{"id":"39f0cbbc-6bdf-4eb5-bee9-91dc8b8d9cff","arxiv_id":"2509.08102","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"RAISOR is a recursive adaptive importance sampler that uses a proven exponential replenishment schedule to balance computational cost and sample quality under model misspecification.","lead":"This paper introduces RAISOR, a Bayesian sampling method that combines recursive importance sampling with periodic sample replenishment, and proves that exponential replenishment schedules are asymptotically optimal. It shows the method parallelizes well and demonstrates speed gains on simulated data and sea surface temperature prediction.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (14) defines θ* via expected density ratio, but the proof invokes the misspecified BvM theorem for the KL minimizer; as written, the expected score need not vanish and the CLT in Lemma A.2 is unjustified.","rationale":"The reader's weakest-assumption pick (i.i.d. observations) is a real limitation, but it is explicitly acknowledged and the theorem is clearly conditional on it. The more load-bearing issue is that the assumptions as written misdefine the very quantities on which Theorem 3.1 depends: θ* is not the parameter to which the misspecified MLE/posterior converges, and Wθ* is not a covariance matrix as typed. The proof's invocation of Kleijn & Van der Vaart (2012) requires the KL minimizer and the score outer product. Because this affects the central theoretical claim, it should be corrected before acceptance. The correction appears straightforward (insert 'log' in (14) and write the outer product in (18)), and with that correction the proof structure seems sound, so I would not reject the paper; I would make acceptance conditional on fixing these definitions and re-verifying the first-order condition.","tokens_in":30864,"tokens_out":17810,"duration_ms":216807,"concrete_test":"Re-derive the first-order condition of Eq. (14): compute ∂/∂θ E*([y1]*/[y1|θ]) and check whether it implies E*ℓ̇(θ*)=0. Then test the CLT in Lemma A.2 by evaluating the mean of δ_{n-n0}ℓ̇^{n}_{n0+1}(θ*) under the stated definition: if E*ℓ̇(θ*)≠0, the centered quantity is not asymptotically normal. As a constructive check, replace (14) with θ* = argmin E* log([y1]*/[y1|θ]) and (18) with Wθ* = E*[ℓ̇(θ*)ℓ̇(θ*)']; if the proof of Lemma A.2 and Theorem 3.1 then goes through unchanged, the issue is a typo in the assumptions rather than a substantive gap.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Theorem 3.1's proof rests on Kleijn & Van der Vaart's Bernstein-von-Mises theorem under misspecification, which requires θ* to be the Kullback-Leibler minimizer, i.e., the maximizer of E* log[y|θ]. However, Eq. (14) defines θ* = argmin E*([y1]*/[y1|θ]), an expected density-ratio (χ²-type) objective. The first-order condition for (14) is E*[(f*/f_θ*) ℓ̇(θ*)] = 0, not E*ℓ̇(θ*) = 0. Without the latter, the centered score sequence used in Lemma A.2, δ_{n-n0} ℓ̇^{n}_{n0+1}(θ*), has mean √(n-n0) E*ℓ̇(θ*) which generally diverges, so its claimed CLT fails and the whole limiting-RESS derivation collapses. Similarly, Wθ* in (18) is written as the expectation of the gradient vector rather than the outer product E*[ℓ̇(θ*)ℓ̇(θ*)']; taken literally, Mθ* is not well-defined. These are likely typographical slips—the subsequent algebra clearly intends the KL minimizer and the score covariance—but they are central assumptions, not peripheral notation. A reader who follows Eq. (14) literally cannot verify Theorem 3.1, and the paper's claim of asymptotic optimality under misspecification depends on that theorem.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RAISOR, a recursive adaptive importance sampling algorithm that alternates fast recursive weight updates with occasional sample replenishment. The central theoretical contribution is Theorem 3.1, which states that, under i.i.d. observations and possible model misspecification, the relative effective sample size RESS(n|n0) with n0 = alpha_n n and alpha_n -> alpha has the non-degenerate limiting distribution RESS_{theta*,alpha}(z) = {alpha(2-alpha)}^{d/2} exp(-((1-alpha)/(2-alpha)) z' M_{theta*} z). This result is used to argue that exponentially growing replenishment times achieve an asymptotically optimal trade-off between O(n^p) Monte Carlo cost and sample