{"id":"de383697-b8bc-43eb-a0ff-95a3c517a121","arxiv_id":"2504.20307","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FigBO augments any myopic acquisition function with a decaying global-information-gain term and claims faster convergence with an unchanged asymptotic rate.","lead":"FigBO adds a decaying exploration bonus, the expected reduction in global posterior variance, to standard Bayesian optimization acquisition rules like expected improvement. The paper claims this tiny extra term yields faster convergence and state-of-the-art results while preserving the convergence rate of the base rule.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Corollary 1's asymptotic-equivalence proof is invalid: the limit in Eq. B9 drops a denominator that also vanishes, and for smooth kernels in d=1 the Γ-term is same order as the EI rate.","rationale":"The reader's weakest assumption already identifies the unstated rate condition and MC fidelity. My independent reading confirms the rate condition is real and finds a more direct defect: Corollary 1's proof in Eq. B9 performs the limit only on the numerator while the denominator tends to zero, so the proof is incomplete regardless of kernel smoothness; the condition (ν∧1)/d < 1 is what would rescue it, but it is absent. For d=1 smooth kernels this condition fails exactly, making the extra term same order as the EI regret. This is load-bearing because the central theoretical claim is that adding Γ does not harm convergence. The empirical results are plausible and the framework may still be a useful heuristic; the concern is about the stated asymptotic equivalence, not about all practical value. I therefore keep the reader's conditional verdict rather than moving to reject: the paper can be accepted if the theory is corrected to a rate-only statement (O(n^{-(ν∧1)/d}(log n)^β) still likely holds) or the condition is stated and the limit recomputed. The MC approximation issue is secondary and could be addressed by an error bound or larger L, but it does not change the verdict.","tokens_in":13100,"tokens_out":5227,"duration_ms":58225,"concrete_test":"Re-derive the limit in Appendix B with m = floor(n/3)+1 and keep the denominator: compute lim_{n→∞} [C0/n] / [Θ(n^{-(ν∧1)/d}(log n)^β) + C0/n], where C0 = τ(R/σ)/τ(-R/σ) η max_x Γ(x). For d=1, ν>1, this limit is a positive constant unless the proof supplies additional cancellation, so Eq. B9's conclusion fails; if a correctly specified scaling condition produces 0, the theorem survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The framework's theoretical guarantee is Corollary 1, but its proof (Appendix B, Eq. B9) is not valid as written. The ratio is upper-bounded by 1 + C′/(Bull bound + C′), and the argument then lets only the numerator C′ = τ(R/σ)/τ(-R/σ) (η/n) max_x Γ(x) go to zero. The denominator also tends to zero because the Bull bound is O(n^{-(ν∧1)/d}(log n)^β); the resulting limit is 0/0 and the conclusion '=1' does not follow. Making the comparison valid requires C′/L_n(EI_n) → 0, i.e. η/n divided by n^{-(ν∧1)/d}(log n)^β → 0, or (ν∧1)/d < 1. This condition is never stated. For Matérn ν>1 or RBF kernel in d=1, (ν∧1)/d = 1 and the extra term is O(1/n), the same order as Bull's EI bound, so Γ can change the leading constant; asymptotic equivalence is unsupported. The MC approximation of Γ with L=100 (Eq. 9) is an additional gap between theory and practice, but the logical flaw in B9 is the more load-bearing issue because the claimed 'no-harm' guarantee rests entirely on it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FigBO, a framework that augments any myopic acquisition function with a look-ahead term: the next query is selected as argmax_x [α(x) + λΓ(x)], where Γ(x) is the integrated reduction of the GP posterior variance over the input space when x is added, λ = η/n is a decaying weight, and the integral is approximated by L uniform Monte Carlo samples. The theoretical section focuses on the case α = expected improvement (EI). Theorem 1 gives an upper bound on the expected regret of the resulting strategy EI_{Γ,n} in terms of Bull's EI bound plus an additive term C' = O(η/n). Corollary 1 claims that as n → ∞, the loss of EI_{Γ,n} is asymptotically equivalent to that of vanilla EI, yielding the rate O(n^{-(ν∧1)/d}(log n)^β). Empirically, the authors test FigBO on GP prior