{"id":"ef475907-0ca7-45cc-9f04-b9c5f44a2175","arxiv_id":"2412.04409","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A stabilized FEM with POD and autoencoder compression plus a learned solution operator reconstructs PDE solutions in unique continuation problems, with a first-order H1 error estimate in the linear case and numerical demonstrations for nonlinear cases.","lead":"Researchers combined data compression, neural networks, and finite element methods to reconstruct solutions of partial differential equations from partial interior measurements when boundary conditions are unknown. The approach turns a notoriously unstable inverse problem into a lower-dimensional optimization and comes with an error bound in the linear case.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Nonlinear experiments are self-referential: reference solutions are generated by the same operator network used in reconstruction, so the numerical claims do not yet demonstrate recovery of a true FE solution of the nonlinear inverse problem.","rationale":"The paper's theoretical core is Proposition 3.2, and I checked the proof in §3.3. The stabilized form m_h supplies a mesh-independent discrete stability estimate (Lemma 3.1), and the error estimate follows by splitting the error into the FE-mode approximation (I-I_h) and the stabilized projection residual. The duality argument in Lemma 3.1 is standard: Galerkin orthogonality, elliptic regularity, and trace inequalities give the h-weighted bound on ||(I-I_h)bv||_omega. I do not find a missing term or a circular step in this part; the constant's exponential dependence on N is acknowledged and consistent with the ill-posedness. Thus the linear contribution is credible as stated. The overall paper, however, makes a second central claim: that composing an autoencoder decoder with a learned FE solution operator solves the nonlinear inverse problem (4.12). The support for this claim is numerical, and all nonlinear inverse experiments use a self-generated reference: the 'reference solution' is the operator network's own output for the chosen coefficient vector, not a FEniCS solution of the PDE. The paper explicitly concedes this in the final remarks of §5.3. This matters because the reference coefficient p14=10 is far outside the N(0,0.09) training distribution of the operator network, so the network's output there is extrapolation; a method that inverts its own extrapolated output tells us little about whether it inverts the true PDE solution operator. The linear experiment (Figure 6) is not affected, since its reference is a genuine FE superposition, but the nonlinear demonstrations are. The reader's weakest assumption—representability of the target boundary data by the decoder—is a related but distinct concern. Table 3 already shows the decoder-based reconstructions are worse than the coefficient-space one, consistent with representability failure for the out-of-distribution target. However, even for a representable target, the self-referential validation would not establish correctness. The concrete test I propose separates the two issues: use a true FEniCS reference (both in-distribution and the out-of-distribution p14=10 case) and compare. This single experiment would tell whether the circularity is benign (network is accurate enough that self-referential and true-FE errors coincide) or whether the numerical claims overstate what is demonstrated. Because the linear theory stands and the practical gap is already flagged by the authors, I do not recommend changing the CONDITIONAL verdict; I agree with the reader's overall assessment but locate the load-bearing concern in the self-referential validation rather than in the representability assumption alone. Hence UNCHANGED and partial agreement.","tokens_in":24349,"tokens_out":14431,"duration_ms":180838,"concrete_test":"Use FEniCS to solve the actual nonlinear FE problem for the energy (5.7) with the same 21-coefficient boundary data used in Figures 7 and 11–13 (p14=10, others zero), producing a true reference solution u_ref and data u0 = u_ref|_omega plus noise. Then run the trained operator-network inverse optimization over coefficients and the decoder latent-space optimization over Z, and report ||u_opt - u_ref|| in H1 and L2. Also report the same errors for a target inside the autoencoder training distribution, e.g., coefficients generated from (5.10) with x ~ U(-2,18). If the reconstruction errors are comparable to the self-referential MSEs in Table 3, the practical claims survive; if they are much larger, the nonlinear demonstrations are not evidence of solving (2.5).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The linear theory in §3 (Proposition 3.2) appears coherent: the stabilized reduced-order projection has an O(h) H1 error bound with constant depending on N and the stability ratio, and I find no gap in the duality argument of Lemma 3.1. The load-bearing weakness is in the claimed practical extension to nonlinear problems. All nonlinear demonstrations (Figures 7 and 11–13, Table 3) use as reference the output of the very operator network that is later used for reconstruction: the reference is obtained by setting p14=10 in the network input and reading out its FE-DoF output, and the data u0 is generated from that output. The paper itself flags this in the final remarks of §5.3: \"taking some output of the method under consideration as the reference solution, as is done here, is typically not a proper choice since it is too idealized.