{"id":"76579be8-4281-49c9-8132-fdc465c4d5f7","arxiv_id":"2501.19073","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A mixture of over- and under-truncated Pareto-frontier distributions, with the weight chosen by variational lower bound maximization, yields a new acquisition function that performs well, particularly for three or more objectives.","lead":"The paper proposes a new acquisition function for multi-objective Bayesian optimization that estimates the information a candidate point provides about the Pareto frontier. It blends an over-truncated and an under-truncated approximation of the frontier distribution and tunes the blend via a variational lower bound, especially helpful for three or more objectives.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported gains may stem from the biased MAP estimator (Eq. 8), not from the variational over/under-truncation mixture; the paper never compares the unbiased estimator against PFES.","rationale":"Read in good faith: the paper presents a clean variational bound, a sensible two-truncation mixture, and broad experiments. The math for the bound and lambda concavity is standard and appears correct given the support condition. Independent support: the paper includes extensive ablations and a separate estimator analysis (Appendix D.3), which is good practice. However, the central empirical claim is that the mixture helps for L>=3, and the evidence for this mechanism is missing because the unbiased estimator (the actual lower bound) is never benchmarked against PFES. The reader's weakest assumption identified the same estimator, so I agree. My concern sharpens it: the bias is not just a possible ordering distortion; it is a confound with the paper's stated novelty. The paper's own ablation shows the choice of estimator dramatically changes performance, so the burden is on the authors to show the mixture, not the estimator, drives the gains. This does not change the conditional verdict but adds a specific missing experiment that should be required before acceptance.","tokens_in":41774,"tokens_out":10958,"duration_ms":102928,"concrete_test":"Reproduce the GP-derived synthetic experiments (d=3, L=3,4,5,6) with three variants: (A) PFEV as published (Eq. 8, K=10), (B) PFEV with the unbiased Monte Carlo estimator (Eq. 6) at K=200 to match variance, and (C) PFES with the same number of GPs, RFM basis, and NSGA-II settings. Also include (D) PFEV with Eq. (8) but lambda fixed to a small value (e.g., 1e-6, effectively over-truncation only). If (B) does not outperform (C) while (A) does, the empirical advantage is attributable to the MAP heuristic rather than the under/over-truncation mixture. If (D) matches (A), the mixture itself is irrelevant.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's core novelty is the mixture of over- and under-truncation with lambda optimized by lower-bound maximization (Section 3.2). However, every main experiment uses the bias-corrected estimator Eq. (8), in which the indicator I(f in A_O) is replaced by theta_MAP = (p_hat + I)/2. This replaces the unbiased MC estimate of the lower bound with a biased heuristic that is no longer guaranteed to be a lower bound on MI; the variational guarantee applies only to Eq. (6). The only ablation, Fig. 7 and Appendix K.1, compares Eq. (6) vs Eq. (8) within PFEV and shows the MAP version performs markedly better. This is exactly the confound: the reported advantage of PFEV over PFES and {PF}2ES at L>=3 could be caused by the MAP smoothing, not by the under-truncation mixture that the paper claims is responsible. Without an experiment comparing PFEV with the unbiased estimator (or with large K) against PFES, the central claim that the variational mixture helps for L>=3 is unsupported. The paper even notes the bias but attributes the benefit to variance reduction, which is a heuristic argument rather than a test of the claimed mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Pareto-frontier Entropy search with Variational lower bound maximization (PFEV), an acquisition function for multi-objective Bayesian optimization. The method bounds the mutual information between the objective vector f(x) and the Pareto frontier F* by constructing a variational distribution that is a mixture of an under-truncated and an over-truncated predictive distribution, with