{"id":"e28fdfd4-2e85-4a88-961d-705f5ad71a3d","arxiv_id":"2412.11610","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A simultaneous geometry-and-field Bayesian inversion method that exploits the domain independence of the Karhunen-Loeve expansion to avoid repeated integral eigenvalue solves and Moore-Penrose inverses.","lead":"This paper introduces a faster Bayesian inversion method that estimates both the shape of buried features and the surrounding material properties at once. It reuses one fixed mathematical expansion instead of recomputing it at every step, cutting gradient computation time by about two orders of magnitude.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The equivalence of the fixed-bounding-domain truncated K-L prior to the intended random field on updated domains is validated only by a single 1D baseline; no quantitative truncation-error bound or 2D comparison supports the central claim.","rationale":"I read the paper as claiming two things: (i) the fixed-domain K-L discretization yields essentially the same posterior as the discrete-K-L simultaneous method while reducing gradient cost to O(ne); (ii) explicit geometry parameterization improves over field-only inversion. Claim (ii) is supported by the 1D and 2D experiments against field-only estimation. Claim (i) is the load-bearing one: the efficiency gain is only valuable if the posterior is not materially changed. The paper proves the gradient formula and measures O(ne) scaling, which is credible and independently supported by Eq. (48) and Fig. 9; that part is not my concern. The vulnerability is that the approximation underlying the fixed-domain expansion is uncontrolled. Section 2.2 explicitly says truncation error exists and is larger for the bounding domain; Section 5.1.1 asserts it is negligible based on one 1D benchmark; Section 5.2 gives no baseline. The truncation threshold and correlation lengths are user-chosen and the paper itself shows posterior sensitivity to l, so a quantitative error check over visited geometries is needed before the central equivalence can be accepted. The reader identified the same weakest assumption; I agree. I do not see an internal inconsistency in Eqs. (46)-(48), and the computational claims are well supported. The verdict should remain CONDITIONAL until the 2D truncation error or a 2D discrete-K-L comparison is provided.","tokens_in":27677,"tokens_out":7804,"duration_ms":78456,"concrete_test":"Draw 20 posterior samples of 2θ from the l=8 m 2D run (Section 5.2). For each, solve the IEVP on the actual domain D(2θ) using the same Nyström quadrature (N=400) and build the M1-term covariance Σ λ_i φ_i φ_i on the FEM mesh; compare it against the fixed-bounding-domain M1-term covariance Σ λ'_i φ'_i φ'_i in relative Frobenius norm. If the maximum relative error over these posterior-visited geometries exceeds about 5%, the 'negligible effect' claim is not supported and the 2D posterior should be re-examined with a domain-specific expansion; if the error is below 5%, the main approximation is quantifiably benign.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the IEVP can be solved once on the bounding domain D' without changing the posterior relies on the truncated K-L expansion on D' being an accurate surrogate for the random field on every updated domain D(2θ). This is precisely where the argument is weakest. The domain independence property stated in Section 2.2 (Eqs. 6-9) is a moment-equality result for the complete expansion; the paper concedes that truncation introduces an error and that this error is larger for the bounding-domain expansion. The truncation threshold λ_{M1}>10^{-3} (1D) or λ_{M1}>λmax×10^{-3} (2D) is a convention and is set on D', not on the actual D(2θ). The only evidence that this error is negligible is the 1D comparison in Section 5.1.1 showing the proposed method and Koch et al. produce nearly identical results. Section 5.2, the 2D demonstration, contains no comparison against the discrete-K-L simultaneous baseline, so the equivalence claim is unverified in the setting where the proposed efficiency gain matters most. Because the induced prior covariance on D(2θ) is Σ_{i≤M1} λ'_i φ'_i(z)φ'_i(z*) rather than the intended target covariance, the posterior and its uncertainty statements could be biased for geometries or correlation lengths not covered by the 1D test. The gradient algebra in Eqs. (46)-(48) is internally consistent given the fixed expansion; the vulnerability is upstream, in the accuracy of that expansion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops an efficient Bayesian inversion method for simultaneous estimation of spatial random fields and geometric parameters. The key idea is to exploit the domain independence property of the Karhunen-Loève expansion: the integral eigenvalue problem (IEVP) is solved once on a fixed bounding domain D′ that contains all possible updated domains D(2θ), and the resulting eigenpairs are reused throughout Hamiltonian Monte Carlo sampling. This avoids repeated IEVP solves and eliminates the Moore-Penrose inverse