{"id":"1b4b5a8d-dc62-44da-87c4-7576dc62ea6b","arxiv_id":"2412.01379","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A deformation-based metric framework lets neural operators learn PDE solution maps on non-diffeomorphic, varying domains, with a convergence proof for Poisson on star domains and a D2E variant that avoids deformation-related regularity artifacts.","lead":"This paper gives a mathematical framework for training neural networks to solve PDEs on any of a family of changing domains, by deforming each domain to a standard one and measuring distances between problems. It provides convergence guarantees for a common case (Poisson equations on star-shaped regions) and shows a version of the method preserves linearity for fast iterative solvers.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5's proof of the continuity of H (the key hypothesis for the convergence theorem) relies on Arzelà–Ascoli for an L2-compact set and on a uniform L∞ bound that dX compactness does not provide; as written, this proof is invalid.","rationale":"The paper's stated milestone is a rigorous convergence analysis; Theorem 5 is the hinge that supplies continuity of H for the one fully worked PDE example. Without a valid proof of Theorem 5, Theorems 2 and 3 cannot be invoked for Poisson on star domains, so the convergence guarantee is conditional on an unproven premise. I focused on this rather than the Lemma 2 algebraic slip (Eq. 100) because Lemma 2's statement is standard and easily repaired by orthogonal projection facts, whereas Theorem 5's gaps sit exactly at the boundary between the metric construction and the PDE estimate. The concern is not that the theorem is obviously false. Indeed, the Poisson map is continuous from L2 to H^1_0 with uniform constants for this family, and L2 composition with uniformly converging diffeomorphisms is continuous without equicontinuity. So there is a plausible repair route. But the repair is substantial: the Arzelà–Ascoli invocation in Eq. (104) is not a harmless typo, and the Mf in Step 3 cannot be inferred from dX compactness. A referee must see the replacement argument before the paper can be called definitive. The numerical experiments provide real support that the method works in practice, but they do not settle the continuity question, because training error and test error at fixed discretization do not test the asymptotic convergence theorem. Thus my read leaves the reader's CONDITIONAL verdict unchanged: the central contribution stands if the proof of Theorem 5 is repaired as described, and the concrete test above determines whether that repair is possible.","tokens_in":23389,"tokens_out":12559,"duration_ms":112598,"concrete_test":"Independently re-derive Step 1 of Appendix F without Arzelà–Ascoli: assume only a_n := f_Ωn∘D[Ωn] → a in L2(Ω0), D[Ωn]→D[Ω] uniformly with bounded Jacobians, and prove a_n∘D[Ωn]^{-1}→a∘D[Ω]^{-1} in L2(Ω) by splitting into (a_n−a)∘D_n^{-1} plus a∘(D_n^{-1}−D^{-1}) and using translation continuity in L2. Then re-derive Step 3 replacing the uniform L∞ bound on f with the H^1_0 stability estimate ∥u_n−u∥_{H^1_0(V)} ≤ C(∥f_n−f∥_{L2(V)} + dU(Ω_n,Ω)) using uniform Poincaré constants. If both replacements succeed, the gap is repairable and the CONDITIONAL publication stands; if either fails, Theorem 5 is false or requires extra hypotheses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that for 2D Poisson on smooth star domains the solution map H is continuous with respect to dX, because Theorems 1–3 turn continuity of H into a neural-network approximation guarantee. The proof of Theorem 5 (Appendix F) is the only place this continuity is verified, and it is not established as written. In Step 1, Eq. (104) asserts that K2 = π2(σ(K)) is uniformly equicontinuous 'by Arzelà–Ascoli' because K2 is compact in L2(Ω0). But L2-compactness never implies equicontinuity: one can put L2-small spikes of arbitrarily high oscillation or amplitude into a compact L2 set. Hence the displayed limit ∥f_Ωn − f_Ωn∘D[Ωn]∘D[Ω]^{-1}∥_{L2(J)}→0 does not follow from the stated assumptions. In Step 3, the pointwise bound |u_Ωn(x)| ≤ C1 + C2|ln dist(x,∂Ωn)| uses a single constant Mf bounding |f_Ωn| over all n. Compactness in dX only bounds L2 norms of the deformed functions, not their L∞ norms, so this dominant can be unbounded and the Vitali-convergence argument collapses. The theorem may be true — an H^1-stability route would likely repair it — but as written the paper's 'rigorous convergence analysis' is missing its most load-bearing premise. This is exactly the premise the reader flagged, and it has not been discharged.