WorksMathematics
Reduced-order modeling for electromagnetic inverse problems: a layered medium benchmark
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.
Introduction
Wave propagation in heterogeneous media plays a central role in many areas of science and engineering, including radar and sonar imaging [3, 13, 14, 20], seismology and geophysical exploration [25, 30] and medical imaging via ultrasound [27]. In these applications, one seeks to recover information about an inaccessible medium from measurements of waves generated by controlled sources and recorded by an array of sensors. This leads to inverse problems for wave equations, where the goal is to estimate spatially varying material parameters such as wave speed, density, or electromagnetic coefficients.
Mathematically, wave propagation is typically modeled by second-order wave equations or, more generally, by first-order hyperbolic systems [19, 23]. The unknown medium properties enter as coefficients in these equations and the measurements correspond to partial observations of the resulting wave field. The associated inverse problem is highly nonlinear and ill-posed, as the mapping from the coefficients to the observed data is complicated and sensitive to perturbations. Fundamental questions of uniqueness and stability have been extensively studied in the literature [1, 24, 33], but these results often rely on idealized data that are not available in practical applications.
In practice, waveform inversion is commonly formulated as a nonlinear least-squares optimization problem, known as full waveform inversion (FWI), where one minimizes the discrepancy between measured and simulated data. Despite its widespread use, FWI suffers from severe limitations due to the non-convexity of the objective function. In particular, for band-limited and high-frequency data, the objective function exhibits numerous local minima, a phenomenon known as cycle skipping, which causes gradient-based optimization methods to fail even when initialized close to the true solution.
Several approaches have been proposed to mitigate these difficulties. One direction consists of replacing the classical data misfit by alternative metrics, such as the Wasserstein distance from optimal transport theory [17, 32]. This approach can improve the convexity of the objective function in simple settings [18], although this property does not hold in general heterogeneous media [10]. Another approach is the so-called modeling operator extension or extended FWI [22, 29, 31], which introduces additional degrees of freedom in the inversion procedure. While effective in certain configurations, its theoretical guarantees are limited to simplified acquisition geometries [26].
A different and promising strategy is based on reduced-order modeling (ROM) [4, 5, 7, 8, 9, 10, 15, 16, 28]. In this framework, one constructs a low-dimensional dynamical system that captures the essential features of wave propagation in the unknown medium. These ROMs are data-driven: they are computed directly from the measurements without requiring knowledge of the wave field inside the medium. The construction relies on Galerkin projection onto spaces spanned by wave snapshots [2, 12, 21], leading to reduced operators that encode the propagation dynamics. Inversion methods based on ROM have been shown to yield objective functions with improved optimization properties compared to classical FWI. Despite these advances, several challenges remain. In particular, understanding the behavior of ROM-based inversion in structured media and under realistic perturbations is still an open problem. Moreover, while the general ROM framework applies to a wide class of wave equations, its concrete implementation and analysis often benefit from simplified settings that retain the essential features of wave propagation, such as multiple scattering and reflection phenomena.
In this work, we investigate reduced-order modeling for inverse problems in layered media, with a focus on Goupillaud-type structures. This setting provides a natural discrete representation of wave propagation, where scattering events occur at discrete times and the wave field evolves according to a dynamical system. This structure allows for an explicit connection between wave propagation, data acquisition and reduced-order modeling. Our main contributions are as follows. The reduced-order modeling framework that we use follows closely the construction introduced in [6, 11]. Our contribution is to adapt and analyze this framework in the specific setting of layered media, where the wave dynamics admit a discrete representation. First, we formulate a ROM-based objective for layered media that compares normalized reduced operators and removes global scaling effects. Then, we establish a precise connection between the Goupillaud structure and a data-driven discrete dynamical system, allowing for an explicit ROM construction. Finally, we provide a systematic numerical study under structured perturbations, identifying regimes in which the ROM-based inversion improves stability and reconstruction accuracy while remaining competitive in all cases.
The paper is organized as follows. In § 2, we introduce the electromagnetic forward problem and derive its first-order hyperbolic formulation. In § 3, we show how this system can be reduced to a scalar wave equation in a layered medium. § 5 is devoted to the analysis of one-dimensional layered media and the connection with the ROM framework. In § 6, we formulate the inverse problem and introduce the ROM-based objective function. § 7 presents numerical experiments comparing the ROM-based inversion with classical approaches under various perturbations. Finally, § 8 concludes the paper with a discussion of the results and perspectives for future work.
Problem formulation
We consider the inverse problem of recovering properties of an electromagnetic medium from measurements recorded by an array of antennas. The propagation of electromagnetic waves is governed by Maxwell’s equations. In this section we formulate the forward problem, introduce the frequency-domain representation, and describe the reduced-order modeling framework that will be used for inversion.