computations needed for shape derivatives in the prior discrete-K-L method of Koch et al. (2021). Gradient formulas for the K-L expansion w.r.t. geometry parameters are derived (Eqs. (46)–(48)) with O(ne) complexity. The method is tested on 1D and 2D steady seepage problems, showing gains in computation time and comparable posterior estimates to the baseline in the 1D case.","tokens_in":28037,"tokens_out":9107,"duration_ms":80455,"significance":"If the central approximation is justified, the paper offers a practically useful computational speedup: the K-L gradient cost is reduced from O(ne^3) to O(ne), and the implementation avoids pseudoinverse computations. The 1D comparison shows that the proposed method can reproduce the discrete-K-L simultaneous inversion results almost exactly, while running much faster. The exposition is clear, and the gradient derivation is internally consistent. The main caveat is that the load-bearing approximation—using a truncated K-L expansion on the bounding domain for all updated subdomains—is not quantitatively controlled, and the 2D demonstration does not compare against the discrete-K-L baseline. Thus the significance is conditional: the method is promising, but the evidence for its central claim is incomplete.","major_comments":[{"comment":"The central approximation of the paper is that a truncated K-L expansion computed on a fixed bounding domain D′ accurately represents the random field on every updated domain D(2θ). Section 2.2 concedes that the domain independence property holds for the complete expansion and that truncation introduces an error, which is larger for the bounding-domain expansion. However, no quantitative error bound is provided, and the only empirical support is the 1D comparison in Section 5.1.1. The 2D study in Section 5.2 contains no comparison against the discrete-K-L simultaneous method, so the equivalence claim is unverified in the setting where the efficiency gain matters most. I would like to see either a rigorous truncation-error estimate in terms of the eigenvalue decay and the geometry of D(2θ), or a 2D baseline comparison, before the central claim can be accepted.","section":"§2.2 and §5.2"},{"comment":"The text states that 'due to the domain independence property, the K-L expansion of u constructed on D′ is equivalent to that on D(2θ)'. This is only true for the complete expansion. The truncated expansion in Eq. (12) induces a prior covariance on D(2θ) given by Σ_{i≤M1} λ′_i φ′_i(z)φ′_i(z∗), which is not the optimal M1-term truncated K-L covariance on D(2θ). The authors should explicitly label this as an approximation and quantify its effect on posterior inference, for example by reporting the captured variance fraction on the actual subdomains for the tested cases, not just on D′.","section":"§2.3 and Eq. (12)"},{"comment":"The comparison with Koch et al. (2021) uses different numbers of K-L terms for the two simultaneous methods (e.g., M=16 vs. M=13 for l=2.5 m). The truncation threshold λ>10^-3 is applied on different domains (D′ for the proposed method, D1 and D2 for the baseline), so the amount of captured variance per domain may differ. To substantiate the claim that the truncation error has negligible effect on the inversion results, the authors should report the variance fraction captured on D(2θ) for each method and perhaps a quantitative posterior discrepancy measure (e.g., Wasserstein distance between marginal posteriors) rather than only visual comparison of mean and credible intervals.","section":"§5.1.1 and Table 1"}],"minor_comments":[{"comment":"There is a typo in the truncated K-L expansion: the right-hand side should involve the mean function X̄(z) (or u(z)), not X(z,ω) itself. As written, the equation is circular.","section":"Eq. (4)"},{"comment":"The notation u(z,ω) = u(z,1θ(ω)) is confusing because u is used both for the random field and for the mean function. Please use a clearer notation, for example u(z,1θ) for the random field and ū(z) for its mean.","section":"Eq. (12) and Section 2.4"},{"comment":"The spatial-field-only estimation method is cited as [5], which is a paper on elastic modulus estimation. Since the numerical examples are seepage flow problems, the authors should cite a more directly relevant spatial-field-only inversion method, or clarify that [5] is used as a general adjoint-HMC framework.","section":"Introduction, reference [5]"},{"comment":"The shape derivative formulas for G and |J_e| are stated without derivation or reference to the exact finite element formulation. While they are plausible and follow [33], a brief explanation of the notation (e.g., the meaning of G as the matrix of global shape-function derivatives) would improve readability.","section":"Section 4, Eqs. (43)–(44)"},{"comment":"The two-orders-of-magnitude improvement is measured for the K-L gradient computation time at a specific parameter setting. It would be helpful to state the wall-clock times or the hardware used, so that the reader can gauge the practical impact.","section":"Section 5.1.2, Fig. 