Introduction
Background and motivation
Neural PDE solvers have emerged as promising mesh-free algorithms for scientific simulations (e.g., [1 , 3 , 29 , 24 ] ). However, learning highly oscillatory, high-frequency functions remains a profound challenge for standard neural architectures. This difficulty is mainly driven by spectral bias, a phenomenon where neural networks inherently prioritize learning low-frequency components and struggle to fit high-frequency details [27 ] . Furthermore, minimizing high-frequency PDE residuals often requires computing second-order spatial derivatives, which can result in highly non-convex, poorly conditioned loss landscapes that severely impede optimizer convergence.
A particularly challenging setting for testing these neural solvers is time-harmonic wave propagation, governed by the Helmholtz equation. As the frequency and medium contrast increase, the exact solutions become highly oscillatory and multiscale. While classical discretizations require prohibitively large, indefinite linear systems to resolve these fine wavelengths, standard deep neural networks equally fail due to spectral bias and the difficulty of resolving sharp localized features in highly heterogeneous velocity models, such as those commonly found in geophysics and medical imaging [6 , 14 ] .
The Lippmann-Schwinger integral formulation eliminates second-order derivatives that appear in the strong form of the Helmholtz equation. As a result, the training loss does not require differentiation of the network with respect to its input, and the associated integral operator is better conditioned than the differential Helmholtz operator. Embedding the outgoing radiation condition into the background Green’s function turns the scattering problem into a Fredholm integral equation of the second kind. The differential Helmholtz operator is unbounded and ill-conditioned, whereas the resulting integral operator is a bounded identity-plus-compact operator. This structural preconditioning gives a smoother loss landscape and helps stabilize training. The action of the Green operator on the trial space is evaluated on a uniform grid via FFT, exploiting the Toeplitz structure of the kernel. This evaluation is far faster than direct evaluation of the Helmholtz operator, which requires differentiating the network at every collocation point.
Abedi et al. [2 ] proposed the Lippmann-Schwinger integral equation as a training loss for neural solvers of scattering problems, combining it with an auxiliary strong-form residual as a regularizer. We adopt the same Green-integral loss and instead improve the architecture and optimization strategy: for the same loss, our design converges faster and reaches lower error in strongly heterogeneous media. The NN architecture of the trial space is designed so that the wave structure is represented directly by the network, and it is paired with an optimization strategy built specifically to make training fast and stable.
In this study, we introduce NAP-Wave, a data-free neural solver built on three design principles.
The wave structure belongs in an adaptive plane-wave trial space. A generic network must discover, from a smooth coordinate input, a field oscillating on the scale of a wavelength. Trefftz and plane-wave methods avoid this by building basis functions that satisfy the governing equation exactly [17 , 8 ] . We adopt the same principle in a neural setting. Each basis function carries a plane-wave that solves the homogeneous Helmholtz equation exactly, and the network represents the varying amplitude that modulates it.
The linear coefficients are solved exactly, while the basis adapts. The Lippmann-Schwinger operator is linear, so the coefficients combining the basis functions admit a closed-form solution for any fixed basis. We solve for them by least squares at every step. Adam updates only the nonlinear parameters that shape the basis functions themselves. This is the classical variable projection principle [13 ] , applied here with every neuron of every hidden layer treated as an individual basis function.
The trial space grows with training, rather than staying fixed. A network initialized at full depth must resolve a high-wavenumber field from a random basis. We instead train a sequence of progressively deeper sub-networks. Each new layer adds a block of basis functions and preserves the first- and second-order moments of the optimizer from the converged parameters of the previous stage, so that the trial space is enriched only once the coarser one has been exploited. Networks are known to learn low-frequency components first [27 ] ; growing depth progressively lets each new layer target the higher-frequency content the shallower network could not yet represent. In practice, we find that this progressive schedule acts as a warm start: once a shallow sub-network has converged, its parameters already place the next, deeper sub-network in a good basis, and adding further training iterations to the shallow network alone yields no additional gain.
Related work
Integral-equation solvers for scattering. The Lippmann-Schwinger equation is a classical tool in scattering theory. It appears across many fields, including quantum mechanics [28 ] , electromagnetics [9 ] , acoustics [22 ] , and seismic wave propagation [18 ] , and we adopt it here for frequency-domain seismic scattering. Its mathematical foundations are well established, including compactness of the volume potential, Fredholm invertibility, and radiation conditions [20 ] . Traditional numerical solvers discretize the integral operator and assemble the resulting dense system. This system is prohibitively expensive to invert directly. Practitioners therefore use iterative solvers instead, in which each matrix-vector product is evaluated efficiently via the FFT. This approach exploits the Toeplitz structure of the Green’s kernel on uniform grids. The Born series [23 ] is a classical instance of this iterative approach. It converges for weak scattering contrasts, but strong contrasts require effective preconditioners to obtain a convergent iteration [12 , 25 ] . The FFT-accelerated evaluation of the Green operator that we adopt in this study relies on the same classical construction where we use it to assemble the action of the operator on the trial space spanned by the plane-wave neural basis functions. This yields a differentiable solver for the scattered field that can be evaluated at arbitrary spatial locations.
Trefftz methods, plane-wave basis, and Herglotz wave functions. Trefftz and quasi-Trefftz methods use basis functions that satisfy the governing equation exactly, or up to a high-order local error, rather than generic polynomials. Generalized plane-wave methods extend this idea to smoothly varying media [17 ] . Our plane-wave neural basis is related in spirit: the plane-wave carriers solve the homogeneous Helmholtz equation exactly, while the neural amplitude captures the spatially varying correction. Our approach differs from plane-wave activation networks [10 ] , in which plane-waves are used as the activation functions of a single-hidden-layer network, whereas we combine plane-wave carriers with a deep neural amplitude and target general heterogeneous, spatially varying wavenumber media rather than a homogeneous background. It also differs from the plane-wave discontinuous Galerkin domain-decomposition method of [31 ] , which is formulated for piecewise-constant wavenumbers, whereas our method addresses smoothly and strongly varying velocity models within a single global trial space.
Our approach is based on Herglotz wave functions, which superpose plane-waves at a fixed wavenumber and integrated over all directions on the unit sphere. Classical results show such superpositions are dense in the solution space of the homogeneous Helmholtz equation, with approximation properties in Sobolev norms studied rigorously [8 ] . Each basis function carries a plane wave at a fixed direction, sampled from a finite set of directions on the unit sphere, together with a trainable wavenumber and a trained amplitude that modulates it. The homogeneous solution space is thus approximated over a finite, discrete set of directions by varying wavenumber and amplitude, rather than over a continuum of directions at a single fixed wavenumber. Classical density results for plane-wave bases, such as [8 ] , are stated for a continuum of directions at fixed wavenumber, so they do not apply to our construction. We prove a separate density result for our finite-direction, trainable-wavenumber basis in Theorem 3.3 .
Neural solvers for Helmholtz and seismic modeling. PINNs and related neural approaches have been applied to frequency-domain wave modeling in heterogeneous media, including Gabor, Fourier-feature, and Green-integral-constrained formulations [1 , 2 , 3 , 11 , 29 ] . These methods train the network end-to-end on the Helmholtz equation or its integral formulation, relying solely on the network’s own representational capacity, without introducing an inductive bias such as the plane-wave basis proposed here. Oscillatory structures are generally incorporated into the network architecture through input encoding techniques, such as positional encoding [1 ] , or by warm-starting the training from models previously trained at lower frequencies [16 ] . In contrast, our approach embeds this structure directly in the trial space through the plane-wave carriers, which operate effectively on the physical coordinates without any feature embedding.
Least-squares for the linear coefficients. The idea of adaptive basis functions with a linear output layer solved by least squares appears independently in several lines of work. Extreme learning machines fix a single hidden layer at random and solve only for the output weights by least squares [15 ] . The variable projection (VarPro) principle for nonlinear least-squares problems formalizes this split more generally: the linear parameters are solved using least squares, and the first-order optimization is performed over the nonlinear parameters alone [13 ] . This principle has been adopted in PINNs (see e.g., [3 , 5 ] ), where the network parameters are optimized by Adam while the weights of the last layer are solved exactly by least squares. Our approach differs in that all neural basis functions, taken from every hidden layer rather than only the last, are connected to the output, treating each neuron as an individual basis function. We then use the exact least-squares solve for the coefficients coupled with gradient-based training of the nonlinear network parameters. Unlike a single hidden layer, a deep network builds hierarchical features across layers, so treating every hidden neuron as a basis function lets us capture local, fine-scale structure in the field that a shallow basis cannot represent.
Contributions and paper organization
We summarize our contributions as follows.
A plane-wave neural basis for time-harmonic wave propagation, in which every neuron of every hidden layer contributes a basis function modulated by an oscillatory plane-wave carrier with trainable wavenumber and amplitude.
An extension of the variable projection scheme to all layers, in which the linear coefficients are obtained at each step from a regularized least-squares solve, while gradient-based training updates only the nonlinear carrier and amplitude parameters.
An adaptive-depth training strategy that grows the architecture layer-by-layer instead of fixing it at full depth from the start. Combined with the exact coefficient solve, it removes the loss spikes that occur when a new layer is added and keeps the error decreasing smoothly, an effect that neither ingredient produces alone.
A GPU implementation in JAX combining an FFT evaluation of the Green operator with a complex-valued direct solve in double precision. It reaches relative L 2 L^2 L 2 errors of order 1% against exact 2D and 3D solutions, and 5%, 9%, and 10% against the reference solution on Marmousi (10 Hz), Overthrust (10 Hz), and Otway (20 Hz), with the 10-Hz Marmousi case solved in roughly three minutes on a single GPU.
The remainder of the paper is organized as follows. In Section 2, we recall the Lippmann-Schwinger equation and its abstract properties. In Section 3, we present the proposed methodology. In Section 4, we describe the algorithmic and computational aspects of the proposed method. In Section 5, we describe the benchmark problems and the reference solutions. In Section 6, we report the numerical experiments. In Section 7, we summarize the main conclusions of this study. Further details are given in Appendices B-F.
Mathematical model: The Lippmann-Schwinger equation
In this section, we recall the Lippmann-Schwinger equation, the mathematical model adopted in this study to describe time-harmonic acoustic wave propagation in heterogeneous media.
Setting and medium contrast
Let Ω ⊂ R d \Omega\subset\mathbb{R}^{d} Ω ⊂ R d be a bounded connected open set with d = 2 , 3 d=2,3 d = 2 , 3 . We consider time-harmonic wave propagation at angular frequency ω = 2 π f 0 > 0 \omega=2\pi f_{0}>0 ω = 2 π f 0 > 0 , where f 0 > 0 f_{0}>0 f 0 > 0 denotes the temporal frequency. The wave speed is denoted by v : Ω → R v:\Omega\to\mathbb{R} v : Ω → R and we assume that there exist constants 0 < v min ≤ v max < ∞ 0<v_{\min}\le v_{\max}<\infty 0 < v m i n ≤ v m a x < ∞ such that v min ≤ v ( x ) ≤ v max v_{\min}\le v(x)\le v_{\max} v m i n ≤ v ( x ) ≤ v m a x for all x ∈ Ω x\in\Omega x ∈ Ω . The local wavenumber is given by k ( x ) = ω / v ( x ) k(x)=\omega/v(x) k ( x ) = ω / v ( x ) , and we denote the squared slowness by m ( x ) = v ( x ) − 2 m(x)=v(x)^{-2} m ( x ) = v ( x ) − 2 . The corresponding background squared slowness and wavenumber are m 0 = v 0 − 2 m_{0}=v_{0}^{-2} m 0 = v 0 − 2 and k 0 = ω / v 0 k_{0}=\omega/v_{0} k 0 = ω / v 0 , respectively, where v 0 > 0 v_{0}>0 v 0 > 0 is the constant background wave speed. The medium contrast, which quantifies the deviation of the true medium from the background, is defined by
δ m ( x ) = m ( x ) − m 0 , (1) \delta m(x)=m(x)-m_{0}, \tag*{(1)} δ m ( x ) = m ( x ) − m 0 , (1)
and satisfies δ m ∈ L ∞ ( Ω ) \delta m\in L^{\infty}(\Omega) δ m ∈ L ∞ ( Ω ) with supp ( δ m ) ⊂ Ω \operatorname{supp}(\delta m)\subset\Omega supp ( δ m ) ⊂ Ω , which ensures that the scatterer is compactly supported within Ω \Omega Ω .
Incident and scattered fields
The total acoustic field u u u is decomposed as
u = u 0 + u s , (2) u=u_{0}+u_{s}, \tag*{(2)} u = u 0 + u s , (2)
where u 0 u_{0} u 0 denotes the incident field, propagating in a prescribed homogeneous background medium with constant wave speed v 0 > 0 v_{0}>0 v 0 > 0 , and u s u_{s} u s is the scattered field arising from the medium heterogeneities.
Let G 0 G_{0} G 0 denote the outgoing free-space Green’s function of the background Helmholtz operator L 0 : = − Δ − ω 2 m 0 L_{0}:=-\Delta-\omega^{2}m_{0} L 0 := − Δ − ω 2 m 0 , that is, the unique solution of L 0 G 0 ( ⋅ , y ) = δ y L_{0}G_{0}(\cdot,y)=\delta_{y} L 0 G 0 ( ⋅ , y ) = δ y satisfying the Sommerfeld radiation condition at infinity. Since v 0 v_{0} v 0 is constant, G 0 G_{0} G 0 is translation invariant and admits the explicit kernel representation
G 0 ( x , y ) = g 0 ( x − y ) , x ≠ y . (3) G_{0}(x,y)=g_{0}(x-y),\qquad x\ne y. \tag*{(3)} G 0 ( x , y ) = g 0 ( x − y ) , x = y . (3)
In this study, we consider the free-space Green’s function
g 0 ( x ) = { i 4 H 0 ( 2 ) ( k 0 ∥ x ∥ ) , d = 2 , exp ( − i k 0 ∥ x ∥ ) 4 π ∥ x ∥ , d = 3 , (4) g_{0}(x)=
\begin{cases}
\frac{\mathrm{i}}{4}H_{0}^{(2)}(k_{0}\lVert x\rVert), & d=2,\\
\frac{\exp(-\mathrm{i}k_{0}\lVert x\rVert)}{4\pi\lVert x\rVert}, & d=3,
\end{cases} \tag*{(4)} g 0 ( x ) = { 4 i H 0 ( 2 ) ( k 0 ∥ x ∥) , 4 π ∥ x ∥ e x p ( − i k 0 ∥ x ∥) , d = 2 , d = 3 , (4)
where H 0 ( 2 ) H_{0}^{(2)} H 0 ( 2 ) is the Hankel function of the second kind and order zero. For a point source located at x s ∈ Ω x_{s}\in\Omega x s ∈ Ω , the incident field is accordingly the outgoing solution of the homogeneous background problem,
u 0 ( x ) = g 0 ( x − x s ) . (5) u_{0}(x)=g_{0}(x-x_{s}). \tag*{(5)} u 0 ( x ) = g 0 ( x − x s ) . (5)
The Green’s integral operator and the governing equation
The Green’s integral operator G : L 2 ( Ω ) → L 2 ( Ω ) \mathcal{G}:L^{2}(\Omega)\to L^{2}(\Omega) G : L 2 ( Ω ) → L 2 ( Ω ) is defined, for φ ∈ L 2 ( Ω ) \varphi\in L^{2}(\Omega) φ ∈ L 2 ( Ω ) and x ∈ Ω x\in\Omega x ∈ Ω , by
( G φ ) ( x ) : = ω 2 ∫ Ω g 0 ( x − y ) δ m ( y ) φ ( y ) d y . (6) (\mathcal{G}\varphi)(x):=\omega^{2}\int_{\Omega}g_{0}(x-y)\,\delta m(y)\,\varphi(y)\,\mathrm{d}y. \tag*{(6)} ( G φ ) ( x ) := ω 2 ∫ Ω g 0 ( x − y ) δ m ( y ) φ ( y ) d y . (6)
The scattered field u s u_{s} u s we seek must satisfy the Lippmann-Schwinger equation: given an incident field u 0 ∈ L 2 ( Ω ) u_{0}\in L^{2}(\Omega) u 0 ∈ L 2 ( Ω ) , find u s ∈ L 2 ( Ω ) u_{s}\in L^{2}(\Omega) u s ∈ L 2 ( Ω ) such that
( I − G ) u s = G u 0 , (7) (I-\mathcal{G})u_{s}=\mathcal{G}u_{0}, \tag*{(7)} ( I − G ) u s = G u 0 , (7)
where I I I denotes the identity operator on L 2 ( Ω ) L^{2}(\Omega) L 2 ( Ω ) .
Abstract formulation of the Lippmann-Schwinger equation
This section outlines the abstract formulation of the Lippmann-Schwinger equation. These theoretical foundations are well-established within scattering theory, and a comprehensive mathematical treatment is available in [20 ] .
