{"id":"dee4067e-1cbe-4002-b8e8-35d3f06f91e3","arxiv_id":"2607.07034","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":12,"one_line_summary":"A two-stage algorithm learns a coordinate chart and a low-rank tensor source for a linear PDE, reducing supervised learning on manifolds to independent 1D integrals with automatic intrinsic dimension discovery.","lead":"The paper introduces a method that learns a target function on a data manifold by fitting the source term of a linear PDE, then integrating it against a Green's kernel. A smart generalist might read it because it offers a principled way to discover low-dimensional structure in data while keeping the fitting step nearly convex.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The low-rank source assumption (Eq. 2) underpins the O(KRd) complexity claim, but the paper's hardest experiment (MNIST) never directly verifies that the learned source is actually low-rank—only that classification accuracy is high.","rationale":"The reader correctly identified the low-rank source assumption as the most load-bearing concern. I agree and sharpen it: the paper's most complex experiment (MNIST) tests classification accuracy and dimension discovery but never directly verifies that the learned source is low-rank. The ρ_r gating mechanism is described (Table 5) and used for rank discovery, but no values are reported for MNIST. This is a critical gap because the entire O(KRd) complexity claim depends on the source being low-rank, and the gap between the theoretical condition (bounded mixed derivatives, Eq. 19) and empirical validation (simple 2D synthetic data only) is wide. The two-stage training mechanism and automatic dimension discovery are sound contributions with clear empirical support (Figure 2, Table 3). The 'near-convex' characterization has a minor inconsistency (γ_k placement in §3.1 vs. Algorithm 1), but this is secondary. The verdict should remain CONDITIONAL: the framework is novel and mathematically coherent, but the central computational advantage rests on an assumption validated only on simple synthetic data. Direct rank reporting on MNIST and validation on harder benchmarks would be needed for ACCEPT. No code availability further limits reproducibility of the empirical claims.","tokens_in":15347,"tokens_out":6245,"duration_ms":266157,"concrete_test":"On the MNIST experiment, report: (1) the effective source rank R_eff via the ρ_r gating mechanism (how many of the 128 ranks are active); (2) linear probe accuracy as R_max is swept from 8 to 256 while holding all other hyperparameters fixed; (3) the relative contribution of the Green's integral term vs. the polynomial/null-space augmentation term (Eq. 5) to the final logits, measured as the ratio of their L2 norms averaged over the test set. If R_eff is small (<32) and accuracy degrades sharply when R_max is reduced below R_eff, the low-rank assumption is validated. If R_eff is large or accuracy is rank-insensitive, the O(KRd) advantage claim weakens significantly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central computational claim—O(KRd) cost linear in intrinsic dimension—depends entirely on the source f admitting low-rank CP decomposition (Eq. 2, §3.2). The paper acknowledges this is 'conditional on coordinates' (Appendix B.1) and invokes bounded mixed derivatives (Eq. 19) as the theoretical justification. This is a strong coordinate-dependent condition requiring ∂²f/∂x∂y to be bounded, and the paper provides no evidence that realistic supervised learning targets satisfy this in any discoverable chart. The examples given (sin(x)sin(y)) are specially constructed; real classification boundaries with complex coordinate coupling may not qualify in any chart. More critically, the MNIST experiment—the paper's most complex validation—reports classification accuracy (0.991 linear probe) and dimension discovery (12 of 64 gates active) but never reports the effective source rank R_eff via the ρ_r gating mechanism (Table 5 lists ρ_r as 'Learned' but no values are given). The paper uses R=128 anchors for MNIST but does not report how many are active, nor does it report accuracy as a function of R. Without this, we cannot distinguish: (a) the source genuinely admits low-rank structure (small R_eff, graceful degradation when R is reduced) versus (b) the polynomial/null-space augmentation terms (Eq. 5) or the CNN encoder are doing the real work, with the Green's integral contributing marginally. The Swiss Roll ablation (Table 2) shows rank-dependent MSE improvement, but only on a 2D manifold with a simple smooth target—exactly the regime where low-rank structure is expected a priori. A secondary issue: §3.1 places γ_k (kernel scale weights) in Stage 2 (θ_2), making the inner problem bilinear in (γ_k, w_r) and thus non-convex, contradicting the 'near-convex linear solve' characterization. Algorithm 1 only solves for w, suggesting γ_k is actually in the outer loop, creating an inconsistency between the parameter split in §3.1 and the algorithm in Appendix C.1.