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a deformation-based framework for learning solution mappings of PDEs defined on varying domains. It equips the union of function spaces on a family of domains with a metric via a reference domain and deformation maps, proves a general approximation theorem for continuous metric-to-Banach mappings, and introduces two subframeworks, D2D and D2E. For 2D Poisson on smooth star domains, the paper claims to verify all standing assumptions and thereby provide a rigorous convergence analysis. Numerical experiments on polygonal domains, smooth star domains, locally deformed domains, and a pipe flow illustrate the method.","tokens_in":23673,"tokens_out":10118,"duration_ms":95143,"significance":"The framework is a useful conceptual contribution: it replaces the Banach-space assumption of classical operator learning with a metric-space setting and cleanly separates the D2D and D2E design choices. The main approximation theorem (Theorem 1) is self-contained, and the star-domain discretization in Section 4.1 is thoughtfully constructed. The numerical results and the public code are also strengths. However, the paper's headline claim of a rigorous convergence analysis depends critically on Theorem 5, whose proof is not valid as written; the same is true for part of the proof of Lemma 2. The underlying ideas are defensible and likely repairable, but the manuscript currently overstates its theoretical guarantees.","major_comments":[{"comment":"The assertion that K2 = π2(σ(K)) is uniformly equicontinuous 'by Arzelà–Ascoli' because it is compact in L2(Ω0) is false: L2-compactness does not imply equicontinuity, since a sequence of narrow spikes with growing height and decreasing support can be L2-compact while having no equicontinuous subsequence. Consequently, the displayed limit ∥fΩn − fΩn∘D[Ωn]∘D[Ω]−1∥_{L2(J)} → 0 does not follow from the stated dX-convergence assumptions. This is load-bearing because Theorem 5 is the only place where continuity of H is verified for the Poisson star-domain example, and that continuity is the hypothesis underlying Theorems 1–3 and the claimed convergence guarantee.","section":"Appendix F, Step 1, Eq. (104)"},{"comment":"The dominant bound |uΩn(x)| ≤ C1 + C2|ln dist(x, ∂Ωn)| uses a single constant Mf bounding |fΩn| for all n. Compactness of K in dX gives only L2 convergence of fΩn∘D[Ωn] on Ω0, not a uniform L∞ bound on the original functions. Without such a bound, the functions gn need not converge in L2(V), and the Vitali argument in Step 3 collapses. The theorem may be true, but the present proof needs a different route, for example uniform H1 estimates followed by Sobolev embedding.","section":"Appendix F, Step 3, around Eq. (111)"},{"comment":"The displayed chain '∥f−Pn(f)∥² = Σ∫(f−avg f)² ≤ Σ∫(f−avg f)² + (avg g−avg f)² = ∥f−Pn(g)∥²' is algebraically incorrect: the final equality omits the factor m(Ωn_i) multiplying (avg_i g − avg_i f)². The intended statement, that the piecewise-constant projection is the closest point in the L2 norm, is true, but this proof is wrong as written. Since Lemma 2 feeds into Theorem 4 and hence into Assumption 3, the proof must be corrected.","section":"Appendix E, Eq. (100)"},{"comment":"The discretization actually implemented in Algorithm 4.1 uses point evaluations ϕ2_q2(fΩi∘D[Ωi]) = (f(x1),...,f(x_q2)), whereas Lemma 2 is proved for the piecewise-constant averaging encoder. The text after Eq. (52) asserts without proof that for piecewise continuous functions point evaluation 'can simply be applied.' This does not follow from the L2-compactness used in Assumption 3: a sequence of narrow spikes of growing height is L2-compact, yet its point values at fixed locations can diverge. The convergence guarantee supplied by Theorems 1–4 therefore does not cover the encoder used in the numerical method; either the