The Green’s operator defined in (6) is a linear map G : L 2 ( Ω ) → L 2 ( Ω ) \mathcal{G}:L^2(\Omega)\to L^2(\Omega) G : L 2 ( Ω ) → L 2 ( Ω ) that possesses several mathematical properties under suitable, physically realistic assumptions. First, Lemma C.1 (Appendix C) establishes that G \mathcal{G} G is a Hilbert-Schmidt operator, which in particular implies that it is compact and bounded. Moreover, G \mathcal{G} G maps L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) into H l o c 2 ( R d ) H^2_{\mathrm{loc}}(\mathbb{R}^d) H loc 2 ( R d ) , where the local Sobolev space is defined by
H l o c 2 ( R d ) = { u : R d → C ∣ u ∣ K ∈ H 2 ( K ) for every bounded open set K ⊂ R d } . (8) H^2_{\mathrm{loc}}(\mathbb{R}^d)=\{u:\mathbb{R}^d\to\mathbb{C}\mid u|_K\in H^2(K)\text{ for every bounded open set }K\subset\mathbb{R}^d\}. \tag*{(8)} H loc 2 ( R d ) = { u : R d → C ∣ u ∣ K ∈ H 2 ( K ) for every bounded open set K ⊂ R d } . (8)
Consequently, G \mathcal{G} G acts as a smoothing operator that elevates the spatial regularity by two orders of differentiation.
The Lippmann-Schwinger equation (7) is formulated as follows. Given a source term f = G u 0 ∈ L 2 ( Ω ) f=\mathcal{G}u_0\in L^2(\Omega) f = G u 0 ∈ L 2 ( Ω ) , we seek the unique scattered field u s ∈ L 2 ( Ω ) u_s\in L^2(\Omega) u s ∈ L 2 ( Ω ) satisfying
B u s = f , (9) \mathcal{B}u_s=f, \tag*{(9)} B u s = f , (9)
where B = I − G \mathcal{B}=\mathcal{I}-\mathcal{G} B = I − G . The following proposition shows that B \mathcal{B} B is boundedly invertible on L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) , its proof, which follows the classical Fredholm-alternative and unique-continuation argument of [20 ] , is given in Appendix C.§ 1 .
Proposition 2.1 (Well-posedness of the scattering operator). Let Ω ⊂ R d \Omega\subset\mathbb{R}^d Ω ⊂ R d be a bounded connected open set with d = 2 , 3 d=2,3 d = 2 , 3 , and let δ m ∈ L ∞ ( Ω ) \delta_m\in L^\infty(\Omega) δ m ∈ L ∞ ( Ω ) . Let G 0 ( x , y ) G_0(x,y) G 0 ( x , y ) be the outgoing free-space Green’s function of the background operator L 0 = − Δ − ω 2 m 0 L_0=-\Delta-\omega^2m_0 L 0 = − Δ − ω 2 m 0 , where k 0 = ω / v 0 k_0=\omega/v_0 k 0 = ω / v 0 . Then the operator B = I − G : L 2 ( Ω ) → L 2 ( Ω ) \mathcal{B}=\mathcal{I}-\mathcal{G}:L^2(\Omega)\to L^2(\Omega) B = I − G : L 2 ( Ω ) → L 2 ( Ω ) is boundedly invertible.
Approximation space and training methodology
In this study, our main objective is the accurate approximation of the scattered wavefield u s u_s u s in heterogeneous media for realistic benchmark scenarios. We develop a data-free numerical approach that employs a NN ansatz to approximate the unknown field u s u_s u s , trained in an unsupervised manner directly by enforcing the Lippmann-Schwinger equation (7) .
Residual minimization
Since closed-form exact solutions of heterogeneous media scattering problems are generally unavailable, we seek an approximate solution via a residual minimization method. Recall that the scattered field u s ∈ L 2 ( Ω ) u_s\in L^2(\Omega) u s ∈ L 2 ( Ω ) satisfies the Lippmann-Schwinger equation B u s = f \mathcal{B}u_s=f B u s = f , where B = I − G \mathcal{B}=\mathcal{I}-\mathcal{G} B = I − G and G \mathcal{G} G is the integral operator defined in (6) . Given an incident source f ∈ L 2 ( Ω ) f\in L^2(\Omega) f ∈ L 2 ( Ω ) , we introduce the least-squares functional J : L 2 ( Ω ) → R ≥ 0 J:L^2(\Omega)\to\mathbb{R}_{\geq0} J : L 2 ( Ω ) → R ≥ 0 defined by
J ( v ) = 1 2 ∥ B v − f ∥ L 2 ( Ω ) 2 . (10) J(v)=\frac{1}{2}\lVert\mathcal{B}v-f\rVert_{L^2(\Omega)}^2. \tag*{(10)} J ( v ) = 2 1 ∥ B v − f ∥ L 2 ( Ω ) 2 . (10)
As shown in Theorem 3.1, J J J is a strictly convex quadratic functional admitting a unique global minimizer, which coincides with the unique solution of the governing integral equation; the proof is given in Appendix C.§ 2 .
Theorem 3.1 (Existence and uniqueness of the global minimizer). The least-squares functional J : L 2 ( Ω ) → R ≥ 0 J:L^2(\Omega)\to\mathbb{R}_{\geq0} J : L 2 ( Ω ) → R ≥ 0 given by (10) is strictly convex and quadratic, and admits a unique global minimizer u s ∈ L 2 ( Ω ) u_s\in L^2(\Omega) u s ∈ L 2 ( Ω ) , which is the unique solution of B u s = f \mathcal{B}u_s=f B u s = f .
A central question is whether minimizing this residual-based loss also controls the true approximation error, e.g., in the L 2 L^2 L 2 sense. We show that this is indeed the case and therefore, minimizing J J J is equivalent to minimizing the true error, as stated precisely in the following theorem; see Appendix C.§ 3 for the proof.
Theorem 3.2 (Error estimate via residual control). Let f ∈ L 2 ( Ω ) f\in L^2(\Omega) f ∈ L 2 ( Ω ) , let u ∈ L 2 ( Ω ) u\in L^2(\Omega) u ∈ L 2 ( Ω ) be the exact solution of B u = f \mathcal{B}u=f B u = f , and let v ∈ L 2 ( Ω ) v\in L^2(\Omega) v ∈ L 2 ( Ω ) be an arbitrary approximation. Define the approximation error e ( v ) = u − v e(v)=u-v e ( v ) = u − v . Then, there exist constants c 1 , c 2 > 0 c_1,c_2>0 c 1 , c 2 > 0 such that
c 1 J ( v ) ≤ ∥ e ( v ) ∥ L 2 ( Ω ) ∥ u ∥ L 2 ( Ω ) ≤ c 2 J ( v ) . (11) c_1\sqrt{J(v)}\leq\frac{\lVert e(v)\rVert_{L^2(\Omega)}}{\lVert u\rVert_{L^2(\Omega)}}\leq c_2\sqrt{J(v)}. \tag*{(11)} c 1 J ( v ) ≤ ∥ u ∥ L 2 ( Ω ) ∥ e ( v ) ∥ L 2 ( Ω ) ≤ c 2 J ( v ) . (11)
The ratio of the equivalence constants in (11) is given by the condition number κ ( B ) = c 2 / c 1 = ∥ B ∥ ∥ B − 1 ∥ \kappa(\mathcal{B})=c_2/c_1=\lVert\mathcal{B}\rVert\lVert\mathcal{B}^{-1}\rVert κ ( B ) = c 2 / c 1 = ∥ B ∥ ∥ B − 1 ∥ . In the high-frequency or strong-contrast regime, ∥ G ∥ \lVert\mathcal{G}\rVert ∥ G ∥ grows and B = I − G \mathcal{B}=\mathcal{I}-\mathcal{G} B = I − G becomes ill-conditioned, so a small residual no longer guarantees a small error.
We now restrict the minimization of J J J to a finite-dimensional trial space constructed from NNs and seek an approximation of the scattered field u s u_s u s within it.
Approximation space
We seek to approximate the unknown scattered field u s ∈ L 2 ( Ω ) u_s \in L^2(\Omega) u s ∈ L 2 ( Ω ) in a space spanned by neural basis functions. For a given feedforward NN, each neuron in each hidden layer defines a real-valued function of x ∈ Ω x \in\Omega x ∈ Ω , and consecutive neuron pairs yield complex-valued basis functions. A schematic overview is given in Figure 1 and the full construction is given in Appendix § 11 .
Figure 1. Construction of the neural basis functions from a feedforward network architecture, illustrated for a network with three hidden layers. (a) Consecutive neurons within each layer are paired to form complex-valued basis functions; each hidden layer therefore has an even number of neurons. Basis functions from every layer are connected directly to the output, so the two constructions shown, standard (using the complex pairs directly) and plane-wave (multiplying each pair by an oscillatory carrier exp ( i k j d j ⋅ x ) \exp(i k_j \mathbf{d}_j \cdot\mathbf{x}) exp ( i k j d j ⋅ x ) ). (b) Detailed construction of a single plane-wave neuron: two real sub-neurons produce the real and imaginary parts a j a_j a j , b j b_j b j , which combine into the complex feature φ j = a j + i b j \varphi_j = a_j + i b_j φ j = a j + i b j ; this is then modulated by the carrier exp ( i k j d j ⋅ x ) \exp(i k_j \mathbf{d}_j \cdot\mathbf{x}) exp ( i k j d j ⋅ x ) to yield the plane-wave neuron ψ j \psi_j ψ j .
Let θ ∈ R p \theta\in\mathbb{R}^p θ ∈ R p denote the vector of network parameters, with p p p the total number of parameters. For a fixed θ \theta θ , let φ j ( ⋅ ; θ ) : Ω → C \varphi_j(\cdot\,;\theta):\Omega\to\mathbb{C} φ j ( ⋅ ; θ ) : Ω → C for j = 1 , … , N j=1,\ldots,N j = 1 , … , N be complex-valued functions generated by a NN. We consider the approximation space
V N ( θ ) = span { φ 1 ( ⋅ ; θ ) , … , φ N ( ⋅ ; θ ) } , (12) V_N(\theta)=\operatorname{span}\{\varphi_1(\cdot\,;\theta),\ldots,\varphi_N(\cdot\,;\theta)\}, \tag*{(12)} V N ( θ ) = span { φ 1 ( ⋅ ; θ ) , … , φ N ( ⋅ ; θ )} , (12)
and we seek u N ∈ V N ( θ ) u_N\in V_N(\theta) u N ∈ V N ( θ ) of the form
u N ( x ; θ , c ) = ∑ j = 1 N c j φ j ( x ; θ ) , x ∈ Ω , (13) u_N(x;\theta,\mathbf{c})=\sum_{j=1}^{N}c_j\varphi_j(x;\theta),\qquad x\in\Omega, \tag*{(13)} u N ( x ; θ , c ) = j = 1 ∑ N c j φ j ( x ; θ ) , x ∈ Ω , (13)
where c = ( c j ) j = 1 N ∈ C N \mathbf{c}=(c_j)_{j=1}^{N}\in\mathbb{C}^N c = ( c j ) j = 1 N ∈ C N .
The nonlinear network parameters θ \theta θ and the linear coefficients c \mathbf{c} c are determined by minimizing J J J defined in (10) , as we detail in § 3.3 . Note that only θ \theta θ requires initialization, and the coefficients c \mathbf{c} c need not be initialized, since for any given θ \theta θ they are determined in closed form. Moreover, since θ \theta θ is updated during training, the space V N ( θ ) V_N(\theta) V N ( θ ) itself adapts, and the method constitutes an adaptive basis strategy. For a fixed number of basis functions N N N , the search space over all admissible parameters is denoted as
V N = ⋃ θ ∈ R p V N ( θ ) . (14) \mathcal{V}_N=\bigcup_{\theta\in\mathbb{R}^p}V_N(\theta). \tag*{(14)} V N = θ ∈ R p ⋃ V N ( θ ) . (14)
We note that the functions φ j \varphi_j φ j are not required to be linearly independent in general. In this study, we consider two types of neural basis functions.
Standard neural basis functions. The basis functions are formed directly from hidden neuron pairs, without any additional structure imposed on the network; see (B.4). This construction relies entirely on training to discover any oscillatory behavior present in the solution.
Plane-wave neural basis functions. Each basis function is formed as the product of a hidden neuron pair and an oscillatory carrier exp ( i k j d j ⋅ x ) \exp(\mathrm{i}k_j\mathbf{d}_j\cdot x) exp ( i k j d j ⋅ x ) , where the propagation direction d j \mathbf{d}_j d j is sampled once at initialization and held fixed, while the wavenumber k j k_j k j is a trainable network parameter determined during training, as described in (B.5). By construction, each carrier is an exact solution of the homogeneous Helmholtz equation, so that the dominant oscillatory structure of u s u_s u s is built directly into the basis rather than left for the network to learn from scratch. The remaining task for the network is then to represent the spatially varying complex amplitude that modulates each carrier, which is typically a much smoother function than u s u_s u s itself and could therefore be easier to approximate.
Note that plane-wave basis functions form a generalization of the standard basis; setting the wave-number samples k j = 0 k_j=0 k j = 0 for all j j j , and the standard basis is recovered as a special case.
In the following, we show that the considered approximation space is dense in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) , a natural result stemming from the density of NNs reformulated as an adaptive basis strategy. We denote by V \mathcal{V} V the class generated by the standard basis and by V w \mathcal{V}_w V w the class generated by the plane-wave basis, each taken over all possible numbers of basis functions N N N , that is V = ⋃ N V N \mathcal{V}=\bigcup_N\mathcal{V}_N V = ⋃ N V N . Since the solution of the Lippmann-Schwinger equation belongs to L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) , we aim to show that both V \mathcal{V} V and V w \mathcal{V}_w V w are dense in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) .
Theorem 3.3 (Universal approximation). Let Ω ⊂ R d \Omega\subset\mathbb{R}^d Ω ⊂ R d be a bounded Lipschitz domain, and let σ \sigma σ be a continuous, non-polynomial activation function. Then the neural approximation classes V \mathcal{V} V and V w \mathcal{V}_w V w are dense in L p ( Ω ) L^p(\Omega) L p ( Ω ) for all p ∈ [ 1 , ∞ ) p\in[1,\infty) p ∈ [ 1 , ∞ ) .
Proof. See Appendix C.4.
Optimization
We now present the optimization procedure adopted to determine θ \theta θ and c \mathbf{c} c . For a fixed number of basis functions N N N , the network parameters θ ∈ R p \theta\in\mathbb{R}^p θ ∈ R p define the approximation space V N ( θ ) V_N(\theta) V N ( θ ) . Within this space, the optimal coefficients c \mathbf{c} c are recovered by minimizing J J J over c \mathbf{c} c alone; since J J J is quadratic in c \mathbf{c} c for fixed θ \theta θ , this subproblem reduces to a linear least-squares problem with a closed-form solution. The parameters θ \theta θ are then updated to refine the space V N ( θ ) V_N(\theta) V N ( θ ) , and the two steps alternate until a convergence criterion is met. This variable projection strategy decouples the linear and nonlinear parts of the optimization, and we give more details in the following sections.
Construction of the loss function on VN
We seek the best approximation of u s u_s u s within V N \mathcal{V}_N V N , which amounts to finding optimal parameters ( θ , c ) ∈ R p × C N (\theta,c) \in\mathbb{R}^p \times\mathbb{C}^N ( θ , c ) ∈ R p × C N minimizing the least-squares functional J J J defined in (3.1). For fixed θ ∈ R p \theta\in\mathbb{R}^p θ ∈ R p , any v ∈ V N ( θ ) v \in V_N(\theta) v ∈ V N ( θ ) takes the form
v ( x ; θ , c ) = ∑ j = 1 N c j φ j ( x ; θ ) , x ∈ Ω , c j ∈ C . (15) v(x;\theta,c) = \sum_{j=1}^{N} c_j \varphi_j(x;\theta), \qquad x \in\Omega,\quad c_j \in\mathbb{C}. \tag*{(15)} v ( x ; θ , c ) = j = 1 ∑ N c j φ j ( x ; θ ) , x ∈ Ω , c j ∈ C . (15)
By linearity of B \mathcal{B} B , the functional J J J evaluated at v v v reads
J ( v ) = 1 2 ∥ ∑ j = 1 N c j B φ j ( ⋅ ; θ ) − f ∥ L 2 ( Ω ) 2 . (16) J(v) = \frac{1}{2}\left\lVert\sum_{j=1}^{N} c_j \mathcal{B}\varphi_j(\cdot;\theta)-f\right\rVert_{L^2(\Omega)}^2. \tag*{(16)} J ( v ) = 2 1 j = 1 ∑ N c j B φ j ( ⋅ ; θ ) − f L 2 ( Ω ) 2 . (16)
To isolate the linear structure in c c c , we define the operator Φ θ : C N → V N ( θ ) \Phi_\theta:\mathbb{C}^N \to V_N(\theta) Φ θ : C N → V N ( θ ) by
Φ θ c : = ∑ j = 1 N c j φ j ( ⋅ ; θ ) , (17) \Phi_\theta c := \sum_{j=1}^{N} c_j \varphi_j(\cdot;\theta), \tag*{(17)} Φ θ c := j = 1 ∑ N c j φ j ( ⋅ ; θ ) , (17)
and J J J takes the compact form
J ( θ , c ) = 1 2 ∥ B Φ θ c − f ∥ L 2 ( Ω ) 2 . (18) J(\theta,c) = \frac{1}{2}\left\lVert\mathcal{B}\Phi_\theta c-f\right\rVert_{L^2(\Omega)}^2. \tag*{(18)} J ( θ , c ) = 2 1 ∥ B Φ θ c − f ∥ L 2 ( Ω ) 2 . (18)
We aim to solve the optimization problem
( θ ∗ , c ∗ ) ∈ arg min θ ∈ R p , c ∈ C N J ( θ , c ) , (19) (\theta^*,c^*) \in\mathop{\arg\min}_{\theta\in\mathbb{R}^p,\ c\in\mathbb{C}^N} J(\theta,c), \tag*{(19)} ( θ ∗ , c ∗ ) ∈ arg min θ ∈ R p , c ∈ C N J ( θ , c ) , (19)
which is equivalently reformulated as the nested problem
min θ ∈ R p ( min c ∈ C N J ( θ , c ) ) . (20) \min_{\theta\in\mathbb{R}^p}\left(\min_{c\in\mathbb{C}^N}J(\theta,c)\right). \tag*{(20)} θ ∈ R p min ( c ∈ C N min J ( θ , c ) ) . (20)
This formulation separates the joint problem into two subproblems: for fixed θ \theta θ , the inner minimization over c c c is a linear least-squares problem, and the outer minimization over θ \theta θ is a nonlinear problem addressed using gradient-based optimization.