quality. The paper also provides detailed implementation guidance, two simulated experiments (conjugate normal and Gaussian process regression with Vecchia approximation), and an application to sea surface temperature data in the Gulf of Mexico. The proof in Appendix A is substantial and explicitly builds on Kleijn & Van der Vaart (2012).","tokens_in":31309,"tokens_out":11177,"duration_ms":123667,"significance":"If Theorem 3.1 is made correct, the paper makes a useful contribution: it gives a concrete asymptotic characterization of degeneracy in a recursive importance sampling scheme, explains why exponential replenishment is the right growth rate, and provides practical heuristics with extensive appendices. The explicit limiting formula, including the dependence on alpha and on the asymptotic covariance structure, is a valuable diagnostic. The paper also carefully states the limitations of the i.i.d. assumption and applies the method to a challenging NNGP spatial model. The main problem is that two definitions in the theorem statement are inconsistent with the proof: the objective defining theta* is not the KL objective required by the cited Bernstein-von-Mises theorem, and W_{theta*} is defined as a vector while used as a matrix. These are fixable locally, but as submitted the theorem cannot be verified.","major_comments":[{"comment":"Eq. (14) defines theta* = argmin E*([y1]*/[y1|theta]), an expected density-ratio (chi-squared type) objective, but the proof relies on Kleijn & Van der Vaart (2012), whose Bernstein-von-Mises theorem requires theta* to be the KL minimizer, i.e., the maximizer of E* log[y1|theta]. The first-order condition for Eq. (14) is E*[(f*/f_{theta*}) ldot(theta*)] = 0, not E*[ldot(theta*)] = 0. Consequently, the score central limit theorem used in Lemma A.2, Eq. (55), is generally false: sqrt(n-n0) times the average score has mean sqrt(n-n0) E*[ldot(theta*)], which diverges unless E*[ldot(theta*)] = 0. Since the limiting RESS derivation rests on this CLT, the theorem as stated is not justified. Please change Eq. (14) to the KL-minimizing definition (or supply a separate proof for the density-ratio objective).","section":"Section 3.1, Eq. (14); Appendix A, Eq. (55)"},{"comment":"W_{theta*} is defined as E*(d/dtheta log[y|theta] | theta=theta*), which is a d-dimensional vector, but it is subsequently used as a positive-definite d x d covariance matrix: the theorem writes W_{theta*}^{1/2}, and Lemma A.2 uses W_{theta*} in normal limits such as Eq. (55) and Eq. (56). The definition should be W_{theta*} = E*[ldot(theta*) ldot(theta*)'], the expected outer product of the score vector. As written, M_{theta*} = (W^{1/2})' V^{-1} (W^{1/2}) is ill-defined and the proof's algebra cannot be checked.","section":"Theorem 3.1, Eq. (18); Appendix A, Lemma A.2"}],"minor_comments":[{"comment":"In the cost discussion, the text says 'the cost per Monte Carlo sample C_W of recalculating the weights throughout the procedure will be C_R(n) = O(sum_{k=1}^n k^p)'. The symbol C_R is used instead of C_W. Please fix the notation.","section":"Section 2.3.2"},{"comment":"The posterior mean formula for the conjugate normal model is written as mu_n = sigma_n^2 {sigma_0^{-2} mu_0}, omitting the data term n sigma^{-2} ybar. If the simulations silently use ybar = 0, please state that; otherwise the formula is incorrect.","section":"Section 4.1.1"},{"comment":"The estimator in Eq. (11) is typeset as '\\RESS' with a literal backslash. It should be '\\widehat{RESS}' or a similar estimator notation.","section":"Eq. (11)"},{"comment":"The notation for replenishment times, n*_k = ceil(alpha^{-k} n) or ceil(alpha_n^{-k}), is ambiguous in the typeset equation. Please clarify whether the multiplier 'n' is part of the argument or a subscript.","section":"Section 2.3.3 and Theorem 3.1"},{"comment":"The definition of z in Example C.2 is garbled: 'z = { n/((n-n0)n0)}^{-1/2} W^{-1/2} V(...)' should be rewritten so that the scaling and the random variable are unambiguous.","section":"Appendix C, Example C.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is promising and the proof strategy is standard once the two definitional issues in Theorem 3.1 are corrected. I do not see a need for reject: the errors appear to be local and fixable, but they are not merely cosmetic because a reader cannot verify the theorem as written. Once Eq. (14) and Eq. (18) are corrected, the proof should be re-checked carefully, especially the CLT in Eq. (55) and the definition of M_{theta*}."