{"id":"8e98f7d5-b368-43ca-be00-793f7e75ec25","arxiv_id":"2608.03996","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"On a five-layer Goupillaud benchmark, a normalized reduced-operator objective outperforms classical data misfit in several regimes and stays competitive in the rest.","lead":"This paper compares a reduced-order-modeling (ROM) based objective with classical data misfit for recovering layered-medium impedance profiles from wave data. In a five-layer Goupillaud benchmark, the ROM objective gives smaller median reconstruction errors in clean, layer-perturbed, and time-origin-error cases, and remains competitive otherwise.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Oracle normalization: clean/layer ROM advantage may be an artifact of setting ζ_ref to the true medium; the data-perturbation circularity is secondary.","rationale":"I read the paper in good faith and find the Goupillaud-based benchmark and the numerical protocol largely transparent. The central empirical claim, however, depends critically on the reference impedance ζ_ref used in the ROM normalization. The most load-bearing issue is not the reader's circularity claim about data perturbations — which is only partially accurate for the time-origin experiment, where errors are measured against ζ⋆ — but the fact that in the clean and layer-perturbation experiments, ζ_ref is chosen to be the true medium. This gives the ROM objective access to the unknown solution in the two experiments that drive the abstract's headline. The paper does not discuss how ζ_ref would be chosen in practice, nor whether the reported advantage persists under a non-oracle choice. A single re-run of Section 6.1 and 6.2 with a fixed non-oracle ζ_ref would settle whether the claimed improvement is a genuine landscape property or an artifact of normalization. Because this is an addressable empirical issue rather than a demonstrated error, the conditional verdict remains appropriate; no reason to reject or accept outright.","tokens_in":21235,"tokens_out":11994,"duration_ms":140556,"concrete_test":"Recompute the clean experiment (Sec. 6.1) and the layer-perturbation experiment (Sec. 6.2) with a fixed, non-oracle reference impedance for all trials, e.g., ζ_ref = (1,1,1,1,1), and also with ζ_ref set to the mean of the random initializations. Keep the same random seeds, optimizer, and 50 trials, and compare the median errors and win fractions to Figures 2 and 3. If the ROM median error remains near 1e-4 in the clean case and the win fractions stay above 0.5 in the layer case, the advantage is robust to the reference choice. If the ROM error rises to the FWI level, the reported advantage is an artifact of oracle normalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (5.5) defines O_ROM via R(ζ)R(ζ_ref)^†_α − I_5. Section 5.2 explicitly sets ζ_ref = ζ⋆ in the clean experiment and ζ_ref = ζ_pert (the ground truth) in the layer-perturbation experiment. Thus the ROM objective in the two experiments where the advantage is strongest is normalized with the true unknown model. The classical FWI objective has no analogous oracle input. The reported clean result (1.34e-4 vs 1.88e-3 median error) may therefore reflect a benefit of true-model normalization rather than a property of the ROM misfit as a deployable inversion objective, since a practitioner would have to choose ζ_ref before knowing the solution. In the data-perturbation experiments, ζ_ref is fitted by minimizing O_FWI (Sec. 5.2, C.5), which anchors the ROM objective at the classical solution. The reader's circularity concern is only partly valid: for time-origin, errors are still measured against ζ⋆ (C.7), so the comparison is not directly circular, but the normalization is nevertheless derived from the classical solution. Either way, the central claim is contingent on an unspecified, experiment-dependent choice of ζ_ref.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a reduced-order-modeling (ROM) inversion objective for electromagnetic inverse problems in one-dimensional layered Goupillaud media. It derives the forward model from Maxwell's equations, reduces it to a scalar layered wave equation, constructs a data-driven ROM, and compares a normalized ROM objective with a classical data misfit in a five-layer benchmark. The numerical experiments report ROM advantages in the clean setting, under layer perturbations, and under a time-origin error, and comparable performance under other structured data perturbations. The mathematical derivations in Sections 2–4 are mostly standard and internally consistent, but the numerical comparison is affected by how the reference impedance entering the ROM normalization