Determination of the linear coefficients c
For fixed θ \theta θ , the map c ↦ J ( θ , c ) c \mapsto J(\theta,c) c ↦ J ( θ , c ) is quadratic, and the optimal coefficients are defined by
c ( θ ) ∈ arg min c ∈ C N J ( θ , c ) . (21) c(\theta) \in\mathop{\arg\min}_{c\in\mathbb{C}^N}J(\theta,c). \tag*{(21)} c ( θ ) ∈ arg min c ∈ C N J ( θ , c ) . (21)
In general, problem (3.12) may admit multiple solutions, and we therefore replace it by the well-posed Tikhonov-regularized problem
c μ ( θ ) = arg min c ∈ C N { J ( θ , c ) + μ 2 ∥ c ∥ C N 2 } , (22) c_\mu(\theta) = \mathop{\arg\min}_{c\in\mathbb{C}^N}\left\{J(\theta,c)+\frac{\mu}{2}\lVert c\rVert_{\mathbb{C}^N}^2\right\}, \tag*{(22)} c μ ( θ ) = arg min c ∈ C N { J ( θ , c ) + 2 μ ∥ c ∥ C N 2 } , (22)
where μ > 0 \mu>0 μ > 0 is the regularization parameter. Since B Φ θ : C N → L 2 ( Ω ) \mathcal{B}\Phi_\theta:\mathbb{C}^N\to L^2(\Omega) B Φ θ : C N → L 2 ( Ω ) is bounded and its adjoint ( B Φ θ ) ∗ : L 2 ( Ω ) → C N (\mathcal{B}\Phi_\theta)^*:L^2(\Omega)\to\mathbb{C}^N ( B Φ θ ) ∗ : L 2 ( Ω ) → C N is well defined, problem (3.13) is equivalent to solving the normal equations
( ( B Φ θ ) ∗ B Φ θ + μ I ) c μ ( θ ) = ( B Φ θ ) ∗ f . (23) \left((\mathcal{B}\Phi_\theta)^*\mathcal{B}\Phi_\theta+\mu I\right)c_\mu(\theta)=(\mathcal{B}\Phi_\theta)^*f. \tag*{(23)} ( ( B Φ θ ) ∗ B Φ θ + μ I ) c μ ( θ ) = ( B Φ θ ) ∗ f . (23)
The normal equations (3.14) have a unique solution for each θ ∈ R p \theta\in\mathbb{R}^p θ ∈ R p and μ > 0 \mu>0 μ > 0 . The regularity of c μ c_\mu c μ as a function of θ \theta θ is inherited from the regularity of the basis functions, and the following proposition establishes that θ ↦ c μ ( θ ) \theta\mapsto c_\mu(\theta) θ ↦ c μ ( θ ) is continuously differentiable.
Proposition 3.4 (Regularity of c μ c_\mu c μ ). Assume that for each j = 1 , … , N j=1,\ldots,N j = 1 , … , N , the basis function φ j ( ⋅ ; θ ) \varphi_j(\cdot;\theta) φ j ( ⋅ ; θ ) is continuously differentiable with respect to θ ∈ R p \theta\in\mathbb{R}^p θ ∈ R p . Then, for each μ > 0 \mu>0 μ > 0 , the map
c μ : R p → C N , θ ↦ c μ ( θ ) = ( ( B Φ θ ) ∗ B Φ θ + μ I ) − 1 ( B Φ θ ) ∗ f , (24) c_\mu:\mathbb{R}^p\to\mathbb{C}^N,\qquad\theta\mapsto c_\mu(\theta)=\left((\mathcal{B}\Phi_\theta)^*\mathcal{B}\Phi_\theta+\mu I\right)^{-1}(\mathcal{B}\Phi_\theta)^*f, \tag*{(24)} c μ : R p → C N , θ ↦ c μ ( θ ) = ( ( B Φ θ ) ∗ B Φ θ + μ I ) − 1 ( B Φ θ ) ∗ f , (24) is well defined and continuously differentiable.
Proof. See Appendix C.5.
Determination of the nonlinear parameters
Substitution of c μ ( θ ) c_\mu(\theta) c μ ( θ ) into (3.9) yields the reduced loss
J μ ( θ ) : = J ( θ , c μ ( θ ) ) = 1 2 ∥ B Φ θ c μ ( θ ) − f ∥ L 2 ( Ω ) 2 , (25) \mathcal{J}_\mu(\theta) := \mathcal{J}(\theta,c_\mu(\theta)) = \frac{1}{2}\left\lVert\mathcal{B}\Phi_\theta c_\mu(\theta)-f\right\rVert_{L^2(\Omega)}^2, \tag*{(25)} J μ ( θ ) := J ( θ , c μ ( θ )) = 2 1 ∥ B Φ θ c μ ( θ ) − f ∥ L 2 ( Ω ) 2 , (25)
which depends on θ \theta θ alone. The outer minimization in (3.11) thus reduces to
θ ∗ ∈ arg min θ ∈ R p J μ ( θ ) . (26) \theta^* \in\underset{\theta\in\mathbb{R}^p}{\arg\min}\ \mathcal{J}_\mu(\theta). \tag*{(26)} θ ∗ ∈ θ ∈ R p arg min J μ ( θ ) . (26)
A key practical advantage of this formulation is that differentiation through the map θ ↦ c μ ( θ ) \theta\mapsto c_\mu(\theta) θ ↦ c μ ( θ ) is not required. Indeed, the following proposition shows that the gradient of the reduced loss equals the partial gradient of the joint loss with respect to θ \theta θ , evaluated at the optimal coefficients.
Proposition 3.5 (Reduced gradient identity). Let μ > 0 \mu>0 μ > 0 and assume that each φ j ( ⋅ ; θ ) \varphi_j(\cdot;\theta) φ j ( ⋅ ; θ ) is continuously differentiable in θ \theta θ . Define the regularized loss
J μ ( θ , c ) : = J ( θ , c ) + μ 2 ∥ c ∥ C N 2 . (27) \mathcal{J}_\mu(\theta,c):=\mathcal{J}(\theta,c)+\frac{\mu}{2}\lVert c\rVert_{\mathbb{C}^N}^2. \tag*{(27)} J μ ( θ , c ) := J ( θ , c ) + 2 μ ∥ c ∥ C N 2 . (27) Then the reduced loss J μ ( θ ) : = J μ ( θ , c μ ( θ ) ) \mathcal{J}_\mu(\theta):=\mathcal{J}_\mu(\theta,c_\mu(\theta)) J μ ( θ ) := J μ ( θ , c μ ( θ )) is differentiable and
∇ J μ ( θ ) = ∇ θ J ( θ , c μ ( θ ) ) . (28) \nabla\mathcal{J}_\mu(\theta)=\nabla_\theta\mathcal{J}(\theta,c_\mu(\theta)). \tag*{(28)} ∇ J μ ( θ ) = ∇ θ J ( θ , c μ ( θ )) . (28)
Proof. See Appendix C.6.
To update the nonlinear parameters of the basis functions, we use Adam [19 ] to solve the outer nonlinear optimization problem (3.17). Adam is a stochastic first-order optimizer that preconditions the gradient descent update with a diagonal approximation of the curvature, assigning each parameter an adaptive learning rate from exponential moving averages of the first and second moments of the gradient. It is the default choice in the machine learning community for its scalability and efficiency in high dimensions. Given g n = ∇ θ J μ ( θ n ) g_n=\nabla_\theta\mathcal{J}_\mu(\theta_n) g n = ∇ θ J μ ( θ n ) , the update reads
θ n + 1 = θ n − η n m ^ n v ^ n + ε , (29) \theta_{n+1}=\theta_n-\eta_n\frac{\widehat{m}_n}{\sqrt{\widehat{v}_n}+\varepsilon}, \tag*{(29)} θ n + 1 = θ n − η n v n + ε m n , (29)
where
m n = β 1 m n − 1 + ( 1 − β 1 ) g n , v n = β 2 v n − 1 + ( 1 − β 2 ) g n ⊙ g n , (30) m_n=\beta_1m_{n-1}+(1-\beta_1)g_n,\qquad v_n=\beta_2v_{n-1}+(1-\beta_2)g_n\odot g_n, \tag*{(30)} m n = β 1 m n − 1 + ( 1 − β 1 ) g n , v n = β 2 v n − 1 + ( 1 − β 2 ) g n ⊙ g n , (30)
m ^ n = m n 1 − β 1 n , v ^ n = v n 1 − β 2 n , (31) \widehat{m}_n=\frac{m_n}{1-\beta_1^n},\qquad\widehat{v}_n=\frac{v_n}{1-\beta_2^n}, \tag*{(31)} m n = 1 − β 1 n m n , v n = 1 − β 2 n v n , (31)
with default hyperparameters ( β 1 , β 2 , ε ) = ( 0.9 , 0.999 , 10 − 8 ) (\beta_1,\beta_2,\varepsilon)=(0.9,0.999,10^{-8}) ( β 1 , β 2 , ε ) = ( 0.9 , 0.999 , 1 0 − 8 ) and the learning rate is decayed by a factor of 0.9 every 2000 steps.
Discretization of the normal equations and the empirical loss
The normal equations (3.14) and the reduced loss (3.16) involve L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) inner products that must be approximated numerically. Let { x k } k = 1 N q ⊂ Ω \{x_k\}_{k=1}^{N_q}\subset\Omega { x k } k = 1 N q ⊂ Ω denote quadrature points forming a uniform grid with equal weights ∣ Ω ∣ / N q |\Omega|/N_q ∣Ω∣/ N q . Approximating the L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) inner products by this quadrature rule turns the continuous normal equations (3.14) into the discrete normal equations
( B θ ∗ B θ + μ I ) c μ ( N q ) ( θ ) = B θ ∗ f , (32) \left(B_\theta^*B_\theta+\mu I\right)c_\mu^{(N_q)}(\theta)=B_\theta^*f, \tag*{(32)} ( B θ ∗ B θ + μ I ) c μ ( N q ) ( θ ) = B θ ∗ f , (32)
where the matrix and vector entries are given by
( B θ ∗ B θ ) i j : = ∑ k = 1 N q B φ i ( x k ; θ ) B φ j ( x k ; θ ) ‾ , ( B θ ∗ f ) i : = ∑ k = 1 N q f ( x k ) B φ i ( x k ; θ ) ‾ , (33) (B_\theta^*B_\theta)_{ij}:=\sum_{k=1}^{N_q}\mathcal{B}\varphi_i(x_k;\theta)\overline{\mathcal{B}\varphi_j(x_k;\theta)},\qquad(B_\theta^*f)_i:=\sum_{k=1}^{N_q}f(x_k)\overline{\mathcal{B}\varphi_i(x_k;\theta)}, \tag*{(33)} ( B θ ∗ B θ ) ij := k = 1 ∑ N q B φ i ( x k ; θ ) B φ j ( x k ; θ ) , ( B θ ∗ f ) i := k = 1 ∑ N q f ( x k ) B φ i ( x k ; θ ) , (33)
with B θ ∈ C N q × N B_\theta\in\mathbb{C}^{N_q\times N} B θ ∈ C N q × N defined entrywise by ( B θ ) k j = B φ j ( x k ; θ ) (B_\theta)_{kj}=\mathcal{B}\varphi_j(x_k;\theta) ( B θ ) k j = B φ j ( x k ; θ ) , adjoint B θ ∗ = B θ ‾ T B_\theta^*=\overline{B_\theta}^{\mathsf{T}} B θ ∗ = B θ T , and f = ( f ( x 1 ) , … , f ( x N q ) ) T f=(f(x_1),\ldots,f(x_{N_q}))^{\mathsf{T}} f = ( f ( x 1 ) , … , f ( x N q ) ) T . Correspondingly, the empirical green-integral (GI) loss takes the vector form
J μ ( N q ) ( θ ) = ∣ Ω ∣ 2 N q ∥ B θ c μ ( N q ) ( θ ) − f ∥ 2 2 ∝ 1 N q ∥ B θ c μ ( N q ) ( θ ) − f ∥ 2 2 . (34) \mathcal{J}_\mu^{(N_q)}(\theta)=\frac{|\Omega|}{2N_q}\left\lVert B_\theta c_\mu^{(N_q)}(\theta)-f\right\rVert_2^2\propto\frac{1}{N_q}\left\lVert B_\theta c_\mu^{(N_q)}(\theta)-f\right\rVert_2^2. \tag*{(34)} J μ ( N q ) ( θ ) = 2 N q ∣Ω∣ B θ c μ ( N q ) ( θ ) − f 2 2 ∝ N q 1 B θ c μ ( N q ) ( θ ) − f 2 2 . (34)
Discretization of the Green integral operator G G G
Since B = I − G B = I - G B = I − G , the main computational task is the evaluation of the Green integral operator G G G defined in (6) . We discretize G G G on the same uniform grid { x k } k = 1 N q ⊂ Ω \{x_k\}_{k=1}^{N_q} \subset\Omega { x k } k = 1 N q ⊂ Ω introduced above, with grid spacings Δ x \Delta x Δ x and Δ z \Delta z Δ z . For clarity, the formulas below are written for d = 2 d = 2 d = 2 ; the three-dimensional case is analogous, with the cell volume Δ x Δ z \Delta x \Delta z Δ x Δ z replaced by Δ x Δ y Δ z \Delta x \Delta y \Delta z Δ x Δ y Δ z . For each basis function φ j \varphi_j φ j , the action of G G G at a grid point x k x_k x k is approximated by the composite midpoint rule
G φ j ( x k ; θ ) ≈ ∑ ℓ = 1 N q g 0 ( x k − x ℓ ) Δ x Δ z ω 2 δ m ( x ℓ ) φ j ( x ℓ ; θ ) ⏟ ( F θ ) ℓ j , k = 1 , … , N q . (35) G\varphi_j(x_k;\theta) \approx\sum_{\ell=1}^{N_q} g_0(x_k-x_\ell) \underbrace{\Delta x \Delta z \, \omega^2 \delta m(x_\ell)\varphi_j(x_\ell;\theta)}_{(F_\theta)_{\ell j}}, \qquad k=1,\ldots,N_q. \tag*{(35)} G φ j ( x k ; θ ) ≈ ℓ = 1 ∑ N q g 0 ( x k − x ℓ ) ( F θ ) ℓ j Δ x Δ z ω 2 δ m ( x ℓ ) φ j ( x ℓ ; θ ) , k = 1 , … , N q . (35)
The underbraced term collects the quadrature weight, the scattering potential, and the basis function into a single weighted-basis matrix
F θ = Δ x Δ z ω 2 diag ( δ m ) Φ θ ∈ C N q × N , (36) F_\theta= \Delta x \Delta z \, \omega^2 \operatorname{diag}(\delta m)\Phi_\theta\in\mathbb{C}^{N_q \times N}, \tag*{(36)} F θ = Δ x Δ z ω 2 diag ( δ m ) Φ θ ∈ C N q × N , (36)
where δ m = ( δ m ( x 1 ) , … , δ m ( x N q ) ) T \delta m = (\delta m(x_1),\ldots,\delta m(x_{N_q}))^\mathsf{T} δ m = ( δ m ( x 1 ) , … , δ m ( x N q ) ) T and Φ θ ∈ C N q × N \Phi_\theta\in\mathbb{C}^{N_q \times N} Φ θ ∈ C N q × N has entries ( Φ θ ) k j = φ j ( x k ; θ ) (\Phi_\theta)_{kj} = \varphi_j(x_k;\theta) ( Φ θ ) k j = φ j ( x k ; θ ) . With this notation, the quadrature rule (35) for all j j j and k k k takes the compact matrix form
G Φ θ = g 0 F θ ∈ C N q × N , (37) G\Phi_\theta= g_0 F_\theta\in\mathbb{C}^{N_q \times N}, \tag*{(37)} G Φ θ = g 0 F θ ∈ C N q × N , (37)
where g 0 ∈ C N q × N q g_0 \in\mathbb{C}^{N_q \times N_q} g 0 ∈ C N q × N q is the discrete Green matrix with entries ( g 0 ) k ℓ = g 0 ( x k − x ℓ ) (g_0)_{k\ell} = g_0(x_k-x_\ell) ( g 0 ) k ℓ = g 0 ( x k − x ℓ ) . Because the grid is uniform, g 0 g_0 g 0 is a Toeplitz matrix, and (37) applies it to each column of F θ F_\theta F θ independently:
( G Φ θ ) ⋅ , j = g 0 ( F θ ) ⋅ , j , j = 1 , … , N . (38) (G\Phi_\theta)_{\cdot,j} = g_0(F_\theta)_{\cdot,j}, \qquad j=1,\ldots,N. \tag*{(38)} ( G Φ θ ) ⋅ , j = g 0 ( F θ ) ⋅ , j , j = 1 , … , N . (38)
A Toeplitz matrix-vector product is precisely a discrete linear convolution. Thus each column of (37) can equivalently be evaluated by convolving g 0 g_0 g 0 with the corresponding column of F θ F_\theta F θ ; applied column by column. This gives
G Φ θ = g 0 ∗ F θ . (39) G\Phi_\theta= g_0 * F_\theta. \tag*{(39)} G Φ θ = g 0 ∗ F θ . (39)
This convolution structure is what enables an FFT-accelerated evaluation of G Φ θ G\Phi_\theta G Φ θ , avoiding the dense O ( N 2 ) O(N^2) O ( N 2 ) cost of forming and applying g 0 g_0 g 0 directly. Once G Φ θ G\Phi_\theta G Φ θ is computed, the discrete operator B θ B_\theta B θ is recovered as
B θ = Φ θ − G Φ θ ∈ C N q × N . (40) B_\theta= \Phi_\theta- G\Phi_\theta\in\mathbb{C}^{N_q \times N}. \tag*{(40)} B θ = Φ θ − G Φ θ ∈ C N q × N . (40)
Remark 3.6. The Green’s function kernel g 0 g_0 g 0 is singular at the origin, but the quadrature formula (35) requires its value there. We regularize this singular value using a circular-cell strategy or a rectangular-cell approximation. Both give a similar accuracy, so we adopt the circular-cell regularization throughout. See Appendix § 13 for details.