the mixture weight λ optimized to maximize the bound. The authors derive the lower bound, provide Monte Carlo estimators, prove concavity of the objective in λ, and demonstrate empirically that PFEV achieves competitive or superior hypervolume performance, especially for L≥3 objectives. Extensions to parallel querying, decoupled observations, joint entropy search, and noisy settings are discussed.","tokens_in":41996,"tokens_out":5552,"duration_ms":52037,"significance":"If the central empirical claim holds, PFEV would be a useful contribution to information-theoretic multi-objective Bayesian optimization, complementing methods like PFES and {PF}2ES by unifying over- and under-truncation in a principled variational framework. The theoretical derivation of the lower bound and the concavity/existence results in Appendices A and C are clean, and the paper is unusually complete in its appendices, including proofs, variance analysis, and extensive synthetic and real-world experiments. However, the main experimental results are obtained with a biased Monte Carlo estimator (Eq. 8), and the paper does not isolate that estimator's contribution from the effect of the variational mixture; the significance is therefore conditional on resolving this confound.","major_comments":[{"comment":"The variational lower-bound guarantee and the KL-minimization interpretation in Remark 3.1 apply to the naive Monte Carlo estimator Eq. (6), but all main experiments (Figs. 3-6 and the appendix results) use the biased MAP estimator Eq. (8), where I(f∈A^O) is replaced by θ_MAP=(p̂+I)/2. Eq. (8) is not guaranteed to be a lower bound on MI, as acknowledged in the text, while the mechanism claimed to improve performance (the mixture of over- and under-truncation) only affects Eq. (6). The only ablation, Fig. 7 and Appendix K.1, compares Eq. (8) with Eq. (6) within PFEV and shows that the MAP version performs considerably better; this makes the reported gains over PFES and {PF}2ES at L≥3 potentially attributable to the smoothing/bias of Eq. (8) rather than to the variational mixture. To support the paper's central claim, the authors should compare PFEV with Eq. (6) (or with a sufficiently large K unbiased estimate) against PFES and {PF}2ES on the same problems, and ideally also report the selected λ values and acquisition ordering for both estimators.","section":"Section 3.3, Eq. (8); Section 6.4, Fig. 7; Appendix K.1"},{"comment":"The beta-prior approximation with mode at p̂=Z_O/Z_U and r=1 is an ad hoc heuristic that is not derived from the model: p̂ is the probability under the under-truncated distribution, not under the true conditioning on F*, so the prior mode can itself be substantially biased when the over- and under-truncated regions differ greatly (which the paper argues is the case for L≥3, see Fig. 2). The paper's justification for Eq. (8) is purely a variance-reduction argument, and the empirical evidence (Fig. 7 and Fig. 21) is based on average hypervolume, not on whether the estimator preserves the ranking of acquisition values across candidate points. A sensitivity analysis with respect to r, together with a direct comparison of the acquisition functions induced by Eq. (6) and Eq. (8), is needed to determine whether the observed benefit is due to the variational mixture or to the specific smoothing procedure.","section":"Appendix D.1; Section 3.3, Eq. (8)"}],"minor_comments":[{"comment":"The notation \"B\" used to denote equality by definition (e.g., \"𝒇 𝒙 B(...)\") appears to be a formatting artifact and should be replaced with a standard symbol such as \"≜\" throughout.","section":"Section 2"},{"comment":"The set A^{U\\O} is used in Eq. (3) but is formally defined only after the equation; it would be clearer to define A^{U\\O} = A^U \\ A^O just before the display.","section":"Section 3.2, Eq. (3)"},{"comment":"The figure caption is dense and does not explain the meaning of the red star points, the dashed lines, or why the mixture is conceptually between the over- and under-truncated regions; a short explanation of each panel would improve readability.","section":"Figure 1"},{"comment":"The reference \"Vlennet, R., Fonteix, C., and Marc, I.\" should be \"Viennet\" to match the standard spelling of the Viennet function.","section":"References"},{"comment":"The paper uses \"hyper-volume\" and \"hypervolume\" interchangeably; please choose one spelling for consistency.