samples, synthetic functions (Branin, Levy, Hartmann), and MLP hyperparameter-tuning tasks, reporting improved average log regret or accuracy relative to myopic and non-myopic baselines, and they include ablations over the MC sample count L and the decay hyperparameter η.","tokens_in":13443,"tokens_out":7448,"duration_ms":71981,"significance":"If the theoretical claim held, FigBO would be an attractive and conceptually simple plug-and-play upgrade to most myopic acquisition functions, with a proof that the extra exploration term does not degrade the asymptotic convergence rate while providing empirical speedups. The idea of adding a decaying, information-based exploration bonus is plausible, and the empirical results, while not accompanied by formal significance tests, are consistent across a range of benchmarks and show a clear advantage over the myopic baselines. The method is considerably cheaper than non-myopic alternatives such as PES. However, the central 'no-harm' guarantee rests on a limit argument that is invalid as written, and the theory applies only to an exact-Γ version while the implementation uses a Monte Carlo approximation with no error control. These issues make the main theoretical contribution unsupported in its current form and need to be fixed before the manuscript can be recommended for acceptance.","major_comments":[{"comment":"The proof of Corollary 1 is not valid as written. The ratio in (B9) is bounded by 1 + C'/(Bull_bound + C'), where C' = (τ(R/σ)/τ(−R/σ)) max_x λΓ(x) and Bull_bound = (τ(R/σ)/τ(−R/σ))[2Rm^{-1} + (R+σ)C m^{-(ν∧1)/d}(log m)^β]. The argument lets only the numerator C' → 0, but the denominator also tends to 0 because Bull_bound → 0 as n → ∞. Consequently the limit is of the form 0/0 and the conclusion '= 1' does not follow. A valid comparison requires C'/Bull_bound → 0, i.e., (η/n)/n^{-(ν∧1)/d}(log n)^β → 0, which is equivalent to (ν∧1)/d < 1 (up to logarithmic factors). This condition is never stated. For Matérn kernels with ν > 1 or the RBF kernel in dimension d = 1, (ν∧1)/d = 1 and the Γ term is of the same order as Bull's EI bound, so the leading constant of the regret can be affected. As stated, Corollary 1 is unsupported.","section":"Appendix B, Eq. (B9); Corollary 1"},{"comment":"The theoretical results (Theorem 1 and Corollary 1) are stated for the exact global-information integral Γ(x), but the deployed algorithm uses the Monte Carlo approximation (9) with L = 100 uniform samples. No bound on the approximation error is given, and the analysis does not show that the error is small enough for the theoretical guarantee to carry over to the implemented acquisition function. This is a load-bearing gap between theory and practice: even if the exact-Γ analysis were correct, it would not justify the empirical behavior of the algorithm that actually selects points using the MC estimate. The authors should either prove a uniform error bound for the MC approximation (e.g., under regularity conditions on the kernel) or explicitly restrict the theoretical claims to the exact version and discuss the approximation as an additional heuristic assumption.","section":"Section 3.2, Eq. (9) and Algorithm 1"}],"minor_comments":[{"comment":"In the first sentence of Section 5.5, 'n low-dimensional tasks' appears to be a typo for 'in low-dimensional tasks'.","section":"Section 5.5"},{"comment":"The notation in Eqs. (7)–(9) is confusing: the integration variable in (7) is also denoted x, while in (9) the candidate point is x and the MC samples are indexed by l. This makes it hard to see that k_{n,l}(x) is the cross-covariance between the l-th MC point and the n+1 points including the candidate. Please clarify the notation, e.g., by renaming the integration variable and defining k_{n,l}(x) explicitly.","section":"Eqs. (7)–(9)"},{"comment":"The paper claims that FigBO 'seamlessly integrates with most myopic acquisition functions,' but the theoretical analysis covers only EI. The empirical section shows that the framework also helps UCB and PI, but there is no theoretical justification for these cases. The authors should state this scope limitation more clearly.","section":"Introduction and Related Work"},{"comment":"The empirical evaluation reports average results with standard errors over 20 or 50 repetitions, but no statistical significance tests (e.g., paired