{"id":"2d20f218-1992-45e9-b77c-a4a378d95701","arxiv_id":"2607.14680","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An operator-split Bayesian posterior, which pushes independent neural-network posteriors for the source and boundary through the elliptic solution map, contracts around the true solution at a near-minimax rate.","lead":"The paper builds a Bayesian method for solving elliptic PDEs from noisy measurements of the source inside the domain and of the boundary values, with unequal sample sizes. It proves that the resulting posterior distribution converges to the true solution at a rate that matches the known lower bound up to logarithmic factors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Near-minimax claim is not established: Theorem 4.2 gives posterior concentration, but Corollary 4.5/abstract conclude a point-estimator upper bound without the second-moment control that Remark 4.3 admits is missing.","rationale":"I read the main contraction theorem as likely correct: applying Theorem 2.5 separately to the interior regression (smoothness beta-2 on Omega) and the boundary-coordinate regression (smoothness beta on the boundary) and then using the exact identity E(S(F,G))=||F-f||^2+lambda||G-g||^2 gives the stated posterior concentration. The single-chart assumption flagged by the reader is real but less severe: for smooth compact boundaries a single chart covering up to measure zero is usually available (e.g., the top cell of a handle decomposition), and the finite-atlas extension in Remark 3.3 removes the issue when it is not; the square experiment lies outside the C^beta-smooth-boundary assumptions but is presented as an illustration. The most load-bearing concern is the near-minimax claim. Theorem 2.7 is a lower bound for estimators, and Corollary 4.5 compares a posterior contraction radius to that lower bound. Without a proof that a specific estimator (such as the posterior mean) attains O(R^2) risk, the phrase 'near-minimax upper bound' in the abstract and Corollary 4.5 overstates what has been shown. Remark 4.3 makes this explicit by conditioning the estimator-risk conclusion on unproved second-moment contractions. The concrete test I propose would settle whether those moments are automatic or whether the claim must be restated. Since the reader already judged the paper CONDITIONAL, my read does not change the verdict.","tokens_in":21770,"tokens_out":36364,"duration_ms":386282,"concrete_test":"Derive a bound on the posterior-mean risk directly from (4.3). Try to prove E_{P_0^{(nΩ)}Q_0^{(n∂)}} E(\\bar u) <= C R^2_{nΩ,n∂} using only the mass-concentration inequality (4.3) and the boundedness |F|,|G|<=B. If the derivation fails, construct a one-dimensional analogue (a bounded regression function with a heavy-tailed BNN prior satisfying Assumptions 2.3/2.4) where the posterior contracts at rate R in the sense of (4.3) but the posterior-mean loss is not O(R^2). If such a counterexample exists, Corollary 4.5's 'near-minimax upper bound' wording must be weakened to a contraction-radius comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 4.2 itself is a valid application of Corollary 4.1 plus the exact split-loss identity (3.5); the single-chart assumption is a removable simplification (Remark 3.3) and for smooth boundaries a single chart covering up to measure zero can typically be obtained by a handle decomposition. The real gap is the advertised near-minimax optimality. Corollary 4.5 and the abstract claim that the contraction radius R^2 matches the minimax lower bound of Theorem 2.7. But Theorem 2.7 is a lower bound on E E(ψ(D)) for point estimators, while Corollary 4.5 concerns posterior mass concentration. One cannot conclude a near-minimax upper bound for point estimation unless one proves, for instance, that the posterior mean \\bar u satisfies E E(\\bar u)=O(R^2). Theorem 4.2 only shows Pi(E>M^2R^2)->0 for every M->infinity; such concentration does not control the posterior first moment. Remark 4.3 explicitly imposes unproved second-moment conditions to conclude E(\\bar u)=O_p(R^2). Thus the near-minimax upper-bound claim is not supported by the present arguments; without those moments the posterior-mean risk could be M_n^2 R^2 with M_n->infinity, which is not the claimed rate.