1}