is chosen.","tokens_in":21508,"tokens_out":4768,"duration_ms":51528,"significance":"If the reported advantages were robust, the paper would provide a useful benchmark supporting normalized reduced-operator misfits as alternatives to classical L2 waveform inversion for layered media. The paper has strengths: the perturbation taxonomy is systematic, the Goupillaud setup is exactly discretizable, and the clean and layer-perturbation experiments are internally consistent. However, the central comparative claim is undermined by the oracle choice of the reference impedance in exactly the experiments where the ROM advantage is strongest, and by the data-perturbation protocol that derives the reference from the classical misfit. The theoretical value of the ROM construction is not in question; what is in question is whether the numerical evidence supports the abstract claim that the ROM objective itself is more favorable.","major_comments":[{"comment":"In the clean experiment, ζ_ref is set to ζ⋆, the true unknown impedance, and in the layer-perturbation experiment, ζ_ref is set to the perturbed ground-truth medium. The classical misfit has no analogous oracle input. The clean-setting advantage reported in §6.1 (median 1.34e-4 vs 1.88e-3) and the layer-perturbation advantage in §6.2 could therefore be an artifact of normalizing the ROM objective with the true model rather than a property of the ROM misfit as a deployable inversion objective. A practitioner would have to choose ζ_ref before knowing the solution. This is load-bearing for the abstract claim. Please re-run the experiments with ζ_ref chosen independently of the true medium (e.g., a fixed reference, a function of the initial guess, or using a separate calibration step), or explicitly demonstrate that the reported advantages are insensitive to a wide range of ζ_ref choices.","section":"§5.2, Eq. (5.5); Appendix C.12 and C.5"},{"comment":"For data perturbations, the manuscript says an effective reference impedance is fitted by minimizing the direct data misfit against the perturbed data, and this fitted ζ_ref is then used to normalize the ROM objective. If the fitted ζ_ref is the classical solution or close to it, the ROM objective is artificially anchored at the classical minimizer, biasing any comparison. The issue applies directly to the multiplicative-noise experiment and potentially to the time-origin and other data-perturbation experiments. The manuscript is also ambiguous: C.5 specifies the fitted-reference procedure only for multiplicative data noise, while C.7 refers to the 'data-noise experiment' without clarifying which experiments use it. Please specify exactly which experiments fit ζ_ref, and use a reference constructed without solving the classical inverse problem.","section":"§5.2, §C.5, §C.7"},{"comment":"The theoretical objective in Eq. (4.23) is defined as ||R(ζ)R^{-1} - I_{2nm}||_F^2, while the numerically implemented objective in Eq. (5.5) is ||R(ζ)R(ζ_ref)^†_α - I_5||_F^2. These are not the same: the former uses an inverse of the data-driven ROM operator, the latter uses a Tikhonov-regularized right inverse of R(ζ_ref), and the dimensions differ. This inconsistency makes it unclear which objective is actually minimized and whether the reported behavior is a property of the proposed ROM misfit or of the particular finite-dimensional regularization used. Please reconcile the definitions and explain the dimension reduction from 2nm to 5 in the layered benchmark.","section":"§4.4.2, Eq. (4.23) vs §5.2, Eq. (5.5)"},{"comment":"The comparison of the ROM and classical objectives is not symmetric with respect to regularization. The ROM objective uses the Tikhonov parameter α=1e-8 in a right inverse, while the classical FWI objective has no analogous regularization term. The sensitivity of the reported win fractions and median errors to α is not examined. Since the normalization step is central to the claimed advantage, please provide a sensitivity study over α and over the number of snapshots n, or justify that the chosen values do not favor either method.","section":"§5.2, §C.5"}],"minor_comments":[{"comment":"Equation (4.21) defines O(ζ)=||R-R(ζ)||_F^2, but the text then says 'Instead of directly minimizing O', which is inconsistent with the subsequent definition of O_ROM. Clarify the relationship between O, O_FWI, and O_ROM.","section":"§4.4.2, Eq. (4.21)"},{"comment":"The optimization protocol uses 20 random initializations for the clean inversion, but the number of initializations per perturbation trial is not stated. Please specify whether each of the 50 Monte Carlo trials uses the same initialization protocol and whether the reported