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an operator-split Bayesian construction for elliptic Dirichlet problems with independent noisy interior-source and boundary observations. Separate Bayesian neural-network priors are placed on the source f and on a local-coordinate representation of the boundary data g; the product posterior is pushed forward through the linear elliptic solution operator. Theorem 4.2 asserts posterior contraction in the physics-informed loss at radius ε_{nΩ}^2 + λ ε_{n∂}^2, with separated d-dimensional interior and (d−1)-dimensional boundary rates. The proof uses the exact identity E(S(F,G)) = ||F−f||^2 + λ||G−g||^2, the external BNN contraction result Theorem 2.5, and a union bound. The paper further claims near-minimax optimality by comparison with the lower bound of [47, Thm 3.5], derives a boundary sampling budget condition, and reports 1D and 2D numerical experiments with finite-element propagation of posterior samples.","tokens_in":22136,"tokens_out":6726,"duration_ms":76241,"significance":"If Theorem 4.2 is read with its explicit single-chart assumption, the core derivation is clean and the separated-dimensional rate is a genuinely useful contribution: the split-loss identity (3.5) is exact, no fitted constants are used, and the proof is a valid combination of an external contraction theorem and a union bound. The numerical experiments usefully illustrate the two-sample budget trade-off and the stabilization of the boundary contribution as n∂ grows. However, two advertised conclusions go beyond what is proved: the near-minimax point-estimator claim and the scope of the theorem for general smooth boundaries. The paper is a solid conditional contribution, but these gaps need to be addressed before it can be accepted as stated.","major_comments":[{"comment":"The near-minimax claim is not supported. Theorem 4.2 is a posterior concentration statement: for every M_n→∞, Π_S(E > M_n^2 R^2 | D) → 0. This does not control the first posterior moment, and in particular it does not imply E[E(\\bar u)] = O(R^2) for the posterior mean. Remark 4.3 explicitly imposes unproved second-moment contractions to conclude E(\\bar u) = O_p(R^2). Without such control, the posterior-mean risk could be M_n^2 R^2 with M_n→∞, which is not the claimed rate. Since Theorem 2.7 is a minimax lower bound for point estimators, Corollary 4.5 and the abstract's 'near-minimax upper bound' overstate the result. Please either prove the second-moment contraction or revise the optimality claim to a statement about the contraction radius only.","section":"Corollary 4.5, Remark 4.3, Abstract"},{"comment":"The main theorem assumes a single boundary chart φ:B⊂R^{d−1}→∂Ω that covers ∂Ω up to surface measure zero. This is false for general smooth compact boundaries: for example, a sphere or torus does not admit such a single chart, and if B is required compact (as in Lemma 3.1), a chart cannot cover a connected closed boundary up to measure zero. The proof and the contraction statement are one-chart only; Remark 3.3 acknowledges the finite-atlas issue but does not supply a multi-chart version of Theorem 4.2. Moreover, the 2D numerical experiment is on the unit square with the piecewise parametrization γ:[0,4]→∂Ω, which is not a smooth chart and has corners where g is not C^β for β>2; the experiment therefore lies outside the theorem's hypotheses. Please either prove the finite-atlas extension or explicitly restrict the claims and experiments to domains satisfying the stated chart condition.","section":"Theorem 4.2, §3.2, Remark 3.3, §5.2"}],"minor_comments":[{"comment":"The notation '(R^d)^{⊗nΩ} × R^{⊗nΩ} × (R^d)^{⊗n∂} × R^{⊗n∂}' is garbled and should be cleaned up.","section":"Theorem 2.7"},{"comment":"The paper alternates between an open set B⊂R^{d−1} in the chart definition and a compact B in Lemma 3.1. Since the single-chart covering assumption is central, the topological assumptions on B should be stated consistently.","section":"§3.2 / Lemma 3.1"},{"comment":"Figure 4 is discussed in terms of 's=2' but the horizontal axis is the boundary parameter t; please unify the notation.","section":"§5.2"},{"comment":"References [27]–[30] appear only in general introduction material and are not related to the theoretical argument; they could be trimmed or moved to a broader discussion.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is an incremental but potentially useful contribution relative to [47]. The central contraction argument is sound under its explicit assumptions, but the headline near-minimax claim is not justified and the single-chart geometry issue means the theorem does not cover the numerical experiment or general smooth boundaries. These are fixable with additional work: a second-moment argument or a softened optimality claim, and a multi-chart proof or restricted statements. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read this carefully. The short version: the construction is genuinely new and the main concentration theorem is probably correct, but it is essentially an application of existing BNN regression contraction twice, and the paper's advertised near-minimax optimality is not actually established. The theorem's core holds up; the overclaim is in the packaging.