Maxwell’s equations
Let be a compact domain with Lipschitz boundary . For and , the electromagnetic fields consist of the electric field and the magnetic flux density . We assume that the medium is described by a dielectric permittivity tensor and a scalar magnetic permeability . The tensor is assumed to be symmetric and positive definite, while is assumed to be strictly positive. Both coefficients are time independent.
The fields satisfy Maxwell’s equations
where and denote the electric and magnetic current sources, respectively. We consider vanishing initial conditions prior to the excitation,
On the boundary of the truncated domain, we use a standard conservative Maxwell realization. For definiteness, we impose perfectly conducting boundary conditions,
Lemma 2.1. Introducing the state vector
Maxwell’s equations can be written as a first-order hyperbolic system of the form
where
and
Here the notation means that the multiplication by is applied before taking the curl.
Lemma 2.2. Using the normalized fields
where denotes the symmetric positive definite square root of , the system (5) becomes
with
The differential operator is
and
In (11), the coefficient is part of the argument of the curl. Thus
Lemma 2.3. The operator matrix is skew-symmetric on the above domain. With the standard perfectly conducting Maxwell realization, it is skew-adjoint.
Frequency-domain formulation
Frequency-domain formulation
Excitation expression
We assume that the data for the inversion are collected by an array of antennas. Let denote the polarizations. Then the excitation of the antenna is given by
and
This choice ensures that the normalized source has the same form in all polarizations. Here are the canonical basis vectors in .
Here, denotes the probing pulses emitted by the antennas. They are supported in a short time interval and have Fourier transform that is non-negligible at frequencies satisfying . The pulse satisfies
Component-wise problem formulation
Using the polarization of each antenna, we write
Substituting (13), (14) and (16) into (9), we get
Let be the canonical basis in and introduce the antenna excitation matrix
Then the component-wise system becomes
The associated initial condition is
The boundary conditions are chosen so that the boundary contribution in the curl Green formula vanishes. For the perfectly conducting realization used here, this corresponds in the normalized variables to
The natural domain of is therefore
Wave decomposition
For the potential representation of the null-space component, we assume for simplicity that is simply connected with connected boundary.
Theorem 2.4 (Helmholtz decomposition). The solution to (18) with the initial and boundary conditions above has the Helmholtz decomposition
where and , for and . The operator is defined by
Proof. See Appendix A.4.
Remark. The propagating component is orthogonal to the null space of and satisfies the corresponding weak divergence constraints. This is the analogue of the standard Maxwell decomposition; see, for instance, Lemma 4.5 in [23].
Let denote the orthogonal projection onto , and let . For
we write
The null-space component satisfies
The range component satisfies
where is the range component of .
Array response matrix
Let us assume that the antenna profiles are spatially localized in a small neighborhood of the receiver position . More precisely, each is supported in a region whose diameter is small compared with the wavelength and satisfies the normalization condition
In this sense, the functions act as approximate Dirac distributions centered at . As a result, the measurements correspond to localized averages of the wave field, and we use the approximation
We define the array response matrix to be the time dependent matrix . In the column of this matrix, we store the recording of at all the antennas. Then,
Let us emphasize that, although the general formulation allows for anisotropic media and variable permeability, we assume that the medium is known, isotropic and homogeneous in the vicinity of the array. In this region, and , so the reference scalar wave speed is
By vicinity, we mean within a distance traveled by the waves over the duration of the probing pulse .
Thus, since the medium is known and homogeneous within a distance from the array, we can compute for and remove the component lying in the null space. Then, we obtain the transformed array response matrix , with entries
Spectral decomposition
Since we have removed the null space components of , let
be the propagating subspace. We denote by
the restriction of the Maxwell operator to this subspace. The operator is skew-adjoint on the projected range space. Hence is positive self-adjoint. Under the Maxwell compactness property for the chosen realization, the operator has compact resolvent on ; see, e.g., [23]. Therefore there exists an orthonormal basis of the considered space and a sequence of strictly positive eigenvalues , possibly with multiplicity, such that
We then define
It follows that for each , the pair spans a two-dimensional invariant subspace of , and satisfies
In the case of multiple eigenvalues, the construction is understood within each eigenspace. For the full proof, we refer to Appendix § 9.§ 6.
For initial data expanded in the components,
the solution to the homogeneous problem
is
where
We now return to the forced problem. The solution can be written as a convolution in time between the source term and the homogeneous propagator. In the same spirit as in [11], this can be interpreted as mapping the source excitation to an effective initial condition.
Remark. Since is skew-adjoint on the projected range space, is positive self-adjoint. We therefore define the source filter through the functional calculus of the positive operator , rather than through itself.
More precisely, we write
where
solves the homogeneous problem. Equivalently, using the spectral decomposition of ,
with
Remark. Since is skew-adjoint, Stone’s theorem implies that generates a strongly continuous unitary group . Thus, the notation is understood in this semigroup sense. On each two-dimensional invariant subspace associated with an eigenfunction of , this unitary evolution is represented by (33).
