{"id":"5840488b-b811-4d37-a6bb-2095275ae9e6","arxiv_id":"2607.10793","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":6,"one_line_summary":"HBQ adaptively partitions the domain into local stationary GPs, recombines leaf integrals by hierarchical conditioning, and beats standard BQ on nonstationary integrands while matching it on stationary ones.","lead":"Hierarchical Bayesian Quadrature splits an integration domain into local stationary Gaussian-process models and recombines their integral estimates through a tree of GP conditioning. It gives a practical way to spend evaluation budget where nonstationary integrands are hard, with quantified uncertainty and no MCMC.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged caveats.","rationale":"The reader's CONDITIONAL verdict already weights the partial theory, geometric restrictions, and weak tree-vs-independent distinction correctly. The experiments (Figs. 3, 6, 1; structure recovery 100%/96%; Supp. C.1 d=3 and C.4 BIC sensitivity) give direct empirical support for the strongest claim under the stated restrictions; the method reduces to ordinary BQ when BIC rejects splits. No additional correctness risk (e.g., inconsistency in Eqs. 16–17 or unacknowledged bias from the log-normal child prior) rises to the level of overturning that assessment. Hence the stress-test leaves the verdict and confidence unchanged.","tokens_in":19232,"tokens_out":569,"duration_ms":6183,"concrete_test":"Re-run the d=2 Genz Corner Peak and Continuous families (C=9, 20 trials, N up to 512) with an ablated HBQ that forces pure independent-sum recombination and freezes kernel order to Matérn-3/2 (no per-leaf ν search), and report median |I−Î| vs full HBQ and vs BQ; if the nonstationary gain on Corner Peak collapses below ~2× while Continuous remains matched, the hierarchical/BIC machinery is load-bearing rather than incidental.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (substantial gains on nonstationary integrands, parity on stationary ones via adaptive local stationary GPs + hierarchical recombination + BIC control) is supported by the Genz suite (esp. Corner Peak), the reaction–diffusion PDE, the SIR evidence ridge, and the on-model piecewise-Matérn recovery experiment, with code and candid limitations. The reader's weakest assumption correctly isolates the softest points: axis-aligned hyperrectangles + closed-form tensor-product Matérn/Wendland embeddings (§3.1), the 5d min-points rule and BIC gate (§4.1), and the deliberately partial theory (Thm. 1 / Cor. 2 are conditional on the final tree and kernels; adaptive construction and hyperparameter adaptation are not covered; §7.1). Those are real scope limits, not internal contradictions. The empirical near-indistinguishability of tree conditioning vs independent sum (Supp. C.3) further softens the hierarchical-recombination novelty claim, but does not overturn the adaptive-partitioning gains. No stronger load-bearing flaw (e.g., broken recombination formulas, mis-specified BIC asymptotics that would systematically over-split, or experimental confounds that erase the nonstationary advantage) is present in the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Hierarchical Bayesian Quadrature (HBQ): an adaptive tree of axis-aligned hyperrectangular subdomains, each carrying a local stationary Matérn (or Wendland) GP with its own lengthscale and smoothness, whose leaf integral estimates are recombined bottom-up by hierarchical GP conditioning that reintroduces parent-kernel cross-subdomain correlations. Tree growth is gated by a BIC approximation to the split Bayes factor (with a 5d minimum points-per-leaf rule), and refinement is driven by a batched integral-variance-reduction score. Theorem 1 / Corollary 2 bound the parent integral error by the worst-case child error under quasi-uniform local designs. Experiments on Genz families (d=2 and d=3), an SIR model-evidence ridge, an on-model piecewise Matérn recovery task with calibration, and a multi-scale reaction–diffusion PDE show substantial gains over standard BQ on nonstationary integrands and parity on stationary ones, with a public Julia implementation.","tokens_in":19629,"tokens_out":1093,"duration_ms":13702,"significance":"If the empirical claims hold, HBQ is a practical Bayesian counterpart to Genz–Malik adaptive cubature that preserves closed-form kernel mean embeddings and calibrated integral uncertainty while automatically allocating budget to local complexity. The combination of BIC-gated splitting, hierarchical recombination of integral functionals, and the partial but clean recombination theory is a genuine contribution to probabilistic numerics for expensive low-to-moderate-dimensional nonstationary integrands (model evidences, multi-scale PDE/simulator responses). Strengths include reproducible code, multiple controlled baselines (MC/QMC/BQ), an on-model structure-recovery experiment with z-score calibration, and an explicit limitations section that correctly flags the partial theory and axis-aligned restriction.","major_comments":[{"comment":"§3.2 Eqs. (15)–(17) and Supp. C.3: tree conditioning is presented as the primary recombination