\n\nWhat is new and good: assigning independent BNN priors to the source and boundary trace, multiplying the posteriors, and pushing them through the linear solution map is a clean, sensible separation. It gives separate d-dimensional and (d-1)-dimensional rates without putting a prior directly on u, and the loss identity E(S(F,G)) = ||F-f||^2 + lambda||G-g||^2 is transparent. The proof of Theorem 4.2 is short: apply the cited BNN contraction to each regression, convert norms, and union-bound. That is legitimate. The boundary-sampling-budget rule is a useful, simple consequence. I credit the paper for stating its limitations honestly.\n\nSoft spots, in order.\n\n1. The near-minimax claim is not supported. Theorem 4.2 gives posterior mass concentration: Pi(E > M^2 R^2) -> 0. Corollary 4.5 and the abstract conclude that the rate matches the minimax lower bound of [47]. But that lower bound is about sup_{u*} E E(psi(D)) for point estimators. Posterior mass concentration does not control the posterior mean. Remark 4.3 admits this: you need unproved second-moment contraction to get E(bar u) = O_p(R^2). Without it, the posterior mean could fail the claimed rate. This is a real gap, but it is localized in Corollary 4.5 and the abstract, and fixable by either proving the second moments or restating the claim as contraction-radius matching the lower bound.\n\n2. The single-chart assumption does not cover the experiments. The theorem assumes one boundary chart covers dOmega up to surface measure zero. That is false for a general smooth boundary, e.g. a torus, and certainly false for the square used in Section 5.2. Remark 3.3 acknowledges finite atlases, and a partition-of-unity fix is plausible, but the main theorem as stated does not cover the numerical example. That needs fixing in the statement, or the experiment needs a smooth domain.\n\n3. Minor: no code or data, and the experiments are illustrative. The timing result is just the expected linear scaling of repeated FEM solves; it is fine but not a strong computational comparison.\n\nVerdict: the core contraction theorem is sound within its assumptions, the construction is worth knowing, and the overclaims are addressable. This deserves peer review; a referee should ask for the near-minimax claim to be either proved or cut. I would send it back for serious revision rather than desk reject.","headline":"The operator-split posterior construction is genuinely new and the main concentration theorem is likely correct as an application of existing BNN regression contraction, but the advertised near-minimax optimality is not proven and the single-chart assumption does not cover the square experiment.","tokens_in":22629,"tokens_out":2738,"would_cite":true,"duration_ms":31532,"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":"Operator-split Bayesian posteriors for elliptic PDEs contract at near-minimax rates when interior and boundary data are unequal.","keywords":["Bayesian neural networks","elliptic PDEs","posterior contraction","operator splitting","uncertainty quantification","Dirichlet boundary conditions","imbalanced data","physics-informed loss"],"falsifier":"Check whether the boundary of the unit square admits a single connected chart φ:B⊂R→∂Ω that is a diffeomorphism onto ∂Ω minus a measure-zero set. It does not: the smooth part of a square boundary has four connected components, while B is connected, so the theorem's single-chart hypothesis is false for the paper's own 2D experiment. A re-run of that experiment therefore cannot be validated by Theorem 4.2 as stated.","tokens_in":1350,"feed_emoji":"📉","tokens_out":1915,"duration_ms":61379,"temperature":0.7,"pith_summary":"The paper proposes a Bayesian way to learn the solution of a second-order elliptic Dirichlet problem when interior source measurements and boundary measurements are noisy, imbalanced in number, and treated as separate statistical problems. It assigns independent neural-network priors to the source term and to the boundary data, combines the two posteriors into a product, and pushes this distribution through the elliptic solution map. The central claim is that the resulting posterior contracts around the true solution at a rate that separates into an interior term depending on the d-dimensional sample size and a boundary term depending on the (d-1)-dimensional boundary sample size, matching the minimax lower bound up to logarithms. A sympathetic reader would care because this gives a rigorous uncertainty-quantification guarantee for a practical split PINN construction and shows how to allocate sampling effort between volume and boundary.","feed_headline":"Split source and boundary posteriors hit the minimax rate","feed_subtitle":"Interior and boundary data concentrate at their own dimensional rates, matching the lower bound up to logs.","key_machinery":"The central object is the elliptic solution map S(F,G)=S₀F+S₁G, where S₀ solves Lv=F with zero boundary data and S₁ lifts boundary data G to a solution of Lw=0. The key identity is the operator-split loss E(S(F,G))=∥F−f∥²_{L²(Ω)}+λ∥G−g∥²_{L²(∂Ω)}, which converts posterior contraction for f and g directly into posterior contraction for u. Boundary observations are