Remark. The representation (35) should be understood as the source-filtered formulation used in the ROM construction, following the data transformation in [11]. It is not meant as a pointwise identity for the raw causal field before processing. Rather, after the source filtering and time symmetrization of the measured response, the probing pulse is represented by the positive spectral multiplier , while denotes the corresponding homogeneous evolution.
The effective initial condition is
Following the data transformation in [11], we now work with the processed array response, obtained from the measured response by source filtering and time symmetrization. This transformation absorbs the effect of the probing pulse into the effective initial condition and allows the data to be written as inner products of homogeneous wave fields. We denote the transformed data matrix by :
With the above definition, the transformed data satisfy
This data transformation follows the ROM framework developed for wave-based inverse scattering in [4, 5, 7, 8, 10].
ROM formulation
ROM formulation
The reduced-order modeling framework used here follows the general approach developed in [11].
Wave snapshots
We are interested in the evolution of the wave fields (36) on a uniform time grid given by
where the step is chosen according to the Nyquist criterion . In order to ease the notation, we define the dimensional fields , called the wave snapshot at time , given by
Since is skew-adjoint on the propagating subspace, Stone’s theorem implies that it generates a unitary group . Hence, from (36),
for and . In particular, we know that
for . Thus, we obtain
where is the propagator operator, given by
Thus, (44) defines a discrete-time dynamical system with initial state .
The ROM is an algebraic discrete dynamical system, obtained via the Galerkin projection of (44) on the space
The end time index is chosen according to the distance from the array at which we wish to image. If the characteristic length of the medium is , then we should have . This construction is standard in data-driven ROM approaches based on wave snapshots and Galerkin projections [11].
Remark. The condition ensures that the observation time window is sufficiently long for waves to propagate across the medium and return to the array. In particular, this guarantees that the collected snapshots contain information about reflections from the entire domain. If this condition is not satisfied, deeper parts of the medium are not adequately probed and the resulting reduced-order model may fail to capture the relevant wave interactions.
Data-driven ROM
The Galerkin approximation of (43) is defined by
where are the Galerkin coefficients, calculated so that when substituting (47) in (44), the residual is orthogonal to the space .
In order to obtain the Galerkin equation, we use an orthonormal basis , given by
The ROM
whose components satisfy
Since our basis is causal, we have
Explicitly, is defined via the Gram-Schmidt orthogonalisation of , i.e.
where is block upper triangular.
The equation for the Galerkin coefficients can now be written as
with initial condition
where we have introduced the mass matrix , given by
and the stiffness matrix , given by
of the Galerkin scheme. This construction is the standard data-driven ROM construction based on snapshot Gram matrices; see, for example, [5, 7, 15, 16]. The following results hold for the mass and stiffness matrices.
Lemma 2.5. It can be shown that
and that the first Galerkin coefficients are the block columns of ,
where denotes the -th block column of .
Proof. The proof follows the same logic as in [11].
Lemma 2.6. We can show that both the mass and the stiffness matrix are data driven, i.e.
and
These identities show that the ROM can be constructed directly from the data, a key feature of the approach introduced in [10].
The ROM
We define the ROM to be the discrete time analog of (44), with states
called the ROM snapshots. The first such snapshots are the block columns of , as follows from (56),
Theorem 2.7. The ROM snapshots evolve as follows
where
is the ROM propagator matrix.
Proof. In order to obtain this, we see from (51) that
and applying (55), we get
which gives the desired result. □
Remark. From (57) and (58), we see that both and are data driven and so, the ROM can be computed directly from , for .
Reduction to scalar wave equation
In this section, we show how the full three-dimensional Maxwell system introduced in § 2 can be reduced to a one-dimensional scalar wave equation under appropriate assumptions on the geometry and polarization of the fields. This reduction provides the connection between the electromagnetic formulation and the one-dimensional layered media model considered later in this work.
In the reduced-order modeling framework introduced in § 2, the effect of the source term can be incorporated into an equivalent initial condition using a Duhamel-type representation. Consequently, in the following, we consider the homogeneous Maxwell system
with initial data determined by the excitation. This allows us to focus on the propagation properties of the wave field.
Let . In this section we specialize to a layered isotropic medium, with scalar coefficients depending only on one spatial variable:
This corresponds to a medium composed of planar layers orthogonal to the -axis. We further consider transverse electromagnetic fields that depend only on and are polarized orthogonally to the direction of variation:
One-dimensional layered media
Theorem 3.1. Under the assumptions (64)–(65), the source-free Maxwell equations (1) reduce to the one-dimensional first-order system
Consequently, satisfies the scalar wave equation in divergence form
Equivalently,
The magnetic flux component satisfies
In a layer where and are constant, (67) reduces to
Proof. See Appendix B.1.