Adaptive-depth strategy
Scattered wave propagation is a highly multiscale problem: spectral bias hinders efficient training of deep networks that must capture the fine scales of the wavefield. Training a deep network of the full target depth L L L from random initialization can be challenging in strong scattering regimes due to the highly non-convex, oscillatory nature of the loss landscape. We instead develop an adaptive-depth training strategy for the plane-wave basis. We train a sequence of sub-networks of increasing depth ℓ = 1 , 2 , … , L \ell= 1,2,\ldots,L ℓ = 1 , 2 , … , L , each initialized from the converged parameters of the previous stage. This preserves the first- and second-order moments of Adam, and the linear coefficients of all neurons forming the basis are recovered exactly at every step, which together avoid spikes in the loss during training. We explain the algorithm as follows.
We start training with a network of a single layer, giving a coarse first basis, and grow the network one layer at a time. At depth ℓ \ell ℓ , only the first ℓ \ell ℓ layers are active and contribute basis functions, so the active basis grows as N ℓ = ∑ i = 1 ℓ q i / 2 N_\ell= \sum_{i=1}^{\ell} q_i/2 N ℓ = ∑ i = 1 ℓ q i /2 . At each depth ℓ \ell ℓ we run a fixed number of optimizer steps T ℓ T_\ell T ℓ , solving the normal equations (32) at every step. When depth increases from ℓ \ell ℓ to ℓ + 1 \ell+1 ℓ + 1 , the parameters of the first ℓ \ell ℓ layers, together with their first- and second-order Adam moments, are carried over unchanged; only the new layer is randomly initialized. Resetting the moments would make Adam treat the converged parameters as freshly initialized, discarding their accumulated curvature information and causing a sharp increase in the loss. This transfer of both parameters and optimizer state acts as a form of transfer learning across depths, and it is what provides a good initialization at every successive depth as the network grows. A schematic overview is given in Figure 2 .
Figure 2. Adaptive-depth strategy. At each stage ℓ \ell ℓ , only the first ℓ \ell ℓ hidden layers are active (violet) and the newly introduced layer (orange) is randomly initialized, while layers 1 , … , ℓ − 1 1,\ldots,\ell-1 1 , … , ℓ − 1 are warm-started from the previous stage. All basis functions from all active layers are connected to the output, and the linear coefficients c ∈ C N ℓ \boldsymbol{c} \in\mathbb{C}^{N_\ell} c ∈ C N ℓ are recomputed by solving the normal equations at each step. The number of basis functions grows as N ℓ = ∑ i = 1 ℓ a i / 2 N_\ell= \sum_{i=1}^{\ell} a_i/2 N ℓ = ∑ i = 1 ℓ a i /2 .
Adaptive-depth training has been explored previously in the data-driven setting [30 , 21 ] . We develop it here for a different setting, with two consequences for how it must be designed. First, each added layer contributes a new set of neural basis functions rather than standard weights: the corresponding linear coefficients are recovered by solving a least-squares problem at every step, not learned directly, so growing the network grows the basis itself. Second, the target is a scientific computing problem governed by a physics-based loss, not a data-driven task, so no ground-truth field is ever seen during training.
This second point drives the central design question of any depth-growing scheme: when to add a new layer, and how long to train at each depth. In a data-driven setting this decision can be guided directly, using a validation set to trigger growth once the model stops improving against the reference solution. In our setting no such reference exists: the only supervisory signal is the GI loss, which is physics-informed and never sees the ground-truth field. Since the decision cannot be made adaptively from a validation signal, we design it in advance instead: we fix the target depth L L L and a training schedule ( T 1 , … , T L ) (T_1,\ldots,T_L) ( T 1 , … , T L ) ahead of training, with ∑ ℓ = 1 L T ℓ = T t o t a l \sum_{\ell=1}^{L} T_\ell= T_{\mathrm{total}} ∑ ℓ = 1 L T ℓ = T total the total iteration budget.
We allocate this budget according to what we call the heavy-tail schedule: the iteration count increases with depth, T 1 ≤ T 2 ≤ ⋯ ≤ T L T_1 \le T_2 \le\cdots\le T_L T 1 ≤ T 2 ≤ ⋯ ≤ T L . Layer 1, the cheapest sub-network to optimize, is allotted the fewest iterations; deeper, harder-to-optimize sub-networks receive proportionally more iterations to converge before the next layer is added. The network parameters θ \theta θ are updated with a learning rate initialized at 10 − 3 10^{-3} 1 0 − 3 and decayed exponentially, with gradient norm clipping applied at a threshold of one.
Algorithmic and computational aspects
In the following, we describe the main implementation aspects of the NAP-Wave solver. The full procedure is summarized in Algorithm 1, and the general workflow is illustrated in Figure 4 . We focus on the following aspects: the efficient assembly of the discrete normal-equation matrices in (3.23), the linear solver, the choice of backend library, and numerical precision.
Figure 4. General workflow of the NAP-Wave solver. Each forward pass evaluates the loss J μ ( θ ) \mathcal{J}_{\mu}(\theta) J μ ( θ ) : for fixed θ \theta θ , the basis functions φ j \varphi_j φ j are evaluated on the quadrature grid, B θ B_{\theta} B θ is assembled via the FFT-accelerated Green operator, and c c c is obtained by solving the Tikhonov-regularized normal equations through a Cholesky factorization. In the backward pass, the reduced gradient ∇ θ J μ \nabla_{\theta}\mathcal{J}_{\mu} ∇ θ J μ is computed without differentiating through c μ ( θ ) c_{\mu}(\theta) c μ ( θ ) and used to update θ \theta θ with Adam. The inner loop repeats this for the T ℓ T_{\ell} T ℓ steps allotted to the current depth ℓ \ell ℓ ; the outer loop then appends a layer and warm-starts the next stage, until the target depth L L L is reached.
Computational domain
One of the main practical challenges is that we only have access to the velocity field on a bounded domain Ω \Omega Ω , whereas the theoretical formulation of § 2 requires the contrast δ m \delta m δ m to have compact support strictly inside Ω \Omega Ω . To resolve this, we embed Ω \Omega Ω inside a larger computational domain Ω f u l l = Ω ∪ Γ w \Omega_{\mathrm{full}} = \Omega\cup\Gamma_w Ω full = Ω ∪ Γ w , where Γ w \Gamma_w Γ w denotes padding (taper) region. Within Γ w \Gamma_w Γ w , the contrast is smoothly damped to zero using a raised-cosine taper (see Appendix F.§ 1 ), such that the tapered contrast δ m ~ \widetilde{\delta m} δ m satisfies supp ( δ m ~ ) ⊂ Ω f u l l \operatorname{supp}(\widetilde{\delta m}) \subset\Omega_{\mathrm{full}} supp ( δ m ) ⊂ Ω full . A schematic overview is given in Figure 3 .
Figure 3. Construction of the full computational domain Ω f u l l \Omega_{\mathrm{full}} Ω full and the tapering of the medium contrast. (a) The tapered contrast δ m ~ \widetilde{\delta m} δ m over Ω f u l l \Omega_{\mathrm{full}} Ω full : the original domain Ω \Omega Ω is embedded in a larger domain surrounded by a border region Γ t a p e r \Gamma_{\mathrm{taper}} Γ taper of width w x w_x w x in the horizontal direction and w z w_z w z in the vertical direction. (b) Cross-section at fixed z z z comparing the untapered contrast δ m \delta m δ m with δ m ~ \widetilde{\delta m} δ m .
We adopt a nondimensionalization form of the Lippmann-Schwinger integral equation. The physical coordinates x x x and z z z , the velocities v v v and v 0 v_0 v 0 , are rescaled by λ 0 = 2 π v 0 / ω \lambda_0 = 2\pi v_0/\omega λ 0 = 2 π v 0 / ω , giving v n d = v / λ 0 v_{\mathrm{nd}} = v/\lambda_0 v nd = v / λ 0 and v 0 , n d = v 0 / λ 0 v_{0,\mathrm{nd}} = v_0/\lambda_0 v 0 , nd = v 0 / λ 0 . The nondimensional contrast then becomes
δ m n d ( x ) = 1 v n d ( x ) 2 − 1 v 0 , n d 2 = λ 0 2 ( 1 v ( x ) 2 − 1 v 0 2 ) = λ 0 2 δ m ( x ) . (41) \delta m_{\mathrm{nd}}(x) = \frac{1}{v_{\mathrm{nd}}(x)^2} - \frac{1}{v_{0,\mathrm{nd}}^2} = \lambda_0^2\left(\frac{1}{v(x)^2} - \frac{1}{v_0^2}\right) = \lambda_0^2\delta m(x). \tag*{(41)} δ m nd ( x ) = v nd ( x ) 2 1 − v 0 , nd 2 1 = λ 0 2 ( v ( x ) 2 1 − v 0 2 1 ) = λ 0 2 δ m ( x ) . (41)
Network architecture
We use a sine-activated feedforward network whose hidden activations are reorganized into complex basis functions, as detailed in Appendix B and shown in Figure 1 . Sine activations provide infinitely differentiable features that are well suited to representing oscillatory wave fields. We initialize the neural parameters θ = { W ( i ) , b ( i ) } i = 1 L \theta= \{W^{(i)}, b^{(i)}\}_{i=1}^{L} θ = { W ( i ) , b ( i ) } i = 1 L with a uniform scheme: each weight matrix is drawn entrywise i.i.d. from
W k l ( i ) ∼ U ( − 6 d i − 1 , 6 d i − 1 ) , i = 1 , … , L , (42) W_{kl}^{(i)} \sim\mathcal{U}\left(-\sqrt{\frac{6}{d_{i-1}}},\sqrt{\frac{6}{d_{i-1}}}\right), \qquad i=1,\ldots,L, \tag*{(42)} W k l ( i ) ∼ U ( − d i − 1 6 , d i − 1 6 ) , i = 1 , … , L , (42)
and the biases are initialized to zero, b ( i ) = 0 ∈ R a i b^{(i)} = \mathbf{0} \in\mathbb{R}^{a_i} b ( i ) = 0 ∈ R a i .
The plane-wave basis (B.5) is built from complex carriers χ i ( x ) = exp ( i k i d i ⋅ x ) \chi_i(x) = \exp(i k_i \mathbf{d}_i \cdot x) χ i ( x ) = exp ( i k i d i ⋅ x ) , each defined by a unit direction vector d i ∈ S d − 1 \mathbf{d}_i \in\mathbb{S}^{d-1} d i ∈ S d − 1 sampled deterministically and uniformly over the sphere
d i = { ( cos α i , sin α i ) , d = 2 , ( ρ i cos β i , ρ i sin β i , ζ i ) , d = 3 , i = 1 , … , N , (43) \mathbf{d}_i =
\begin{cases}
(\cos\alpha_i,\sin\alpha_i), & d=2,\\
(\rho_i \cos\beta_i,\rho_i \sin\beta_i,\zeta_i), & d=3,
\end{cases}
\qquad i=1,\ldots,N, \tag*{(43)} d i = { ( cos α i , sin α i ) , ( ρ i cos β i , ρ i sin β i , ζ i ) , d = 2 , d = 3 , i = 1 , … , N , (43)
where, for d = 2 d=2 d = 2 , α i = 2 π ( i − 1 ) / N \alpha_i = 2\pi(i-1)/N α i = 2 π ( i − 1 ) / N , and for $d=3,
ζ i = 1 − 2 i − 1 N , ρ i = 1 − ζ i 2 , β i = 2 π ( i − 1 ) ( 1 + 5 2 ) − 2 , (44) \zeta_i = 1-\frac{2i-1}{N}, \qquad\rho_i = \sqrt{1-\zeta_i^2}, \qquad\beta_i = 2\pi(i-1)\left(\frac{1+\sqrt{5}}{2}\right)^{-2}, \tag*{(44)} ζ i = 1 − N 2 i − 1 , ρ i = 1 − ζ i 2 , β i = 2 π ( i − 1 ) ( 2 1 + 5 ) − 2 , (44)
and a positive wavenumber
k i = softplus ( k ~ i ) + k min , k min > 0 , (45) k_i = \operatorname{softplus}(\widetilde{k}_i)+k_{\min}, \qquad k_{\min}>0, \tag*{(45)} k i = softplus ( k i ) + k m i n , k m i n > 0 , (45)
where k ~ i ∈ R \widetilde{k}_i \in\mathbb{R} k i ∈ R is trainable. We adopt this softplus-plus-k min k_{\min} k m i n parametrization to guarantee k i > k min > 0 k_i > k_{\min} > 0 k i > k m i n > 0 for any k ~ i ∈ R \widetilde{k}_i \in\mathbb{R} k i ∈ R , so the wavenumber remains strictly positive and bounded away from zero throughout training. We initialize each k ~ i \widetilde{k}_i k i from the true local wavenumbers k ( x ) k(x) k ( x ) over the velocity grid, restricted to the quantile range [ q 0.05 , q 0.95 ] [q_{0.05},q_{0.95}] [ q 0.05 , q 0.95 ] to exclude outliers.
Efficient solving of the normal equations
The dominant cost at each iteration of the NAP-Wave solver is the construction of the matrix G Φ θ n \mathcal{G}\Phi_{\theta_n} G Φ θ n given by (39) . To implement this efficiently, we exploit the convolution theorem via the fast Fourier transform, which yields
G Φ θ = F − 1 ( F [ g 0 ] ⊙ F [ F θ ] ) , (46) \mathcal{G}\Phi_{\theta}=\mathcal{F}^{-1}\left(\mathcal{F}[g_0]\odot\mathcal{F}[F_{\theta}]\right), \tag*{(46)} G Φ θ = F − 1 ( F [ g 0 ] ⊙ F [ F θ ] ) , (46)
where F \mathcal{F} F denotes the discrete Fourier transform and F − 1 \mathcal{F}^{-1} F − 1 its inverse. This reduces the cost of computing (39) from O ( N N q 2 ) O(NN_q^2) O ( N N q 2 ) to O ( N N q log N q ) O(NN_q\log N_q) O ( N N q log N q ) per iteration. Implementation details are given in Appendix F.§ 2 . The transform F [ g 0 ] \mathcal{F}[g_0] F [ g 0 ] is computed once, at initialization.