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the asymptotic analysis of the RESS under misspecification: Theorem 3.1 provides a limit distribution and shows that exponential replenishment is asymptotically optimal in the cost/quality trade-off. The algorithm itself is a clear synthesis of ideas that are already in the literature, and the paper is honest about its scope, explicitly flagging the i.i.d. assumption and the case-by-case behavior when it fails. The cost calculation in Section 2.3.3 is straightforward and correct, and the simulated examples do illustrate the qualitative behavior of the limiting RESS.\n\nBut the central theorem has a problem that a careful referee will catch quickly. Equation (14) defines θ* as the minimizer of E*([y1]*/[y1|θ]), a density-ratio objective. The proof of Lemma A.2 and the appeal to Kleijn & Van der Vaart (2012) require θ* to be the KL minimizer, i.e., the maximizer of E* log[y|θ]. Those two parameters are generally different under misspecification. The first-order condition for (14) is E*[(f*/f_θ*) ℓ̇(θ*)] = 0, not E*ℓ̇(θ*) = 0; without the latter, the centered score sequence in (55) has a mean that diverges and the CLT fails. Similarly, (18) defines Wθ* as the expectation of the gradient vector, a d-vector, but the proof and examples treat it as the score covariance matrix E*[ℓ̇ℓ̇']. These look like typographical slips rather than conceptual errors, but they sit at the load-bearing point of the paper. As printed, the proof of Theorem 3.1 does not go through. The authors need to either restate θ* as the KL minimizer (which is what the surrounding algebra already assumes) or justify why the density-ratio minimizer satisfies the needed score condition, and redefine Wθ* as the outer product.\n\nThe other limitations are more minor and are mostly acknowledged: the theorem assumes perfect replenishment and i.i.d. data, while the method uses approximate proposals and stopping times; the validation is limited to two examples, one of which (spatial GP) is outside the theorem's assumptions. The paper gives a fair discussion of these gaps and provides enough implementation detail to reproduce the experiments.\n\nWho is this for? People working on parallel-friendly Bayesian computation for low-dimensional posteriors. It deserves a serious referee, but the referee should send it back for a corrected theorem statement or a proof patch before the central claim is trusted. I would accept for peer review.","headline":"The RAISOR algorithm is a sensible integration of recursive Bayesian updating with adaptive importance sampling, and the exponential-replenishment cost story is clean, but the theorem as written has a load-bearing mismatch between the definition of θ* and the BvM theorem it invokes — fix that and the paper is solid.","tokens_in":31746,"tokens_out":2655,"would_cite":true,"duration_ms":34293,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","65C05","62M30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Exponential replenishment keeps recursive importance sampling both fast and stable.","keywords":["adaptive importance sampling","recursive Bayesian inference","effective sample size","sample replenishment","model misspecification","parallel computing","Gaussian process regression","sea surface temperature"],"falsifier":"Run the method on repeated i.i.d. datasets under the paper's regularity conditions, record the RESS at exponentially spaced replenishment times with a fixed α, and compare the empirical distribution to {α(2−α)}^{d/2} exp(−((1−α)/(2−α)) z' M_{θ*} z); a systematic mismatch for large n would refute the claimed convergence. Separately, under dependent observations generated by, say, a Gaussian process or an AR(1) model, check whether increasing the replenishment growth rate beyond exponential still maintains a non-degenerate RESS at O(n^p) cost; a collapse would confirm the paper's conjecture that","tokens_in":30803,"feed_emoji":"⚖️","tokens_out":3646,"duration_ms":48788,"temperature":0.7,"pith_summary":"The paper proposes RAISOR, a recursive adaptive importance sampling algorithm for Bayesian inference that alternates cheap recursive