Assuming time-harmonic solutions of the form
we obtain the one-dimensional Helmholtz equation
Inside a homogeneous layer, where and are constant, this becomes
or equivalently
This reduction leads to the class of one-dimensional layered media, where the wave speed is piecewise constant:
At interfaces, the Maxwell transmission conditions correspond to continuity of the tangential electric field and of the tangential magnetic field. In the scalar formulation, this gives continuity of and of across interfaces. In each layer, the scalar wave equation admits explicit solutions, with reflection and transmission generated by these interface conditions. A particularly important example is the Goupillaud medium, in which layer thicknesses are proportional to the local wave speed. This model provides a convenient framework for both analytical calculations and numerical simulations.
One-dimensional layered media
Layered media provide a classical setting in which multiple scattering and wave reflections can be described explicitly [19]. Related reduced-order approaches in layered media include [6]. In § 3, we showed that, under suitable assumptions on the polarization and spatial variability of the fields, Maxwell’s equations reduce to a one-dimensional scalar wave equation with variable electromagnetic coefficients. We now specialize this reduced Maxwell model to layered media and introduce an electromagnetic analogue of the Goupillaud construction. The goal is to obtain a benchmark in which propagation, reflection, transmission and multiple scattering are represented explicitly, while remaining directly connected to the electromagnetic formulation.
The classical Goupillaud model is formulated for acoustic waves in layered media. Here we adapt the same idea to the reduced Maxwell setting by working with electromagnetic impedance variables and imposing an equal travel-time condition across layers. In these variables, the right- and left-going wave amplitudes satisfy explicit interface scattering relations, and the equal travel-time assumption leads to an exact discrete-time propagation model. Thus, the model considered below should be understood as a Maxwell-adapted Goupillaud medium: it preserves the ROM-relevant mechanisms of two-way propagation, reflection and transmission at interfaces, multiple scattering, and data-driven reduced dynamics.
Layered medium and wave propagation
We consider the scalar wave equation obtained in § 3, written here in the form
where the wave speed is piecewise constant. This corresponds to a layered medium composed of layers in which the physical parameters are constant. To make the connection with the underlying physical model explicit, we recall that this equation can be derived from the one-dimensional acoustic system
where is the density and the bulk modulus. The wave speed and impedance are given by
In particular, we observe for the impedance that it can be written as follows:
Introducing the right- and left-going modes
we obtain the transport equations
In addition, since our medium is composed of layers, we have that
and so, we can write the solutions of the wave equation (75) in the following form:
where represents right-going waves and represents left-going waves, as described in (79), for . At each interface , continuity conditions give rise to reflection and transmission of the right- and left-going waves. In the impedance-normalized variables used here, the corresponding coefficients are
for .
Goupillaud medium
We now introduce a particular class of layered media, known as Goupillaud media, defined by the condition that each layer has the same travel time:
This assumption plays a crucial role as it implies that waves propagate from one interface to the next in a fixed time . As a consequence, scattering events occur only at discrete times , , which induces a natural discrete-time dynamical system. Between these discrete times, the right- and left-going waves propagate freely, while reflections and transmissions occur instantaneously at the interfaces. The wave field at time can therefore be written as
where and represent the amplitudes of right- and left-going waves at interface . These amplitudes satisfy the recurrence relation
where and are the reflection and transmission coefficients.
Connection to ROM formulation
We consider measurements of the wave field at discrete times . The sequence of states defines a collection of snapshots of the system. The observed data are given by
as also described by (41), which depends on the reflection and transmission coefficients through multiple scattering interactions. Each data point corresponds to a superposition of contributions associated with different scattering paths. The discrete evolution of the wave field takes the form
where is the propagator over one time step. This structure matches exactly the reduced-order modeling framework introduced in § 2. In particular, from Lemma 2.6, we see that the reduced-order model can be constructed directly from the observed data.
This explicit connection between wave propagation, discrete dynamics, and data-driven reduced-order modeling makes the layered medium an ideal test case for analyzing the stability and performance of ROM-based inversion methods.
Inverse problem and objective functions
We now turn to the inverse problem associated with the layered medium. The goal is to recover the impedance profile
from measurements of the reflected wave.
Forward map
As described above, the wave propagation through the layered medium gives rise to a sequence of observations
which depend on the impedance parameters through multiple scattering effects. This defines a nonlinear forward map
This map is highly nonlinear since the waves interact with multiple interfaces, which lead to products of reflection and transmission coefficients.
A natural approach consists of minimizing the difference between modeled and observed data:
This objective directly compares the data in the observation space. However, it is sensitive to amplitude errors and scaling effects, which may arise from modeling inaccuracies or measurement noise.
Remark. The quadratic data misfit can be interpreted in a Bayesian framework as a maximum a posteriori estimator under the assumption of additive white Gaussian noise on the data. In this setting, minimizing the discrepancy between observed and modeled data is statistically optimal. This observation explains why the classical misfit is expected to perform well when the perturbations are predominantly unstructured and additive.