9"}],"recommendation":"major_revision","confidential_remarks":"The paper's central algorithmic idea is sound, but the evidence for the key truncation-error assumption is limited to a single 1D baseline. The 2D study, which is the more relevant setting, does not compare against the discrete-K-L simultaneous method; adding such a comparison or a quantitative error bound is essential. The self-referential nature of the baseline (Koch is a co-author) is not by itself a problem, but the reader should be aware that the 1D comparison is between methods developed within the same group. The paper fits the applied-statistics scope reasonably well, though it is closer to computational geophysics/engineering; that is not an obstacle for this journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honestly, this is a solid methods paper with one real gap. The core trick is legitimate: use the domain independence property of the K-L expansion to solve the integral eigenvalue problem once on a fixed bounding box, then evaluate the fixed eigenfunctions at moving nodes during HMC. That removes two computational bottlenecks—repeated IEVP solves and the Moore-Penrose inverse for shape derivatives—and the gradient formula in Eq. (48) checks out. The scaling study shows the K-L gradient cost drops from roughly O(ne^3) to O(ne), about two orders of magnitude at large ne. That is a real win.\n\nWhere it works, it works well. The 1D seepage benchmark shows the proposed method gives posterior estimates almost identical to Koch et al.'s discrete-K-L simultaneous method, across three correlation lengths. The 2D cavity example demonstrates good reconstruction and sensible uncertainty bands, and the paper is clear about how the geometry prior and correlation length affect results. The authors also honestly state in Section 2.2 that truncation error is larger for the bounding-domain expansion.\n\nThe soft spot is the one the authors themselves flag but don't close: the truncated expansion on the bounding domain is only an approximation of the random field on updated subdomains. The theory guarantees moment equivalence only for the complete expansion. The only evidence that the truncation error is negligible for the posterior is the single 1D comparison. The 2D demonstration, where the efficiency gain matters most, has no comparison against the discrete-K-L baseline. So the central load-bearing approximation rests on one empirical test. I'd like to see either a 2D baseline run or a quantitative truncation-error analysis with respect to correlation length, geometry range, and domain-to-boundary size ratio. This is not a fatal flaw—nothing suggests the method fails—but it's an under-supported claim in its current form.\n\nAlso, no code or data is provided, which makes it harder to reproduce the scaling and posterior results. Given the paper's practical target audience, shipping code would move this from a good paper to a very useful one.\n\nOverall: worth a serious referee. The method is new in its specific combination, the gradient derivation is clean, and the empirical evidence, while incomplete, supports the procedure. I'd send it to review with a request for a 2D baseline comparison and ideally a truncation-error study.","headline":"A genuine computational improvement for K-L-based simultaneous geometry and field inversion, with one under-tested truncation approximation that needs a 2D baseline.","tokens_in":28540,"tokens_out":1992,"would_cite":true,"duration_ms":19401,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","65N21"],"pacs":[],"model":"deepseek-v4-flash","headline":"One Karhunen-Loève solve on a fixed bounding domain is enough for simultaneous estimation of an evolving geometry and its spatial field, cutting the K-L gradient cost from cubic to linear in the mesh size.","keywords":["Bayesian inverse problems","Karhunen-Loève expansion","domain independence property","simultaneous geometry and field estimation","interface detection","Hamiltonian Monte Carlo","integral eigenvalue problem","seepage flow"],"falsifier":"Take a 2D problem with a small cavity whose radius is comparable to the prior correlation length, run the proposed fixed-bounding-domain inversion and a reference inversion that recomputes the integral eigenvalue problem on each updated domain, and compare the posterior distributions of the cavity center and radius: if the 95% credible intervals for any geometry parameter differ by more than the width of the reference interval, or if the posterior mean of the conductivity field deviates systematically near the boundary, the domain-independence truncation error is not negligible for that configuration.","tokens_in":27485,"feed_emoji":"🌊","tokens_out":11740,"duration_ms":93124,"temperature":0.7,"pith_summary":"This paper argues that the domain independence property of the Karhunen-Loève expansion removes the two computational bottlenecks in simultaneous estimation of a spatial random field and an embedded geometric feature: the integral eigenvalue problem needs to be solved only once on a fixed bounding domain, and the shape derivatives