For the regularization parameter μ n > 0 \mu_n>0 μ n > 0 , we adopt an exponential decay schedule:
μ n = μ 0 ( μ f μ 0 ) n / N f , (47) \mu_n=\mu_0\left(\frac{\mu_f}{\mu_0}\right)^{n/N_f}, \tag*{(47)} μ n = μ 0 ( μ 0 μ f ) n / N f , (47)
where n n n the current iteration and N f N_f N f the total number of iterations, where with μ 0 = 10 − 3 \mu_0=10^{-3} μ 0 = 1 0 − 3 and μ f = 10 − 6 \mu_f=10^{-6} μ f = 1 0 − 6 .
After assembling B θ n ∗ B θ n + μ n I B_{\theta_n}^{*}B_{\theta_n}+\mu_n I B θ n ∗ B θ n + μ n I , the system is solved via Cholesky factorization followed by two triangular solves. A direct solver is preferred because the system is solved exactly once per iteration, so the O ( N 3 ) O(N^3) O ( N 3 ) factorization cost is amortized against the gradient computation.
The linear system (32) is solved directly using complex-valued arithmetic, rather than converting it into an equivalent real-valued block system, which incurs additional computational cost as was implemented in [3 ] . We note that we also provide a real-valued mode for devices or libraries that do not offer native support for complex-valued operations. We found that solving directly in complex-valued arithmetic is faster than using the equivalent real-valued block system, by a factor of approximately 2.7.
Backend library and float precision
The implementation is built on JAX, selected over PyTorch and TensorFlow after extensive preliminary benchmarks. Two features of JAX are particularly relevant to our algorithm. First, the inner training loop consists of a fixed sequence of dense linear-algebra operations (basis evaluation, FFT, Cholesky factorization, triangular solves, matrix-vector products). Wrapping each iteration in jax.jit fuses these operations into a single optimized XLA computational graph, removing Python-level overhead. Second, the neural basis { φ j ( ⋅ ; θ ) } j = 1 N \{\varphi_j(\cdot;\theta)\}_{j=1}^{N} { φ j ( ⋅ ; θ ) } j = 1 N is expressed as a single network evaluated over N N N output channels and N q N_q N q quadrature nodes. We use jax.vmap to vectorize both axes without explicit batched code, keeping the implementation close to the mathematical formulation while delivering the throughput of a hand-written batched kernel.
All computations are carried out in double precision (float64). Double precision is required for the Cholesky factorization where a small μ \mu μ or nearly linearly dependent basis functions render the system matrix ill-conditioned, causing the factorization to lose positive definiteness in single precision.
Case studies
In this section, we present numerical experiments to assess the accuracy and computational efficiency of the proposed NAP-Wave solver. We consider five benchmarks. The first two, scattering by a fluid cylinder in 2D and a linear radial velocity gradient in 3D, admit closed-form analytical solutions and serve to verify the method against exact references.
The remaining three benchmarks, Marmousi, Overthrust, and Otway, are highly heterogeneous, multiscale velocity models originally derived from real geophysical structural data. These datasets are standard in the seismic imaging community, and we adopt them here as rigorous tests for neural PDE solvers. They exhibit sharp, non-smooth property contrasts, complex stratigraphic layering, and strong multiscale heterogeneities. These features cause severe spectral bias in standard neural networks and create poorly conditioned loss landscapes, making them difficult benchmark datasets for machine learning solvers to fit accurately without specialized architectures. The velocity field and source location for each of these three benchmarks are illustrated in Figure 5 ; a summary of the numerical parameters is given in Table 1 , and detailed descriptions are provided in the following subsections.
Table 1. Summary of the numerical benchmarks considered in this paper.
Figure 5. Velocity fields v ( x , z ) v(x,z) v ( x , z ) of the three benchmark models: Marmousi (left), Overthrust (middle), and Otway (right). The corresponding parameters are summarized in Table 1 .
Exact solution benchmarks
We consider two benchmark problems with closed-form exact solutions.
2D fluid-cylinder scattering. We consider a disk of radius a = 0.3 a = 0.3 a = 0.3 centered at ( 1.5 , 1.0 ) (1.5, 1.0) ( 1.5 , 1.0 ) , with constant velocity v 1 = 2.25 v_1 = 2.25 v 1 = 2.25 , embedded in a homogeneous background v 0 = 1.5 v_0 = 1.5 v 0 = 1.5 over the domain [ 0 , 3 ] × [ 0 , 2 ] [0, 3] \times[0, 2] [ 0 , 3 ] × [ 0 , 2 ] , excited at frequency f = 6 f = 6 f = 6 Hz by the line-source incident field u 0 u_0 u 0 located at x s = ( 0.3 , 0.1 ) x_s = (0.3, 0.1) x s = ( 0.3 , 0.1 ) . The exact solution, expressed in terms of the
Algorithm 1 NAP-Wave solver with adaptive-depth training strategy
Require: Initial parameters θ 0 ∈ R p \theta^{0} \in\mathbb{R}^{p} θ 0 ∈ R p , target network depth L L L , iterations per depth { T ℓ } ℓ = 1 L \{T_{\ell}\}_{\ell=1}^{L} { T ℓ } ℓ = 1 L , regularization schedule μ n > 0 \mu_{n} > 0 μ n > 0 , quadrature points ( x i ) i = 1 N q (x_{i})_{i=1}^{N_{q}} ( x i ) i = 1 N q , source vector f ∈ C N q f \in\mathbb{C}^{N_{q}} f ∈ C N q
Ensure: Approximate scattered field u N ∈ V N u_{N} \in V_{N} u N ∈ V N
Precompute and store F [ g 0 ] \mathcal{F}[g_{0}] F [ g 0 ] on the zero-padded grid
Compute the contrast vector δ m \delta m δ m on the quadrature points
for ℓ = 1 , … , L \ell= 1,\ldots,L ℓ = 1 , … , L do ▷ Adaptive-depth loop
Set N ℓ = ∑ i = 1 ℓ a i / 2 N_{\ell} = \sum_{i=1}^{\ell} a_{i}/2 N ℓ = ∑ i = 1 ℓ a i /2 basis functions
if ℓ > 1 \ell> 1 ℓ > 1 then
Warm-start layers 1 , … , ℓ − 1 1,\ldots,\ell-1 1 , … , ℓ − 1 from the previous stage, along with their first- and second-order Adam moments
Initialize layer ℓ \ell ℓ randomly
end if
for n = 0 , … , T ℓ − 1 n = 0,\ldots,T_{\ell}-1 n = 0 , … , T ℓ − 1 do
Evaluate Φ θ n ∈ C N q × N ℓ \Phi_{\theta_{n}} \in\mathbb{C}^{N_{q}\times N_{\ell}} Φ θ n ∈ C N q × N ℓ on the grid
Assemble B θ n = Φ θ n − G Φ θ n B_{\theta_{n}} = \Phi_{\theta_{n}} - G\Phi_{\theta_{n}} B θ n = Φ θ n − G Φ θ n via FFT (46)
Factorize B θ n ∗ B θ n + μ n I = C C ∗ B_{\theta_{n}}^{*}B_{\theta_{n}} + \mu_{n}I = CC^{*} B θ n ∗ B θ n + μ n I = C C ∗ (Cholesky)
Solve for c μ n ( N q ) ( θ n ) c_{\mu_{n}}^{(N_{q})}(\theta_{n}) c μ n ( N q ) ( θ n )
Evaluate J μ n ( N q ) ( θ n ) = 1 N q ∥ B θ n c μ n ( N q ) ( θ n ) − f ∥ 2 2 \mathcal{J}_{\mu_{n}}^{(N_{q})}(\theta_{n}) = \frac{1}{N_{q}}\left\lVert B_{\theta_{n}}c_{\mu_{n}}^{(N_{q})}(\theta_{n}) - f\right\rVert_{2}^{2} J μ n ( N q ) ( θ n ) = N q 1 B θ n c μ n ( N q ) ( θ n ) − f 2 2
Compute ∇ θ J μ n ( N q ) ( θ n ) \nabla_{\theta}\mathcal{J}_{\mu_{n}}^{(N_{q})}(\theta_{n}) ∇ θ J μ n ( N q ) ( θ n )
Update: θ n + 1 = Adam ( θ n , ∇ θ J μ n ( N q ) ( θ n ) ) \theta_{n+1} = \operatorname{Adam}\left(\theta_{n}, \nabla_{\theta}\mathcal{J}_{\mu_{n}}^{(N_{q})}(\theta_{n})\right) θ n + 1 = Adam ( θ n , ∇ θ J μ n ( N q ) ( θ n ) )
end for
end for
return u N ( x ) = ∑ j = 1 N ( c μ n ( N q ) ( θ f i n a l ) ) j φ j ( x ; θ f i n a l ) u_{N}(x) = \sum_{j=1}^{N}\left(c_{\mu_{n}}^{(N_{q})}(\theta^{\mathrm{final}})\right)_{j}\varphi_{j}(x;\theta^{\mathrm{final}}) u N ( x ) = ∑ j = 1 N ( c μ n ( N q ) ( θ final ) ) j φ j ( x ; θ final )
total field reads
u ( r , ϕ ) = { ∑ n ∈ Z c n [ J n ( k 0 r ) + a n H n ( 2 ) ( k 0 r ) ] e i n ϕ , r > a , ∑ n ∈ Z c n b n J n ( k 1 r ) e i n ϕ , r < a . (48) u(r,\phi) =
\begin{cases}
\sum_{n\in\mathbb{Z}} c_{n}\left[J_{n}(k_{0}r)+a_{n}H_{n}^{(2)}(k_{0}r)\right]e^{in\phi}, & r>a,\\
\sum_{n\in\mathbb{Z}} c_{n}b_{n}J_{n}(k_{1}r)e^{in\phi}, & r<a.
\end{cases} \tag*{(48)} u ( r , ϕ ) = { ∑ n ∈ Z c n [ J n ( k 0 r ) + a n H n ( 2 ) ( k 0 r ) ] e in ϕ , ∑ n ∈ Z c n b n J n ( k 1 r ) e in ϕ , r > a , r < a . (48)
where ( r , ϕ ) (r,\phi) ( r , ϕ ) are polar coordinates centered at the disk, k 0 = ω / v 0 k_{0} = \omega/v_{0} k 0 = ω / v 0 , k 1 = ω / v 1 k_{1} = \omega/v_{1} k 1 = ω / v 1 , and c n c_{n} c n is the incident field’s local multipole amplitude at the disk. The coefficients a n , b n a_{n}, b_{n} a n , b n follow from matching u u u and its normal derivative at r = a r = a r = a .
3D radial velocity gradient. The velocity increases linearly with distance from the source x s = ( 0 , 0 , 0 ) x_{s} = (0,0,0) x s = ( 0 , 0 , 0 ) ,
v ( r ) = v 0 + α r , r = ∥ x − x s ∥ , α > 0 , (49) v(r) = v_{0} + \alpha r,\qquad r = \lVert x-x_{s}\rVert,\quad\alpha> 0, \tag*{(49)} v ( r ) = v 0 + α r , r = ∥ x − x s ∥ , α > 0 , (49)
with v 0 = 1.5 v_0 = 1.5 v 0 = 1.5 and α = 0.6 \alpha= 0.6 α = 0.6 , over the domain [ − 1 , 1 ] 3 [-1, 1]^3 [ − 1 , 1 ] 3 at frequency f = 2 H z f = 2\,\mathrm{Hz} f = 2 Hz . The exact solution, expressed in terms of the total field reads
u ( r ) = 1 4 π r v ( r ) v 0 ( v 0 v ( r ) ) i μ ~ , μ ~ = ( ω α ) 2 − 1 4 , (50) u(r) = \frac{1}{4\pi r}\sqrt{\frac{v(r)}{v_0}}\left(\frac{v_0}{v(r)}\right)^{i\tilde{\mu}}, \qquad\tilde{\mu} = \sqrt{\left(\frac{\omega}{\alpha}\right)^2 - \frac{1}{4}}, \tag*{(50)} u ( r ) = 4 π r 1 v 0 v ( r ) ( v ( r ) v 0 ) i μ ~ , μ ~ = ( α ω ) 2 − 4 1 , (50)
Marmousi benchmark
We consider the Marmousi model, a synthetic acoustic velocity model used as a benchmark for seismic imaging and full-waveform inversion in heterogeneous media [7 ] . The geometric and physical parameters of this benchmark are summarized in Table 1 , and the corresponding velocity field v ( x , z ) v(x, z) v ( x , z ) is shown in Figure 5 . The model is representative of complex sedimentary structures, with a strongly heterogeneous, multi-layered velocity field exhibiting stratigraphic complexity, dipping reflectors, and sharp contrasts across geological interfaces. These features make it a challenging configuration for wave-scattering solvers, as the resulting wavefield combines high-frequency oscillations with localized reflections and refractions.
Overthrust benchmark
We consider the Overthrust model [4 ] , a synthetic acoustic benchmark released by the Society of Exploration Geophysicists (SEG) and the European Association of Geoscientists and Engineers (EAGE), used as a reference for seismic imaging in structurally complex, faulted geological settings. The geometric and physical parameters of this benchmark are summarized in Table 1 , and the corresponding velocity field v ( x , z ) v(x, z) v ( x , z ) is shown in Figure 5 . The model represents a thrust-belt geological setting, characterized by a layered sedimentary sequence disrupted by overthrust faulting, producing offset velocity discontinuities and locally steep, non-conformable interfaces. Compared to the Marmousi model, the Overthrust velocity field exhibits smoother lateral variations within each layer but sharper, structurally-driven discontinuities across fault planes, making it a complementary test for assessing the sensitivity of wave-scattering solvers to abrupt structural offsets rather than fine-scale stratigraphic heterogeneity.
Otway benchmark
As a last benchmark, we consider the Otway model, a synthetic acoustic velocity model built from geological and well-log data of the Otway Basin, Australia, a sedimentary basin widely studied in the context of geological carbon storage and C O 2 \mathrm{CO}_2 CO 2 sequestration monitoring. The geometric and physical parameters of this benchmark are summarized in Table 1 , and the corresponding velocity field v ( x , z ) v(x, z) v ( x , z ) is shown in Figure 5 . The model features a comparatively smooth, gently layered sedimentary sequence with moderate velocity contrasts between formations, in contrast with the strong heterogeneity and stratigraphic complexity of the Marmousi and Overthrust models. This benchmark is therefore representative of near-surface, moderately heterogeneous sedimentary basins, and allows us to assess the accuracy and efficiency of the proposed solver in a regime dominated by smoother, large-scale velocity variations rather than sharp interfaces.
Reference solutions, reproducibility, and hardware
For velocity models admitting a closed-form solution, we compare directly against the exact solution. For the realistic velocity models, we generate a numerical reference using an optimal nine-point finite-difference discretization of the scattered-field Helmholtz equation with a perfectly matched layer (PML). We compute the relative L 2 L^2 L 2 error on a uniform grid as
Relative L 2 ( i n % ) = 100 × ( ∑ i ∣ u p r e d ( x i ) − u r e f ( x i ) ∣ 2 ∑ i ∣ u r e f ( x i ) ∣ 2 ) 1 / 2 , (51) \operatorname{Relative}\ L^2\ (\mathrm{in}\ \%) = 100 \times\left(\frac{\sum_i \left|u_{\mathrm{pred}}(x_i) - u_{\mathrm{ref}}(x_i)\right|^2}{\sum_i \left|u_{\mathrm{ref}}(x_i)\right|^2}\right)^{1/2}, \tag*{(51)} Relative L 2 ( in % ) = 100 × ( ∑ i ∣ u ref ( x i ) ∣ 2 ∑ i ∣ u pred ( x i ) − u ref ( x i ) ∣ 2 ) 1/2 , (51)
computed over the interior (taper-excluded) region.
All experiments are performed on a single NVIDIA H100 80GB HBM3 GPU. The implementation uses JAX 0.9.1 with Flax 0.12.6 and Optax 0.2.4 for automatic differentiation and optimization, and NumPy 2.1.1 and SciPy 1.17.0 for numerical routines, running on Python 3.11.5 with CUDA 12 support. The source code is publicly available at https://github.com/Hamza-Kamil/NAP-WAVE.
Numerical results
Qualitative validation
For each benchmark, we depict the real and imaginary parts of the predicted and reference scattered field, together with the pointwise error. All fields are displayed in the nondimensional coordinates. The network architecture and optimizer settings are summarized in Table 2 .
Table 2. Network architecture, training, and evaluation settings used to produce Figure 6 -Figure 10 and Table 3 . The quadrature grid and test grid are given as N x × N z N_x \times N_z N x × N z in 2D and N x × N y × N z N_x \times N_y \times N_z N x × N y × N z in 3D, where N x N_x N x , N y N_y N y , and N z N_z N z denote the number of points along the x x x -, y y y -, and z z z -directions, respectively. The total number of iterations is the sum of the iterations allocated to each depth.