Inverse problem
We formulate the inverse problem as an optimization problem. Given observed data corresponding to an unknown impedance profile , we seek to recover by solving
Following the reduced-order modeling framework, the inversion can be formulated in terms of the internal wavefield. Denoting by the true wavefield snapshots and by their ROM-based reconstruction, we define the objective
As in the general ROM framework, this objective can be expressed in terms of a reduced operator. In particular, one obtains
where is the reduced operator constructed from the observed data and is the one associated with the model.
This normalized operator comparison is motivated by ROM-based waveform inversion methods, where the inverse problem is reformulated at the level of reduced propagators or reduced operators rather than directly at the level of the measured traces [5, 7, 8, 9, 10]. Instead of directly minimizing , we consider the modified objective
where
Remark. The objective measures discrepancies between reconstructed and true internal wavefields through their reduced representation, while compares the associated reduced operators. The latter formulation removes global scaling effects and leads to improved stability properties, as will be demonstrated in the numerical experiments.
Numerical protocol and stability metrics
We now describe the numerical protocol used to compare the direct data misfit and the ROM-based objective. The goal of this section is to specify the benchmark medium, the perturbation models and the evaluation metrics before presenting the numerical results in § 7. Further implementation details, including the exact construction of the data vector, the ROM matrix and the perturbation grids, are given in Appendix § 11.
Benchmark configuration
All numerical experiments are performed on a five-layer Goupillaud medium. The reference impedance vector is
The exterior impedance is kept fixed, and the unknown parameter is
For each impedance vector , the Goupillaud recurrence generates a discrete data vector
where the odd entries vanish by the parity structure of the model. The corresponding ROM matrix is denoted by .
Compared inversion objectives
We compare two inversion strategies. The first one is the direct data misfit
The second one is the normalized ROM objective
where, since is rectangular, we use the Tikhonov-regularized right inverse
Here denotes the reference impedance used to build the ROM normalization. In the clean experiment, this is and in the layer-perturbation experiment, this is the perturbed medium generating the data. 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. The precise numerical construction is described in Appendix § 11. This form is a finite-dimensional analogue of the normalized ROM objectives used in [8, 9, 10].
Perturbation models
We consider several structured perturbation models. These perturbations are intended to represent different sources of uncertainty in either the medium or the measurements.
First, we consider the clean inverse problem, where the observed data are generated exactly from . This experiment isolates the effect of the optimization landscape, without any additional perturbation.
Second, we perturb the layer impedances themselves. In this case, the perturbed medium is the actual ground truth for the trial, and the goal is to recover this perturbed layered structure.
Third, we consider perturbations acting directly on the data. These include multiplicative data noise, propagation-speed uncertainty, receiver-position errors and systematic acquisition errors. Propagation-speed uncertainty and receiver-position errors are modeled as structured perturbations whose effect can grow with the arrival time. Systematic errors include global gain errors, additive bias and time-origin shifts.
The purpose of these experiments is not to show that the ROM objective is uniformly better for every possible perturbation. Rather, we aim to identify the regimes in which the normalization built into the ROM formulation gives a genuine advantage, and to distinguish them from regimes where the two objectives behave similarly.
Evaluation metrics
For each experiment, we compare the reconstructed impedances obtained from the two objectives. We denote these reconstructions by
The reconstruction error is measured with respect to a comparison impedance :
In the clean experiment, . In the layer-perturbation experiment, is the perturbed medium. For data perturbations, the comparison impedance is chosen according to the procedure described in Appendix § 11.
For each perturbation level, we report the median reconstruction error over the Monte Carlo trials. We also report the win fraction
estimated empirically over the trials. A win fraction larger than 0.5 indicates that the ROM-based reconstruction is more accurate than the direct data-misfit reconstruction in a majority of trials.
Numerical results
We now present the numerical results obtained with the protocol described in § 6. We focus on the perturbation regimes that are most informative for comparing the two objectives. The remaining perturbation models are summarized at the end of the section.
Clean inversion
We first consider the unperturbed inverse problem. Both methods are initialized from the same family of random initial guesses and are applied to the exact synthetic data generated by . Thus, the experiment tests the optimization behavior of the two objectives in the absence of measurement or modeling errors.
In the clean setting, the median reconstruction error is
for the direct data misfit and for the ROM-based objective.
Thus, the ROM-based objective yields a substantially smaller median reconstruction error than the direct data misfit. Since no perturbation is present in this experiment, this improvement cannot be attributed to noise filtering. It indicates instead that, for this benchmark, the ROM objective has a more favorable numerical landscape for the reconstruction of the impedance profile.
Layer perturbations
We next perturb the layer impedances themselves. In each Monte Carlo trial, the perturbed medium is treated as the ground truth, and both inversion methods are asked to recover it from the corresponding data. This experiment tests robustness with respect to changes in the layered structure, rather than perturbations added only at the measurement level.
Figure 2a. Median reconstruction error.
Figure 2b. Win fraction .
Figure 2. Comparison of the two inversion strategies under simultaneous perturbations of all layers.