medians are over trials or over all initializations.","section":"Appendix C.4"},{"comment":"The axis labels in the plotted figures appear garbled (e.g., '1005 × 10−1 6 × 10−1' in Figure 2). Please check the figure rendering and ensure the σ values match the grids in Appendix C.6.","section":"Figures 2–4"},{"comment":"The proof of Lemma 2.5 is deferred to 'the same logic as in [11]'. Since the paper's contribution includes adapting the ROM framework to the layered Goupillaud setting, a short self-contained proof would improve the manuscript's usefulness.","section":"Lemma 2.5"},{"comment":"The phrase 'normalized ROM objective' is used for Eq. (5.5), but the normalization depends on a reference impedance that changes per experiment. Consider using a term such as 'reference-normalized ROM objective' to avoid implying a parameter-free normalization.","section":"§5.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is not suitable for publication in its present form. The central comparative claim rests on an oracle choice of ζ_ref in the clean and layer-perturbation experiments, and on a fitted reference derived from the classical method in the data-perturbation experiments. These issues are fixable in principle by redesigning the experimental protocol, but the current evidence does not support the abstract claim that the ROM objective is intrinsically more favorable. If the authors cannot demonstrate a non-oracle advantage, the paper should be reframed as a benchmark without comparative superiority claims. I would also ask the editor to ensure that the final version reconciles Eq. (4.23) with Eq. (5.5)."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's main claim does not survive contact with Section 5.2. In the clean experiment and the layer-perturbation experiment, the ROM objective is normalized using R(ζ_ref) with ζ_ref set to the true medium that generated the data. That is an oracle input. The classical FWI misfit gets no analogous help, so the reported 1.34e-4 versus 1.88e-3 median errors do not support the abstract's claim that the ROM objective is more favorable in clean and structured perturbation regimes. They support the weaker statement that an objective normalized by the unknown solution can have a more favorable landscape, which is not something a practitioner can exploit.\n\nWhat is actually good here: the paper is transparent about what it adapts. It explicitly says the framework follows [6,11] and that the contribution is the layered benchmark, the normalized operator objective, and the perturbation study. The Maxwell-to-Goupillaud derivation is standard and clean, and Appendix C is detailed enough to reproduce the experiments. The taxonomy of perturbations — layer, time-origin, gain, bias, noise — is genuinely useful, and the authors resist overclaiming in the closing sections by reporting regimes where the two methods behave comparably.\n\nThe data-perturbation experiments have a related but distinct problem. For several of them, ζ_ref is fitted by minimizing the classical data misfit against the perturbed data, then used to build the ROM normalization. The errors are still measured against ζ⋆ in most cases, so the reader's \"circular\" label is only partly right. But the normalization is still derived from the classical solution, which means the comparison does not isolate the ROM objective as a standalone misfit. You would need to know the classical answer to set up the ROM objective the way they do.\n\nMinor issues: no error bars, no code or data release, and the medians over 50 trials come without confidence intervals. Those are addressable.\n\nWho gets value from this: someone working on ROM-based inversion will find the benchmark and the perturbation setup worth reading, mainly as a caution about normalization choices. The math is fine and the writing is honest. But the central numerical conclusion needs to be re-benchmarked with a non-oracle reference — for example, a data-driven ROM from the observed data or a reference from a reasonable initial model, with sensitivity to that choice reported.