re-expressed in local coordinates through a chart φ:B⊂R^{d−1}→∂Ω, so the boundary regression lives on a (d−1)-dimensional space. The product posterior factorization and pushforward through S carry the two independent contraction guarantees to the solution space.","core_discovery":"Let u* solve Lu=f in Ω with u=g on ∂Ω. Theorem 4.2 asserts that, when f is β−2-Hölder and g is β-Hölder and a single boundary chart covers ∂Ω up to measure zero, the operator-split posterior satisfies Π_S({u : E(u) > M^2 R^2} | D) → 0 in true-data probability, where R^2 = n_Ω^{−2(β−2)/(d+2(β−2))} (log n_Ω)^{2γ} + λ n_∂^{−2β/(d−1+2β)} (log n_∂)^{2γ}. Because every posterior draw has E(u)=∥F−f∥²_{L²(Ω)}+λ∥G−g∥²_{L²(∂Ω)}, this means the learned source and boundary functions contract separately at their natural dimensional rates. Up to logarithmic factors, the radius matches the cited minimax lower bound, so the split construction is claimed to be near-optimal.","pith_inferences":["The single-chart assumption is the main constraint: for a square, a torus, or any boundary whose smooth part has several connected components, no one connected chart covers ∂Ω up to measure zero, so the theorem as stated does not cover the paper's own square experiment or general domains without a finite-atlas extension.","One testable extension would be to track chart-indexed error terms in a partition-of-unity proof; the rate structure suggests the same separated interior/boundary exponents would survive, but the proof would need explicit treatment of chart overlaps.","The separated loss components suggest an adaptive data-collection strategy: estimate the two empirical errors online and allocate new interior versus boundary samples according to the predicted rate imbalance, an idea the paper mentions but does not develop.","Because the solution map is linear, the same split-posterior construction may extend to Bayesian inverse problems where source or boundary data are recovered from noisy solution observations; the paper lists this as future work."],"forward_implications":["If the theorem is correct, one can learn elliptic solutions without placing a prior directly on the solution, instead splitting the regression into source and boundary components and propagating uncertainty through a deterministic solver.","The two-term contraction radius yields an explicit boundary sampling condition n_∂ ≳ n_Ω^{κ(d,β)} under which the boundary contribution does not dominate, guiding allocation of unequal sampling budgets.","Under additional second-moment conditions, the posterior mean contracts at the same radius, and the physics-informed loss controls H^{1/2}(Ω) error via the cited stability estimate.","The reuse of one stiffness-matrix factorization across posterior draws makes the propagation step linear in ensemble size, as the numerical experiments report.","Up to logarithmic factors, the upper bound matches the two-sample minimax rate, so no statistical price is paid for the operator-split construction in the smooth setting considered."],"fun_headline_variants":["Operator-split BNNs near minimax for elliptic PDEs","Split posteriors hit near-optimal elliptic rates","Unequal data rates: source vs boundary convergence","Two-rate Bayesian contraction matches lower bound","Near-optimal split learning for Dirichlet problems"],"cache_read_input_tokens":23936,"weakest_assumption_plain":"The theorem assumes one smooth boundary chart maps an open subset of R^{d−1} onto the entire boundary up to a measure-zero set; for the square in the experiments and for topologically nontrivial or cornered boundaries no such single connected chart exists, so the central contraction guarantee does not literally cover those cases without a finite-atlas argument.","fun_headline_variants_meta":{"raw":{"variants":["Operator-split BNNs near minimax for elliptic PDEs","Split posteriors hit near-optimal elliptic rates","Unequal data rates: source vs boundary convergence","Two-rate Bayesian contraction matches lower bound","Near-optimal split learning for Dirichlet problems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000187,"raw_usage":{"total_tokens":1172,"prompt_tokens":756,"completion_tokens":416,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":343}},"tokens_in":500,"tokens_out":416,"duration_ms":5238,"temperature":1.0,"reasoning_tokens":343,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T01:23:35.061134+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check whether the boundary of the unit square admits a single connected chart φ:B⊂R→∂Ω that is a diffeomorphism onto ∂Ω minus a measure-zero set. It does not: the smooth part of a square boundary has four connected components, while B is connected, so the theorem's single-chart hypothesis is false for the paper's own 2D experiment. A re-run of that experiment therefore cannot be validated by Theorem 4.2 as stated.","supporting_citations":[],"review_version":1}