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"The paper introduces Intrinsic Green's Learning (IGL), a framework for supervised learning on manifolds that models the target function as the solution to a linear PDE with a learned source term. The key technical contribution is a two-stage algorithm: Stage 1 discovers a coordinate chart via an encoder where the source and Green's kernel admit low-rank tensor decompositions, and Stage 2 solves for source weights via a near-convex linear solve. The tensor structure allows a d-dimensional integral to be factored into d independent 1D integrals via Fubini's theorem, yielding O(KRd) complexity. The paper validates the approach on synthetic manifolds (Swiss Roll, rotated planes) and MNIST, demonstrating topology preservation, sample efficiency, and automatic intrinsic dimension discovery.","tokens_in":15755,"tokens_out":1730,"duration_ms":1964204,"significance":"The paper presents a novel and mathematically grounded approach to learning on manifolds. The core insight—inverting the PDE perspective to learn a source term rather than the solution directly, and using tensor decomposition to make the resulting integral tractable—is elegant and well-motivated. The two-stage variable projection algorithm is a principled solution to the dimensional collapse problem, and the envelope theorem argument (Appendix C.1) provides a clean theoretical justification for why the separation prevents collapse. The Fubini factorization proofs (Appendix A.1-A.3) are algebraically correct. The MNIST results, showing 0.991 linear probe accuracy with 12 of 64 gates active, are promising if the underlying assumptions hold. The framework's modularity (operator choice, encoder architecture) is a practical strength.","major_comments":[{"comment":"The central computational claim of O(KRd) complexity depends entirely on the source term f admitting a low-rank CP tensor decomposition (Eq. 2, §3.2). The paper acknowledges this is 'conditional on coordinates' (Appendix B.1) and invokes bounded mixed derivatives (Eq. 19) as the theoretical justification. However, the paper's most complex experiment—MNIST (Table 3, §4)—does not directly verify that the learned source is actually low-rank. Table 5 lists the rank importance parameter ρ_r as 'Learned' via softplus gating, but no values of the effective rank R_eff are reported for MNIST. The experiment uses R=128 anchors but does not report how many are active, nor does it report classification accuracy as a function of R. Without this, the reader cannot distinguish between (a) the source genuinely admitting low-rank structure (small R_eff, graceful degradation when R is reduced) and (b) the","section":null},{"comment":"polynomial/null-space augmentation terms (Eq. 5) or the CNN encoder doing the real work, with the Green's integral contributing marginally. The Swiss Roll ablation (Table 2) shows rank-dependent MSE improvement, but only on a 2D manifold with a smooth target. The authors should report R_eff and accuracy-vs-R curves for MNIST to substantiate that the low-rank assumption holds in the paper's hardest setting.","section":null},{"comment":"The claim that Stage 2 is 'near-convex' (§3.1, §5) requires qualification. The linearity holds only for fixed encoder parameters θ_Ψ and fixed kernel parameters {γ_k, σ_k,j} (Appendix C.1). However, Table 5 and Algorithm 1 indicate that the anchor positions μ_{r,j} and kernel widths σ_{k,j} are updated in the outer loop (Stage 1), meaning the design matrix Φ (Eq. 24) depends nonlinearly on these parameters. The paper should clarify which parameters are truly fixed in Stage 2 (only w_r and c_β?) versus which are jointly optimized in Stage 1, and soften the 'near-convex' language accordingly. As stated, the claim overstates the convexity of the actual optimization problem being solved.","section":null},{"comment":"The two-stage vs. joint training comparison (Figure 3, §4) attributes the performance gap to optimization structure alone, stating 'architectures are identical.' However, the two-stage method solves Stage 2 exactly via least squares at each outer step, which is computationally equivalent to an infinite number of inner gradient steps. The comparison is therefore