{"id":"c4cbdc3a-6559-42da-837d-612de3f65a14","arxiv_id":"2501.08210","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A primal-dual splitting method solves TV-constrained full-waveform inversion without inner loops, with experiments on the SEG/EAGE Salt model.","lead":"This paper applies a standard primal-dual splitting algorithm to solve full-waveform inversion with a total-variation constraint, replacing inner-loop and approximation-based methods. If it works, it offers a faster way to reconstruct subsurface images from seismic data while preserving sharp boundaries.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'accurately solves' claim rests on PDS convergence conditions that the non-convex FWI misfit does not satisfy; no convergence or step-size analysis is provided, and the experiments do not report feasibility or fixed-point residuals.","rationale":"The paper's most defensible contribution is the observation that the TV- and box-constrained FWI problem (12) has the additive structure of the PDS template (8), with f=E, g=iota_Bbox, h=iota_Bell, and L=D. The proximal and projection formulas (14)-(16) are standard and correctly stated. The gap is that the PDS convergence theorem requires convexity and a Lipschitz gradient for f, and E is neither verified nor convex in general. Because the paper explicitly claims to 'accurately solve' the problem by 'strictly following' PDS, the missing verification is not cosmetic: it undermines the theoretical foundation of the central claim. The numerical experiments cannot substitute because they do not report feasibility or fixed-point residuals, and the hyper-parameters are selected on the test model. I therefore agree with the reader's conditional verdict: the algorithm may work as a heuristic, but the paper must either supply a convergence analysis for non-convex E or weaken the claim, and should add comparisons with prior TV-FWI methods. The proposed experiment would check whether the algorithm actually reaches a feasible fixed point in the paper's own setting, which is a direct test of the 'accurately solves' assertion.","tokens_in":7547,"tokens_out":5756,"duration_ms":59670,"concrete_test":"Using the paper's exact experimental setup (Devito, 20 sources, 5000 iterations), run Algorithm 1 from two different initial models (the smoothed model and a constant-velocity model). At each iteration, record the TV value ||D m^(k)||_{1,2}, the box feasibility, and the PDS fixed-point residual max(||m^(k+1)-m^(k)||, ||y^(k+1)-y^(k)||). If the two runs converge to substantially different final models, or if the final residual does not decrease to a small value while the TV constraint remains active, then the algorithm is not demonstrably solving (12), and the 'accurately solves' claim should be weakened to 'empirically effective heuristic.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the identification of problem (12) with the PDS template (8) and the assertion in Section III that 'by strictly following the PDS framework, our algorithm accurately solves the optimization problem without relying on approximations or relaxations.' Condat's theorem (cited [31]) requires f to be convex with a Lipschitz-continuous gradient. Here f = E(m) = (1/2)||u_obs - u_cal(m)||^2, where u_cal is the nonlinear wave-propagation map; E is non-convex and its gradient, computed by the adjoint-state method, is not shown to be Lipschitz on the box [l,u]^N. The paper provides no convergence analysis, local or global, and no step-size condition relating gamma1, gamma2, the Lipschitz constant of E, and the operator norm of D. The experiments tune alpha, gamma1, and gamma2 on the test model and report RMSE/SSIM against alpha, but never report whether the final iterate satisfies ||Dm||_{1,2} <= alpha or whether the PDS fixed-point residual decreases. Thus the central claim is an extrapolation from a theorem whose hypotheses are not verified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes to solve TV- and box-constrained full-waveform inversion by applying Condat's primal-dual splitting (PDS) directly to problem (12), where the objective is the least-squares FWI misfit E(m) and the constraints are a TV-ball constraint and a box constraint. Algorithm 1 is presented as an inner-loop-free, approximation-free solver, with the claimed advantage that the TV constraint is enforced through a single proximal projection per iteration (Remark 1). The paper reports experiments on the SEG/EAGE Salt model, comparing the proposed method with standard gradient-descent FWI over a range of TV bound values α, and concludes that the method removes wave-like artifacts, improves RMSE/SSIM, and is robust to