Figure 2(a) shows that the ROM-based inversion has a smaller median error over the perturbation levels considered. Figure 2(b) supports the same conclusion: the win fraction remains above 0.5, showing that the ROM reconstruction is more accurate in a majority of trials. This suggests that the ROM normalization is beneficial when the uncertainty acts directly on the layered medium.
Time-origin error
We now consider a systematic time-origin error. This perturbation shifts the temporal alignment of the measured data and therefore affects the organization of the arrivals. It is a structured perturbation rather than an independent random perturbation of each data component.
Figure 3a. Median reconstruction error.
Figure 3b. Win fraction .
As shown in Figure 3, the ROM-based inversion gives smaller median errors over most of the tested range. The win fraction is also frequently well above 0.5. This is one of the clearest regimes in which the ROM objective improves stability. The result is consistent with the fact that the ROM formulation compares normalized reduced operators, and is therefore less sensitive to certain coherent distortions of the data than the direct signal misfit.
Figure 3. Comparison of the two inversion strategies under a time-origin error.
Receiver-position amplitude perturbation
We also consider receiver-position errors modeled as amplitude perturbations of the measured arrivals. This type of perturbation represents errors in coupling or geometric spreading that modify the strength of the recorded signal without explicitly shifting its arrival time.
In Figure 4, the two methods are closer than in the clean and time-origin experiments. The ROM-based inversion still shows a mild advantage for small to moderate perturbations, visible both in the median errors and in the win fraction, but the improvement is less pronounced. This provides an intermediate regime: the ROM objective remains competitive and slightly preferable, but the perturbation does not produce as clear a separation between the two methods.
Figure 4. Comparison of the two inversion strategies under receiver-position errors modeled as amplitude perturbations.
Perturbations with comparable behavior
We also tested additional structured perturbations, including multiplicative data noise, propagation-speed uncertainty, receiver-position errors modeled as time shifts, constant gain bias and additive bias. In these cases, the two inversion procedures behave comparably. The median errors remain close, and the win fraction typically fluctuates around 0.5, without revealing a persistent dominance of either method.
These experiments are important because they show that the ROM objective is not uniformly superior under all perturbations. Its advantage appears most clearly in the clean inversion, in layer perturbations and in coherent temporal distortions such as the time-origin error. For other perturbation mechanisms, the normalization in the ROM objective does not significantly change the reconstruction accuracy, but it also does not lead to a systematic degradation.
Summary of numerical observations
The numerical experiments support a balanced conclusion. In the clean deterministic inverse problem, the ROM-based objective produces a more favorable reconstruction than the direct data misfit. This advantage persists when the layered medium itself is perturbed, and it is especially visible under time-origin errors, where the perturbation has a coherent temporal structure.
At the same time, the ROM formulation does not dominate the direct data misfit in every perturbation regime. For several structured data perturbations, the two methods perform similarly. Thus, the main conclusion is not that ROM-based inversion is universally better, but that it can provide a clear advantage in regimes where the reduced operator normalization captures the relevant structure of the wave propagation more effectively than a direct comparison of the measured signals.
Conclusion
In this work, we investigated reduced-order modeling for wave-based inverse problems in layered media within the framework of Goupillaud media. This setting provides a natural discrete representation of wave propagation and allows for a direct comparison between classical data misfit and a ROM-based objective defined at the level of reduced operators. In this sense, the numerical study complements the ROM-based inversion program developed in [5, 7, 8, 9, 10], by testing the behavior of a normalized ROM objective in an exactly discretizable layered benchmark.
The numerical experiments highlight a nuanced but consistent picture. In the clean inversion setting, the ROM-based formulation yields a significantly improved reconstruction, indicating a more favorable optimization landscape. This advantage persists in certain structured perturbation regimes, such as perturbations of the layered medium and time-origin errors, where the ROM-based inversion achieves smaller median errors and higher win fractions. These results suggest that the normalization inherent in the ROM formulation is particularly effective in handling global or structurally coherent distortions of the data. At the same time, the ROM-based approach does not uniformly outperform the classical misfit. For several other perturbations, including additive noise and propagation-related distortions, the two methods exhibit comparable performance, with no systematic dominance of either approach. This indicates that the benefit of the ROM formulation depends on the nature of the perturbation and how it interacts with the underlying wave dynamics.
Overall, the results show that reduced-order modeling provides a meaningful and, in some regimes, advantageous alternative to standard inversion approaches. Its strength lies in capturing structural features of the problem that are not directly emphasized by the data misfit, while remaining competitive across a broad range of perturbations.
Proofs of Section 2
Proof of Lemma 2.1
From (1), we have
and
Thus, for
we obtain
Equivalently,
where
This proves the result.
Proof of Lemma 2.2
From (8), we have
Since and are independent of time,
Similarly,
Therefore,
Thus
with
This proves the normalized formulation.