mechanism that reintroduces cross-subdomain correlations, yet on the Genz families the independent-sum ablation matches both median absolute error and posterior standard deviation to within 0.05% at every budget. The hierarchical novelty claim therefore rests almost entirely on adaptive partitioning rather than on the recombination hierarchy. Either demonstrate a setting (e.g., the piecewise Matérn or PDE) where tree conditioning materially improves calibration or accuracy, or reframe the contribution so that independent sum is the default and tree conditioning is optional.","section":"§3.2 / Supp. C.3"},{"comment":"§3.3 Theorem 1 / Corollary 2 and §7.1: the convergence statements are deliberately conditional on a fixed final tree and fixed local kernels; they do not cover the adaptive construction, BIC decisions, or hyperparameter re-fitting that constitute the algorithm. The central claim that HBQ “achieves substantial gains … while matching … on stationary ones” is therefore supported only empirically. A short discussion of what would be required for a full adaptive rate (or an explicit statement that the theory is only a recombination sanity check) would make the theoretical contribution proportionate to what is proved.","section":"§3.3 / §7.1"}],"minor_comments":[{"comment":"§4.1 / Eq. (23): the BIC penalty uses Δk = 3 (shared lengthscale) or d+2 (ARD) and a hard 5d minimum; Supp. C.4 shows robustness on the piecewise task, but a one-sentence justification for the 5d rule (beyond “reliable hyperparameter estimation”) would help readers who wish to retune.","section":"§4.1"},{"comment":"Fig. 1(c) and §6.2: SIR results use independent-sum recombination while the PDE and Genz main figures use tree conditioning; a brief note in the caption or text would avoid confusion about which recombination is active.","section":"Fig. 1 / §6.2"},{"comment":"§5 Related work: Nested kriging and BART-Int are cited appropriately; a short clause distinguishing HBQ’s closed-form integral functionals from Nested kriging’s second-level GP aggregation would sharpen the positioning.","section":"§5"},{"comment":"Typographical: arXiv header and abstract both use “model evidence computations”; ensure consistent hyphenation (“model-evidence”) and that the Genz–Malik citation appears with the same spelling throughout.","section":"Abstract / §5"}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is a solid UAI-level contribution. The only soft spot is that Supp. C.3 undercuts the hierarchical-recombination novelty more than the main text acknowledges; once that is reframed, the paper is ready. Scope (axis-aligned, closed-form embeddings, low-to-moderate d) is already candidly stated and matches the journal’s probabilistic-numerics audience."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean engineering paper in probabilistic numerics. The punchline is right: on nonstationary integrands HBQ beats standard BQ by adaptive axis-aligned partitioning into local stationary Matérn GPs, with BIC-gated splits and variance-reduction refinement; on stationary ones it matches BQ. That claim is backed by Genz (especially Corner Peak, also d=3), a reaction–diffusion PDE, an SIR evidence ridge, and an on-model piecewise-Matérn recovery study with calibration. Code is linked, limitations are stated honestly, and there is no MCMC.\n\nWhat is actually new is the combination for integration: closed-form subdomain BQ under tensor-product half-integer Matérn/Wendland kernels, the adaptive loop that chooses split vs refine, and the tree-conditioning recombination of integral functionals. Nested kriging, treed GPs, BIC splits, and Genz–Malik are prior art; the paper is explicit about that. Theorem 1 / Corollary 2 give a clean bound: parent error is controlled by the worst child, conditional on the final tree and kernels. That is useful framing even if incomplete.\n\nSoft spots, in proportion. Theory does not cover adaptive tree growth or hyperparameter adaptation—the authors say so. Axis-aligned hyperrectangles plus closed-form embeddings keep everything tractable but restrict geometry; that is a scope choice, not a hidden flaw. The supplement shows tree conditioning and independent sum are essentially identical on the Genz families they tried, so the hierarchical-recombination novelty is thinner than the abstract suggests; the real win is adaptive budget allocation. Free parameters (BIC penalty, 5d min points, lengthscale priors, Matérn candidates) are standard GP knobs and they sensitivity-check BIC. Citations look fair.