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's real contribution is the variational lower bound for Pareto-frontier mutual information, built from a mixture of over- and under-truncated distributions. That is new, and the math is solid: the bound is a genuine lower bound for lambda in (0,1], concavity in lambda and the existence of a maximizer are proven, and the geometric intuition that over- and under-truncation diverge quickly as L grows is plausible. If the mechanism works, it would be a clean fix for the over-truncation bias in PFES. I also give credit for the broad experiments, including GP-synthetic benchmarks, DTLZ, a LightGBM hyperparameter task, and extensions to parallel, decoupled, and noisy settings.\\n\\nThe soft spot is exactly what the stress-test note says. All main experiments use Eq. (8), the beta-MAP estimator, which replaces an indicator with the average of that indicator and a probability ratio. This estimator has no variational guarantee. The paper's own ablation, Fig. 7, shows the MAP version beats the unbiased version within PFEV, and it never compares PFEV-with-unbiased-estimator against PFES. So the headline claim that the mixture helps for L>=3 is not actually tested. The variance-reduction story is plausible but post hoc; the paper acknowledges the bias but does not isolate the mechanism.\\n\\nSecondary issues: no code is released, and the fixed r=1 prior concentration is a free parameter that changes the estimator; the authors discuss a decaying r but do not use it in the main experiments. These are not fatal, but they add to the underdetermination.\\n\\nWho should read this: people working on information-theoretic MOBO and on variational bounds for entropy search. The derivation in Section 3 and the appendices are worth a careful look. A referee should require one missing experiment: PFEV with the unbiased estimator at large K versus PFES on the same L>=3 settings. That would separate the contribution of the mixture from the contribution of the MAP smoothing.\\n\\nI would send this to peer review—the math is real and the variational idea is useful—but I would not accept it in this form. Major revision, with the isolating comparison as a condition.","headline":"Clean variational lower bound, but the main empirical claim is confounded: the reported gains at L>=3 could come from the beta-MAP estimator, not the mixture.","tokens_in":42530,"tokens_out":2633,"would_cite":true,"duration_ms":26369,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C29"],"pacs":[],"model":"deepseek-v4-flash","headline":"PFEV approximates Pareto-frontier information gain by a variational mixture of over- and under-truncated distributions, optimizing the mixture weight.","keywords":["multi-objective Bayesian optimization","Pareto-frontier entropy search","variational lower bound","truncated normal distribution","mutual information acquisition function","over-truncation","under-truncation","hypervolume"],"falsifier":"Run PFEV on the $L \\geq 3$ GP-derived synthetic problems with the unbiased estimator (6) instead of the MAP estimator (8) under the same $K=10$ sampling; if the relative-hypervolume advantage over PFES disappears or reverses, the reported gains come from the estimator's prior, not from the variational mixture itself.","tokens_in":41531,"feed_emoji":"🎯","tokens_out":7057,"duration_ms":59164,"temperature":0.7,"pith_summary":"Multi-objective Bayesian optimization needs to score how much a candidate evaluation would reveal about the unknown Pareto frontier, a quantity that requires the distribution of objective values conditioned on that frontier. Since the true frontier is a continuous set, the paper replaces the impossible exact truncation with two computable approximations: an over-truncation, which assumes the frontier is only a finite sampled set, and an under-truncation, which removes only points that dominate that finite set. The proposed acquisition function, PFEV, mixes the two truncated distributions and optimizes the mixing weight so that the variational lower bound on the mutual information is maximized, which is equivalent to minimizing the average KL divergence to the true conditioned distribution. The paper claims this yields higher or comparable