Exact solutions
We first validate the proposed NAP-Wave solver against the closed-form scattered-field solutions of the 2D fluid-cylinder scattering and 3D linear radial velocity gradient benchmarks. The results are displayed in Figure 6 -Figure 7 . In the 2D case, the predicted field is visually indistinguishable from the reference solution: the wavefronts, amplitude decay, and phase structure are reproduced across the full domain, and the pointwise error remains several orders of magnitude smaller than the field itself, concentrated mainly near the source and scatterer, where the field gradients are largest, and decaying rapidly at larger radii. The line profiles along fixed x x x , fixed z z z , and the diagonal cut confirm this agreement, with predicted and reference curves overlapping almost exactly, including in the low-amplitude oscillatory tails. The 2D benchmark achieves a relative L 2 L^2 L 2 error of 0.97% and GI residual loss of order 10 − 9 10^{-9} 1 0 − 9 with the plane-wave basis function, confirming that the network accurately resolves both the oscillatory structure and the smoothly decaying envelope.
Figure 6. Fluid-cylinder scattering benchmark: real part (top row) and imaginary part (middle row) of the scattered field u s u_s u s , comparing the reference, the prediction, and the pointwise error. Bottom row: cross-sections along fixed x x x , fixed z z z , and the domain diagonal. Relative L 2 L^2 L 2 error: 0.97%.
Figure 7. 3D radial velocity gradient benchmark: real part of the scattered field u s u_s u s mapped onto the outer surfaces of a cube enclosing the domain (top row) and via internal cross-section slices (bottom row). Comparison shows the prediction (left), the exact reference solution (center), and the pointwise error (right). Relative L 2 L^2 L 2 error: 1.53%.
In the 3D case, the solver reproduces the alternating positive and negative lobes of the reference field closely on the domain boundary ∂ Ω \partial\Omega ∂ Ω , shown here on the six faces of the cubic domain. Unlike in the 2D case, the pointwise error is not concentrated near the source but is instead spread more broadly over the boundary, with local maxima near the cube edges and corners. The 3D benchmark attains a relative L 2 L^2 L 2 error of 1.53%, showing the scalability of the NAP-Wave solver.
Marmousi benchmark
The results obtained for the Marmousi benchmark are shown in Figure 8 . The predicted scattered field reproduces the reference wavefronts across the domain, including their pronounced curvature and the local wavefront distortions induced by the strong, finely stratified velocity heterogeneity of the Marmousi model. The pointwise error of the real and imaginary parts remains low over most of the domain, with the largest discrepancies concentrated in the region of strongest lateral velocity contrast (roughly x < − 2.5 x < -2.5 x < − 2.5 ). The cross-sections in Figure 8 confirm close agreement between the predicted and reference curves across the full oscillatory range.
Figure 8. Marmousi benchmark: real part (top), imaginary part (middle) of u s u_s u s , and cross-sections (bottom).
Overthrust benchmark
We train the proposed NAP-Wave model on the Overthrust benchmark and the results are shown in Figure 9 . The predicted field correctly reproduces the sharp wavefront pattern characteristic of this benchmark, which arises from the abrupt fault-plane offsets in the Overthrust velocity model acting as strong, spatially localized reflectors. Both the real and imaginary parts are in good quantitative agreement with the reference solution over the bulk of the domain. The largest errors are concentrated near the top of the domain, close to the source location and along the outer edge of the computational domain, where the truncated finite-difference reference itself is most sensitive to residual reflections from the perfectly matched layer. The cross-sections in Figure 9 further confirm that the model accurately resolves the multiscale character of the field, capturing both the long-wavelength envelope and the fine oscillatory modes near the diagonal cut, although a modest mismatch is visible at the very start of the diagonal profile.
Figure 9. Overthrust benchmark: real part (top), imaginary part (middle) of u s u_s u s , and cross-sections (bottom).
Otway benchmark
The results obtained after training the NAP-Wave model on the Otway benchmark are shown in Figure 10 . Because the Otway velocity model is comparatively smooth and gently layered relative to Marmousi and Overthrust, the resulting scattered field exhibits smoother, more regular wavefronts, and the network reproduces both the real and imaginary parts, as well as the amplitude, with pointwise errors remaining close to zero over most of the domain. The residual error that does appear is distributed as a fine, speckle-like pattern of low amplitude rather than being localized to a specific region, consistent with a well-conditioned, mildly heterogeneous scattering problem. The cross-sections in Figure 10 confirm this good agreement, with the predicted and reference curves nearly indistinguishable, including in the regions of most rapid oscillation.
Figure 10. Otway benchmark: real part (top), imaginary part (middle) of u s u_s u s , and cross-sections (bottom).
Comparison with the state of the art
We compare the performance of NAP-Wave against published results from the literature in Table 3 . The results show that NAP-Wave achieves the lowest relative L 2 L^2 L 2 error on all three benchmarks compared to representative PINN methods.
Table 3. Comparison with representative PINN methods on Marmousi, Overthrust, and Otway benchmarks.
Ablation study
After we established that NAP-Wave produces accurate scattered fields across all benchmarks, we now isolate the contribution of each design choice.
Effect of the quadrature and basis size
We perform this ablation on the fluid-cylinder scattering benchmark. We study how the accuracy of the NAP-Wave solver depends on (i) the number of quadrature points used to approximate the Green’s-integral loss, and (ii) the number of plane-wave neural basis functions. Accuracy is measured by the relative L 2 L^2 L 2 error against the exact solution and by the Green’s-integral loss during training. We also check whether the theoretical bound relating the Green’s-integral loss to the true error, given in Theorem 3.2 , holds in practice at the discrete level as training progresses. Figure 11 summarizes these three results.
Figure 11. Fluid-cylinder scattering benchmark. Left: relative L 2 L^2 L 2 error versus optimization step with the theoretical bounds from Theorem 3.2 ; vertical dotted lines mark each depth transition. Middle: error versus N q N_q N q , with the number of basis functions fixed. Right: error versus N N N , with the number of quadrature points fixed.
Error bound versus training progress. The left panel in Figure 11 tracks the relative L 2 L^2 L 2 error (black) over the course of training, together with the theoretical lower and upper bounds c 1 J ( u θ ) c_1\sqrt{J(u_\theta)} c 1 J ( u θ ) and c 2 J ( u θ ) c_2\sqrt{J(u_\theta)} c 2 J ( u θ ) derived from the Green’s-integral loss J ( u θ ) J(u_\theta) J ( u θ ) . The shaded region is the range allowed by the theory, and the vertical dotted lines mark the steps at which a new layer is added under the adaptive-depth training strategy. The true error stays inside the theoretical band throughout training, including across every depth transition, and both quantities flatten out together once the network has converged at depth 5, confirming the bound in this setting.
Effect of the number of quadrature points. The middle panel in Figure 11 shows the relative L 2 L^2 L 2 error as the number of Green’s-integral evaluation points N q N_q N q is increased, with the number of basis functions held fixed. The error decreases steadily, close to log-linearly, as the integral is evaluated more accurately, with no sign of saturation over the range tested.
Effect of the number of basis functions. The right panel in Figure 11 shows the relative L 2 L^2 L 2 error as the number of basis functions increases, with the number of quadrature points held fixed. The basis count is increased by growing the depth of the network. We observe that the relative L 2 L^2 L 2 error drops sharply at first and then plateaus beyond roughly 200 basis functions. One explanation is that, beyond a certain network size, the error becomes limited by other factors, such as the discretization of the integral, rather than by the number of basis functions. Another possible explanation is that as more basis functions are added, they become increasingly linearly dependent and start to carry redundant information, so that further increases no longer improve the approximation.
Effect of the depth schedule, the coefficient solve, and the basis choice
We consider the Marmousi model to test the robustness of the proposed NAP-Wave solver through three ablations: the depth-training schedule, the linear-coefficient optimization strategy, and the choice of basis function.
Depth-training schedule. We compare the adaptive-depth training strategy against training with a fixed depth from the start, for final depths ranging from 2 to 6, using the plane-wave basis with c c c solved by least squares, under an equal total iteration budget T t o t a l ( L ) = 1000 × L T_{\mathrm{total}}(L) = 1000 \times L T total ( L ) = 1000 × L for fairness, and averaging each configuration over n = 5 n = 5 n = 5 independent seeds. Results are reported in Table 4 and Figure 12 . Table 4 shows that adaptive-depth is more accurate than fixed depth at every depth in relative L 2 L^2 L 2 error, and more accurate in GI loss as well from depth 3 onward. Moreover, adaptive depth is also faster in total training time at every depth, by an average of about 1.21 × 1.21\times 1.21 × across depths 2-6. Figure 12 extends this comparison to widths 80 and 128, up to depth 5. The ranking is consistent across all four settings: at both widths, the adaptive schedule (solid) reaches a given loss or error level at fewer optimization steps than the fixed schedule (dashed), and the gap opens immediately after the first layer is added. The narrower network (80, green) plateaus at a visibly higher GI loss and relative L 2 L^2 L 2 error than the wider one (128, orange) under either schedule.
Table 4. Marmousi benchmark: effect of network depth and the adaptive-depth schedule. Adaptive models follow the per-stage schedule ( T 1 , … , T L ) (T_1, \ldots, T_L) ( T 1 , … , T L ) , and fixed models train in one stage for T t o t a l ( L ) = ∑ ℓ T ℓ T_{\mathrm{total}}(L) = \sum_{\ell} T_\ell T total ( L ) = ∑ ℓ T ℓ . Mean ± \pm ± standard deviation over n = 5 n = 5 n = 5 seeds. Bold marks the better of Adaptive/Fixed per metric.
Figure 12. Marmousi benchmark: effect of the depth-training schedule (adaptive vs. fixed) and network width on convergence, for the plane-wave basis, with c c c solved by least squares throughout. The GI loss (left) and relative L 2 L^2 L 2 error (right) are plotted against optimization step on a log scale. Vertical dashed lines mark the steps at which a new layer is added.
Linear-coefficient optimization strategy. We now fix the depth-training schedule to adaptive and study how the strategy used for the linear coefficients c c c affects convergence. Two strategies are compared: c-LS, which solves the normal equations (32) exactly at each step for the current θ \theta θ , and c-trained, which treats c c c as a trainable parameter updated jointly with θ \theta θ by Adam. In the c-trained setting, coefficients introduced by a new layer are initialized to zero so the augmented network initially reproduces the previous depth’s output. Results are reported in Figure 13 -Figure 14 . Figure 13 tracks the GI loss and the relative L 2 L^2 L 2 error against the optimization step for the plane-wave basis. The c-LS variant attains a consistently lower GI loss and relative L 2 L^2 L 2 error than c-trained throughout training, at both widths and at every depth transition. The GI-loss trajectory of c-trained shows spikes at several depth transitions, most visibly at width 128, before resuming its downward trend. The c-LS trajectory shows no such spike at any of the depth transitions. We attribute this to two mechanisms. The coefficients c c c are solved exactly by least squares at each step, and each new
Figure 13. Marmousi benchmark: effect of the c c c -update strategy on convergence, using the plane-wave basis and the adaptive-depth schedule. GI loss (left) and relative L 2 L^2 L 2 error (right) against optimization step, both on a log scale. Vertical dashed lines mark the steps at which a new layer is added.
Figure 14. Marmousi benchmark. Effect of the c c c -update strategy (closed-form least-squares solve / trained jointly with θ \theta θ ) on the predicted scattered field Re ( u s ) \operatorname{Re}(u_s) Re ( u s ) , using the adaptive-depth schedule with per-stage iteration table ( T 1 , … , T 5 ) = ( 100 , 1000 , 2000 , 3000 , 3000 ) (T_1,\ldots,T_5)=(100,1000,2000,3000,3000) ( T 1 , … , T 5 ) = ( 100 , 1000 , 2000 , 3000 , 3000 ) . Left: reference field. Middle: pointwise error for c-trained. Right: pointwise error for c-LS, shared color scale with the middle panel.
layer is warm-started from the previous one, inheriting the optimizer’s first- and second-moment estimates. In Figure 14 , we provide a qualitative visualization of this gap at the pointwise level. The c c c -trained variant concentrates visibly larger pointwise error in the strongly heterogeneous region near x < − 5 x < -5 x < − 5 , whereas c c c -LS keeps a smaller and more uniformly distributed residual across the domain.
Basis function choice. We retain the c c c -LS strategy together with the adaptive-depth schedule, and compare the plane-wave and standard bases. Figure 15 reports the results for widths 80 and 128. The plane-wave basis converges markedly faster: at width 128 it reaches its final accuracy level after roughly 10,000 epochs, whereas the standard basis needs the full budget of 35,000 epochs to attain a comparable relative L 2 L^2 L 2 error. The two bases therefore reach a similar final L 2 L^2 L 2 accuracy, and the benefit of the plane-wave carrier is a large reduction in the number of iterations required. The GI loss behaves differently and remains consistently lower for the plane-wave basis throughout training, at both widths. We interpret this acceleration as a mitigation of spectral bias: the oscillatory component of the field is supplied by the carrier rather than learned, so the network only has to fit the slowly varying amplitude, whereas the standard basis must build the oscillations from scratch.
Figure 15. Marmousi benchmark, adaptive-depth schedule with c c c updated by least squares (c c c -LS): effect of the plane-wave basis on convergence. GI loss (left) and relative L 2 L^2 L 2 error (right) against optimization step (Epoch), both on a log scale. Vertical dashed lines mark the steps at which a new layer is added.
Combining all factors. We conclude with a quantitative ablation over all configurations, which combines the choice of basis, the coefficient strategy, and the depth schedule at two network widths. Table 5 reports, for each configuration, the optimization step and the wall-clock time at which the relative L 2 L^2 L 2 error first drops below a threshold of 10%. This measure isolates the cost of reaching an accurate solution rather than the accuracy attained at the end of a fixed budget. For every basis and every depth schedule, c c c -LS crosses the threshold within about 20,000 steps, whereas c c c -trained requires at least 78,297 steps, an acceleration of about four times. Under c c c -LS, the plane-wave basis reaches the threshold before the standard basis at both widths and under both schedules, with a speedup of 1.5× in wall-clock time. The adaptive-depth schedule adds a further speedup over fixed depth, which completes the picture: the three ingredients act in the same direction and their combination is the fastest configuration tested. Adaptive depth improves on fixed depth in almost every setting, c-LS outperforms c-trained in both GI loss and relative L 2 L^2 L 2 error regardless of basis, and the plane-wave basis reaches a given accuracy in fewer steps than a basis built from the network neurons alone. Together they reach 10% error in 7,853 steps and 2.30 min.
Table 5. Marmousi benchmark. Optimization step and wall-clock time at which each method’s relative L 2 L^2 L 2 error first drops to or below a 10% success threshold. All runs use depth 5; the adaptive schedule spends 100, 1000, 2000, 3000 iterations at layers 1-4 respectively before growing, then trains the final (depth-5) layer until the threshold is hit. Every run is checked every epoch and stopped the instant its relative L 2 L^2 L 2 error crosses 10%. Bold marks the better of Adaptive/Fixed on each metric.
Conclusion
We introduced NAP-Wave, a data-free physics-informed neural solver for time-harmonic scattering in heterogeneous media. The solver is trained by minimizing the residual of the Lippmann-Schwinger integral equation, which requires no differentiation in space of the network and makes training more stable and faster than strong-form PINN losses. Our main contributions are: (a) a neural adaptive plane-wave basis, where each basis function is the product of a neural amplitude and a plane-wave carrier with trainable wavenumber and fixed direction that solves the homogeneous Helmholtz equation exactly, encoding the oscillatory structure of the field directly into the trial space rather than learning it from scratch; (b) a decoupled optimization scheme, where the linear coefficients c c c are solved exactly at each step by least squares while the nonlinear basis parameters θ \theta θ are updated by a gradient-based optimizer, yielding a better-conditioned optimization than training all parameters jointly; and (c) an adaptive-depth training strategy, where the network starts shallow and grows deeper as training progresses, which makes training stable and efficient.
We validated the proposed solver against exact 2D and 3D solutions and on three realistic geophysical benchmarks (Marmousi 10 Hz, Overthrust 10 Hz, Otway 20 Hz), achieving relative L 2 L^2 L 2 errors of order 1% on the 2D and 3D exact solutions, and 5%, 9%, and 10% on Marmousi, Overthrust, and Otway respectively. An ablation study isolates the effect of the plane-wave basis and the exact least-squares coefficient solve: solving for c c c by least squares is markedly more accurate than training c c c and θ \theta θ jointly, and the plane-wave basis leads to faster convergence. The adaptive-depth strategy makes training faster and more reliable than training at full depth from the start.