point-evaluation discretization must be proved to satisfy Assumption 3 under explicit extra conditions, or the algorithm should be reformulated with the averaged encoder.","section":"Section 4.1 and Algorithm 4.1"}],"minor_comments":[{"comment":"The abstract says the framework has 'three important features', while Section 1 and Section 6 list four; the D2D-versus-D2E comparison item is missing from the abstract.","section":"Abstract"},{"comment":"The statement that K is 'a compact set in F_{Ω∈U} C(Ω)' should specify the topology: since C(Ω) is not given its own metric, compactness should be understood with respect to the metric dX from Definition 1 restricted to the union, and this should be stated explicitly.","section":"Theorem 5, Section 4.1"},{"comment":"The averaging discretization in Eq. (51) and the point-evaluation discretization in Eq. (52) are both presented as valid choices, but the paper gives no regularity condition under which the point-evaluation version is guaranteed to satisfy Assumption 3; this connection between theory and implementation should be clarified.","section":"Section 4.1, Eq. (51) and Algorithm 4.1"}],"recommendation":"major_revision","confidential_remarks":"The technical gaps are substantial but appear repairable; I see no evidence of circularity or inappropriate citation in the central derivation. The paper's claim to provide the first rigorous convergence analysis for varying-domain operator learning should be softened until Theorem 5 is repaired."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nBottom line: the metric-space formulation is a real contribution and the D2E variant is worth having, but the headline claim of a \"rigorous convergence analysis\" isn't delivered. The proof of the load-bearing continuity theorem (Theorem 5) is invalid as written, and there's a smaller algebraic error in Lemma 2.\n\nWhat's actually new: treating the union of function spaces over varying domains as a metric space via a deformation system, then representing the PDE solution map as a continuous metric-to-Banach map approximated by the tensor-product argument of Theorem 1. That subsumes Geo-FNO, DIMON, and their earlier [40] under one umbrella, and it's cleanly modular: metric, deformation, discretization, and continuity of H are separated. D2E — learning the extended solution on a fixed box and restricting — is a genuinely practical fix: the surrogate never composes with D[Ω]^{-1}, so it doesn't inherit the deformation's nondifferentiability. The numerics support that: D2D and Geo-FNO show kinks along lines from vertices to centroid, D2E doesn't; errors are comparable between D2D and D2E (~3.5%) and clearly better than Geo-FNO (6.5%). The linearity-preservation application to HIM with 8.4x speedup is a nice extra. Citation pattern is honest; the paper credits DIMON and [40] for the D2D-style idea, and self-citations are ancillary. Code is public.\n\nSoft spots. The big one: Theorem 5. Continuity of H w.r.t. dX is what makes Theorems 1–3 yield convergence, and the proof doesn't go through. Step 1 claims K2 is uniformly equicontinuous by Arzelà–Ascoli because it's compact in L2(Ω0). L2 compactness implies no such thing; oscillatory spikes can live in an L2-compact set. Step 3 needs a uniform L∞ bound on sources, but dX compactness gives only L2 bounds of the deformed functions. So u_{Ωn} → u_Ω in L2(V) is not established. The theorem may be true — an H^1-stability route would likely fix it — but as written the \"rigorous\" convergence for the star-domain flagship example is unproven. That's load-bearing, and the abstract overstates what's shown.\n\nLemma 2's proof has a false algebraic identity (Eq. 100); the lemma statement is true and easy to repair. Minor by itself, but combined with Theorem 5 it says the proofs need careful checking.\n\nThe repeated claim that D2E is \"comprehensively better\" is stronger than the evidence: same accuracy, better regularity. Say \"avoids a regularity issue at modest cost\" and stop there.\n\nThe reader's worry about an approximation property for Y in Theorem 1 is a non-issue; the injective tensor product argument works for any Banach Y. Don't send them chasing that.