\n\nWho it is for: people doing expensive low-to-moderate-d model evidence or multi-scale simulators who already use BQ or Genz–Malik and want uncertainty with adaptive effort. A serious referee should see it. I would engage: cite when I need nonstationary BQ, and bring it to reading group for the adaptive design and the recombination idea.","headline":"Solid adaptive BQ for nonstationary integrands: real empirical gains, usable algorithm, partial theory and soft hierarchical novelty.","tokens_in":20251,"tokens_out":531,"would_cite":true,"duration_ms":8124,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Hierarchical Bayesian Quadrature adapts stationary GP models to nonstationary integrands by growing a tree of local surrogates and recombining their integrals with hierarchical conditioning.","keywords":["Bayesian quadrature","Gaussian processes","nonstationary integrands","adaptive domain partitioning","hierarchical conditioning","model evidence","numerical integration"],"falsifier":"Run the method on a nonstationary integrand whose iso-complexity contours are strongly diagonal or curved (so that axis-aligned rectangles cannot isolate them) and check whether the adaptive tree still reduces integral error relative to a single stationary GP at matched evaluation budgets; if it does not, the partition geometry is insufficient.","tokens_in":20099,"feed_emoji":"🌳","tokens_out":652,"duration_ms":8758,"temperature":0.7,"pith_summary":"Standard Bayesian quadrature puts a single stationary Gaussian process prior on an integrand and turns the posterior into a closed-form integral estimate with uncertainty. That prior is misspecified when the integrand’s roughness or length scale changes across the domain, so evaluations are wasted on flat regions while peaks or ridges stay unresolved. This paper shows that the same closed-form machinery can be kept if the domain is split, on the fly, into a binary tree of axis-aligned hyperrectangles, each carrying its own stationary Matérn GP. Local integral posteriors are recombined by treating each child estimate as a noisy observation of a linear functional of the parent GP, thereby restoring cross-boundary correlations. BIC model selection decides when a split is worth making, so the method collapses to ordinary Bayesian quadrature when a single stationary model already fits. On Genz benchmarks, an SIR model-evidence integral, and a reaction–diffusion PDE, the adaptive scheme concentrates evaluations where complexity is highest and reports substantially lower error on nonstationary targets while matching ordinary BQ on stationary ones.","feed_headline":"Tree of local GPs beats single-GP quadrature on nonstationary integrals","feed_subtitle":"Adaptive splits and hierarchical recombination cut error while matching ordinary BQ when the integrand is stationary.","key_machinery":"Tree conditioning: each branch node retains the stationary GP that was fitted before the split; the children’s integral estimates are treated as noisy observations of the corresponding sub-integrals under that parent prior, yielding a joint Gaussian update that reintroduces cross-subdomain correlations and propagates upward to the global integral.","core_discovery":"An adaptively grown tree of local stationary Gaussian processes, whose leaf integral estimates are recombined by hierarchical GP conditioning and whose growth is gated by a BIC split test, recovers the accuracy of ordinary Bayesian quadrature on stationary integrands and substantially improves it on nonstationary ones, all without MCMC and while preserving closed-form kernel mean embeddings.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Tree of local GPs recombines integrals to beat BQ on nonstationary problems","Adaptive GP partitions with hierarchical conditioning improve nonstationary quadrature","BIC-gated tree of stationary GPs matches BQ when stationary and beats it otherwise","Hierarchical local GP recombination adapts Bayesian quadrature without MCMC","Leafwise stationary GPs recombined by hierarchy cut error on nonstationary integrands"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That axis-aligned rectangular splits and stationary product kernels with closed-form mean embeddings are flexible enough to capture the nonstationarity that matters, and that the partial error bounds that assume a finished tree still justify the adaptive construction.","fun_headline_variants_meta":{"raw":{"variants":["Tree of local GPs recombines integrals to beat BQ on nonstationary problems","Adaptive GP partitions with hierarchical conditioning improve nonstationary quadrature","BIC-gated tree of stationary GPs matches BQ when stationary and beats it otherwise","Hierarchical local GP recombination adapts Bayesian quadrature without MCMC","Leafwise stationary GPs recombined by hierarchy cut error on nonstationary integrands"]},"model":"grok-4.5","effort":"low","cost_usd":0.007612,"raw_usage":{"total_tokens":1800,"prompt_tokens":701,"num_sources_used":0,"completion_tokens":98,"cost_in_usd_ticks":76120000,"prompt_tokens_details":{"text_tokens":701,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1001,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":701,"tokens_out":98,"duration_ms":11036,"temperature":1.0,"reasoning_tokens":1001,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T09:12:40.540808+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the method on a nonstationary integrand whose iso-complexity contours are strongly diagonal or curved (so that axis-aligned rectangles cannot isolate them) and check whether the adaptive tree still reduces integral error relative to a single stationary GP at matched evaluation budgets; if it does not, the partition geometry is insufficient.","supporting_citations":[],"review_version":1}