{"id":"2c376718-84b9-4c2f-932a-18d60c88c611","arxiv_id":"2607.23396","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Distributed and recursive INLA with data-and-latent-field partitioning recovers latent-field posteriors close to full-data inference while cutting cost for large spatio-temporal models.","lead":"The paper builds a practical Bayesian toolkit that splits big spatio-temporal models across machines or time steps using INLA, so posteriors can be updated without pooling raw data. It matters for privacy-sensitive, streaming, and memory-heavy environmental and policy analyses where full-data INLA is too costly.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The log-determinant surrogates behind latent-field partitioning (Eqs. 43–47) are uncontrolled zeroth-order truncations that bias the hyperparameter posterior even for Gaussian likelihoods; the exactness results cover only data partitioning.","rationale":"The reader's weakest assumption targets the non-Gaussian product approximation (Eq. 13) and recursive mode updates (Eqs. 35–38). That is a real soft spot and the paper is commendably honest about it (PM2.5 CCD divergence; explicit acknowledgment of recursive non-invariance under reordering). My concern lives in the same neighborhood — hyperparameter-layer accuracy — but identifies a different and in one way sharper mechanism: the determinant surrogates of §5 are approximations that operate even when the likelihood is Gaussian, so they break the exactness story for the latent-field-partitioning half of the strongest claim, not merely degrade the non-Gaussian half. The paper gives credit to exact algebra where it exists (Eq. 9), which makes the absence of any error control or empirical isolation for Eqs. 43–47 the least secure point in the argument. I stop short of moving the verdict because: (a) the neglected Schur term may well be small for coarse partitions with thin interfaces, which is testable at low cost; (b) latent-field marginals empirically survive even when hyperparameters drift, consistent with hyperparameter error being partially absorbed by integration over θ; (c) the reader's CONDITIONAL already demands quantified cost/accuracy trade-offs, and this concern is best read as making that condition concrete rather than escalating it. Recommended verdict: UNCHANGED (CONDITIONAL), with the added explicit condition that the θ-dependence and partition-count scaling of the Eq. 47 surrogate error be quantified — ideally by evaluating the interface Schur complement log|BQ~^{-1}B^T| directly, which is cheap relative to full factorization.","tokens_in":24787,"tokens_out":3697,"duration_ms":107842,"concrete_test":"In the §6.1 simulated model (Gaussian likelihood, known θ), compute log|Q(θ)| and log|Qx|y,θ(θ)| exactly and via the block-correlated surrogate Σ_i log|Qi(θ)| (Eq. 47) over a grid of hyperparameter values, for 2, 4, and 8 temporal partitions. Then run distributed vs. full-data INLA on the Gaussian simulated data at each partition count. If the surrogate log-det error varies with θ (rather than being a constant offset that cancels in the normalized posterior of Eq. 14) or grows with partition count/interface size, and the Gaussian-case hyperparameter posteriors drift correspondingly from the full-data posteriors, the accuracy-preservation claim for latent-field partitioning fails independently of the non-Gaussian issues; if the error is θ-independent and small, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim bundles two operations: (i) data/likelihood partitioning and (ii) partitioning of the structured latent field. For (i) with Gaussian likelihoods the identity in Eq. 9 is genuinely exact; no objection there. But (ii) rests on determinant approximations the paper itself derives as truncations: Eq. 45 truncates the expansion of log|D(I+D^{-1}R)|, and Eq. 47 drops the interface Schur-complement term log|BQ~^{-1}B^T| from the exact identity in Eq. 48. The stated justification is that inter-partition constraints are \"weakly coupled perturbations\" — an assumption with no bound, no scaling analysis in partition count or interface size, and no empirical isolation in §6. This is load-bearing because both |Q(θ)| and |Qx|y,θ(θ)| enter the hyperparameter posterior through Eq. 14; any θ-dependent error in either log-determinant tilts the hyperparameter geometry, and through it the CCD support points and everything downstream (Eqs. 5, 22). So \"preserving accuracy\" for latent-field partitioning is unestablished even in the Gaussian-likelihood regime where the rest of the machinery is exact. The §6.1 example cannot settle this: only 2 partitions with a small temporal interface, where the neglected term is plausibly minor. And the PM2.5 divergence (Figs. 10, 13) is attributed to non-Gaussianity and hyperparameter heterogeneity, but the determinant surrogate is an unexamined confound in that attribution. Notably, the neglected term is computable at interface dimension (shared/halo nodes only), so the error is directly measurable without any full factorization.