of the K-L terms can be computed from the derivative of the autocovariance function without any Moore-Penrose inverse. If this is right, Bayesian inversion with Hamiltonian Monte Carlo can track moving boundaries and field realizations at a K-L gradient cost that grows linearly with the number of finite elements rather than cubically. The paper reports about two orders of magnitude speedup in the K-L gradient computation for a one-dimensional seepage problem and posterior estimates nearly identical to the earlier simultaneous method, while showing that explicit geometry parameterization captures abrupt conductivity jumps that spatial-field-only inversion misses. It also finds that introducing geometry parameters lowers the number of unknown parameters, because the boundary, rather than a short correlation length, carries the discontinuity.","feed_headline":"One fixed-domain eigen-solve replaces repeated geometry-update solves","feed_subtitle":"New gradient formulas make field-plus-boundary Bayesian inversion scale with the mesh instead of its cube.","key_machinery":"The load-bearing object is the domain independence property of the K-L expansion: for two overlapping domains $D$ and $D'$, the complete K-L expansions built on each domain have the same first and second moments at all points in the overlap, even though the individual eigenpairs differ. The paper exploits this by choosing a bounding domain $D'$ and solving the integral eigenvalue problem once by the Nyström method, then treating the eigenfunctions as fixed shapes whose values at the moving points $z_l(2\\theta)$ change only through the geometry-dependent coordinates. The gradient machinery is then Eqs. (46)-(48): the derivative with respect to the field coefficients is the fixed eigenpair term $\\sqrt{\\lambda'_i}\\phi'_i(z(2\\theta))$, and the derivative with respect to the geometry parameters is a sum over the already computed eigenpairs times $\\partial\\phi'_i/\\partial 2\\theta$, which Eq. (48) reduces to a weighted sum of derivatives of the Gaussian autocovariance kernel. This avoids both repeated eigenvalue solves and pseudo-inverse computations, leaving a per-element cost that is independent of the mesh size.","core_discovery":"The central claim is that the K-L expansion constructed on any fixed bounding domain $D'$ that contains every visited domain $D(2\\theta)$ can stand in for the domain-dependent expansion on each $D(2\\theta)$ during inversion. Because the autocovariance function is the same on the overlap, the first- and second-order moments of the corresponding random fields coincide on $D(2\\theta)$, so the truncated field $\\hat u(z,{}^1\\theta) = \\bar u(z)+\\sum_{i=1}^{M_1}\\sqrt{\\lambda'_i}\\phi'_i(z)\\,{}^1\\theta_i$ remains a valid representation as the geometry parameters $2\\theta$ move the evaluation points. The gradient formulas (46)-(48) express the spatial-field derivative directly as $\\sqrt{\\lambda'_i}\\phi'_i(z(2\\theta))$ and the geometry derivative through $\\partial C(z(2\\theta),z'_j)/\\partial 2\\theta$, quantities available from the single Nyström solve on $D'$, eliminating the repeated eigenvalue solves and the Moore-Penrose inverse required by the earlier discrete-K-L method. The paper demonstrates in a 1D three-layer seepage problem that the proposed scheme yields posteriors almost identical to the earlier simultaneous method while reducing the K-L gradient complexity from $O(n_e^3)$ to $O(n_e)$, and in 1D and 2D seepage problems that it identifies the interface location and its uncertainty more accurately than spatial-field-only estimation.","pith_inferences":["The same fixed-eigenpair gradient strategy should carry over to level-set or immersed-boundary parameterizations, because the field representation itself never depends on the evolving domain once the bounding-domain eigenpairs are fixed.","A direct test of the truncation-error assumption would be to repeat the 2D cavity experiment with a solver that recomputes the IEVP on each updated domain; the paper does not include that comparison in 2D.","If the correlation length is inferred during inversion, a parameterized K-L expansion or a nested eigenvalue solve on the bounding domain would be needed, because the IEVP changes with the correlation length.","The reported asymptotic gain is for the K-L gradient only; in large 2D and 3D problems the forward solver's linear-algebra cost will typically determine whether the overall speedup approaches two orders of magnitude."],"forward_implications":["For any inverse problem where a bounding domain can enclose all geometry realizations, the K-L gradient is no longer the computational bottleneck: its cost per HMC leapfrog step is $O(n_e)$ rather than $O(n_e^3)$.","Simultaneous estimation of the interface and the field needs fewer unknown parameters than field-only inversion, because the boundary parameterization carries the discontinuity and the prior correlation length can be larger.","The posterior for the boundary location and the conductivity field can be reported together, with boundary uncertainty