Pith. sign in

REVIEW 4 major objections 5 minor 33 references

Reduced-order modeling for electromagnetic inverse problems: a layered medium benchmark

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read 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

desk verdict 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. read the letter →

arxiv 2608.03996 v1 pith:ZXO3U7TS submitted 2026-08-04 math.NA cs.NAmath-phmath.APmath.MPphysics.comp-ph

Konstantinos Alexopoulos , Josselin Garnier This is my paper
classification math.NAcs.NAmath-phmath.APmath.MPphysics.comp-ph MSC 65M3235R3078A46
keywords reduced-ordermodelingfullwaveforminversionlayeredmediaGoupillaudmediumimpedancerecoveryMaxwellequationsdatamisfitinversescattering
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

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.

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 (4)
  1. [§5.2, Eq. (5.5); Appendix C.12 and C.5] 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.
  2. [§5.2, §C.5, §C.7] 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.
  3. [§4.4.2, Eq. (4.23) vs §5.2, Eq. (5.5)] 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.
  4. [§5.2, §C.5] 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.
minor comments (5)
  1. [§4.4.2, Eq. (4.21)] 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.
  2. [Appendix C.4] 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.
  3. [Figures 2–4] 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.
  4. [Lemma 2.5] 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.
  5. [§5.2] 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.

Circularity Check

1 steps flagged · score 4.0 of 10

Data-perturbation benchmark is partially circular: the fitted reference impedance defines both the ROM normalizer and the error target; the clean/layer experiments use an oracle reference, a separate limitation.

  1. fitted input called prediction [Section 5.2 and Appendix C.5 (multiplicative data noise); see also C.7]
    "In perturbation models that act directly on the data and may not correspond exactly to an admissible layered medium, an effective reference impedance is first fitted from the perturbed data. ... we first compute an effective reference impedance ζ_ref by minimizing the direct data misfit against D_obs. The ROM objective is then formed using R(ζ_ref), and the reconstruction errors are measured relative to ζ_ref."

    The quantity used to normalize the ROM objective and to define the reconstruction-error metric is itself the minimizer of the classical data misfit O_FWI against the same perturbed data. Thus in the data-perturbation experiments the classical method is being evaluated against its own fitted solution, and the ROM method is anchored at that solution. The reported errors and win fractions in these regimes therefore measure agreement with the classical least-squares fit, not independent recovery of the physical medium; the comparison is partially circular by construction.

full rationale

The paper's central claim is an empirical benchmark, not a derivation. The clean and layer-perturbation experiments are self-contained in the sense that the data are generated from a known Goupillaud medium and both objectives are minimized from random initializations; no fitted parameter is reused as the error target. However, those experiments set the ROM normalization reference ζ_ref to the true unknown medium (Section 5.2, C.3), which is an oracle choice: the objective is given the answer as an input. This is a serious limitation of the benchmark as a test of a deployable inversion objective, but it is not a circularity in the strict sense because the minimizer is not logically forced to equal ζ_ref. The genuine circular step is in the data-perturbation experiments: Appendix C.5 fits ζ_ref by minimizing the same direct data misfit that defines the classical method, then forms the ROM objective with R(ζ_ref) and measures both methods' errors relative to ζ_ref. That makes the comparison in those regimes self-referential. The time-origin experiment (Section 6.3), one of the claimed advantage regimes, is affected by the same ambiguity: Section 5.2 says data perturbations use a fitted effective reference, but Appendix C.5 only details this for multiplicative data noise, and C.7 states the error target is ζ⋆ for remaining models, leaving the actual ζ_ref used in the time-origin ROM objective unspecified. Overall, the clean/layer results have independent content but are oracle-normalized; the data-perturbation results are partially circular. I therefore assign a score of 4 rather than higher, because the main clean/layer comparison is not a reduction by construction and the fitted-reference circularity affects the supporting perturbation experiments more than the headline clean result.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The ledger is dominated by implementation choices and by the assumed validity of the source-filtered data processing. The key fitted quantity is zeta_ref in data-perturbation experiments, which is obtained by minimizing the classical misfit and then used to normalize the ROM objective; this is the main circularity burden in the numerical comparison.