\" Consequently, a successful latent-space or coefficient-space fit can look perfect even if the operator network is systematically wrong on that input, which is plausible here since p14=10 lies far outside the training distribution N(0,0.09) used in §5.2. The experiments therefore do not establish that the ML pipeline solves the original problem (2.5) for a true FE solution of the nonlinear PDE. The representability limitation noted by the reader is real (Table 3 shows decoder reconstructions miss the reference), but the more fundamental gap is that no nonlinear test measures error against a ground-truth FE solution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a computational framework for the unique continuation inverse problem in which, in addition to a bulk measurement u0 in a subdomain ω, one has access to a large set of boundary data GS from a low-dimensional family G. The authors combine POD and autoencoders to parametrize the boundary data in a latent space, train an operator network that approximates the FE solution map from boundary coefficients to finite element degrees of freedom, and then solve the inverse problem by unconstrained optimization in the latent space. For a linear model problem, they prove H1 error estimates for a stabilized reduced-order projection onto a space spanned by approximately computed harmonic basis functions, with first-order convergence in the mesh size h. For nonlinear problems, they present numerical demonstrations for a nonlinear diffusion-type operator with linear and nonlinear boundary-data models.","tokens_in":24644,"tokens_out":3633,"duration_ms":38077,"significance":"The linear theory is the main rigorous contribution: Proposition 3.2 appears to provide a clean O(h) H1 error bound for the stabilized reduced-order projection, extending earlier finite-dimensional-trace unique continuation results to a practical reduced-order setting. The paper also gives a concrete and reproducible numerical pipeline built on publicly available code, and it is honest about several limitations, including the fact that the nonlinear inverse tests use network-generated reference solutions. The practical claim that the ML framework solves the nonlinear inverse problem (2.5) is plausible but is not established by the current experiments, because the validation is partly self-referential and no error bound links the autoencoder/operator-network approximation to the inverse reconstruction error.","major_comments":[{"comment":"The nonlinear inverse problems are validated against a reference solution generated by the same operator network used in the reconstruction: the reference is the network output for the input p14 = 10, all other pn = 0, and the data u0 is produced from that output. The authors explicitly acknowledge in the final remarks of Section 5.3 that 'taking some output of the method under consideration as the reference solution, as is done here, is typically not a proper choice since it is too idealized.' This is a load-bearing weakness: the optimizer only needs to invert the operator network, not to find a true FE solution of the nonlinear PDE, so the numerical results in Figures 7 and 11–13 do not demonstrate that the composed map (ϕu,N,h ◦ ϕde) solves the original inverse problem (2.5). The nonlinear experiments should be repeated with reference solutions computed by an independent FE solver and with errors measured against those solutions.","section":"Section 5.3, Figures 7 and 11–13, Table 3"},{"comment":"The latent-space formulation assumes that the POD coefficients of the true boundary data lie in the image of the decoder, or at least within a small δ-neighborhood of it. The paper provides no bound linking the autoencoder reconstruction error and the operator-network approximation error to the reconstruction error of the inverse problem. If the target boundary trace is not representable by the decoder, the objective (4.12) has no minimizer near the true field and the logarithmic ill-posedness of unique continuation returns. This representability is an unverified engineering assumption, and Table 3 shows that it can fail substantially (the 'sd = 0.15' decoder gives MSE 1.99e-2 on ω). A numerical test with a reference solution generated from boundary data outside the decoder's training distribution, or an error decomposition along the lines of data approximation + autoencoder error + operator-learning error + optimization error, is needed to support the nonlinear practical claims.","section":"Eq. (4.12) and Section 4.1"}],"minor_comments":[{"comment":"The first inequality in the chain should read ∥bv∥_{R^N} ≲ ∥Ibv∥_ω rather than ∥Ibv∥^2_ω; as written the displayed chain is dimensionally inconsistent and does not match the subsequent argument.","section":"Eq. (3.28)"},{"comment":"The inequality attributed to 'Galigliardo-Nirenberg' should be 'Gagliardo-Nirenberg'.","section":"Section 3.2"},{"comment":"The statement that the constant in Proposition 3.2 depends 'on higher order Sobolev spaces of g' is vague; the authors should state the precise norm on g and the explicit dependence on N and λ_min, since this constant is the quantity that controls the ill-posedness of the problem.","section":"Eq. (3.52)"},{"comment":"The term I b v in the boundary stabilization uses the coefficient extraction