hypervolume than existing methods, with the clearest gains when there are three or more objectives, where the two truncations differ most.","feed_headline":"Blend under- and over-truncation to beat Pareto-search baselines","feed_subtitle":"PFEV tunes the blend by maximizing a mutual-information lower bound, gaining hypervolume when L >= 3.","key_machinery":"The load-bearing object is the variational mixture $q_\\lambda(\\mathbf{f}(\\mathbf{x})\\mid\\mathcal{F}^*_S) = \\lambda q_U + (1-\\lambda) q_O$, where $q_U$ is the under-truncated normal (PFTN-U) and $q_O$ the over-truncated normal (PFTN-O) built from a finite sampled subset $\\mathcal{F}^*_S$ of the Pareto frontier. Its role is to define the lower bound $\\mathrm{LB}(\\mathbf{x},\\lambda)=\\mathbb{E}[\\log((\\lambda/Z_U+(1-\\lambda)/Z_O)\\mathbb{I}(\\mathbf{f}\\in A^O)+(\\lambda/Z_U)\\mathbb{I}(\\mathbf{f}\\in A^{U\\setminus O}))]$, whose maximization over $\\lambda$ is equivalent, by Remark 3.1, to minimizing the expected KL divergence from the true truncated distribution. The machinery also includes the variational distribution condition $\\mathrm{supp}(q) \\supseteq \\mathrm{supp}(p)$, which forces $\\lambda>0$, and a MAP Monte-Carlo estimator (Eq. 8) that replaces the indicator with $\\theta_{\\mathrm{MAP}}=(\\hat{p}+\\mathbb{I})/2$ to reduce variance.","core_discovery":"The central claim is that the mutual information $\\mathrm{MI}(\\mathbf{f}(\\mathbf{x});\\mathcal{F}^*)$ can be lower-bounded by a term that uses a mixture $q_\\lambda = \\lambda q_U + (1-\\lambda) q_O$ of under- and over-truncated Pareto-frontier truncated normals, and that choosing $\\lambda$ to maximize this lower bound at each candidate $\\mathbf{x}$ is a sound acquisition rule for multi-objective Bayesian optimization. Because the bound's gap equals the expected KL divergence between the true truncated posterior $p(\\mathbf{f}(\\mathbf{x})\\mid\\mathcal{F}^*)$ and the mixture, optimizing $\\lambda$ directly minimizes approximation error, unlike prior over-truncation-only approaches whose truncation is always too strong. The paper further shows the optimized bound is positive and bounded below by a probability-of-improvement term, and it introduces a bias-corrected Monte-Carlo estimator that pools the indicator with a normalizing-constant ratio to cut variance. The empirical claim is that on GP-generated functions, benchmarks, and LightGBM class-weight optimization, PFEV matches or outperforms PFES, $\\{\\mathrm{PF}\\}^2$ES, JES, EHVI, and other baselines, with the advantage concentrated at $L \\geq 3$ objectives.","pith_inferences":["If the mixture family is too restrictive, a single scalar $\\lambda$ cannot represent the shape mismatch between over- and under-truncation, so the lower bound may remain loose even at the optimal $\\lambda$; richer variational families with more parameters would be a natural next test.","The observed advantage at $L \\geq 3$ likely tracks the geometric gap between $A^O$ and $A^U$, and the paper's Appendix L suggests that gap only closes with exponentially many frontier points, so gains should persist on many-objective problems but may depend on frontier shape.","The MAP estimator's bias from $\\hat{p}$ could be eliminated at larger $K$ by decaying $r$ with sample size (as the paper notes), and a fair fixed-compute comparison should count the extra hypervolume-decomposition evaluations PFEV spends on the under-truncation region."],"forward_implications":["Selecting points by PFEV should yield higher or comparable relative hypervolume than PFES, {PF}2ES, JES, EHVI, and random search, with the margin largest when the number of objectives $L \\geq 3$.","The mixture weight $\\lambda$ can be optimized per candidate and per iteration without heuristics, so the acquisition function adapts to how much under- or over-truncation the current posterior suggests.","Parallel querying, decoupled observations, noisy observations, and a joint-entropy variant all inherit the same lower-bound construction, extending the method beyond the basic sequential setting.","The lower bound is provably positive and bounded below by the probability of improvement, so the acquisition score carries a built-in improvement signal rather than only an information signal.","Because the bound's tightness is the expected KL divergence to the