free parameters (5)
  • Tikhonov regularization alpha = 1e-8
    Chosen by hand for the right inverse in O_ROM (Eq. C.11); objective depends on it.
  • time step tau = 1
    Normalized in the Goupillaud construction (Eq. C.4); sets the discrete grid and snapshot times.
  • number of snapshots n = 9
    Chosen so that the observation window covers the five-layer medium (Sections 5.1 and C.2); ROM size depends on it.
  • effective reference impedance zeta_ref = fitted by minimizing O_FWI against Dobs
    Sections 5.2 and C.5: fitted from perturbed data for data-perturbation experiments, then used to build the ROM normalization and as the error target; embeds a fitted value into the benchmark.
  • initialization perturbation scale = 0.3
    Random starts are theta_star .* (1 + 0.3 xi), Eq. (C.15); optimization results may depend on this choice.
assumptions (4)
  • domain assumption Goupillaud equal travel-time condition L_j - L_{j-1} = c_j tau for all j
    Section 4.2, Eq. (4.10): makes scattering occur on a discrete time grid and is required for the exact discrete dynamical system and ROM/data connection.
  • domain assumption Source-filtered and time-symmetrized data transformation represents the probe pulse as |f_hat(Lambda)| and lets measured data equal inner products of homogeneous wavefields
    Sections 2.2.5 to 2.3, especially the remark after Eq. (2.35): the paper states this is not a pointwise identity for the raw causal field but a processing step following [11]; the benchmark uses the transformed data.
  • domain assumption Reduction of Maxwell equations to a scalar 1D wave equation for transverse fields in a layered isotropic medium
    Section 3, assumptions (3.2)-(3.3) and Theorem 3.1; the numerical study only covers this 1D setting.
  • standard math Compactness of the Maxwell resolvent and spectral decomposition of -A^2
    Section 2.2.5, Eq. (2.29), relies on a standard Maxwell compactness property (e.g., [23]); used to define the snapshot basis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reduced-order modeling for electromagnetic inverse problems: a layered medium benchmark." pith.science (2026). https://pith.science/paper/ZXO3U7TS

@misc{pith2026260803996,
  author       = {Pith},
  title        = {Pith review of: Reduced-order modeling for electromagnetic inverse problems: a layered medium benchmark},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZXO3U7TS}},
  note         = {Machine review of arXiv:2608.03996}
}
read the original abstract

We study reduced-order modeling for inverse problems in layered media, focusing on the recovery of impedance profiles from time-domain measurements. Using the Goupillaud structure, we formulate the forward problem as a discrete dynamical system and introduce a ROM-based objective defined at the level of reduced operators. Through numerical experiments on a 5-layer medium under various structured perturbations, we compare the ROM-based objective with a classical data misfit. The results show that the ROM-based inversion provides a more favorable reconstruction in the clean setting and in certain structured perturbation regimes, while remaining consistently competitive with the classical approach across all cases considered.

Figures

Figures reproduced from arXiv: 2608.03996 by the authors.

Figure 2
Figure 2. Comparison of the two inversion strategies under simultaneous perturbations of all layers. [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Comparison of the two inversion strategies under a time-origin error. [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Comparison of the two inversion strategies under receiver-position errors modeled as amplitude [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figures from the paper (1 more)
Figure 1
Figure 1. Figure 1: Propagation in a Goupillaud medium. The equal travel-time condition ensures that waves move [PITH_FULL_IMAGE:figures/full_fig_p031_1.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 30 canonical work pages

  1. [1]

    M. I. Belishev. Recent progress in the boundary control method.Inverse problems, 23(5):R1–R67, 2007

    work page 2007

  2. [2]

    Benner, S

    P. Benner, S. Gugercin, and K. Willcox. A survey of projection-based model reduction methods for parametric dynamical systems.SIAM review, 57(4):483–531, 2015

    work page 2015

  3. [3]

    Blondel and B

    P. Blondel and B. J. Murton.Handbook of seafloor sonar imagery, volume 7. Wiley Chichester, 1997

    work page 1997

  4. [4]

    Borcea, V

    L. Borcea, V. Druskin, A. V. Mamonov, and M. Zaslavsky. Untangling the nonlinearity in inverse scattering with data-driven reduced order models.Inverse Problems, 34(6):065008, 2018

    work page 2018

  5. [5]

    Borcea, V

    L. Borcea, V. Druskin, A. V. Mamonov, M. Zaslavsky, and J. Zimmerling. Reduced order model approach to inverse scattering.SIAM Journal on Imaging Sciences, 13(2):685–723, 2020

    work page 2020

  6. [6]

    Borcea, V

    L. Borcea, V. Druskin, and J. Zimmerling. A reduced order model approach to inverse scattering in lossy layered media.Journal of Scientific Computing, 89(1):1, 2021

    work page 2021

  7. [7]

    Borcea, J

    L. Borcea, J. Garnier, A. V. Mamonov, and J. Zimmerling. Reduced order model approach for imaging with waves.Inverse Problems, 38(2):025004, 2022

    work page 2022

  8. [8]

    Borcea, J

    L. Borcea, J. Garnier, A. V. Mamonov, and J. Zimmerling. Waveform inversion via reduced order modeling.Geophysics, 88(2):R175–R191, 2023

    work page 2023

Show all 33 references
  1. [9]

    Borcea, J

    L. Borcea, J. Garnier, A. V. Mamonov, and J. Zimmerling. Waveform inversion with a data driven estimate of the internal wave.SIAM Journal on Imaging Sciences, 16(1):280–312, 2023. 28

    2023

  2. [10]

    Borcea, J

    L. Borcea, J. Garnier, A. V. Mamonov, and J. Zimmerling. When data driven reduced order modeling meets full waveform inversion.SIAM Review, 66(3):501–532, 2024