noise.","tokens_in":7835,"tokens_out":4557,"duration_ms":49249,"significance":"If the central claim were established, the paper would offer a practically valuable simplification of TV-constrained FWI: eliminating inner loops in constraint enforcement is attractive for large-scale seismic inversion, and the proximal operators for the TV ball and box are correctly derived. The problem formulation is clear, the algorithm is simple and reproducible, and the experiments show a qualitative improvement over unconstrained gradient descent on a standard benchmark. However, the main mathematical claim—that Algorithm 1 'accurately solves' problem (12) by strictly following the PDS framework—is not supported, because the PDS convergence theorem invoked is for convex objectives and the FWI misfit is nonconvex. The experimental section is also too narrow to substantiate the efficiency and accuracy claims. No circular reasoning was found; the self-citations are motivational rather than load-bearing.","major_comments":[{"comment":"The identification of problem (12) with the PDS template (8) requires f = E to be convex with a Lipschitz-continuous gradient. Here E(m) = (1/2)||u_obs - u_cal(m)||^2, where u_cal is a nonlinear wave-propagation map; E is nonconvex in m, and no Lipschitz constant for its gradient is provided. Condat's theorem [31] therefore does not apply, and the assertion that the algorithm 'accurately solves the optimization problem without relying on approximations or relaxations' is unsupported. The paper needs either a convergence analysis for the nonconvex setting (e.g., local convergence, subsequential stationarity, or a monotone-operator argument) or an explicit reframing of Algorithm 1 as a heuristic with correspondingly weakened claims.","section":"Section III, Eq. (12) and Algorithm 1"},{"comment":"The step sizes γ1 = 1e-4 and γ2 = 1e2 are set without any relation to the Lipschitz constant of ∇E or the norm of D, and no step-size condition is derived. Since no convergence theorem is established for the nonconvex objective, there is no guarantee that the iterates approach a solution of (12). The paper should report how these step sizes were selected, provide a step-size sensitivity study, and, if possible, state a valid range based on the problem data.","section":"Section IV-A, Algorithm 1 step sizes"},{"comment":"The optimal value α = 350 is selected post-hoc on the test model: the paper varies α and reports the best RMSE/SSIM, so the quantitative advantage of the proposed method includes oracle parameter choice. This does not demonstrate that the method can be applied without access to the true model. Please add a validation or model-selection procedure, report performance across an interval of α values, or explicitly restrict the claim to the tuned setting.","section":"Section IV-B, Figs. 6-7"},{"comment":"The experiments compare only against standard gradient-descent FWI. There is no comparison with existing TV-regularized or TV-constrained FWI methods [21]-[24], and no wall-clock time or iteration-count comparison against an inner-loop method, so the efficiency claim in Remark 1 is not empirically validated. Additionally, the paper never reports whether the final iterates satisfy ||Dm||_{1,2} ≤ α or whether the PDS fixed-point residual decreases. Without such feasibility and convergence diagnostics, the figures do not establish that Algorithm 1 actually solves problem (12) rather than merely producing a useful reconstruction.","section":"Section IV, experimental comparison"}],"minor_comments":[{"comment":"The abstract and introduction mention the SEG/EAGE Salt and Overthrust Models, but the experiments in Section IV describe only the Salt model; clarify which benchmarks were actually run.","section":"Abstract and Section I"},{"comment":"The input line lists γ0 and γ1, but the iteration uses γ1 and γ2; correct the pseudo-code to match the text.","section":"Algorithm 1"},{"comment":"The notation P_{||·||_1≤α} is used without a formal definition; define the projection onto the ℓ1 ball and specify the zero-norm case in Eq. (15) explicitly.","section":"Eq. (16)"},{"comment":"There are typos in the text and figure captions, including 'sinthesized' in the Fig. 2 caption and 'archive accurate' in Section IV-B; both should be corrected.","section":"Section IV-B"},{"comment":"The phrase 'iteration-dependent factor is eliminated' is misleading because each PDS iteration still requires the gradient computation ∇E, which is the dominant cost; rephrase to say that no inner loop is used within the constraint-enforcement step.","section":"Remark 1"},{"comment":"The figures would benefit from clearer axis labels and error bars or multiple noise