Proof of Lemma 2.3
Let
belong to the domain of . The boundary condition (3) implies, in the normalized variables, that
Indeed, if , then
Thus (A.15) is the normalized form of the perfectly conducting condition
Then,
Equivalently,
Using the curl integration-by-parts formula
the boundary terms vanish because of the tangential boundary conditions above. Hence
Therefore,
By the definition of , this is
Thus is skew-symmetric. The skew-adjointness of the corresponding perfectly conducting Maxwell realization is a standard result; see, for instance, [23].
Proof of Theorem 2.4
Let
Then implies
On a simply connected domain, there exist scalar potentials and such that
Therefore
This gives
where
Using the orthogonal decomposition
which follows from the general identity
for skew-adjoint operators, see for instance [23], the solution decomposes as
This proves the decomposition.
Claim in spectral decomposition
We justify the construction of the family associated with the orthonormal eigenbasis , and prove the identities
Let be the propagating subspace and let
We work with the standard skew-adjoint Maxwell realization on . Hence is positive self-adjoint. Assuming the corresponding Maxwell compactness property, has compact resolvent on . Therefore there exists an orthonormal eigenbasis of and positive eigenvalues , with , such that
We do not assume that the eigenvalues are simple. In general, they may have multiplicity. In that case, the construction above is performed within each eigenspace and an orthonormal basis of eigenfunctions is chosen.
Now, for each , we define
This is well-defined since and . We will first show that is an orthonormal family. Using the skew-symmetry of , we compute
Thus, . Then, let . It holds that
Hence,
which gives the orthogonality.
Let us also show the orthogonality between and , for . Using skew-symmetry,
which implies
In general, the functions belong to the span of the family , so orthogonality between and does not hold for , with .
Finally, let us prove (A.28). By definition,
Furthermore,
This concludes the proof.
Proof of Lemma 2.6
Proof. For , it holds from (53) that
which, from (41), becomes
In addition, we have from (55), that the mass matrix is symmetric, which implies that for , we have
Similarly, for , we have from (54) and (44) that
which gives
Then, for , with , we have
which gives
This concludes the proof.
Proofs of Section 3
Proof of Theorem 3.1
Proof. We compute the curl operators explicitly under the assumptions (64)–(65). Since
we obtain
Similarly, since
and , we have
Therefore
Substituting into the source-free Maxwell equations
we obtain
This proves (66).
We now derive the second-order equation for . Differentiating the second equation with respect to time gives
Using
we obtain
Hence
We also derive the equation for . Differentiating the first equation with respect to time gives
From the second equation,
Therefore
Finally, if and are constant inside a layer, the equation for becomes
Thus the local wave speed is
This concludes the proof.
Details of the numerical implementation
This appendix describes the numerical procedure used in § 7. The purpose is to make the construction of the data, the ROM matrices and the perturbation experiments reproducible.
Reference medium and unknown parameters
All experiments are performed on a five-layer Goupillaud medium. The reference impedance vector is
The first impedance is kept fixed throughout the inversion, in order to remove the global scaling ambiguity. The unknown parameter vector is therefore
The admissible set used in the numerical optimization is
The time step is normalized to
We use nine snapshots, indexed by .
Forward data
For each impedance vector , the reflection and transmission coefficients at the interfaces are
The right- and left-going amplitudes are then propagated by the Goupillaud recurrence described in § 5. This gives the snapshot coefficients
The discrete data vector is computed from the snapshot inner products
By the parity structure of the Goupillaud model, the odd entries vanish:
Thus, the data used in the experiments are
ROM matrix and normalized objective
The ROM matrix is assembled from the nonzero Goupillaud amplitudes. In the five-layer experiment it has size and sparsity pattern
The entries are obtained from the Goupillaud amplitudes computed by the forward recurrence. Since is rectangular, we use the Tikhonov-regularized right inverse
The ROM objective used in the computations is
Here denotes the reference impedance for the corresponding experiment. In the clean and layer-perturbation experiments it is the true medium used to generate the data. In data-perturbation experiments, it is the effective impedance fitted from the perturbed data. The direct data misfit is
Optimization procedure
All minimizations are performed with the L-BFGS-B algorithm under the box constraints
For the clean inversion, we use 20 random initializations of the form
with clipping to the admissible interval.
For each perturbation level , the reported statistics are computed over 50 independent Monte Carlo trials. The random seed for the perturbation in trial is , and the random seed for the initialization is .
Perturbation models
We use the following perturbation models.
Layer perturbation. The physical medium itself is perturbed. We set
where are independent. The exterior impedance is kept fixed. Negative or zero values are clipped to . The observed data are
The reconstruction error is measured with respect to .
Perturbation grids
Multiplicative data noise. The physical medium is not changed. The observed data are
Since the perturbed data do not necessarily correspond exactly to an admissible layered medium, we first compute an effective reference impedance by minimizing the direct data misfit against . The ROM objective is then formed using , and the reconstruction errors are measured relative to .