expressed as highest-posterior-density regions that envelop the true interface in the tested 1D and 2D seepage problems.","The method preserves the posterior accuracy of the earlier simultaneous method; the truncation error introduced by using the bounding-domain expansion had a negligible effect on the 1D inversion results.","Because the forward solve remains the dominant cost, applying the method to higher-dimensional or transient problems depends mainly on the forward solver's scaling rather than on the K-L gradient computation."],"supporting_citations":[{"why":"Supplies the domain independence property of the K-L expansion and the observation that truncation error is larger for the bounding-domain expansion.","marker":"[42]"},{"why":"Defines the previous simultaneous field-and-geometry method based on the discrete K-L expansion; its Moore-Penrose gradient is the baseline the paper replaces.","marker":"[33]"},{"why":"Provides the reversible mesh-moving update from a fixed reference configuration used to keep HMC reversibility when geometry parameters change.","marker":"[19]"},{"why":"Defines the spatial-field-only adjoint HMC estimation method used as the comparison baseline.","marker":"[5]"},{"why":"Describes the Nyström numerical treatment of the integral eigenvalue problem used to compute eigenpairs on the bounding domain.","marker":"[31]"},{"why":"Gives the differentiation formulas for eigenvalues and eigenvectors whose Moore-Penrose inverse the proposed method avoids.","marker":"[41]"},{"why":"Documents the HMC adaptation scheme with dual averaging and mass-matrix tuning used in the sampler.","marker":"[49]"},{"why":"Documents the truncation-error behaviour of the K-L expansion on overlapping domains, used to justify the approximation on the bounding domain.","marker":"[44]"}],"fun_headline_variants":["One fixed-domain K-L solve replaces repeated geometry solves","Bayesian geometry+field inversion: gradient cost cubic to linear","K-L domain independence gives 100x faster gradient computation","Avoid repeated eigen-solves in Bayesian inversion of buried objects","Single eigen-solve on bounding domain speeds field-geometry inversion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that carries the method is that the truncated K-L expansion computed once on the fixed bounding domain still represents the random field accurately on every updated subdomain visited during sampling; the paper notes the truncation error is larger for the bounding-domain expansion and verifies it is negligible only in the 1D comparison.","fun_headline_variants_meta":{"raw":{"variants":["One fixed-domain K-L solve replaces repeated geometry solves","Bayesian geometry+field inversion: gradient cost cubic to linear","K-L domain independence gives 100x faster gradient computation","Avoid repeated eigen-solves in Bayesian inversion of buried objects","Single eigen-solve on bounding domain speeds field-geometry inversion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000611,"raw_usage":{"total_tokens":2915,"prompt_tokens":1090,"completion_tokens":1825,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":706,"completion_tokens_details":{"reasoning_tokens":1743}},"tokens_in":706,"tokens_out":1825,"duration_ms":15839,"temperature":1.0,"reasoning_tokens":1743,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:46:53.061398+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a 2D problem with a small cavity whose radius is comparable to the prior correlation length, run the proposed fixed-bounding-domain inversion and a reference inversion that recomputes the integral eigenvalue problem on each updated domain, and compare the posterior distributions of the cavity center and radius: if the 95% credible intervals for any geometry parameter differ by more than the width of the reference interval, or if the posterior mean of the conductivity field deviates systematically near the boundary, the domain-independence truncation error is not negligible for that configuration.","supporting_citations":[{"cited_title":"Pranesh, D","cited_arxiv_id":null,"evidence_quote":"Supplies the domain independence property of the K-L expansion and the observation that truncation error is larger for the bounding-domain expansion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the previous simultaneous field-and-geometry method based on the discrete K-L expansion; its Moore-Penrose gradient is the baseline the paper replaces."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the reversible mesh-moving update from a fixed reference configuration used to keep HMC reversibility when geometry parameters change."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the spatial-field-only adjoint HMC estimation method used as the comparison baseline."},{"cited_title":"Papaioannou, Non-intrusive finite element reliability analysis meth- ods, Ph.D","cited_arxiv_id":null,"evidence_quote":"Documents the truncation-error behaviour of the K-L expansion on overlapping domains, used to justify the approximation on the bounding domain."}],"review_version":1}