weight updates with occasional sample replenishment steps. Its central claim is that making the gaps between replenishments grow exponentially is provably optimal: it keeps the computational cost per Monte Carlo sample at the same order as a single posterior evaluation while preventing the weighted sample from degenerating. The proof holds even when the model is misspecified, and the paper also shows the method works in practice on Gaussian process regression and sea surface temperature prediction, with large gains in effective samples per minute over MCMC.","feed_headline":"Exponential resampling keeps Bayesian sampling fast and stable","feed_subtitle":"A recursive sampler proves that replenishment gaps growing exponentially hit the optimal cost-versus-quality trade-off, even for misspecifie","key_machinery":"The relative effective sample size (RESS), defined as {1 + χ²-divergence([θ|y1:n] ∥ [θ|y1:n0])}^{-1}, is the object that tracks sample degeneracy. The argument shows that RESS(n|n0) can be rewritten as a ratio of integrals of a likelihood-ratio power against the partial posterior density, then uses the Bernstein-von Mises theorem under misspecification to replace the partial posterior by a normal density. This reduces the asymptotic problem to a product of Gaussian integrals, yielding the closed-form limiting RESS identity and the exponential-replenishment optimality conclusion.","core_discovery":"The paper's central result is Theorem 3.1: under i.i.d. observations with mild regularity conditions, the relative effective sample size RESS(n|n0) of a sample that was replenished at time n0 converges in distribution to a closed-form random variable, RESS_{θ*,α}(z) = {α(2−α)}^{d/2} exp(−((1−α)/(2−α)) z' M_{θ*} z), where z is standard normal, α = lim n0/n, and M_{θ*} is a sandwich-type matrix built from the score and the prior-expected information at the limiting parameter θ*. This limit is non-degenerate for α in (0,1) and collapses to zero as α→0. Consequently, replenishment times that grow exponentially achieve an asymptotically optimal trade-off: per-Monte-Carlo-sample cost stays at O(n^","pith_inferences":["The paper's stochastic-threshold replenishment rule (trigger on estimated RESS falling below r) is its practical default, yet the optimality theorem is proven only for deterministic exponential schedules; a natural extension is to prove that the threshold rule inherits an exponentially growing sequence of replenishment times in an appropriate asymptotic regime.","The RESS trajectory itself could be repurposed as a model-checking diagnostic: the paper shows that misspecification with W_{θ*} > V_{θ*} accelerates sample degeneracy, so a fitting run that repeatedly hits low RESS values may flag overconfident posteriors, not merely bad sampler settings.","The method's reliance on mixture-of-normals approximations of the partial posterior suggests it could extend to online learning and multimodal targets through annealed bridging, but its effectiveness in high dimensions is expected to degrade, since the dimension d enters exponentially in the deterministic RESS bound.","A concrete testable extension: in correlated or non-i.i.d. data (where Theorem 3.1 does not apply), the paper conjectures the RESS limit must be assessed case by case; simulations with Matérn spatial dependence could reveal whether exponential replenishment still keeps a bounded, healthy RESS empirically."],"forward_implications":["If Theorem 3.1 is right, a practitioner can set replenishment times to grow exponentially and obtain a provable balance: O(n^p) cost per Monte Carlo sample with a non-degenerate RESS, matching the asymptotic cost of Metropolis-Hastings while remaining parallel-friendly.","The limiting distribution gives a quantitative way to choose replenishment thresholds: the deterministic factor {α(2−α)}^{d/2} yields the bound n ≤ c(r_min, d) n0 for guaranteed quality loss, which the paper turns into concrete heuristics for batch sizes and thresholds.","Because the theorem holds under model misspecification, the method is advertised as robust to wrong models, with the sandwich matrix M_{θ*} governing the speed of sample degeneration.","In the Gaussian-process regression and sea-surface-temperature experiments, RAISOR achieves a substantially higher effective-sample-size-per-minute than the