\n\nWho it's for: people working on geometry-aware neural operators and surrogates for PDEs on varying domains. The framework deserves to be read and cited; it's a genuine step, and a serious referee should engage. Recommendation: send to review with major revision — repair Theorem 5 or explicitly demote it to an assumption, fix Lemma 2, soften the D2E claim.","headline":"A genuinely useful unifying framework and a practical D2E variant, but the 'rigorous convergence analysis' rests on a continuity proof (Theorem 5) that does not work as written, so the flagship claim is unproven until repaired.","tokens_in":24236,"tokens_out":4487,"would_cite":true,"duration_ms":38838,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N99","68T07","41A65"],"pacs":[],"model":"deepseek-v4-flash","headline":"PDE solution maps on varying domains can be represented as metric-to-Banach maps and approximated by neural networks with a proven convergence bound, provided the domains admit a continuous deformation and a vanishing-error discretization.","keywords":["operator learning","PDEs on varying domains","deformation-based framework","metric-to-Banach approximation","convergence analysis","neural operators","star domains","Poisson equation"],"falsifier":"Take a sequence of smooth star domains $\\Omega_n$ and source terms $f_{\\Omega_n}$ converging to $(\\Omega, f_{\\Omega})$ in $d_X$, solve each Poisson problem with a high-accuracy finite-element method, and check whether $u_{\\Omega_n}\\circ D[\\Omega_n] \\to u_{\\Omega}\\circ D[\\Omega]$ in $L^2(\\Omega_0)$. Any convergent sequence whose solutions fail to converge in $L^2$ would disprove the continuity claim; alternatively, with a fixed wide network, increasing only the encoder resolution $n$ should drive the test error toward zero, so a persistent positive error floor would contradict the convergence guarantee.","tokens_in":23155,"feed_emoji":"📐","tokens_out":12444,"duration_ms":91540,"temperature":0.7,"pith_summary":"Learning solution operators of PDEs usually assumes a fixed domain. This paper removes that assumption by turning the collection of functions defined over a family of varying domains into a metric space: each domain $\\Omega$ is pulled back to a standard reference domain $\\Omega_0$ by a deformation map $D[\\Omega]$, and the distance between two functions on different domains combines the domain distance with the $L^2$ distance of their pulled-back representatives. Within this metric, the PDE solution map $f_\\Omega \\mapsto u_\\Omega$ becomes a continuous metric-to-metric map, which the authors represent as a continuous metric-to-Banach map and approximate by neural networks with an error bound (Theorem 1). For 2D Poisson on smooth star domains, the paper verifies all assumptions, so learning the operator by either of two subframeworks, D2D (learning the deformed solution $u_\\Omega \\circ D[\\Omega]$) or D2E (learning a zero-extended solution on a large box), converges as the discretization and network size grow. The paper presents this as a rigorous convergence analysis for geometrically varying-domain operator learning that does not require diffeomorphic domains.","feed_headline":"New proof: solution maps for PDEs on changing domains are learnable","feed_subtitle":"One deformation metric turns varying-domain PDE learning into a provably convergent approximation problem.","key_machinery":"The load-bearing object is the deformation system $D:\\mathcal{U}\\to B(\\Omega_0,\\mathbb{R}^d)$, $\\Omega\\mapsto D[\\Omega]$, a bijective (not necessarily continuous) map from a fixed standard domain $\\Omega_0$ to each $\\Omega$, together with the metric $d_X(f_{\\Omega_1}, f_{\\Omega_2}) = d_{\\mathcal{U}}(\\Omega_1,\\Omega_2) + \\|f_{\\Omega_1}\\circ D[\\Omega_1] - f_{\\Omega_2}\\circ D[\\Omega_2]\\|_{L^2(\\Omega_0)}$. This metric identifies the union $X$ with $\\mathcal{U}\\times B(\\Omega_0)$ isometrically, so the solution map $H$ can be pulled back to a metric-to-Banach map $\\hat{H}$ and approximated via Theorem 1. The second key mechanism is the discretization assumption (Assumption 3): encoders $\\phi_n$ and reconstructors $\\psi_n$ with projection error $\\sup_{x\\in K} d_X(x, \\psi_n(\\phi_n(x))) \\to 0$; this replaces the Schauder-basis step in the MIONet approximation theorem and is what makes the