between joint training with a finite (and unspecified) number of inner steps versus two-stage training with exact inner solves. To isolate the effect of optimization structure from the effect of inner solve quality, the authors should either (a) report the number of inner gradient steps used for joint training, or (b) include a baseline where joint training uses a large number of inner steps to approximate the exact solve. Without this, the 'phase transition' in Figure 3 may reflect under-optimized joint training rather than afund","section":null}],"minor_comments":[{"comment":"§3.1: The loss function includes a Group Lasso penalty λ ||w_r γ||_2, but the role of γ_k in this penalty is unclear. Should this be ||w_r||_2 or ||w_r γ_k||_2? The notation is ambiguous.","section":null},{"comment":"Table 2: The d_eff column reports values of 3 for all operators, but the text refers to 'true d=2' with 'embedding curvature overhead.' It would help to clarify whether d_eff=3 is a consistent finding or an artifact of the greedy knockout procedure.","section":null},{"comment":"Figure 2: The convergence curves (columns 3-4) are difficult to read due to axis labeling. Consider using log-scale for the y-axis and clarifying what 'MSE' refers to (training or test).","section":null},{"comment":"Appendix B.1: The statement 'functions with bounded mixed derivatives do admit efficient tensor approximations' cites Khoromskij (2012), but the connection between bounded mixed derivatives and the specific CP decomposition used in Eq. 2 could be made more explicit. The current treatment is somewhat hand-wavy.","section":null},{"comment":"§4, Exp. 7: The MNIST experiment uses a 'shared CNN encoder' but the encoder architecture is not specified. The paper should state the number of layers, hidden dimensions, and whether the encoder is pre-trained or trained from scratch.","section":null},{"comment":"Table 3: The 'Smooth' metric is not defined in the main text. The reader must infer its meaning from context. A brief definition would improve readability.","section":null},{"comment":"Appendix C.2, Eq. 28: The distinction between the particular solution (computed by IGL) and the homogeneous solution (free parameters) is clear, but the paper should note that for the Laplacian on a bounded domain, the null space includes constants and linear functions, which are absorbed by the polynomial augmentation (Eq. 5). This connection is mentioned but not made explicit enough.","section":null},{"comment":"The paper uses both 'EDP' (Figure 1 caption) and 'PDE' (main text). 'EDP' appears to be a typo or non-standard abbreviation and should be corrected to 'PDE'.","section":null}],"recommendation":"major_revision","confidential_remarks":"The paper is mathematically elegant and the core idea is sound, but the gap between theory and empirical validation is concerning. The low-rank source assumption is the linchpin of the entire framework, and the fact that the hardest experiment (MNIST) does not verify it directly is a significant weakness. The skeptic's concern about whether the Green's integral or the encoder/augmentation terms are doing the work is well-founded and should be addressed. If the authors can show that R_eff is small and that accuracy degrades gracefully with reduced R on MNIST, this would substantially strengthen the paper. The two-stage vs. joint comparison also needs tightening to rule out the confound of inner solve quality. I believe these issues are addressable within the scope of a revision, but they are load-bearing for the central claims."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The paper you sent — Intrinsic Green's Learning (IGL) — has a genuinely new idea: instead of fitting a target function directly, learn a source term f and integrate it against a Green's kernel. The source and kernel both get CP tensor decompositions, so a d-dimensional integral collapses to d one-dimensional integrals via Fubini. A two-stage algorithm (variable projection) separates coordinate discovery from source fitting, and learnable gates discover intrinsic dimension. On MNIST, the two-stage variant gets 0.991 linear probe accuracy with 12 of 64 gates active, consistent with prior intrinsic dimension estimates. The Fubini factorization proofs in Appendix A are algebraically correct, and the two-stage vs. joint training comparison on Swiss Roll (Figure 2) is a clean demonstration that the split prevents dimensional collapse. The conceptual framing — PDE as architectural prior, not physical claim — is well-articulated and distinguishes the work clearly from PINNs, KANs, and neural operators. The parameter