t-tests or Wilcoxon tests) are provided. Adding such tests would strengthen the claim that FigBO's improvements are not due to random variation. In addition, the code is not released for review, which limits reproducibility of the results.","section":"Section 5"},{"comment":"Figure 6 shows that the performance of FigBO in the 6D Hartmann task improves substantially when L increases from 5 to 500. This is an important practical consideration, since the default L=100 may be insufficient for higher-dimensional problems; the paper would benefit from a concrete guidance on choosing L as a function of dimension.","section":"Section 5.5"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the validity of Corollary 1. The limit argument in Eq. (B9) is genuinely flawed, and the missing condition (ν∧1)/d < 1 is needed for the claimed asymptotic equivalence. This is not a presentation issue but a technical gap in the main theoretical contribution. The authors can likely fix it by adding the condition and either weakening the claim to an upper bound of the same asymptotic order or proving the equivalence under the condition; however, as written, the manuscript's central 'does not harm the convergence rate' statement is not established for common kernels in d=1. The empirical section is reasonable but would benefit from significance tests and code release. Overall, I would support a major revision and a careful re-review of the theory."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes FigBO: add a decaying global-variance-reduction term Γ(x) to any myopic acquisition function, with weight λ=η/n. The idea is simple and useful, and the experiments suggest it helps early convergence without hurting later performance. The Γ term itself is essentially the integrated mean squared error / integrated variance reduction criterion from spatial statistics and active learning, so the novelty is the plug-and-play combination with a decay schedule—modest, but real. I like the Sherman-Morrison trick to keep the added cost low.\n\nThe soft spot is the theory, and it's load-bearing. Corollary 1 claims asymptotic equivalence between FigBO-EI and vanilla EI, i.e., no harm to the convergence rate. The proof in Appendix B (Eq. B9) is not valid: it bounds the ratio by 1 + C′/(Bull bound + C′), then lets C′ → 0 because λ → 0. But the denominator also goes to zero because the Bull bound is O(n^{-(ν∧1)/d}(log n)^β). The limit is 0/0 unless C′/L_n(EI_n) → 0, which requires (ν∧1)/d < 1. That condition is never stated, and it fails for Matérn ν > 1 or RBF kernels in one dimension—where the added term is the same order as the EI rate and can change the leading constant. So the headline 'asymptotically equivalent' is unsupported in exactly the smooth-kernel low-D setting that matters most. This is a genuine flaw, not a nitpick.\n\nThere's also the MC approximation gap: the theory is for exact Γ(x), but the algorithm uses L=100 uniform samples, and no error analysis connects the two. That may be minor if L is large enough, but it's unaddressed. Empirically, the results look plausible—FigBO often converges faster—but there's no released code and no significance tests, only standard errors. For a paper claiming 'significantly faster', I'd want paired comparisons or confidence intervals.\n\nWho is this for? Practitioners who want a cheap way to boost a myopic acquisition function. The empirical phenomenon may survive even if the theory is cut back to a conditional statement. The paper deserves a serious referee, but the B9 issue should be flagged as a major revision: either state the (ν∧1)/d < 1 condition and restrict the claim, or prove a weaker no-harm result that doesn't drop a vanishing denominator. I'd send it to peer review, not desk reject, but the authors need to fix the math before publication.","headline":"The plug-and-play idea is reasonable, but the main no-harm convergence claim rests on an invalid limit argument that fails exactly in the smooth-1D regime; the paper is worth refereeing but needs substantial theory revision.","tokens_in":13912,"tokens_out":2233,"would_cite":false,"duration_ms":26769,"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":"A decaying global-information term adds look-ahead to myopic acquisition functions without changing EI's convergence rate.","keywords":["Bayesian