operator b·, but the domain and action of b· are only defined implicitly after (3.13); this should be stated explicitly before the stabilized form is introduced.","section":"Eq. (3.24)"},{"comment":"Since the reference solution uses p14 = 10 while the operator network was trained with inputs from N(0, 0.09), it would be useful to report the training distribution's range for p14 and to explain why this out-of-distribution input was chosen for the reference.","section":"Figures 7 and 11–13"}],"recommendation":"major_revision","confidential_remarks":"The linear analysis is solid and the paper is honest about its own limitations, which is commendable. However, the nonlinear validation is currently self-referential in a way that materially weakens the central practical claim. I would encourage the editor to invite a revision in which the nonlinear experiments are rerun against independently computed FE reference solutions, and in which the representability/error-propagation issue for the decoder is addressed either theoretically or by more probing numerical tests."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the linear analysis is the real contribution, and it holds up on inspection; the nonlinear ML experiments are a pipeline demo whose numerical validation is self-referential, and the paper itself admits this. I agree with the stress-test note, and I would add that the same caveat applies to Figure 7, not just the decoder tests.\n\nWhat is new: Proposition 3.2 gives an O(h) H1 error bound for a stabilized reduced-order projection onto a space spanned by approximate FE solutions of the basis functions, under the finite-dimensional trace assumption. That extends the earlier finite-dimensional-trace results [12,13] to a reduced-order setting with an approximate basis and a stabilization term that makes the discrete stability constant independent of the mesh. I read the proofs of Lemma 3.1 and Propositions 3.1–3.2 and did not find a gap; the duality argument is standard, and the notation gaps are cosmetic. The exponential stability constant example with the unit disc is illuminating and useful for intuition.\n\nThe paper also does a reasonable job of assembling a pipeline: POD for linear compression, autoencoder for nonlinear compression, operator network for the FE solution map, then latent-space optimization for the inverse problem. The data-compression discussion in Section 4.1 is sensible, and the experimental setup is described in enough detail to reproduce the broad strokes.\n\nWhere it is soft: the nonlinear experiments do not validate the method against a true FE solution. The reference solution in Figures 7 and 11–13 is the operator network output for p14 = 10; the data are generated from that output; and the same network is used for reconstruction. The paper says in Section 5.3 that using the method's own output as reference is \"typically not a proper choice since it is too idealized.\" That is exactly right. It means the reported MSEs do not demonstrate recovery of a solution to the original problem. The fact that p14 = 10 lies far outside the training distribution N(0, 0.09) makes the network error at that input unquantified and potentially large. The representability issue the reader flags is real but secondary: if the latent decoder cannot represent the target boundary data, the latent-space objective has no good solution, and the paper provides no learned error bound connecting training error to reconstruction error. On top of that, there are no error bars and no released code. None of this undermines the linear theory, but it does undercut the abstract's implication that the nonlinear approach is fully validated.\n\nWho this is for: people working on stabilized FEM for unique continuation, finite-dimensional trace stability, and learned or reduced-order solvers. The linear section deserves serious referee time. The nonlinear section needs new experiments with a ground-truth FE reference generated independently, plus code or data release.\n\nRecommendation: send it to peer review. Ask the authors to rerun the nonlinear tests against true FE solutions and to either quantify the operator network error at the reference coefficients or soften the claims.","headline":"The linear reduced-order error analysis is solid and new; the nonlinear experiments are self-referential and need to be re-run against true FE solutions before the practical claims can be trusted.","tokens_in":25176,"tokens_out":2381,"would_cite":true,"duration_ms":26149,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","65N21","35R30"],"pacs":[],"model":"deepseek-v4-flash","headline":"A stabilized finite element scheme makes unique continuation first-order accurate when boundary data are low-dimensional, and a learned decoder-plus-operator network extends the idea to nonlinear PDEs.","keywords":["unique continuation","inverse problems","stabilized finite element method","proper orthogonal decomposition","autoencoder","operator learning","reduced order model","ill-posed problems"],"falsifier":"Pick a target solution whose boundary trace is a known linear combination of POD modes with coefficients inside the training support, generate its bulk observation $u_0$ in $\\omega$, and run the latent-space minimization; the linear theory predicts error $O(h\\|u_0\\|_\\omega + \\delta)$, so if reconstruction error is large even when the decoder's reconstruction