Propagation-speed uncertainty. Propagation-speed uncertainty is modeled as a multiplicative perturbation of the nonzero even-time data entries, with larger perturbations at later times:
The odd entries are kept equal to zero. The reconstruction error is measured with respect to .
Receiver-position error: time-shift model. A receiver-position error is first modeled as a coherent time-shift distortion. A single Gaussian variable is drawn and we set
The odd entries are kept equal to zero.
Receiver-position error: amplitude model. We also model receiver-position errors as amplitude perturbations of the recorded arrivals:
with weights
Thus the first datum is kept fixed, while later arrivals are perturbed more strongly. The odd entries remain zero.
Constant gain bias. A calibration error is modeled by multiplying the full data vector by a single random gain:
Additive bias. A baseline error is modeled by adding a constant offset to all entries of the data vector:
where
Time-origin error. A time-origin error is modeled by shifting the data vector by an integer number of time steps. We draw
and shift by entries, using zero padding at the boundaries. Here denotes the projection of onto the interval .
Perturbation grids
The perturbation amplitudes used in the numerical experiments are as follows:
Reported statistics
For each perturbation level, we compute the reconstruction errors
Here is the comparison impedance: it is the perturbed medium for the layer-perturbation experiment, the fitted effective reference medium for the data-noise experiment, and the original reference medium for the remaining perturbation models.
The figures report the median of these errors over the Monte Carlo trials, together with the win fraction
estimated empirically over the 50 trials.
Figure 1. Propagation in a Goupillaud medium. The equal travel-time condition ensures that waves move from one interface to the next in a single time step , leading to a discrete dynamical system for the amplitudes and .
References
- [1]M. I. Belishev. Recent progress in the boundary control method. Inverse problems, 23(5):R1–R67, 2007.DOI
- [2]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.DOI
- [3]P. Blondel and B. J. Murton. Handbook of seafloor sonar imagery, volume 7. Wiley Chichester, 1997.
- [4]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.DOI
- [5]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.arxiv.org/abs/1910.13014
- [6]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.arxiv.org/abs/2012.00861
- [7]L. Borcea, J. Garnier, A. V. Mamonov, and J. Zimmerling. Reduced order model approach for imaging with waves. Inverse Problems, 38(2):025004, 2022.DOI
- [8]L. Borcea, J. Garnier, A. V. Mamonov, and J. Zimmerling. Waveform inversion via reduced order modeling. Geophysics, 88(2):R175–R191, 2023.arxiv.org/abs/2202.01824
- [9]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.arxiv.org/abs/2208.11051
- [10]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.arxiv.org/abs/2302.05988
- [11]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.arxiv.org/abs/2403.03844
- [12]S. L. Brunton and J. N. Kutz. Data-driven science and engineering: Machine learning, dynamical systems, and control. Cambridge University Press, 2022.
- [13]M. Cheney and B. Borden. Fundamentals of radar imaging. SIAM, 2009.DOI
- [14]J. C. Curlander and R. N. McDonough. Synthetic aperture radar, volume 11. Wiley, New York, 1991.
- [15]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.DOI
- [16]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.DOI
- [17]B. Engquist and B. D. Froese. Application of the wasserstein metric to seismic signals. arXiv preprint arXiv:1311.4581, 2013.arxiv.org/abs/1311.4581
- [18]B. Engquist and Y. Yang. Optimal transport based seismic inversion: Beyond cycle skipping. Communications on Pure and Applied Mathematics, 75(10):2201–2244, 2022.arxiv.org/abs/2002.00031
- [19]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.
- [20]M. Gilman, E. Smith, and S. Tsynkov. Transionospheric synthetic aperture imaging. Springer, 2017.DOI
- [21]J. S. Hesthaven, C. Pagliantini, and G. Rozza. Reduced basis methods for time-dependent problems. Acta Numerica, 31:265–345, 2022.DOI
- [22]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.DOI
- [23]P. Monk. Finite element methods for Maxwell’s equations. Oxford university press, 2003.
- [24]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.
- [25]W. W. Symes. Migration velocity analysis and waveform inversion. Geophysical prospecting, 56(6):765–790, 2008.DOI
- [26]W. W. Symes. Error bounds for extended source inversion applied to an acoustic transmission inverse problem. Inverse Problems, 38(11):115002, 2022.DOI
- [27]T. L. Szabo. Diagnostic ultrasound imaging: inside out. Academic press, 2013.
- [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.
- [29]T. Van Leeuwen and F. J. Herrmann. Mitigating local minima in full-waveform inversion by expanding the search space. Geophysical Journal International, 195(1):661–667, 2013.DOI
- [30]J. Virieux and S. Operto. An overview of full-waveform inversion in exploration geophysics. Geophysics, 2010.DOI
- [31]M. Warner and L. Guasch. Adaptive waveform inversion: Theory. Geophysics, 81(6):R429–R445, 2016.
- [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.arxiv.org/abs/1612.05075
- [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.