    2024

  3. [11]

    Borcea, Y

    L. Borcea, Y. Liu, and J. Zimmerling. Electromagnetic inverse wave scattering in anisotropic media via reduced order modeling.Journal of Computational Physics, 515:113272, 2024

    2024

  4. [12]

    S. L. Brunton and J. N. Kutz.Data-driven science and engineering: Machine learning, dynamical systems, and control. Cambridge University Press, 2022

    2022

  5. [13]

    Cheney and B

    M. Cheney and B. Borden.Fundamentals of radar imaging. SIAM, 2009

    2009

  6. [14]

    J. C. Curlander and R. N. McDonough.Synthetic aperture radar, volume 11. Wiley, New York, 1991

    1991

  7. [15]

    Druskin, A

    V. Druskin, A. V. Mamonov, A. E. Thaler, and M. Zaslavsky. Direct, nonlinear inversion algorithm for hyperbolic problems via projection-based model reduction.SIAM Journal on Imaging Sciences, 9(2):684–747, 2016

    2016

  8. [16]

    Druskin, A

    V. Druskin, A. V. Mamonov, and M. Zaslavsky. A nonlinear method for imaging with acoustic waves via reduced order model backprojection.SIAM Journal on Imaging Sciences, 11(1):164–196, 2018

    2018

  9. [17]

    Engquist and B

    B. Engquist and B. D. Froese. Application of the wasserstein metric to seismic signals.arXiv preprint arXiv:1311.4581, 2013

    2013 arXiv

  10. [18]

    Engquist and Y

    B. Engquist and Y. Yang. Optimal transport based seismic inversion: Beyond cycle skipping. Communications on Pure and Applied Mathematics, 75(10):2201–2244, 2022

    2022

  11. [19]

    Fouque, J

    J.-P. Fouque, J. Garnier, G. Papanicolaou, and K. Solna.Wave propagation and time reversal in randomly layered media, volume 56. Springer Science & Business Media, 2007

    2007

  12. [20]

    Gilman, E

    M. Gilman, E. Smith, and S. Tsynkov.Transionospheric synthetic aperture imaging. Springer, 2017

    2017

  13. [21]

    J. S. Hesthaven, C. Pagliantini, and G. Rozza. Reduced basis methods for time-dependent problems. Acta Numerica, 31:265–345, 2022

    2022

  14. [22]

    Huang, R

    G. Huang, R. Nammour, and W. W. Symes. Source-independent extended waveform inversion based on space-time source extension: Frequency-domain implementation.Geophysics, 83(5):R449–R461, 2018

    2018

  15. [23]

    Monk.Finite element methods for Maxwell’s equations

    P. Monk.Finite element methods for Maxwell’s equations. Oxford university press, 2003

    2003

  16. [24]

    Stefanov and G

    P. Stefanov and G. Uhlmann. Stable determination of generic simple metrics from the hyperbolic dirichlet-to-neumann map.International Mathematics Research Notices, 2005(17):1047–1061, 2005

    2005

  17. [25]

    W. W. Symes. Migration velocity analysis and waveform inversion.Geophysical prospecting, 56(6):765–790, 2008

    2008

  18. [26]

    W. W. Symes. Error bounds for extended source inversion applied to an acoustic transmission inverse problem.Inverse Problems, 38(11):115002, 2022

    2022

  19. [27]

    T. L. Szabo.Diagnostic ultrasound imaging: inside out. Academic press, 2013

    2013

  20. [28]

    A. Tether. Construction of minimal linear state-variable models from finite input-output data.IEEE Transactions on Automatic Control, 15(4):427–436, 1970

    1970

  21. [29]

    Van Leeuwen and F

    T. Van Leeuwen and F. J. Herrmann. Mitigating local minima in full-waveform inversion by ex- panding the search space.Geophysical Journal International, 195(1):661–667, 2013

    2013

  22. [30]

    Virieux and S

    J. Virieux and S. Operto. An overview of full-waveform inversion in exploration geophysics.Geo- physics, 2010. 29

    2010

  23. [31]

    Warner and L

    M. Warner and L. Guasch. Adaptive waveform inversion: Theory.Geophysics, 81(6):R429–R445, 2016

    2016

  24. [32]

    Y. Yang, B. Engquist, J. Sun, and B. F. Hamfeldt. Application of optimal transport and the quadratic wasserstein metric to full-waveform inversion.Geophysics, 83(1):R43–R62, 2018

    2018

  25. [33]

    O. Yu. Imanuvilov and M. Yamamoto. Global uniqueness and stability in determining coefficients of wave equations.Communications in Partial Differential Equations, 26(7-8):1409–1425, 2001. 30 t = 0 t =τ t = 2τ t = 3τ t = 4τ A0 0B0 0 A1 1 B1 1 A2 0B2 0 A2 2 B2 2 A3 1 B3 1 A4 0B4...

    2001

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.