We note some limitations of this study and potential improvements. First, the Lippmann-Schwinger formulation assumes a constant background wave speed v 0 v_0 v 0 , which enables the FFT-accelerated evaluation of the Green operator; a spatially varying background would break this structure and require a different formulation. Second, although the method was validated on a 3D exact solution, scaling to realistic 3D field-scale models proved difficult in practice due to the rapid growth of the memory footprint and the number of integration points and basis functions required; memory-efficient strategies, such as multi-GPU implementations, are needed to scale to full 3D models. Third, the method relies on a uniform quadrature grid and does not support local refinement; adaptive discretizations would address this, at the cost of losing the FFT structure. Fourth, the GI loss and the relative L 2 L^2 L 2 error do not decay at the same rate. The loss can reaches order 10 − 9 10^{-9} 1 0 − 9 while the error stagnates around 10 − 2 10^{-2} 1 0 − 2 , a gap controlled by the conditioning of B \mathcal{B} B . An efficient preconditioner for B \mathcal{B} B , would align the two rates and may improve the final accuracy. Fifth, in the proposed adaptive-depth training strategy, the iteration budget allocated to each depth stage is fixed in advance rather than triggered adaptively from the loss; automating the depth schedule from a more faithful convergence signal is a promising direction. Sixth, NAP-Wave is a single-instance solver: any change of frequency or the velocity, model requires retraining from scratch; an extension toward operator learning would remove this restriction and, combined with the differentiability of the solver, would open the way to coupling it with inversion frameworks.
Acknowledgments
The authors thank the DeepWave Consortium sponsors for supporting this research. This research was supported by the following Research Projects/Grants: European Union’s Horizon Europe research and innovation programme under the Marie Sklodowska-Curie Actions MSCA-DN-101119556 (IN-DEEP), and MSCA-SE-101342459 (MOST-PINN). PID2023-146678OB-I00 funded by MICIU/AEI /10.13039/501100011033 and by FEDER, EU; BCAM Severo Ochoa accreditation of excellence CEX2021-001142-S funded by MICIU/AEI/10.13039/501100011033; Basque Government through the BERC 2022-2025 program; EGOKIA (KK-2026/00050), funded by the Basque Government through ELKARTEK; Consolidated Research Group MATHMODE (IT1866-26) of the UPV/EHU given by the Department of Sciences, Universities, and Innovation of the Basque Government; BCAM-IKUR-UPV/EHU, funded by the Basque Government IKUR Strategy and by the European Union NextGenerationEU/PRTR.
Code availability
The source code is publicly available at https://github.com/Hamza-Kamil/NAP-WAVE.
Notation
Table 6 collects the symbols used throughout the paper.
Table 6. Summary of notation used throughout the paper.
Neural network approximation
We present the NN architecture used to approximate the unknown scattered field u s ∈ L 2 ( Ω ) u_s \in L^2(\Omega) u s ∈ L 2 ( Ω ) , together with the approximation spaces it generates.
Forward pass
Let L ∈ N L\in\mathbb{N} L ∈ N denote the network depth and let a = ( a 0 , a 1 , … , a L ) a=(a_0,a_1,\ldots,a_L) a = ( a 0 , a 1 , … , a L ) be the architecture vector, where a 0 a_0 a 0 is the input dimension and a ℓ ∈ 2 N a_\ell\in2\mathbb{N} a ℓ ∈ 2 N is the width of the ℓ \ell ℓ -th hidden layer. The parameter space associated with a a a is
Θ a = ∏ ℓ = 1 L ( R a ℓ × a ℓ − 1 × R a ℓ ) , (52) \Theta_a=\prod_{\ell=1}^{L}\left(\mathbb{R}^{a_\ell\times a_{\ell-1}}\times\mathbb{R}^{a_\ell}\right), \tag*{(52)} Θ a = ℓ = 1 ∏ L ( R a ℓ × a ℓ − 1 × R a ℓ ) , (52)
which is identified with R p \mathbb{R}^p R p where p = ∑ ℓ = 1 L a ℓ ( a ℓ − 1 + 1 ) p=\sum_{\ell=1}^{L}a_\ell(a_{\ell-1}+1) p = ∑ ℓ = 1 L a ℓ ( a ℓ − 1 + 1 ) is the total number of trainable parameters. Fixing the architecture a a a also fixes the number of basis functions, given by N = ∑ ℓ = 1 L a ℓ / 2 N=\sum_{\ell=1}^{L}a_\ell/2 N = ∑ ℓ = 1 L a ℓ /2 , consistently with the notation of Section 3.2.
For fixed θ = { W ℓ , b ℓ } ℓ = 1 L ∈ Θ a \theta=\{W_\ell,b_\ell\}_{\ell=1}^{L}\in\Theta_a θ = { W ℓ , b ℓ } ℓ = 1 L ∈ Θ a , the NN realization u ( ⋅ ; θ ) : Ω → R a L u(\cdot\,;\theta):\Omega\to\mathbb{R}^{a_L} u ( ⋅ ; θ ) : Ω → R a L is defined by the composition
u ( x ; θ ) = h L ∘ ⋯ ∘ h 1 ( x ) , x ∈ Ω , (53) u(x;\theta)=h_L\circ\cdots\circ h_1(x),\qquad x\in\Omega, \tag*{(53)} u ( x ; θ ) = h L ∘ ⋯ ∘ h 1 ( x ) , x ∈ Ω , (53)
where each layer map h ℓ : R a ℓ − 1 → R a ℓ h_\ell:\mathbb{R}^{a_{\ell-1}}\to\mathbb{R}^{a_\ell} h ℓ : R a ℓ − 1 → R a ℓ is given by
h ℓ ( z ) = σ ( W ℓ z + b ℓ ) , z ∈ R a ℓ − 1 , (54) h_\ell(z)=\sigma(W_\ell z+b_\ell),\qquad z\in\mathbb{R}^{a_{\ell-1}}, \tag*{(54)} h ℓ ( z ) = σ ( W ℓ z + b ℓ ) , z ∈ R a ℓ − 1 , (54)
with W ℓ ∈ R a ℓ × a ℓ − 1 W_\ell\in\mathbb{R}^{a_\ell\times a_{\ell-1}} W ℓ ∈ R a ℓ × a ℓ − 1 , b ℓ ∈ R a ℓ b_\ell\in\mathbb{R}^{a_\ell} b ℓ ∈ R a ℓ , and σ : R → R \sigma:\mathbb{R}\to\mathbb{R} σ : R → R an activation function applied componentwise.
Neural approximation space
Since u s u_s u s is complex-valued, we extract complex basis functions from the real-valued hidden features of u ( ⋅ ; θ ) u(\cdot\,;\theta) u ( ⋅ ; θ ) by pairing consecutive neurons. For each layer ℓ = 1 , … , L \ell=1,\ldots,L ℓ = 1 , … , L and each index j = 1 , … , a ℓ / 2 j=1,\ldots,a_\ell/2 j = 1 , … , a ℓ /2 , two constructions are considered.
Standard basis. The basis functions are defined directly from the hidden features:
φ ℓ , j ( x ; θ ) = h 2 j − 1 ( ℓ ) ( x ; θ ) + i h 2 j ( ℓ ) ( x ; θ ) . (55) \varphi_{\ell,j}(x;\theta)=h_{2j-1}^{(\ell)}(x;\theta)+\mathrm{i}h_{2j}^{(\ell)}(x;\theta). \tag*{(55)} φ ℓ , j ( x ; θ ) = h 2 j − 1 ( ℓ ) ( x ; θ ) + i h 2 j ( ℓ ) ( x ; θ ) . (55)
Plane-wave basis. Given a unit direction d j ∈ R d d_j\in\mathbb{R}^d d j ∈ R d and wavenumber k j > 0 k_j>0 k j > 0 , the basis functions are
φ ℓ , j ( x ; θ ) = ( h 2 j − 1 ( ℓ ) ( x ; θ ) + i h 2 j ( ℓ ) ( x ; θ ) ) exp ( i k j d j ⋅ x ) . (56) \varphi_{\ell,j}(x;\theta)=\left(h_{2j-1}^{(\ell)}(x;\theta)+\mathrm{i}h_{2j}^{(\ell)}(x;\theta)\right)\exp(\mathrm{i}\,k_jd_j\cdot x). \tag*{(56)} φ ℓ , j ( x ; θ ) = ( h 2 j − 1 ( ℓ ) ( x ; θ ) + i h 2 j ( ℓ ) ( x ; θ ) ) exp ( i k j d j ⋅ x ) . (56)
For fixed θ ∈ Θ a \theta\in\Theta_a θ ∈ Θ a , the finite-dimensional neural space is
V a ( θ ) = span { φ ℓ , j ( ⋅ ; θ ) : ℓ = 1 , … , L , j = 1 , … , a ℓ 2 } . (57) V_a(\theta)=\operatorname{span}\left\{\varphi_{\ell,j}(\cdot\,;\theta):\ell=1,\ldots,L,\ j=1,\ldots,\frac{a_\ell}{2}\right\}. \tag*{(57)} V a ( θ ) = span { φ ℓ , j ( ⋅ ; θ ) : ℓ = 1 , … , L , j = 1 , … , 2 a ℓ } . (57)
which has dimension at most ∑ ℓ = 1 L a ℓ / 2 \sum_{\ell=1}^{L} a_{\ell}/2 ∑ ℓ = 1 L a ℓ /2 . As θ \theta θ ranges over Θ a \Theta_a Θ a , the union
V a = ⋃ θ ∈ Θ a V a ( θ ) (58) \mathcal{V}_a = \bigcup_{\theta\in\Theta_a} V_a(\theta) \tag*{(58)} V a = θ ∈ Θ a ⋃ V a ( θ ) (58)
forms the search space for a fixed architecture. The full approximation class over all admissible architectures is
V = ⋃ a ∈ A V a , (59) \mathcal{V} = \bigcup_{a\in\mathcal{A}} \mathcal{V}_a, \tag*{(59)} V = a ∈ A ⋃ V a , (59)
where A ⊂ ⋃ L ∈ N N L \mathcal{A} \subset\bigcup_{L\in\mathbb{N}} \mathbb{N}^{L} A ⊂ ⋃ L ∈ N N L denotes the set of all architecture vectors with even widths.
Proofs of theoretical results
Proof of Proposition 2.1
To establish a proof of Proposition 2.1, we first need to show the following lemma.
Lemma C.1 (Hilbert-Schmidt property of G \mathcal{G} G ). Let Ω ⊂ R d \Omega\subset\mathbb{R}^{d} Ω ⊂ R d , d ∈ { 2 , 3 } d \in\{2,3\} d ∈ { 2 , 3 } , be a bounded domain with δ m ∈ L ∞ ( Ω ) \delta m \in L^{\infty}(\Omega) δ m ∈ L ∞ ( Ω ) . Then G : L 2 ( Ω ) → L 2 ( Ω ) \mathcal{G}:L^{2}(\Omega)\to L^{2}(\Omega) G : L 2 ( Ω ) → L 2 ( Ω ) is a Hilbert-Schmidt operator with kernel K ( x , y ) = ω 2 g 0 ( x − y ) δ m ( y ) K(x,y)=\omega^{2}g_{0}(x-y)\delta m(y) K ( x , y ) = ω 2 g 0 ( x − y ) δ m ( y ) . In particular, G \mathcal{G} G is compact and bounded on L 2 ( Ω ) L^{2}(\Omega) L 2 ( Ω ) .
Proof. It suffices to show that K ∈ L 2 ( Ω × Ω ) K\in L^{2}(\Omega\times\Omega) K ∈ L 2 ( Ω × Ω ) . Using the definition of the kernel and the boundedness of the contrast function,
∬ Ω × Ω ∣ K ( x , y ) ∣ 2 d x d y ≤ ω 4 ∥ δ m ∥ L ∞ ( Ω ) 2 ∫ Ω ( ∫ Ω ∣ g 0 ( x − y ) ∣ 2 d x ) d y . (60) \iint_{\Omega\times\Omega} |K(x,y)|^{2}\,\mathrm{d}x\,\mathrm{d}y \leq\omega^{4}\|\delta m\|_{L^{\infty}(\Omega)}^{2}\int_{\Omega}\left(\int_{\Omega}|g_{0}(x-y)|^{2}\,\mathrm{d}x\right)\,\mathrm{d}y. \tag*{(60)} ∬ Ω × Ω ∣ K ( x , y ) ∣ 2 d x d y ≤ ω 4 ∥ δ m ∥ L ∞ ( Ω ) 2 ∫ Ω ( ∫ Ω ∣ g 0 ( x − y ) ∣ 2 d x ) d y . (60)
For any fixed y ∈ Ω y\in\Omega y ∈ Ω , the domain Ω \Omega Ω is contained in a closed ball B R ( y ) B_R(y) B R ( y ) of radius R : = sup x , x ′ ∈ Ω ∥ x − x ′ ∥ < ∞ R:=\sup_{x,x'\in\Omega}\|x-x'\|<\infty R := sup x , x ′ ∈ Ω ∥ x − x ′ ∥ < ∞ , so that
∫ Ω ∣ g 0 ( x − y ) ∣ 2 d x ≤ ∫ B R ( y ) ∣ g 0 ( x − y ) ∣ 2 d x = ∫ B R ( 0 ) ∣ g 0 ( z ) ∣ 2 d z , (61) \int_{\Omega}|g_{0}(x-y)|^{2}\,\mathrm{d}x \leq\int_{B_R(y)}|g_{0}(x-y)|^{2}\,\mathrm{d}x = \int_{B_R(0)}|g_{0}(z)|^{2}\,\mathrm{d}z, \tag*{(61)} ∫ Ω ∣ g 0 ( x − y ) ∣ 2 d x ≤ ∫ B R ( y ) ∣ g 0 ( x − y ) ∣ 2 d x = ∫ B R ( 0 ) ∣ g 0 ( z ) ∣ 2 d z , (61)
after the change of variables z = x − y z=x-y z = x − y . Passing to spherical coordinates (r = ∥ z ∥ r=\|z\| r = ∥ z ∥ , d z = r d − 1 d r d S d − 1 \mathrm{d}z=r^{d-1}\,\mathrm{d}r\,\mathrm{d}S_{d-1} d z = r d − 1 d r d S d − 1 , with d S d − 1 \mathrm{d}S_{d-1} d S d − 1 the surface measure on the unit sphere S d − 1 S^{d-1} S d − 1 ), and writing ∣ S d − 1 ∣ |S^{d-1}| ∣ S d − 1 ∣ for its total measure,
∫ B R ( 0 ) ∣ g 0 ( z ) ∣ 2 d z = ∣ S d − 1 ∣ ∫ 0 R ∣ g 0 ( r ) ∣ 2 r d − 1 d r , (62) \int_{B_R(0)}|g_{0}(z)|^{2}\,\mathrm{d}z = |S^{d-1}|\int_{0}^{R}|g_{0}(r)|^{2}r^{d-1}\,\mathrm{d}r, \tag*{(62)} ∫ B R ( 0 ) ∣ g 0 ( z ) ∣ 2 d z = ∣ S d − 1 ∣ ∫ 0 R ∣ g 0 ( r ) ∣ 2 r d − 1 d r , (62)
where, with a slight abuse of notation, g 0 ( r ) : = g 0 ( x ) g_{0}(r):=g_{0}(x) g 0 ( r ) := g 0 ( x ) for ∥ x ∥ = r \|x\|=r ∥ x ∥ = r , since g 0 g_{0} g 0 depends only on ∥ x ∥ \|x\| ∥ x ∥ in both cases considered. We split this radial integral into a singular neighborhood of the origin and a regular exterior part, for fixed ε ∈ ( 0 , R ) \varepsilon\in(0,R) ε ∈ ( 0 , R ) :
∫ 0 R ∣ g 0 ( r ) ∣ 2 r d − 1 d r = ∫ 0 ε ∣ g 0 ( r ) ∣ 2 r d − 1 d r + ∫ ε R ∣ g 0 ( r ) ∣ 2 r d − 1 d r . (63) \int_{0}^{R}|g_{0}(r)|^{2}r^{d-1}\,\mathrm{d}r = \int_{0}^{\varepsilon}|g_{0}(r)|^{2}r^{d-1}\,\mathrm{d}r+\int_{\varepsilon}^{R}|g_{0}(r)|^{2}r^{d-1}\,\mathrm{d}r. \tag*{(63)} ∫ 0 R ∣ g 0 ( r ) ∣ 2 r d − 1 d r = ∫ 0 ε ∣ g 0 ( r ) ∣ 2 r d − 1 d r + ∫ ε R ∣ g 0 ( r ) ∣ 2 r d − 1 d r . (63)
The second integral is finite since g 0 g_{0} g 0 is continuous (hence bounded) on the compact interval [ ε , R ] [\varepsilon,R] [ ε , R ] . It remains to check that the first integral is finite, which depends on the dimension through the singularity of g 0 g_{0} g 0 at the origin.