optimization","acquisition function","look-ahead","global information gain","expected improvement","Gaussian process","regret bound","Monte Carlo approximation"],"falsifier":"Recompute the proof with the exact integral in place of the Monte Carlo approximation and check whether $C' = \\frac{\\tau(R/\\sigma)}{\\tau(-R/\\sigma)}\\max_{x\\in\\mathcal{X}}\\lambda\\Gamma(x)$ divided by $L_n(EI_n)$ tends to 0 as $n\\to\\infty$ under $\\lambda=\\eta/n$; if it does not, the asymptotic equivalence in Corollary 1 fails. A direct experiment: run FigBO and vanilla EI on a one-dimensional squared-exponential GP sample for many iterations and compare the tail slope of log regret; a worse slope for FigBO would contradict the claimed rate.","tokens_in":12887,"feed_emoji":"🎯","tokens_out":8463,"duration_ms":82253,"temperature":0.7,"pith_summary":"This paper tries to show that a cheap additive term can give myopic Bayesian optimization acquisition functions a look-ahead capability. Concretely, it proposes choosing queries by maximizing α(x) + λΓ(x), where α is any myopic acquisition function and Γ measures how much a candidate point would reduce the Gaussian process's predictive variance over the whole input space, with λ = η/n decaying to zero. For expected improvement, the paper proves that this modified policy has the same asymptotic regret rate as vanilla EI, so the extra exploration does not hurt convergence. If true, practitioners could upgrade almost any existing acquisition function without replacing their surrogate model or paying the full cost of non-myopic methods. The paper also reports faster convergence on GP prior samples, synthetic functions, and MLP hyperparameter tuning tasks.","feed_headline":"A decaying uncertainty term gives myopic optimizers a look-ahead","feed_subtitle":"FigBO adds a global-information term to any acquisition function and keeps EI's convergence rate.","key_machinery":"The central object is the global information-gain function $\\Gamma(x)$, the integrated reduction in posterior variance over the whole input space that would result from adding candidate point $x$ to the data. It is approximated by Monte Carlo as $\\Gamma(x) \\approx \\frac{1}{L}\\sum_{l=1}^L k_{n,l}(x)^\\top (K_{n,l}+\\sigma_\\epsilon^2 I)^{-1} k_{n,l}(x)$, and it carries the look-ahead: maximizing $\\alpha(x)+\\lambda\\Gamma(x)$ with $\\lambda=\\eta/n$ spends early iterations reducing global uncertainty, then reverts to the base acquisition function. The proof machinery is the regret analysis for the vanilla EI convergence rate, and the computational machinery is a rank-one update of the inverse Gram matrix that keeps the per-candidate cost at $O(n^2)$.","core_discovery":"The paper's central claim is that the query rule $x_{n+1} \\in \\arg\\max_{x\\in\\mathcal{X}} [\\alpha(x) + \\lambda\\Gamma(x)]$, with $\\lambda = \\eta/n$ and $\\Gamma$ the global posterior-variance reduction integral, endows any myopic acquisition function with look-ahead capability. For expected improvement, Theorem 1 bounds the expected regret of the combined strategy by the regret of vanilla EI plus a constant $C'$, and Corollary 1 states the two are asymptotically equivalent, giving the same rate $O(n^{-(\\nu\\wedge 1)/d}(\\log n)^\\beta)$. Empirically, the paper reports that FigBO-EI, FigBO-UCB, and FigBO-PI converge faster and reach lower log regret than their base functions and than entropy-search baselines on GP prior samples, Branin, Levy, Hartmann, and four MLP classification tasks.","pith_inferences":["A batched version could compute $\\Gamma$ over a whole candidate set rather than one point; the paper does not explore this.","Since the 12D experiments show a smaller advantage, dimension-aware Monte Carlo sampling that grows $L$ with dimension is a natural extension the paper does not test.","The asymptotic equivalence is proved for EI only; the UCB and PI speedups are empirical, and extending the proof to those base functions is an open question."],"forward_implications":["Any myopic acquisition function can be upgraded by adding the decaying term $\\lambda\\Gamma(x)$, and in the EI case the combined policy keeps the same asymptotic convergence rate as vanilla EI.","Early in the run, the global-information term dominates and steers queries toward high-uncertainty regions; later $\\lambda \\to 0$, so the policy reverts to the