\n\nRecommendation: send it to peer review, but with a clear request to fix the normalization protocol. The paper deserves referee time; the current version's headline claim does not.","headline":"Oracle normalization undercuts the central claim: the clean and layer-perturbation advantages are measured against a ROM objective that has been handed the true medium as its reference, so the headline comparison is not a deployable one.","tokens_in":21994,"tokens_out":2687,"would_cite":false,"duration_ms":31531,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M32","35R30","78A46"],"pacs":[],"model":"deepseek-v4-flash","headline":"For layered-medium inverse problems, an objective that compares normalized data-driven reduced operators recovers the impedance profile more accurately than the classical data misfit, with median error 1.34e-4 versus 1.88e-3 in the clean ca","keywords":["reduced-order modeling","full waveform inversion","layered media","Goupillaud medium","impedance recovery","Maxwell equations","data misfit","inverse scattering"],"falsifier":"Re-run the time-origin and data-noise benchmarks with the reference impedance held fixed at the true medium instead of fitted to the perturbed data, and recompute the win fractions; if the ROM advantage vanishes or reverses, the reported advantage in those regimes is an artifact of reference fitting rather than a property of the objective.","tokens_in":21112,"feed_emoji":"📡","tokens_out":5610,"duration_ms":60158,"temperature":0.7,"pith_summary":"This paper argues that, for impedance recovery in layered media, a reduced-order-model (ROM) misfit—comparing normalized reduced operators built directly from the measured data—is a better objective than the classical least-squares data misfit. The claim is tested on a five-layer Goupillaud medium, where the equal-travel-time structure turns wave propagation into an exact discrete dynamical system. In clean synthetic data the ROM objective reaches a median reconstruction error of 1.34e-4 versus 1.88e-3 for the data misfit, and it also wins a majority of trials under layer perturbations and time-origin errors. Under multiplicative noise, speed uncertainty, gain bias, and other structured data errors the two objectives perform comparably, so the paper's conclusion is that the ROM objective is advantageous in coherent or structurally organized regimes and never systematically worse. A careful reader would care because it points to a practical alternative misfit for waveform inversion that may mitigate cycle skipping.","feed_headline":"ROM misfit cuts clean layered-inversion error 14-fold","feed_subtitle":"A normalized reduced-operator objective wins on layer and time-origin perturbations and stays competitive on noise.","key_machinery":"Goupillaud medium: a layered medium with equal travel time tau per layer, so reflection and transmission recurrences produce exact discrete-time scattering events. The data-driven ROM matrix R(zeta), whose entries are inner products of wave snapshots, is assembled directly from measured traces; the normalized objective O_ROM(zeta) = || R(zeta) R(zeta_ref)^+_alpha - I_5 ||_F^2 with Tikhonov-regularized right inverse compares reduced operators and removes global scaling. This object carries the argument: the paper claims its optimization landscape is more favorable and its normalization is more stable under coherent perturbations.","core_discovery":"The paper's central claim is that the inverse problem for a layered electromagnetic medium can be reformulated at the level of the reduced propagator rather than the raw traces, and that this reformulation has better optimization behavior. Using the Goupillaud equal-travel-time assumption, the forward map becomes a discrete recurrence, and the data define a 5x9 ROM matrix R(zeta) whose entries are snapshot inner products. The paper proposes minimizing O_ROM(zeta)=||R(zeta) R(zeta_ref)^+ - I_5||_F^2, where the Tikhonov right inverse normalizes away global scaling. In the clean benchmark, this objective gives median impedance error 1.34e-4 versus 1.88e-3 for the data misfit; it also gives smal","pith_inferences":["The time-origin and data-noise advantage may partly reflect the reference-fitting step: fitting zeta_ref by minimizing the classical misfit against the perturbed data could already encode the classical solution, so holding zeta_ref fixed would be a cleaner test.","The Goupillaud benchmark is exactly discretizable; extending to non-Goupillaud or continuous layered media would test whether the normalized-operator advantage survives approximate discretization and partial travel-time mismatch.","One could combine the ROM misfit with the classical data term as a regularizer, aiming to inherit both the stable landscape of the reduced-operator comparison and the statistical optimality of the data misfit under uncorrelated noise.","The fitted effective reference could be replaced by an estimate independent of the classical objective, such as a direct data-driven calibration, to remove the potential circularity in the perturbation experiments."],"forward_implications":["If the claim holds, the ROM-based objective is a drop-in alternative misfit for layered waveform inversion, requiring no knowledge of the internal wave field.","Clean-data