split between stages is also a real insight: solving Stage 2 to optimality forces the encoder gradient to reflect coordinate quality rather than source capacity shortcuts. That said, there are real gaps. The stress-test concern about the low-rank source assumption lands. The paper acknowledges (Appendix B.1) that low tensor rank is coordinate-dependent and invokes bounded mixed derivatives, but provides no theorem guaranteeing the encoder converges to such coordinates. More concretely, the MNIST experiment never reports the effective source rank R_eff — Table 5 lists ρ_r as learned but gives no values. With R=128 anchors, we can't tell whether the Green's integral is doing real work or whether the CNN encoder and polynomial augmentation carry the load. The Swiss Roll ablation (Table 2) shows rank-dependent MSE improvement, but only on a 2D smooth target where low-rank structure is expected a priori. The stress-test also catches a real inconsistency: §3.1 places γ_k in Stage 2, but Algorithm 1 only solves for w, suggesting γ_k is actually in the outer loop. This makes the inner problem bilinear and the 'near-convex linear solve' characterization slightly overstated. Minor but worth fixing. No code is provided, which limits reproducibility. Validation is limited to MNIST and synthetic data — no CIFAR or harder benchmarks. The two-stage training insight and automatic dimension discovery are the strongest contributions and deserve a serious referee. The framework is novel enough and the math sound enough on the factorization side to warrant careful review, but the reviewer should demand: (1) effective rank reporting on MNIST, (2) accuracy vs. R curves, (3) code, and (4) at least one harder benchmark to test whether the low-rank assumption holds for realistic targets.","headline":"Novel inverse-PDE framework with sound factorization math, but validation is thin and the low-rank source assumption is unverified on the hardest experiment.","tokens_in":16562,"tokens_out":650,"would_cite":false,"duration_ms":111090,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Learn the source, not the solution: inverse-PDE learning on manifolds","keywords":[],"falsifier":"Construct a target function on a manifold that does not admit low-rank tensor structure in any discoverable coordinate chart (e.g., a function with irreducible cross-coordinate coupling), and show that IGL's effective rank grows unboundedly or its accuracy degrades relative to direct approximation methods regardless of encoder capacity or training procedure.","tokens_in":15590,"feed_emoji":"🌊","tokens_out":1073,"duration_ms":86609,"temperature":0.7,"pith_summary":"The paper proposes that instead of directly fitting a target function on a high-dimensional manifold, one should learn a source term whose convolution against a Green's kernel reproduces the target. The source and kernel are both expressed as low-rank tensor decompositions in coordinates discovered by a neural encoder, which collapses a d-dimensional integral into d independent one-dimensional integrals with cost linear in the intrinsic dimension d. A two-stage algorithm—alternating between coordinate discovery (outer, nonlinear) and source fitting (inner, exact linear least squares)—prevents the dimensional collapse that plagues joint optimization, and learnable gates on each coordinate automatically zero out unnecessary dimensions, recovering the manifold's intrinsic dimension. On MNIST, the method achieves 0.991 linear-probe accuracy while automatically selecting 12 of 64 latent dimensions, consistent with prior estimates of MNIST's intrinsic dimension.","feed_headline":"Learn the source, not the solution: inverse-PDE learning on manifolds","feed_subtitle":"Two-stage tensor factorization of Green's kernel collapses high-dimensional integrals to linear cost while auto-discovering intrinsic dim","key_machinery":"Three components: (1) an encoder Ψ: R^D → R^d discovering coordinates where the source admits CP tensor decomposition of rank R; (2) a rank-K separable approximation of the Green's function via the exponential sum trick, which converts the additive eigenvalue coupling 1/(Σ λ_j) into a product of exponentials; (3) a two-stage variable-projection optimizer that solves the inner linear system (source weights w via least squares) to optimality before each outer encoder update. Learnable gates g_j ∈ [0,1] with Group Lasso discover effective dimension; rank gates ρ_r discover effective source rank. Polynomial augmentation handles null-space modes (constants, linear trends) that would otherwise要求高秩","core_discovery":"The central mechanism is the separation of