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript develops a framework for distributed, recursive, and federated Bayesian inference for latent Gaussian models built on INLA/R-INLA. For data partitioning it derives exact identities for the conditional latent-field posterior under Gaussian likelihoods and conditional independence (Eqs. 9–11, 30–31), and Gaussian-approximation combinations for non-Gaussian likelihoods (Eq. 13, recursive updates Eqs. 35–38). It then constructs distributed approximations to the hyperparameter posterior (Eqs. 16–21) and several strategies for hyperparameter marginals (interpolation, asymmetric Gaussian, Laplace, integration-free, consensus MC). Section 5 partitions the latent field itself via block-independent (Eqs. 43–45) and block-correlated/halo (Eqs. 46–48) schemes with log-determinant surrogates. Two examples are given: a Gaussian simulated spatio-temporal case with 2 partitions, and a 873,762-observation PM2.5 Gamma model with space–time interaction. The simulation matches full-data inference well; the PM2.5 case recovers latent fields but shows substantial divergence in the CCD and hyperparameter marginals (Figs. 10, 13), attributed to non-Gaussianity and hyperparameter heterogeneity.","tokens_in":25242,"tokens_out":3384,"duration_ms":108013,"significance":"If the approximation layers hold up, this is a practically significant contribution: it makes INLA viable for massive spatio-temporal datasets, streaming updates, and privacy-constrained (federated) settings within the widely used R-INLA stack. Explicit strengths: (i) the data-partitioning and recursive identities (Eqs. 9–11, 30–31) are exact and parameter-free for Gaussian likelihoods under conditional independence; (ii) the authors are candid about non-Gaussian error accumulation (Eq. 13) and recursive non-invariance under partition reordering (§4.1); (iii) the lifted-space determinant identity (Eq. 48) makes the latent-field partitioning surrogate falsifiable, since the neglected term is computable; (iv) a real 873k-observation PM2.5 case study is attempted. The main limitation is that the two ingredients most central to the headline claim — the determinant surrogates and the complexity reduction — are asserted rather than measured.","major_comments":[{"comment":"§5.1–5.2, Eqs. (43)–(48): the latent-field partitioning rests on two uncontrolled log-determinant surrogates: truncation of the expansion in Eq. (45) after noting tr(D^{-1}R)=0, and dropping the interface Schur-complement term log|B Q~^{-1} B^T| from the exact identity in Eq. (48). These determinants enter the hyperparameter posterior via Eq. (14), so any θ-dependent error tilts the hyperparameter geometry, the CCD support points (Eq. 21), and everything downstream. No bound, no scaling analysis in partition count or interface size, and no empirical isolation is given. This is testable: the neglected term in Eq. (48) is computable at interface dimension; please report its magnitude relative to log|Q| across θ, or run the block-correlated scheme with exact determinants to isolate the surrogate's contribution. As stated, 'preserving accuracy' for latent-field partitioning is unestablished","section":"§5, Eqs. (43)–(48)"},{"comment":"§6.1 and §6.2: the empirical section cannot currently support the partitioning claims. §6.1 uses only 2 temporal partitions with a minimal interface, precisely the regime where the neglected Schur term is small; §6.2 shows substantial divergence in the CCD and hyperparameter marginals (Figs. 10, 13), which the text attributes to non-Gaussianity and hyperparameter heterogeneity — but the determinant surrogate is an unexamined confound in that attribution. Please add a study varying partition count/interface size (e.g., 2, 4, 8 temporal cuts) in the Gaussian simulation where the rest of the