realizations; as written, the results appear to come from a single synthetic realization, which limits the strength of the robustness claims.","section":"Figs. 5-7"}],"recommendation":"major_revision","confidential_remarks":"The central problem is a mismatch between the claimed rigorous status of the algorithm and the actual analysis: the paper invokes a convex-convergence theorem for a nonconvex objective and provides no convergence result. This is fixable by reframing the contribution as an empirical heuristic or by adding a genuine nonconvex convergence analysis, but the current wording overstates the contribution. The experimental section also needs substantial strengthening before the efficiency claim can be assessed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'll give you the short version first. This is a conditional pass, not a reject. The genuinely new piece is the direct application of Condat's primal-dual splitting to TV- and box-constrained FWI, with the TV ball projection handled through the known reduction to ℓ1-ball projection. That combination is new in the cited literature, which mostly uses TV as a penalty or enforces constraints with inner loops. The proximal formulas in Algorithm 1 are correct; the box projection is componentwise clipping, and the TV ball projection via (15)-(16) is standard but implemented cleanly. The computational complexity argument is also right: the per-iteration cost is dominated by the wavefield solves, with the TV constraint update at O(N log N).\n\nWhere it gets soft is the central claim. The paper says that by strictly following PDS, Algorithm 1 accurately solves (12) without approximations or relaxations. Condat's theorem requires f to be convex and ∇f Lipschitz. Here f is the FWI misfit, which is nonlinear and non-convex. No convergence analysis, local or global, is supplied, and no step-size condition connects γ1 and γ2 to the Lipschitz constant or to the norm of D. The chosen values (1e-4 and 1e2) are simply stated. That makes the 'accurately solves' statement an unverified extrapolation. This is a real gap, but not disqualifying: non-convex PDS-type methods are routinely used in FWI as heuristics, and the paper would be fine if it said heuristic solver and backed that with better experiments.\n\nThe experimental section is the weaker half. The only baseline is vanilla gradient descent. TV-regularized FWI and the earlier constrained methods (Esser, Yong, Peters) are cited but not compared. α is selected on the test data, and the RMSE/SSIM-vs-α curves show the expected U-shape. More importantly, the paper never reports whether the final iterate actually satisfies ∥Dm∥1,2 ≤ α, nor any fixed-point residual, so we cannot tell how well the algorithm enforces the constraint. No runtime comparison supports the efficiency claim, only an asymptotic argument.\n\nCredit where it's due: the noise-robustness observation, shown by SSIM not degrading after many iterations, is interesting and worth keeping. The writing is clear and the symbols are mostly consistent.\n\nBottom line: useful for practitioners who want a simple, inner-loop-free TV-constrained FWI solver. It deserves a serious referee, but the authors need to soften the 'accurately solves' language, add a convergence or step-size discussion, test against real prior art, and report feasibility/residuals.","headline":"A correct and useful PDS wrapper for TV-constrained FWI, but the 'accurately solves' claim is not backed by convergence theory or by baseline experiments.","tokens_in":8342,"tokens_out":2531,"would_cite":false,"duration_ms":25887,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65K10","90C25","86A22"],"pacs":[],"model":"deepseek-v4-flash","headline":"A primal-dual splitting algorithm solves the TV- and box-constrained full-waveform inversion problem accurately by reducing the TV-constraint update to a single projection per iteration, eliminating inner loops and approximations.","keywords":["full-waveform inversion","total variation constraint","primal-dual splitting","box constraint","seismic velocity reconstruction","constrained optimization","inverse problems"],"falsifier":"Run Algorithm 1 on a velocity model where standard FWI is known to stall in a local minimum, and record the constraint violation $\\|Dm^{(k)}\\|_{1,2} - \\alpha$ together with the misfit $E(m^{(k)})$ over iterations; if the iterates fail to converge to a feasible point, or oscillate, for step sizes within the PDS-admissible range, the claim that the algorithm solves (12) is refuted.","tokens_in":7352,"feed_emoji":"🌊","tokens_out":16608,"duration_ms":135901,"temperature":0.7,"pith_summary":"This paper presents an algorithm for full-waveform