error bound quantitative. For star domains, the deformation is the radial map $D[\\Omega](x)=c_{\\Omega}+b_{\\Omega}(x/\\|x\\|)x$, and the discretization samples the boundary radius function and the pulled-back source term.","core_discovery":"The paper's central claim is that a PDE solution mapping $H:K\\subset X\\to X$, where $X$ is the union of bounded Borel function spaces over a domain class $\\mathcal{U}$, can be represented as a continuous metric-to-Banach mapping whenever $\\mathcal{U}$ admits a continuous deformation system and $X$ admits a discretization with vanishing projection error on compact sets. Theorem 1 then gives an $\\epsilon$-approximation by a neural network of the form $u\\circ \\tilde{G}\\circ \\phi_q$, where $\\phi_q$ is the encoder, $\\tilde{G}$ is a continuous map from $\\mathbb{R}^{\\kappa(q)}$ to $\\mathbb{R}^p$, and $u$ is a continuous map from $\\mathbb{R}^p$ into the target Banach space ($L^2(\\Omega_0)$ for D2D, $L^2(V)$ for D2E). Theorems 2 and 3 convert this into two representation formulas for $H$: learn $u_{\\Omega}\\circ D[\\Omega]$ and pull back, or learn the zero-extended $u_{\\Omega}$ and restrict. For 2D Poisson on smooth star domains the paper verifies the metric, deformation, discretization, and continuity assumptions (Theorem 5), and supports the theory with experiments showing relative $L^2$ errors near 3% on polygonal and smooth domains, plus a linearity-preserving variant that accelerates Gauss-Seidel iteration by a factor of 8.4.","pith_inferences":["The modular structure suggests an immediate test bed: verify the continuity step for other PDE families (for example, the Navier-Stokes pipe flow shown numerically) and the same convergence guarantee would follow without changing the metric or discretization machinery.","Because the metric $d_X$ is built from the $L^2$ norm on the reference domain, the guarantees are for $L^2$ convergence of solutions; pointwise or $H^1$ convergence would require a stronger metric or extra regularity assumptions, which the paper does not address.","Replacing the generic discretization in Assumption 3 by one with a known interpolating error rate would sharpen Theorem 1 from an existential error bound to an explicit convergence rate; the paper leaves this refinement open.","The D2E zero-extension on a bounding box suggests that the framework could transfer to geometry-agnostic encoders (point clouds or implicit neural representations) if those encoders satisfy the same vanishing projection error condition."],"forward_implications":["A single model can cover all homeomorphic domains in a class, not only diffeomorphic ones, because only continuity of the deformation system is required.","The deformation map itself need not be continuous, so locally deformed domains can be handled by stitching an identity map on the main domain to a localized deformation.","D2D predictions can inherit singularities from the inverse deformation map, while D2E predicts solutions directly and avoids these regularity artifacts.","For linear PDEs, a linearity-preserving neural operator such as MIONet keeps the surrogate linear in the source term, so it can act as a preconditioner inside a hybrid iterative solver.","Once all assumptions are verified for a problem class, the test error is guaranteed to tend to zero as discretization resolution and network size grow."],"supporting_citations":[{"why":"Supplies the MIONet multiple-input operator architecture and its tensor-product approximation theory, which Theorem 1 generalizes from Banach spaces to metric spaces.","marker":"[15]"},{"why":"Provides the injective tensor product representation $C(K,Y)=C(K)\\hat{\\otimes}_\\varepsilon Y$ used in the proof of Theorem 1 to decompose the operator into scalar functions times elements of $Y$.","marker":"[32]"},{"why":"Dugundji's extension theorem is used in the proof of Theorem 1 to extend the approximating scalar functions from the compact set $K$ to the whole metric space $X$.","marker":"[3]"},{"why":"Gives the uniform convergence of Green's functions on planar domains that is the key analytic input in the proof of Theorem 5's continuity of the Poisson solution map.","marker":"[16]"},{"why":"Provides the Green's function existence and