machinery is exact, so the surrogate's error can be isolated, and disentangle the two error sources in §6.2.","section":"§6, Figs. 3, 10, 13"},{"comment":"Eqs. (18)–(21): the global Gaussian approximation sums per-partition Hessians H_{θ*_i} evaluated at partition-specific modes θ*_i (Eq. 19). A Hessian of the global log-posterior evaluated at the global mode is not the sum of Hessians at differing local modes unless the modes coincide; the product-of-Gaussians combination in Eqs. (20)–(21) is a consensus-style heuristic, not an identity. This is load-bearing because the CCD integration design (and hence the marginals in Eq. 22) is built from it. At minimum, state the regularity condition (proximity of the θ*_i) under which this is controlled, quantify the error when it fails (Fig. 10 suggests §6.2 is such a case), and consider evaluating all partition Hessians at a common reference point.","section":"§3.3, Eqs. (18)–(21)"},{"comment":"Abstract and §7: the central claim 'reducing computational complexity while preserving accuracy' is not quantified on either axis. No runtimes, memory footprints, or factorization costs are reported anywhere in §6, despite the complexity reduction being the paper's stated motivation; and §6.2 itself exhibits clear accuracy loss in the hyperparameter posterior (Figs. 10, 13). Please (i) report timing/memory comparisons against full-data R-INLA for both examples, and (ii) temper the abstract/conclusions to match the demonstrated scope: latent-field marginals recover well, hyperparameter geometry does not in the non-Gaussian case.","section":"Abstract, §6, §7"},{"comment":"Eq. (29): consensus Monte Carlo with weights w_i = π̃(y_i) (per-partition marginal likelihoods) is nonstandard — the Scott et al. (2016) algorithm uses precision (inverse-covariance) weights, which is what makes it exact for Gaussian subposteriors. Marginal-likelihood weighting lacks justification here and can be dominated by partition size. Please either justify this choice or use standard precision weighting, and clarify how Eq. (29) interacts with the asymmetric-Gaussian subposteriors from the integration-free algorithm.","section":"§3.4, Eq. (29)"}],"minor_comments":[{"comment":"Broken cross-references: §6 first paragraph cites 'Section ??' and §6.2 cites 'Algorithm ??' for the partitioning procedure; the Supplementary Material reference target must be restored.","section":"§6"},{"comment":"Eq. (12): the canonical-form notation is unclear and the line 'The mean is computed as b_i = w_i Q_i^{-1}(θ)µ_i' appears dimensionally inconsistent — b_i should be the linear coefficient vector (canonical potential), not a scaled mean. Please rewrite Eq. (12) with explicit canonical parameters and check the sign of the linear term.","section":"§3.1, Eq. (12)"},{"comment":"Notation: n denotes both the number of observations (Eq. 1) and the number of partitions (Eq. 6 onward); |y| in the weights below Eq. (10) is undefined; the asymmetric-Gaussian variances in Eqs. (24) and (26) are written ambiguously ((σ^-_j)^{-2} z^2 vs (z/σ)^2) and lack normalization constants.","section":"§2–3"},{"comment":"Typos and garbled text: 'previosly' (§3.2), 'hyperparamters' (§3.2), 'block-independeent', 'propoerties' (§5), 'Dendogram', 'matriz', 'highlited' (Fig. 1 caption), 'could themselves varying' (§7); the sentence defining the simulation model in §6.1 ('βcat us is a spatial component...') is broken.","section":"Throughout"},{"comment":"Figures: Figs. 3 and 10 are difficult to read at print size and captions are sparse (no indication of which hyperparameter axis is which in Fig. 3). Quantitative accuracy metrics (e.g., distances between marginal posteriors, coverage comparisons) would strengthen §6 beyond visual overlay.","section":"§6, Figs. 3–13"},{"comment":"Positioning: the block-correlated scheme is essentially a Schwarz/domain-decomposition approximation for GMRF determinants; there is existing literature on parallel and domain-decomposed inference for latent Gaussian models that should be cited and contrasted, beyond the Toselli–Widlund reference.","section":"§5.2"},{"comment":"§4.3 honestly acknowledges that integration support points cannot in general be reused across recursive steps; a brief summary of what is actually reused and the resulting per-step cost would help readers assess the practical value of the recursive scheme relative to a fresh R-INLA run.","section":"§4.