inversion (FWI) that enforces a total-variation (TV) constraint on the reconstructed subsurface velocity model as a hard constraint rather than as a penalty term. The authors' claim is that a direct application of the primal-dual splitting method (PDS) to the TV- and box-constrained FWI problem yields an update that needs no inner loop for the constraint and no linear or quadratic approximation of the nonsmooth term. If that claim holds, the main computational bottleneck of earlier constrained-FWI solvers disappears: the per-iteration cost is dominated by the wave-equation gradient evaluation, and the TV projection itself costs $O(N \\log N)$. Experiments on standard salt and overthrust benchmark models show the method suppressing the wave-like artifacts and noise sensitivity of standard FWI, with better RMSE and SSIM over the tested range of the TV bound $\\alpha$.","feed_headline":"One projection per step removes inner loops from TV-constrained FWI","feed_subtitle":"Primal-dual splitting enforces the TV bound in full-waveform inversion with one projection per step.","key_machinery":"The central object is the primal-dual splitting update applied to the constrained problem (12), with the TV constraint encoded as the indicator function of the $\\ell_{1,2}$-ball composed with the discrete difference operator D. The key identity is the conjugate-prox relation $\\operatorname{prox}_{\\gamma h^*}(y) = y - \\gamma \\operatorname{prox}_{h/\\gamma}(y/\\gamma)$, which turns the dual update into one metric projection onto the TV ball; that projection is itself computed by projecting the per-pixel gradient magnitudes onto the one-dimensional $\\ell_1$-ball. This construction is what removes inner loops and keeps the TV-constraint cost at $O(N \\log N)$ per iteration, leaving the wave-equation gradient evaluation as the dominant step.","core_discovery":"The paper's central claim is that problem (12) — the FWI misfit minimized subject to the TV-ball constraint $\\|Dm\\|_{1,2} \\le \\alpha$ and a box constraint $m \\in [l,u]^N$ — can be solved accurately by the primal-dual splitting iteration written out in Algorithm 1. Identifying the misfit E with the differentiable function f of the PDS template, the box indicator with g, the TV-ball indicator with h, and the discrete difference operator D with the linear operator L, the authors obtain a closed-form update: one gradient step, one projection onto the box, and one projection onto the $\\ell_{1,2}$-ball for the dual variable. They argue that because the iteration strictly follows the PDS framework, it solves the constrained problem without approximations, and because the constraint handling is a single projection per iteration, there is no inner loop. On the standard salt and overthrust velocity benchmarks, the claim is supported by reconstructions whose RMSE and SSIM beat standard FWI for the tested $\\alpha$ values, with the best result near $\\alpha = 350$, and by SSIM-versus-iteration curves that stay high even after many iterations with noisy data.","pith_inferences":["The paper does not compare against earlier TV-regularized or TV-constrained FWI solvers directly; a head-to-head test would separate the benefit of the constraint formulation from the benefit of the inner-loop-free solver.","Because the convergence argument is inherited from PDS and no step-size rule is derived for the nonconvex misfit, the step sizes $\\gamma_1$ and $\\gamma_2$ in the experiments are heuristic; line-search or adaptive step-size variants are a natural testable extension.","The same splitting structure would survive replacing the isotropic $\\ell_{1,2}$ TV ball with a weighted or direction-dependent TV ball, which could encode geological dip information without changing the one-projection-per-step architecture.","The empirical claims are demonstrated on 50×100 two-dimensional models; scaling to three-dimensional surveys would stress both the $O(S T N)$ gradient cost and the memory of the dual variable, and this scaling is not addressed in the paper."],"forward_implications":["Per-iteration cost of enforcing the TV constraint is $O(N \\log N)$, so the wave-equation gradient evaluation $O(S T N)$ becomes the dominant cost, making constrained FWI practical at larger scales.","The TV bound $\\|Dm\\|_{1,2} \\le \\alpha$ is enforced exactly, not approximately, at every iteration, so every reconstructed model satisfies the stated prior.","With noisy data, the method keeps SSIM high at large iteration counts, unlike standard FWI whose SSIM degrades, indicating reduced overfitting to noise.","The parameter $\\alpha$ acts as a predictable smoothness dial: smaller $\\alpha$ over-smooths the model and larger $\\alpha$ approaches standard FWI, so $\\alpha$ can be chosen from prior knowledge of subsurface structure rather than tuned against the misfit."],"supporting_citations":[{"why":"Defines the standard least-squares FWI objective and the gradient-descent baseline compared in the experiments.","marker":"[1]"},{"why":"Introduces the total-variation functional that defines the constraint $\\|Dm\\|_{1,2} \\le \\alpha$.","marker":"[14]"},{"why":"Represents the conventional TV-regularized FWI approach with inner iterations that the proposed method claims to outperform.","marker":"[21]"},{"why":"Represents prior constrained seismic inversion via projection methods with inner loops, the contrast behind the inner-loop-free claim.","marker":"[24]"},{"why":"Supplies the primal-dual splitting iteration that Algorithm 1 applies directly to the TV- and box-constrained problem.","marker":"[31]"},{"why":"Provides the adjoint-state formula used to compute the gradient $\\nabla E$ of the FWI misfit in every iteration.","marker":"[33]"},{"why":"Gives the efficient projection onto the $\\ell_1$-ball used to evaluate the proximity operator of the TV-ball indicator.","marker":"[34]"}],"fun_headline_variants":["TV-constrained FWI without inner loops: one projection per step","One-step TV projection eliminates FWI inner loops","Primal-dual splitting makes TV-FWI loop-free","Single projection enforces TV bound in FWI","Inner-loop-free TV-FWI: fast and accurate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the nonconvex FWI misfit $E(m)$ can be treated as the convex differentiable term required by the primal-dual splitting convergence theorem; the paper gives no convergence proof and no step-size condition showing that this requirement holds.","fun_headline_variants_meta":{"raw":{"variants":["TV-constrained FWI without inner loops: one projection per step","One-step TV projection eliminates FWI inner loops","Primal-dual splitting makes TV-FWI loop-free","Single projection enforces TV bound in FWI","Inner-loop-free TV-FWI: fast and accurate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000332,"raw_usage":{"total_tokens":1865,"prompt_tokens":984,"completion_tokens":881,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":813}},"tokens_in":600,"tokens_out":881,"duration_ms":7940,"temperature":1.0,"reasoning_tokens":813,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:28:54.361218+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 on a velocity model where standard FWI is known to stall in a local minimum, and record the constraint violation $\\|Dm^{(k)}\\|_{1,2} - \\alpha$ together with the misfit $E(m^{(k)})$ over iterations; if the iterates fail to converge to a feasible point, or oscillate, for step sizes within the PDS-admissible range, the claim that the algorithm solves (12) is refuted.","supporting_citations":[{"cited_title":"Inversion of seismic reflection data in the acoustic approximation,","cited_arxiv_id":null,"evidence_quote":"Defines the standard least-squares FWI objective and the gradient-descent baseline compared in the experiments."},{"cited_title":"Nonlinear total variation based noise removal algorithms,","cited_arxiv_id":null,"evidence_quote":"Introduces the total-variation functional that defines the constraint $\\|Dm\\|_{1,2} \\le \\alpha$."},{"cited_title":"Total variation regularization strategies in full- waveform inversion,","cited_arxiv_id":null,"evidence_quote":"Represents the conventional TV-regularized FWI approach with inner iterations that the proposed method claims to outperform."},{"cited_title":"Projection methods and applications for seismic nonlinear inverse problems with multiple constraints,","cited_arxiv_id":null,"evidence_quote":"Represents prior constrained seismic inversion via projection methods with inner loops, the contrast behind the inner-loop-free claim."},{"cited_title":"A primal–dual splitting method for convex opti- mization involving lipschitzian, proximable and linear compos- ite terms,","cited_arxiv_id":null,"evidence_quote":"Supplies the primal-dual splitting iteration that Algorithm 1 applies directly to the TV- and box-constrained problem."},{"cited_title":"A review of the adjoint-state method for comput- ing the gradient of a functional with geophysical applications,","cited_arxiv_id":null,"evidence_quote":"Provides the adjoint-state formula used to compute the gradient $\\nabla E$ of the FWI misfit in every iteration."},{"cited_title":"Effi- cient projections onto the l 1-ball for learning in high dimen- sions,","cited_arxiv_id":null,"evidence_quote":"Gives the efficient projection onto the $\\ell_1$-ball used to evaluate the proximity operator of the TV-ball indicator."}],"review_version":1}