base acquisition function.","The extra computational cost is mild: a rank-one update keeps the per-candidate evaluation of $\\Gamma$ at $O(n^2)$, far cheaper than entropy-search methods.","The reported experiments show faster convergence and lower log regret than the base myopic functions and non-myopic baselines on GP prior samples, synthetic benchmarks, and MLP classification tasks."],"supporting_citations":[{"why":"Supplies the vanilla EI convergence rate and the lemmas that the FigBO-EI regret bound extends.","marker":"[19]"},{"why":"Provides the assumptions and scale-invariance treatment under which the regret analysis is carried out.","marker":"[5]"},{"why":"Defines expected improvement, the base acquisition function that FigBO upgrades and analyzes.","marker":"[14]"},{"why":"Motivates information gain as a measure of uncertainty reduction and supplies the UCB baseline used in experiments.","marker":"[15]"},{"why":"Supplies the rank-one inverse update that keeps the per-candidate cost of computing $\\Gamma$ at $O(n^2)$.","marker":"[33]"},{"why":"Provides the GP prior sample experimental setup and the joint entropy search baseline that FigBO is compared against.","marker":"[18]"}],"fun_headline_variants":["A decaying global term adds look-ahead to any acquisition function","Plug-and-play look-ahead for Bayesian optimization","Look-ahead for myopic BO with guaranteed convergence rate","FigBO: a single term that gives BO foresight","Add a global-information term, keep EI's regret bound"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the fixed-sample Monte Carlo estimate of the global uncertainty integral being a faithful stand-in for the exact integral, and on the decaying weight $\\lambda=\\eta/n$ shrinking fast enough that the added term becomes negligible next to EI's regret; for smooth kernels in one dimension the required decay condition can fail.","fun_headline_variants_meta":{"raw":{"variants":["A decaying global term adds look-ahead to any acquisition function","Plug-and-play look-ahead for Bayesian optimization","Look-ahead for myopic BO with guaranteed convergence rate","FigBO: a single term that gives BO foresight","Add a global-information term, keep EI's regret bound"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000821,"raw_usage":{"total_tokens":3563,"prompt_tokens":888,"completion_tokens":2675,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":2594}},"tokens_in":504,"tokens_out":2675,"duration_ms":20563,"temperature":1.0,"reasoning_tokens":2594,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:33:37.629425+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the proof with the exact integral in place of the Monte Carlo approximation and check whether $C' = \\frac{\\tau(R/\\sigma)}{\\tau(-R/\\sigma)}\\max_{x\\in\\mathcal{X}}\\lambda\\Gamma(x)$ divided by $L_n(EI_n)$ tends to 0 as $n\\to\\infty$ under $\\lambda=\\eta/n$; if it does not, the asymptotic equivalence in Corollary 1 fails. A direct experiment: run FigBO and vanilla EI on a one-dimensional squared-exponential GP sample for many iterations and compare the tail slope of log regret; a worse slope for FigBO would contradict the claimed rate.","supporting_citations":[{"cited_title":"Journal of Machine Learning Research 12(10) (2011)","cited_arxiv_id":null,"evidence_quote":"Supplies the vanilla EI convergence rate and the lemmas that the FigBO-EI regret bound extends."},{"cited_title":"Journal of Global Optimization 13, 455–492 (1998) 18","cited_arxiv_id":null,"evidence_quote":"Defines expected improvement, the base acquisition function that FigBO upgrades and analyzes."},{"cited_title":"IEEE Transactions on Information Theory 58(5), 3250–3265 (2012)","cited_arxiv_id":null,"evidence_quote":"Motivates information gain as a measure of uncertainty reduction and supplies the UCB baseline used in experiments."},{"cited_title":"The Annals of Mathematical Statistics 21(1), 124–127 (1950)","cited_arxiv_id":null,"evidence_quote":"Supplies the rank-one inverse update that keeps the per-candidate cost of computing $\\Gamma$ at $O(n^2)$."},{"cited_title":"Advances in Neural Information Processing Systems 35, 11494–11506 (2022)","cited_arxiv_id":null,"evidence_quote":"Provides the GP prior sample experimental setup and the joint entropy search baseline that FigBO is compared against."}],"review_version":1}