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as an early-stage preprint: journal placeholder \"TBA\", two broken cross-references (Section ??, Algorithm ??), and a substantial density of typos. The authors' prior-work citations are background tools rather than load-bearing, so no novelty concern there. The technical core (data-partition identities) is sound; the open questions are about the latent-field partitioning surrogates and the gap between the abstract's claims and what §6.2 demonstrates. If the journal's scope emphasizes validated methodology, the revision requested is substantial but well within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The usable core is an INLA-native construction for distributed and recursive updates: weighted prior splits, product-of-partition Gaussians for the hyperparameter mode and CCD, recursive precision updates, and two graph-partition schemes for the latent field. For Gaussian likelihoods the data-partition identities (Eqs. 9–11, 30–31) are exact under conditional independence; that part is clean and immediately useful to anyone already inside R-INLA.\n\nWhat is new is the worked pipeline plus the block-independent and block-correlated (halo) latent-field splits, not the high-level idea of divide-and-conquer Bayes. The paper is honest about non-Gaussian error accumulation and about the PM2.5 case: latent-field marginals stay close while CCD and hyperparameter marginals diverge. Self-citations supply background tools; the distributed identities themselves are derived, not fitted. Citation pattern is appropriate.\n\nSoft spot that matters: the strongest claim bundles data partitioning with latent-field partitioning. The latter rests on zeroth-order log-determinant surrogates (Eqs. 43–47) that drop the interface Schur term (Eq. 48) or the higher terms in the expansion of log|D(I+D^{-1}R)|. No bound, no scaling with partition count or interface size, and §6 never isolates the error (two partitions, small temporal interface in the sim; non-Gaussian confound in PM2.5). Because |Q| and |Q_{x|y,θ}| enter the hyperparameter posterior directly, θ-dependent bias can tilt the whole integration scheme even when the likelihood is Gaussian. The neglected term is cheap to compute on the halo, so the gap is fixable.\n\nMinor: free choices (prior weights, cut locations, CCD design) are left to the user; no code/data in the preprint.\n\nThis is for the R-INLA / latent-Gaussian crowd who need streaming, privacy, or memory relief on spatio-temporal models. It deserves a serious referee. I would engage, cite the Gaussian data-split formulas, and ask for a controlled check of the determinant error before trusting the “preserving accuracy” claim for latent-field partitions.","headline":"Solid INLA-native distributed/recursive pipeline with exact Gaussian data-split identities; latent-field partition log-det surrogates are uncontrolled truncations that the examples do not isolate.","tokens_in":25822,"tokens_out":553,"would_cite":true,"duration_ms":9605,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62M30","65C60"],"pacs":[],"model":"grok-4.5","headline":"Partitioning data and structured latent fields inside INLA yields distributed and recursive Bayesian posteriors that stay close to full-data inference while cutting cost.","keywords":["distributed inference","recursive inference","latent Gaussian models","INLA","spatio-temporal models","Big Data","federated learning","GMRF partitioning"],"falsifier":"On a non-Gaussian spatio-temporal model whose full-data INLA solution is known, increase the number or heterogeneity of partitions until the recombined hyperparameter CCD or the latent-field credible intervals diverge beyond a pre-chosen tolerance from the centralized answer.","tokens_in":25653,"feed_emoji":"📊","tokens_out":894,"duration_ms":16323,"temperature":0.7,"pith_summary":"Massive spatio-temporal datasets, streaming observations, and privacy rules that block raw-data sharing make ordinary Bayesian updates impractical. This paper shows how to run those updates inside the INLA framework by splitting both the observations and the structured pieces of the latent field (spatial, temporal, or graph components) into smaller blocks, computing local Gaussian approximations, and recombining them. For Gaussian likelihoods the recombination is exact; for non-Gaussian likelihoods it is an approximation whose quality the authors monitor. The same machinery supports recursive (online) updating when new batches arrive and federated analysis when sites cannot exchange raw records. Case studies on a simulated spatio-temporal field and on U.S. PM2.5 monitoring data show that latent-field summaries remain close to the centralized answer even when hyperparameter geometry drifts, giving practitioners a practical route to scalable, privacy-aware Bayesian inference without rewriting their models from scratch.","feed_headline":"INLA posteriors split across machines without losing the answer","feed_subtitle":"Partition data and latent fields, recombine local Gaussians, keep full-data accuracy at lower cost","key_machinery":"Block-independent and block-correlated partitions of the latent-field precision matrix (Eqs. 42–48), together with the product-of-Gaussians recombination for the conditional latent posterior (Eq. 13) and the recursive precision/mean updates (Eqs. 35–38).","core_discovery":"Inside INLA, both the data likelihood and the structured components of the latent Gaussian field can be partitioned so that the product (or recursive update) of the resulting local Gaussian approximations recovers, to usable accuracy, the same latent-field marginals and hyperparameter summaries that full-data INLA would produce, while the smaller sparse matrices sharply reduce memory and factorization cost.","pith_inferences":["The same partition-and-recombine algebra could be applied to other sparse latent-Gaussian engines beyond INLA, provided they expose the conditional precision and mode.","When hyperparameter posteriors already differ across blocks (as seen in the PM2.5 example), a hierarchical hyperprior that itself varies by partition may reduce the observed CCD drift.","Federated settings with non-overlapping latent structures (different covariates or meshes) are a direct extension once the mean and precision are expanded to a common super-field."],"forward_implications":["Large spatio-temporal models whose precision matrices no longer fit in memory can be analyzed by automatic graph partitioning without redesigning the likelihood.","Streaming environmental or epidemiological series can be updated recursively by feeding each new batch into the previous latent posterior as prior.","Privacy-sensitive consortia can exchange only local posterior summaries (means, precisions, marginal likelihoods) and still obtain a joint INLA analysis.","Meta-analysis of published studies becomes a special case of the same recursive product of approximate posteriors."],"fun_headline_variants":["Partition INLA data and latents, recombine Gaussians for full-data accuracy","Distributed INLA keeps latent marginals while cutting sparse-matrix cost","Recursive local Gaussians recover full INLA posteriors at lower memory","Split likelihood and latent field in INLA, product matches centralized result","Federated INLA updates stream data without sharing raw observations"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"For non-Gaussian likelihoods the product of the local Gaussian approximations stays accurate enough that hyperparameter geometry and latent-field marginals do not degrade materially when the blocks are recombined.","fun_headline_variants_meta":{"raw":{"variants":["Partition INLA data and latents, recombine Gaussians for full-data accuracy","Distributed INLA keeps latent marginals while cutting sparse-matrix cost","Recursive local Gaussians recover full INLA posteriors at lower memory","Split likelihood and latent field in INLA, product matches centralized result","Federated INLA updates stream data without sharing raw observations"]},"model":"grok-4.5","effort":"low","cost_usd":0.004022,"raw_usage":{"total_tokens":1190,"prompt_tokens":725,"num_sources_used":0,"completion_tokens":98,"cost_in_usd_ticks":40224000,"prompt_tokens_details":{"text_tokens":725,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":367,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":725,"tokens_out":98,"duration_ms":7410,"temperature":1.0,"reasoning_tokens":367,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T23:15:41.338748+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a non-Gaussian spatio-temporal model whose full-data INLA solution is known, increase the number or heterogeneity of partitions until the recombined hyperparameter CCD or the latent-field credible intervals diverge beyond a pre-chosen tolerance from the centralized answer.","supporting_citations":[],"review_version":1}