inversions initialized randomly terminate closer to the true impedance, indicating a shallower or less oscillatory objective landscape.","Win fractions above 0.5 under layer and time-origin perturbations imply the normalization is robust to coherent distortions that shift or rescale the recorded arrivals.","Comparable performance under noise and gain errors means the ROM objective can be used without fear of systematic degradation in typical data-quality scenarios.","The explicit Goupillaud connection gives a small exact benchmark for testing other misfits, regularizers, or initialization strategies for layered inverse problems."],"supporting_citations":[{"why":"Introduces data-driven reduced-order models for inverse scattering, the foundation for building ROMs from measured traces.","marker":"[4]"},{"why":"Develops the reduced-order-model approach to inverse scattering that the paper's operator-comparison objective extends.","marker":"[5]"},{"why":"Applies ROM inversion to lossy layered media, the closest prior layered setting to this benchmark.","marker":"[6]"},{"why":"Shows waveform inversion via reduced-order modeling, motivating the reduced-operator misfit.","marker":"[8]"},{"why":"Connects data-driven ROMs to full-waveform inversion and supplies the normalized objective used here.","marker":"[10]"},{"why":"Provides the electromagnetic ROM construction in anisotropic media that this paper adapts to Goupillaud layers.","marker":"[11]"},{"why":"Supplies the Galerkin projection-based model reduction for hyperbolic systems used to derive the ROM.","marker":"[15]"}],"fun_headline_variants":["ROM misfit cuts layered inversion error 14x","Layered EM inversion error drops 14x with ROM misfit","ROM misfit beats data misfit in layered EM inversion","Reduced-operator misfit cuts layered EM error 14x","ROM objective trims layered EM inversion error 14-fold"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"For the perturbed-data experiments, the comparison assumes that fitting the effective reference impedance by minimizing the classical data misfit against the perturbed data does not bias the comparison toward the ROM objective; if it does, the time-origin and noise advantages are partly circular, while the clean and layer-perturbation results stand separately.","fun_headline_variants_meta":{"raw":{"variants":["ROM misfit cuts layered inversion error 14x","Layered EM inversion error drops 14x with ROM misfit","ROM misfit beats data misfit in layered EM inversion","Reduced-operator misfit cuts layered EM error 14x","ROM objective trims layered EM inversion error 14-fold"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000639,"raw_usage":{"total_tokens":2727,"prompt_tokens":642,"completion_tokens":2085,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":386,"completion_tokens_details":{"reasoning_tokens":2001}},"tokens_in":386,"tokens_out":2085,"duration_ms":16094,"temperature":1.0,"reasoning_tokens":2001,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T04:20:49.248497+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the time-origin and data-noise benchmarks with the reference impedance held fixed at the true medium instead of fitted to the perturbed data, and recompute the win fractions; if the ROM advantage vanishes or reverses, the reported advantage in those regimes is an artifact of reference fitting rather than a property of the objective.","supporting_citations":[{"cited_title":"Borcea, V","cited_arxiv_id":null,"evidence_quote":"Introduces data-driven reduced-order models for inverse scattering, the foundation for building ROMs from measured traces."},{"cited_title":"Borcea, V","cited_arxiv_id":null,"evidence_quote":"Develops the reduced-order-model approach to inverse scattering that the paper's operator-comparison objective extends."},{"cited_title":"Borcea, V","cited_arxiv_id":null,"evidence_quote":"Applies ROM inversion to lossy layered media, the closest prior layered setting to this benchmark."},{"cited_title":"Borcea, J","cited_arxiv_id":null,"evidence_quote":"Shows waveform inversion via reduced-order modeling, motivating the reduced-operator misfit."},{"cited_title":"Borcea, J","cited_arxiv_id":null,"evidence_quote":"Connects data-driven ROMs to full-waveform inversion and supplies the normalized objective used here."},{"cited_title":"Borcea, Y","cited_arxiv_id":null,"evidence_quote":"Provides the electromagnetic ROM construction in anisotropic media that this paper adapts to Goupillaud layers."},{"cited_title":"Druskin, A","cited_arxiv_id":null,"evidence_quote":"Supplies the Galerkin projection-based model reduction for hyperbolic systems used to derive the ROM."}],"review_version":1}