error $\\delta$ is tiny, the learned decoder-operator composition is the point of failure. A more direct check is to compute the decoder's boundary reconstruction error for the specific target trace and the operator network's output error; the paper's claim is stable reconstruction follows once those errors are small, so a target with small decoder error and small operator error that still yields large interior error would falsify the method's premise.","tokens_in":24094,"feed_emoji":"🔍","tokens_out":5697,"duration_ms":54361,"temperature":0.7,"pith_summary":"The paper tries to establish that the severely ill-posed problem of reconstructing a PDE solution from a bulk interior measurement, when the missing boundary data are known to lie on a low-dimensional set, can be made stably and optimally convergent. It proves for a linear Laplace model that a stabilized projection onto the span of finitely many precomputed modes converges in $H^1$ with error $O(h)$ times the data norm, independent of the near-instability of the continuous inverse problem. It then builds a composition of an autoencoder decoder and a learned finite element solution operator, so the same inverse problem becomes an unconstrained minimization over a latent space. If the method is right, it offers a template for turning abundant historical boundary observations into a stable prior for otherwise ill-posed data-assimilation problems.","feed_headline":"Low-dimensional data tames ill-posed PDE reconstruction","feed_subtitle":"POD, autoencoders, and learned FE operators turn missing boundary data into a latent-space optimization.","key_machinery":"The load-bearing object is the stabilized bilinear form $m_h(v,w) = (v,w)_\\omega + h^{-1}(v - I\\hat v, w - I\\hat w)_{\\partial\\Omega} + h(\\nabla_T(v - I\\hat v), \\nabla_T(w - I\\hat w))_{\\partial\\Omega} + \\sum_{F\\in\\mathcal{F}_h} h([\\nabla v],[\\nabla w])_F$, where $I$ maps coefficient vectors onto the continuous harmonic modes and $I_h$ onto their Nitsche-computed finite element approximations. The first stabilization term controls the distance of the discrete solution to the finite-dimensional trace set $G$ in the $H^{1/2}$ norm, and the jump-penalty term supplies enough coercivity that the coefficient bound $\\|\\hat v\\| \\lesssim \\|v\\|_{m_h}$ holds uniformly in $h$, which is exactly what the plain $L^2(\\omega)$ projection lacks. This stabilizes the reduction step in the linear theory. In the nonlinear pipeline, the same idea is replaced by a learned composition of decoder and operator network, so the constraint 'boundary data in $G$' is enforced structurally rather than by a penalty.","core_discovery":"On its own terms, the paper's central result is Proposition 3.2: for boundary data drawn from an $N$-dimensional set $G$, the stabilized reduced-order projection $u_{N,h}$ defined by $m_h(u_{N,h}, v) = (u_0, v)_\\omega$ satisfies $\\|u_N - u_{N,h}\\|_{H^1(\\Omega)} \\lesssim h\\|u_0\\|_\\omega$, with constants growing with $N$ but not with the mesh size. The mechanism is a stabilized bilinear form that adds a tangential boundary control and a face-jump penalty, which restores the discrete stability that the naive Galerkin projection lacks because finite element functions do not satisfy unique continuation. The paper further claims that in practice the dimension reduction can be learned: POD gives a linear basis of collective boundary data, an autoencoder finds a nonlinear latent parametrization, and an operator network approximates the FE solution map from coefficients to degrees of freedom, so the inverse problem is solved by optimizing over the latent space. The nonlinear examples show the composition reconstructs the reference solution, and that a decoder trained on unperturbed data outperforms one trained on noisy data.","pith_inferences":["The theoretical guarantee stops at the linear reduced-order projection; extending Proposition 3.2 to the nonlinear learned pipeline would require a bound linking autoencoder and operator training errors to the reconstruction error, which the paper does not provide.","A testable consequence: for targets whose boundary trace lies in $G$ but is poorly reconstructible by the decoder, the method should fail even though the linear theory would succeed; comparing the decoder-free coefficient-space optimization with latent-space optimization on the same reference would isolate the representability loss.","The stabilization analysis suggests an alternative design for the learned pipeline: instead of decoding to coefficients and then applying the operator network, one could train the operator network on stabilized FE solutions generated by $m_h$, which might transfer the uniform-in-$h$ stability to the nonlinear case.","Because the stability constant grows exponentially with the latent dimension, the autoencoder's success in finding low-dimensional structure is not just a computational convenience but the main driver of accuracy; measuring the effective latent dimension from reconstruction-error plateaus could be used as a model-selection criterion."],"forward_implications":["Linear unique continuation with finite-dimensional boundary trace is solved with optimal first-order $H^1$ accuracy; the constant grows exponentially in $N$, so the benefit of dimension reduction is quantitative, not just computational.","Using collective data to identify a lower-dimensional manifold