MCMC baseline, suggesting the theoretical cost-quality trade-off translates into practical speedups.","The recursive weight update can be batched, so communication between parallel processors is reduced, making the method more efficient in distributed computing environments."],"supporting_citations":[{"why":"Supplies the Bernstein-von Mises theorem under model misspecification, the engine used to replace the partial posterior by a normal density and to control the likelihood-ratio expansion in the proof of Theorem 3.1.","marker":"Kleijn & Van der Vaart (2012)"},{"why":"Provides the consistency and central-limit-theorem theory for self-normalized importance sampling that justifies the IS estimators RAISOR uses.","marker":"Geweke (1989)"},{"why":"Establishes the recursive Bayesian (prior-proposal) framework that the paper's importance PP-RB baseline and RAISOR recursion build on.","marker":"Hooten et al. (2021)"},{"why":"Source of the SMC-style adaptive resampling and threshold-based ESS tracking that motivates RAISOR's stochastic replenishment rule.","marker":"Del Moral et al. (2012)"},{"why":"Gives convergence results for adaptive importance sampling that support the paper's claim that AIS-style adaptation can be made asymptotically sound.","marker":"Delyon & Portier (2018)"},{"why":"Supplies the regularity conditions and MLE consistency/asymptotic-normality results invoked as sufficient conditions in Section 3.1.","marker":"Van der Vaart (2000)"},{"why":"Defines the Matérn covariance used in the geostatistical model for the main applications.","marker":"Matérn (1986)"},{"why":"Provide the Vecchia approximation framework that makes the large-spatial-data likelihood evaluations tractable in both the MCMC baseline and RAISOR experiments.","marker":"Vecchia (1988) and Katzfuss & Guinness (2021)"},{"why":"Supplies the nearest-neighbor Gaussian process (NNGP) approximation used to fit the Gaussian process regression model in the applications.","marker":"Datta et al. (2016)"},{"why":"Supplies the in-situ NOAA iQuam sea surface temperature dataset used in the real-data application.","marker":"Xu & Ignatov (2014)"}],"fun_headline_variants":["Exponential resampling proves optimal in recursive Bayesian sampling","Optimal replenishment grows exponentially in recursive importance sampling","Recursive sampler: exponential replenishment hits optimal cost-quality trade-off","Closed-form result: exponential resampling is optimal for Bayesian samplers","Theorem shows exponential replenishment beats linear in recursive sampling"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The i.i.d. assumption on the observations is load-bearing: the proof of Theorem 3.1 breaks without independence, and the paper itself says the RESS behavior must then be assessed case by case, even though the method is applied to dependent spatial data.","fun_headline_variants_meta":{"raw":{"variants":["Exponential resampling proves optimal in recursive Bayesian sampling","Optimal replenishment grows exponentially in recursive importance sampling","Recursive sampler: exponential replenishment hits optimal cost-quality trade-off","Closed-form result: exponential resampling is optimal for Bayesian samplers","Theorem shows exponential replenishment beats linear in recursive sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1063,"prompt_tokens":721,"completion_tokens":342,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":465,"completion_tokens_details":{"reasoning_tokens":259}},"tokens_in":465,"tokens_out":342,"duration_ms":4981,"temperature":1.0,"reasoning_tokens":259,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T21:15:00.108991+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on repeated i.i.d. datasets under the paper's regularity conditions, record the RESS at exponentially spaced replenishment times with a fixed α, and compare the empirical distribution to {α(2−α)}^{d/2} exp(−((1−α)/(2−α)) z' M_{θ*} z); a systematic mismatch for large n would refute the claimed convergence. Separately, under dependent observations generated by, say, a Gaussian process or an AR(1) model, check whether increasing the replenishment growth rate beyond exponential still maintains a non-degenerate RESS at O(n^p) cost; a collapse would confirm the paper's conjecture that","supporting_citations":[],"review_version":1}