true frontier-conditioned distribution, any future distribution family closer to that target would tighten the bound and potentially improve search."],"supporting_citations":[{"why":"Introduces PFES and the Pareto-frontier truncated normal (PFTN) used to define over-truncation; PFEV extends this MI-based search.","marker":"Suzuki et al., 2020"},{"why":"Supplies the variational lower-bound identity for mutual information that PFEV adopts and optimizes.","marker":"Takeno et al., 2022"},{"why":"Defines {PF}2ES, an over-truncation-based baseline whose heuristic shift PFEV replaces with a principled lambda.","marker":"Qing et al., 2022"},{"why":"Defines JES, the joint-entropy baseline that also relies only on over-truncation and is compared in the experiments.","marker":"Tu et al., 2022"},{"why":"Provides max-value entropy search and the K=10 Monte-Carlo sampling convention PFEV inherits.","marker":"Wang and Jegelka, 2017"},{"why":"Provides random-feature approximations of GP sample paths used to generate F*_S samples.","marker":"Rahimi and Recht, 2008"},{"why":"NSGA-II is the MOO solver that produces the finite sampled Pareto frontier F*_S for each MC sample.","marker":"Deb et al., 2002"},{"why":"Quick hypervolume supplies the cell decomposition used to compute the normalizing constants Z_U and Z_O.","marker":"Russo and Francisco, 2014"}],"fun_headline_variants":["Mix truncations to lower-bound Pareto information gain","Optimize truncation mix via variational bound for MOBO","Pareto search with variational-tuned truncation mixture","Trade truncations to maximize Pareto-frontier mutual information","Variational bound picks truncation blend for Pareto MI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The main experiments use the biased Monte-Carlo estimator (8) that replaces the indicator of the over-truncated region with a prior-smoothed probability estimate, so the method's edge rests on that bias not distorting which candidate points are ranked best.","fun_headline_variants_meta":{"raw":{"variants":["Mix truncations to lower-bound Pareto information gain","Optimize truncation mix via variational bound for MOBO","Pareto search with variational-tuned truncation mixture","Trade truncations to maximize Pareto-frontier mutual information","Variational bound picks truncation blend for Pareto MI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000979,"raw_usage":{"total_tokens":4174,"prompt_tokens":976,"completion_tokens":3198,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":3121}},"tokens_in":592,"tokens_out":3198,"duration_ms":22568,"temperature":1.0,"reasoning_tokens":3121,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T21:26:03.507701+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run PFEV on the $L \\geq 3$ GP-derived synthetic problems with the unbiased estimator (6) instead of the MAP estimator (8) under the same $K=10$ sampling; if the relative-hypervolume advantage over PFES disappears or reverses, the reported gains come from the estimator's prior, not from the variational mixture itself.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces PFES and the Pareto-frontier truncated normal (PFTN) used to define over-truncation; PFEV extends this MI-based search."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the variational lower-bound identity for mutual information that PFEV adopts and optimizes."},{"cited_title":"$\\{\\text{PF}\\}^2$ES: Parallel Feasible Pareto Frontier Entropy Search for Multi-Objective Bayesian Optimization","cited_arxiv_id":"2204.05411","evidence_quote":"Defines {PF}2ES, an over-truncation-based baseline whose heuristic shift PFEV replaces with a principled lambda."},{"cited_title":"and Jegelka, S","cited_arxiv_id":null,"evidence_quote":"Provides max-value entropy search and the K=10 Monte-Carlo sampling convention PFEV inherits."},{"cited_title":"and Recht, B","cited_arxiv_id":null,"evidence_quote":"Provides random-feature approximations of GP sample paths used to generate F*_S samples."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"NSGA-II is the MOO solver that produces the finite sampled Pareto frontier F*_S for each MC sample."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Quick hypervolume supplies the cell decomposition used to compute the normalizing constants Z_U and Z_O."}],"review_version":1}