representation formula for the Poisson problem used to write $u_\\Omega$ and compare solutions on nearby star domains.","marker":"[18]"},{"why":"Geo-FNO, the deformation-based Fourier neural operator baseline the paper positions within the D2D subframework and compares against numerically.","marker":"[19]"},{"why":"Describes the hybrid iterative method that motivates the paper's linearity-preserving property for MIONet and is used in the one-million-mesh experiment.","marker":"[14]"},{"why":"The authors' earlier version of this work that established the D2D subframework with MIONet for homeomorphic domains and whose theory this paper extends.","marker":"[40]"},{"why":"DIMON, another deformation-based method that the paper classifies as D2D and uses as a point of comparison for the proposed D2E subframework.","marker":"[41]"}],"fun_headline_variants":["Proof: PDE solution maps learnable on non-diffeomorphic domains","Deformation framework makes varying-domain PDE learning convergent","Learn PDE solutions on any homeomorphic domains with one model","Varying-domain PDEs: deformation turns solution maps into neural nets","No diffeomorphism needed: PDE maps learnable via deformation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is continuity of the PDE solution map with respect to the deformation metric: if $f_{\\Omega}\\mapsto u_{\\Omega}$ is not continuous when both the domain and the data are close in $d_X$, the approximation theorem has nothing to act on; the paper's proof of that continuity for 2D Poisson relies on a compactness-to-equicontinuity step that goes beyond what the stated $L^2$-based metric alone provides.","fun_headline_variants_meta":{"raw":{"variants":["Proof: PDE solution maps learnable on non-diffeomorphic domains","Deformation framework makes varying-domain PDE learning convergent","Learn PDE solutions on any homeomorphic domains with one model","Varying-domain PDEs: deformation turns solution maps into neural nets","No diffeomorphism needed: PDE maps learnable via deformation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000283,"raw_usage":{"total_tokens":1779,"prompt_tokens":1160,"completion_tokens":619,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":776,"completion_tokens_details":{"reasoning_tokens":535}},"tokens_in":776,"tokens_out":619,"duration_ms":5539,"temperature":1.0,"reasoning_tokens":535,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:25:11.847781+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a sequence of smooth star domains $\\Omega_n$ and source terms $f_{\\Omega_n}$ converging to $(\\Omega, f_{\\Omega})$ in $d_X$, solve each Poisson problem with a high-accuracy finite-element method, and check whether $u_{\\Omega_n}\\circ D[\\Omega_n] \\to u_{\\Omega}\\circ D[\\Omega]$ in $L^2(\\Omega_0)$. Any convergent sequence whose solutions fail to converge in $L^2$ would disprove the continuity claim; alternatively, with a fixed wide network, increasing only the encoder resolution $n$ should drive the test error toward zero, so a persistent positive error floor would contradict the convergence guarantee.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the MIONet multiple-input operator architecture and its tensor-product approximation theory, which Theorem 1 generalizes from Banach spaces to metric spaces."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the injective tensor product representation $C(K,Y)=C(K)\\hat{\\otimes}_\\varepsilon Y$ used in the proof of Theorem 1 to decompose the operator into scalar functions times elements of $Y$."},{"cited_title":"Dugundji","cited_arxiv_id":null,"evidence_quote":"Dugundji's extension theorem is used in the proof of Theorem 1 to extend the approximating scalar functions from the compact set $K$ to the whole metric space $X$."},{"cited_title":"Kalmykov and L","cited_arxiv_id":null,"evidence_quote":"Gives the uniform convergence of Green's functions on planar domains that is the key analytic input in the proof of Theorem 5's continuity of the Poisson solution map."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Green's function existence and representation formula for the Poisson problem used to write $u_\\Omega$ and compare solutions on nearby star domains."}],"review_version":1}