improves stability from logarithmic to linear under the representability assumption, so abundant historical data can rescue an otherwise ill-posed reconstruction problem.","Autoencoders can find a lower latent dimension than PCA when the boundary-data manifold is nonlinear, as shown in the quadratic and Gaussian examples, giving a smaller optimization space for the inverse problem.","Composing the decoder with the operator network yields a differentiable mapping from latent variables to approximate FE solutions, so the inverse problem reduces to unconstrained minimization over the latent space.","A decoder trained on clean data reconstructs more accurately than one trained on noisy data, indicating that the quality of the collective data used for training the reduction map matters for the final reconstruction."],"supporting_citations":[{"why":"Supplies the finite-dimensional-trace stability theory and optimal error bound for unique continuation that Proposition 3.2 extends to the reduced-order stabilized projection.","marker":"[12]"},{"why":"Establishes that computational unique continuation rates are limited by the continuous stability, motivating the finite-dimensional improvement.","marker":"[11]"},{"why":"Documents the logarithmic stability of the Cauchy problem, the baseline that the finite-dimensional data assumption improves upon.","marker":"[1]"},{"why":"Introduces the stabilized finite element method for ill-posed elliptic problems on which the stabilized form $m_h$ is built.","marker":"[7]"},{"why":"Defines the energy-minimization MLP operator-learning method that the paper's operator network uses.","marker":"[30]"},{"why":"Provides the general data-free reduced-order kinematics framework and the public implementation used for the numerical examples.","marker":"[43]"},{"why":"Supplies the earlier idea of enriching unique continuation with a database of boundary observations, which this work extends with autoencoders and operator learning.","marker":"[5]"}],"fun_headline_variants":["Latent-space optimization for inverse PDE problems","Autoencoder-parametrized boundary data solves inverse PDE","Learned FE operators stabilize unique continuation","Data-driven dimension reduction for ill-posed PDEs","Stabilized FEM and autoencoders tame inverse PDEs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything downstream depends on the true missing boundary trace lying inside, or within a small distance $\\delta$ of, the low-dimensional set $G$ that the collective data and the autoencoder manage to capture; if the decoder cannot represent the target trace, the latent-space objective has no solution near the true field and the logarithmic ill-posedness of unique continuation returns.","fun_headline_variants_meta":{"raw":{"variants":["Latent-space optimization for inverse PDE problems","Autoencoder-parametrized boundary data solves inverse PDE","Learned FE operators stabilize unique continuation","Data-driven dimension reduction for ill-posed PDEs","Stabilized FEM and autoencoders tame inverse PDEs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000617,"raw_usage":{"total_tokens":2884,"prompt_tokens":987,"completion_tokens":1897,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":1823}},"tokens_in":603,"tokens_out":1897,"duration_ms":14697,"temperature":1.0,"reasoning_tokens":1823,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:25:02.836082+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pick a target solution whose boundary trace is a known linear combination of POD modes with coefficients inside the training support, generate its bulk observation $u_0$ in $\\omega$, and run the latent-space minimization; the linear theory predicts error $O(h\\|u_0\\|_\\omega + \\delta)$, so if reconstruction error is large even when the decoder's reconstruction error $\\delta$ is tiny, the learned decoder-operator composition is the point of failure. A more direct check is to compute the decoder's boundary reconstruction error for the specific target trace and the operator network's output error; the paper's claim is stable reconstruction follows once those errors are small, so a target with small decoder error and small operator error that still yields large interior error would falsify the method's premise.","supporting_citations":[{"cited_title":"Burman, M","cited_arxiv_id":null,"evidence_quote":"Establishes that computational unique continuation rates are limited by the continuous stability, motivating the finite-dimensional improvement."},{"cited_title":"Alessandrini, L","cited_arxiv_id":null,"evidence_quote":"Documents the logarithmic stability of the Cauchy problem, the baseline that the finite-dimensional data assumption improves upon."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the stabilized finite element method for ill-posed elliptic problems on which the stabilized form $m_h$ is built."},{"cited_title":"Learning Nonlinear Finite Element Solution Operators using Multilayer Perceptrons and Energy Minimization","cited_arxiv_id":"2412.04596","evidence_quote":"Defines the energy-minimization MLP operator-learning method that the paper's operator network uses."},{"cited_title":"Boulakia, C","cited_arxiv_id":null,"evidence_quote":"Supplies the earlier idea of enriching unique continuation with a database of boundary observations, which this work extends with autoencoders and operator learning."}],"review_version":1}