geometry from fitting. By parameterizing the target as u = G * f (Green's kernel convolved with a learned source), the fitting problem for source weights becomes linear given fixed coordinates and kernel scales. Solving this linear problem to optimality at each encoder step (variable projection) forces the encoder gradient to reflect only coordinate quality, not source-capacity shortcuts. This is what prevents dimensional collapse: the encoder cannot compensate for bad coordinates by inflating source rank, because the source is always optimally fit. The tensor factorization of both source and kernel then makes the integral tractable via Fubini's定理,O","pith_inferences":["If the two-stage principle generalizes, it suggests that dimensional collapse in representation learning is primarily an optimization-architecture mismatch, not a fundamental capacity limitation—joint training allows the encoder to find degenerate shortcuts that exact inner solves eliminate.","The gauge symmetry (invariance of the integral under coordinate diffeomorphisms with inverse source transformation) implies the encoder is free to discover any chart that diagonalizes tensor rank, not necessarily an isometric embedding—this is a weaker and potentially more achievable goal than manifold isometry.","The reliance on bounded mixed derivatives for low-rank tensor approximation suggests IGL will struggle on targets with strong cross-coordinate coupling (e.g., radial functions in Cartesian coordinates) unless the encoder finds the diagonalizing chart—empirically demonstrated on simple manifolds but unproven for complex topologies."],"forward_implications":["Any supervised learning problem where data lies on a low-dimensional manifold could benefit from replacing direct function approximation with source-term learning, gaining automatic dimension discovery and topology preservation as byproducts.","The two-stage variable-projection principle—solving the linear inner problem to optimality before updating the nonlinear outer parameters—could be applied to other architectures with linear heads over nonlinear embeddings, wherever joint training causes dimensional collapse.","Operator choice (Laplacian, Helmholtz, harmonic oscillator, fractional) becomes a tunable inductive bias knob: smoothness, locality, oscillation, or long-range correlations, each selected per problem class.","The framework naturally extends to meta-learning: amortize Stage 1 (coordinate discovery) across a task distribution, adapting only the lightweight Stage 2 (O(KR) parameters) per new task."],"fun_headline_variants":["Separate geometry from fitting: inverse-PDE learning on manifolds avoids dimensional colla","Learn the source term, integrate the Green's kernel: linear-cost PDE solver on manifolds","Variable projection prevents dimensional collapse in intrinsic Green's function learning","Two-stage factorization of Green's kernel collapses manifold integrals to linear cost","Encoder discovers coordinates; linear source fit prevents dimensional collapse"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The source term must admit a low-rank tensor decomposition in the coordinates the encoder discovers. The paper acknowledges this is conditional on the encoder finding coordinates where the target has bounded mixed derivatives, but provides no theorem guaranteeing convergence to such coordinates—the evidence is empirical, limited to Swiss Roll and MNIST.","fun_headline_variants_meta":{"raw":{"variants":["Separate geometry from fitting: inverse-PDE learning on manifolds avoids dimensional collapse","Learn the source term, integrate the Green's kernel: linear-cost PDE solver on manifolds","Variable projection prevents dimensional collapse in intrinsic Green's function learning","Two-stage factorization of Green's kernel collapses manifold integrals to linear cost","Encoder discovers coordinates; linear source fit prevents dimensional collapse"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":606,"prompt_tokens":510,"completion_tokens":96,"prompt_tokens_details":null},"tokens_in":510,"tokens_out":96,"duration_ms":47985,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T21:16:03.437105+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Construct a target function on a manifold that does not admit low-rank tensor structure in any discoverable coordinate chart (e.g., a function with irreducible cross-coordinate coupling), and show that IGL's effective rank grows unboundedly or its accuracy degrades relative to direct approximation methods regardless of encoder capacity or training procedure.","supporting_citations":[],"review_version":1}