Case d = 2 d=2 d = 2 . By (2.4), g 0 ( r ) = i 4 H 0 ( 2 ) ( k 0 r ) g_{0}(r)=\frac{i}{4}H_{0}^{(2)}(k_{0}r) g 0 ( r ) = 4 i H 0 ( 2 ) ( k 0 r ) , and the asymptotic behavior of the zero-order Hankel function of the second kind near the origin satisfies ∣ H 0 ( 2 ) ( k 0 r ) ∣ 2 ∼ 4 π 2 ln 2 ( k 0 r ) |H_{0}^{(2)}(k_{0}r)|^{2}\sim\frac{4}{\pi^{2}}\ln^{2}(k_{0}r) ∣ H 0 ( 2 ) ( k 0 r ) ∣ 2 ∼ π 2 4 ln 2 ( k 0 r ) as r → 0 + r\to0^{+} r → 0 + . Hence,
∫ 0 ε ∣ g 0 ( r ) ∣ 2 r d r ∼ 1 4 π 2 ∫ 0 ε r ln 2 ( k 0 r ) d r . (64) \int_{0}^{\varepsilon}|g_{0}(r)|^{2}r\,\mathrm{d}r \sim\frac{1}{4\pi^{2}}\int_{0}^{\varepsilon}r\ln^{2}(k_{0}r)\,\mathrm{d}r. \tag*{(64)} ∫ 0 ε ∣ g 0 ( r ) ∣ 2 r d r ∼ 4 π 2 1 ∫ 0 ε r ln 2 ( k 0 r ) d r . (64)
Substituting u = k 0 r u=k_{0}r u = k 0 r , this scales as 1 4 π 2 k 0 2 ∫ 0 ε k 0 u ln 2 ( u ) d u \frac{1}{4\pi^{2}k_{0}^{2}}\int_{0}^{\varepsilon k_{0}}u\ln^{2}(u)\,\mathrm{d}u 4 π 2 k 0 2 1 ∫ 0 ε k 0 u ln 2 ( u ) d u . Since lim u → 0 + u 2 ln 2 ( u ) = 0 \lim_{u\to0^{+}}u^{2}\ln^{2}(u)=0 lim u → 0 + u 2 ln 2 ( u ) = 0 , the antiderivative
∫ 0 ε k 0 u ln 2 ( u ) d u = [ 1 2 u 2 ln 2 ( u ) − 1 2 u 2 ln ( u ) + 1 4 u 2 ] 0 ε k 0 < ∞ (65) \int_{0}^{\varepsilon k_{0}}u\ln^{2}(u)\,\mathrm{d}u = \left[\frac{1}{2}u^{2}\ln^{2}(u)-\frac{1}{2}u^{2}\ln(u)+\frac{1}{4}u^{2}\right]_{0}^{\varepsilon k_{0}}<\infty \tag*{(65)} ∫ 0 ε k 0 u ln 2 ( u ) d u = [ 2 1 u 2 ln 2 ( u ) − 2 1 u 2 ln ( u ) + 4 1 u 2 ] 0 ε k 0 < ∞ (65)
confirms the integral is bounded.
Case d = 3 d=3 d = 3 . By (2.4), g 0 ( r ) = e − i k 0 r 4 π r g_{0}(r)=\frac{e^{-ik_{0}r}}{4\pi r} g 0 ( r ) = 4 π r e − i k 0 r , so ∣ g 0 ( r ) ∣ 2 = 1 16 π 2 r 2 |g_{0}(r)|^{2}=\frac{1}{16\pi^{2}r^{2}} ∣ g 0 ( r ) ∣ 2 = 16 π 2 r 2 1 , and
∫ 0 ε ∣ g 0 ( r ) ∣ 2 r 2 d r = 1 16 π 2 ∫ 0 ε d r = ε 16 π 2 < ∞ . (66) \int_{0}^{\varepsilon}|g_{0}(r)|^{2}r^{2}\,\mathrm{d}r = \frac{1}{16\pi^{2}}\int_{0}^{\varepsilon}\mathrm{d}r = \frac{\varepsilon}{16\pi^{2}}<\infty. \tag*{(66)} ∫ 0 ε ∣ g 0 ( r ) ∣ 2 r 2 d r = 16 π 2 1 ∫ 0 ε d r = 16 π 2 ε < ∞. (66)
In this case the apparent singularity of g 0 g_0 g 0 at the origin is entirely compensated by the volume factor r d − 1 = r 2 r^{d-1}=r^2 r d − 1 = r 2 , so no logarithmic correction is needed.
In either case, ∫ 0 R ∣ g 0 ( r ) ∣ r 2 d − 1 d r = : C ( k 0 , R ) < ∞ \int_0^R |g_0(r)|r^{2d-1}\,\mathrm{d}r=:C(k_0,R)<\infty ∫ 0 R ∣ g 0 ( r ) ∣ r 2 d − 1 d r =: C ( k 0 , R ) < ∞ , and substituting back into (C.1) gives
∥ K ∥ L 2 ( Ω × Ω ) 2 ≤ ω 4 ∥ δ m ∥ L ∞ ( Ω ) 2 ∣ S d − 1 ∣ C ( k 0 , R ) ∣ Ω ∣ < ∞ , (67) \lVert K\rVert_{L^2(\Omega\times\Omega)}^2\leq\omega^4\lVert\delta m\rVert_{L^\infty(\Omega)}^2\lvert S^{d-1}\rvert C(k_0,R)\lvert\Omega\rvert<\infty, \tag*{(67)} ∥ K ∥ L 2 ( Ω × Ω ) 2 ≤ ω 4 ∥ δ m ∥ L ∞ ( Ω ) 2 ∣ S d − 1 ∣ C ( k 0 , R ) ∣ Ω ∣ < ∞ , (67)
where ∣ Ω ∣ \lvert\Omega\rvert ∣ Ω ∣ denotes the Lebesgue measure of Ω \Omega Ω . Consequently K ∈ L 2 ( Ω × Ω ) K\in L^2(\Omega\times\Omega) K ∈ L 2 ( Ω × Ω ) , so G \mathcal{G} G is a Hilbert-Schmidt operator, and in particular compact and bounded on L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) .
We now establish the proof of Proposition 2.1 .
By Lemma C.1 , the operator G : L 2 ( Ω ) → L 2 ( Ω ) \mathcal{G}:L^2(\Omega)\to L^2(\Omega) G : L 2 ( Ω ) → L 2 ( Ω ) is Hilbert-Schmidt, hence compact. Therefore B = I − G \mathcal{B}=\mathcal{I}-\mathcal{G} B = I − G is a Fredholm operator of index zero. It follows that B \mathcal{B} B is invertible if and only if it is injective. To show injectivity, suppose B u = 0 \mathcal{B}u=0 B u = 0 for some u ∈ L 2 ( Ω ) u\in L^2(\Omega) u ∈ L 2 ( Ω ) , i.e., u = G u u=\mathcal{G}u u = G u . Define the volume potential
w ( x ) = ω 2 ∫ Ω g 0 ( x − y ) δ m ( y ) u ( y ) d y , x ∈ R d . (68) w(x)=\omega^2\int_{\Omega}g_0(x-y)\,\delta m(y)\,u(y)\,\mathrm{d}y,\qquad x\in\mathbb{R}^d. \tag*{(68)} w ( x ) = ω 2 ∫ Ω g 0 ( x − y ) δ m ( y ) u ( y ) d y , x ∈ R d . (68)
Set φ = ω 2 δ m ⋅ u ∈ L 2 ( Ω ) \varphi=\omega^2\delta m\cdot u\in L^2(\Omega) φ = ω 2 δ m ⋅ u ∈ L 2 ( Ω ) . By Theorem 7.11 in [20 ] , w ∈ H l o c 2 ( R d ) w\in H^2_{\mathrm{loc}}(\mathbb{R}^d) w ∈ H loc 2 ( R d ) , w w w satisfies the Sommerfeld radiation condition, and w w w is the unique radiating solution of Δ w + k 0 2 w = − φ \Delta w+k_0^2w=-\varphi Δ w + k 0 2 w = − φ in R d \mathbb{R}^d R d . Since φ = ω 2 δ m ⋅ u \varphi=\omega^2\delta m\cdot u φ = ω 2 δ m ⋅ u is supported in Ω \Omega Ω , we have φ ≡ 0 \varphi\equiv0 φ ≡ 0 in R d ∖ Ω ‾ \mathbb{R}^d\setminus\overline{\Omega} R d ∖ Ω , and therefore
Δ w + k 0 2 w = 0 in R d ∖ Ω ‾ . (69) \Delta w+k_0^2w=0\qquad\text{in }\mathbb{R}^d\setminus\overline{\Omega}. \tag*{(69)} Δ w + k 0 2 w = 0 in R d ∖ Ω . (69)
That is, w w w is a radiating solution of the homogeneous Helmholtz equation in the exterior of Ω ‾ \overline{\Omega} Ω . The Sommerfeld radiation condition implies that lim R → ∞ ∫ ∥ x ∥ = R ∣ w ∣ 2 d s = 0 \lim_{R\to\infty}\int_{\lVert x\rVert=R}|w|^2\,\mathrm{d}s=0 lim R → ∞ ∫ ∥ x ∥ = R ∣ w ∣ 2 d s = 0 (see Proof of Theorem 7.9 in [20 ] ). By Rellich’s lemma (Theorem 7.5 in [20 ] ),
w ( x ) = 0 for all ∥ x ∥ ≥ b , (70) w(x)=0\qquad\text{for all }\lVert x\rVert\geq b, \tag*{(70)} w ( x ) = 0 for all ∥ x ∥ ≥ b , (70)
where b > 0 b>0 b > 0 is chosen so that Ω ‾ ⊂ B ( 0 , b ) \overline{\Omega}\subset B(0,b) Ω ⊂ B ( 0 , b ) .
The restriction of w w w to Ω \Omega Ω satisfies w ∣ Ω = G u = u w|_{\Omega}=\mathcal{G}u=u w ∣ Ω = G u = u . Substituting u = w ∣ Ω u=w|_{\Omega} u = w ∣ Ω into the equation Δ w + k 0 2 w = − ω 2 δ m ⋅ u \Delta w+k_0^2w=-\omega^2\delta m\cdot u Δ w + k 0 2 w = − ω 2 δ m ⋅ u on Ω \Omega Ω gives Δ w + k 0 2 w + ω 2 δ m ⋅ w = 0 \Delta w+k_0^2w+\omega^2\delta m\cdot w=0 Δ w + k 0 2 w + ω 2 δ m ⋅ w = 0 on Ω \Omega Ω , which can be written as Δ w + k 0 2 n ( x ) w = 0 \Delta w+k_0^2n(x)w=0 Δ w + k 0 2 n ( x ) w = 0 , where n ( x ) = 1 + ω 2 δ m ( x ) / k 0 2 n(x)=1+\omega^2\delta m(x)/k_0^2 n ( x ) = 1 + ω 2 δ m ( x ) / k 0 2 . Since δ m ≡ 0 \delta m\equiv0 δ m ≡ 0 outside Ω \Omega Ω , we have n ( x ) = 1 n(x)=1 n ( x ) = 1 for x ∈ R d ∖ Ω x\in\mathbb{R}^d\setminus\Omega x ∈ R d ∖ Ω , and therefore w w w satisfies Δ w + k 0 2 n ( x ) w = 0 \Delta w+k_0^2n(x)w=0 Δ w + k 0 2 n ( x ) w = 0 in all of R d \mathbb{R}^d R d with n ∈ L ∞ ( R d ) n\in L^\infty(\mathbb{R}^d) n ∈ L ∞ ( R d ) . By (C.11), w w w has compact support, so w ∈ H 2 ( R d ) w\in H^2(\mathbb{R}^d) w ∈ H 2 ( R d ) . The hypotheses of Theorem 7.7 in [20 ] are therefore satisfied, and the conclusion of that theorem gives w ≡ 0 w\equiv0 w ≡ 0 on R d \mathbb{R}^d R d . In particular, u = w ∣ Ω = 0 u=w|_{\Omega}=0 u = w ∣ Ω = 0 , which establishes the injectivity of B \mathcal{B} B .
Proof of Theorem 3.1
We first expand J J J into its quadratic form. For any v ∈ L 2 ( Ω ) v\in L^2(\Omega) v ∈ L 2 ( Ω ) ,
J ( v ) = 1 2 ⟨ B ∗ B v , v ⟩ L 2 ( Ω ) − Re ⟨ v , B ∗ f ⟩ L 2 ( Ω ) + 1 2 ∥ f ∥ L 2 ( Ω ) 2 , (71) J(v)=\frac{1}{2}\langle\mathcal{B}^*\mathcal{B}v,v\rangle_{L^2(\Omega)}-\operatorname{Re}\langle v,\mathcal{B}^*f\rangle_{L^2(\Omega)}+\frac{1}{2}\lVert f\rVert_{L^2(\Omega)}^2, \tag*{(71)} J ( v ) = 2 1 ⟨ B ∗ B v , v ⟩ L 2 ( Ω ) − Re ⟨ v , B ∗ f ⟩ L 2 ( Ω ) + 2 1 ∥ f ∥ L 2 ( Ω ) 2 , (71)
where B ∗ \mathcal{B}^* B ∗ denotes the Hilbert-space adjoint of B \mathcal{B} B . The Hessian of J J J is therefore B ∗ B \mathcal{B}^*\mathcal{B} B ∗ B . Since B \mathcal{B} B is invertible (Proposition 2.1 ), it has trivial kernel, so B ∗ B \mathcal{B}^*\mathcal{B} B ∗ B is strictly positive definite:
⟨ B ∗ B v , v ⟩ L 2 ( Ω ) = ∥ B v ∥ L 2 ( Ω ) 2 > 0 ∀ v ≠ 0. (72) \langle\mathcal{B}^*\mathcal{B}v,v\rangle_{L^2(\Omega)}=\lVert\mathcal{B}v\rVert_{L^2(\Omega)}^2>0\qquad\forall\,v\neq0. \tag*{(72)} ⟨ B ∗ B v , v ⟩ L 2 ( Ω ) = ∥ B v ∥ L 2 ( Ω ) 2 > 0 ∀ v = 0. (72)
Hence J J J is strictly convex.
It remains to identify the minimizer. By Proposition 2.1 , there exists a unique u s ∈ L 2 ( Ω ) u_s\in L^2(\Omega) u s ∈ L 2 ( Ω ) satisfying B u s = f \mathcal{B}u_s=f B u s = f , and evaluating (3.1) at u s u_s u s gives J ( u s ) = 0 J(u_s)=0 J ( u s ) = 0 . Since J ( v ) ≥ 0 J(v)\geq0 J ( v ) ≥ 0 for all v ∈ L 2 ( Ω ) v\in L^2(\Omega) v ∈ L 2 ( Ω ) , the element u s u_s u s is a global minimizer. Uniqueness follows from strict convexity: if v ∈ L 2 ( Ω ) v\in L^2(\Omega) v ∈ L 2 ( Ω ) also satisfies J ( v ) = 0 J(v)=0 J ( v ) = 0 , then ∥ B v − f ∥ L 2 ( Ω ) = 0 \lVert\mathcal{B}v-f\rVert_{L^2(\Omega)}=0 ∥ B v − f ∥ L 2 ( Ω ) = 0 , hence B v = f \mathcal{B}v=f B v = f , and the injectivity of B \mathcal{B} B forces v = u s v=u_s v = u s . Furthermore, because B \mathcal{B} B is a Fredholm operator, its range Im ( B ) \operatorname{Im}(\mathcal{B}) Im ( B ) is a closed subspace of L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) . This topological closure guarantees that the residual function J J J defined in (3.1) possesses a unique global minimum, which corresponds exactly to the unique solution of the Lippmann-Schwinger equation.
Proof of Theorem 3.2
Since B : L 2 ( Ω ) → L 2 ( Ω ) B : L^2(\Omega) \to L^2(\Omega) B : L 2 ( Ω ) → L 2 ( Ω ) is bounded, linear, and invertible, the open mapping theorem guarantees that its inverse B − 1 : L 2 ( Ω ) → L 2 ( Ω ) B^{-1} : L^2(\Omega) \to L^2(\Omega) B − 1 : L 2 ( Ω ) → L 2 ( Ω ) is also linear and bounded. We first observe that the residual and the error are related through B B B by
r ( v ) = f − B v = B u − B v = B ( u − v ) = B e ( v ) . (73) r(v) = f - Bv = Bu - Bv = B(u-v) = Be(v). \tag*{(73)} r ( v ) = f − B v = B u − B v = B ( u − v ) = B e ( v ) . (73)
Upper bound. Applying B − 1 B^{-1} B − 1 to both sides of (C.14) and using the boundedness of B − 1 B^{-1} B − 1 ,
∥ e ( v ) ∥ L 2 ( Ω ) = ∥ B − 1 r ( v ) ∥ L 2 ( Ω ) ≤ ∥ B − 1 ∥ L ( L 2 ( Ω ) ) ∥ r ( v ) ∥ L 2 ( Ω ) . (74) \lVert e(v)\rVert_{L^2(\Omega)} = \lVert B^{-1}r(v)\rVert_{L^2(\Omega)} \leq\lVert B^{-1}\rVert_{\mathcal{L}(L^2(\Omega))}\lVert r(v)\rVert_{L^2(\Omega)}. \tag*{(74)} ∥ e ( v ) ∥ L 2 ( Ω ) = ∥ B − 1 r ( v ) ∥ L 2 ( Ω ) ≤ ∥ B − 1 ∥ L ( L 2 ( Ω )) ∥ r ( v ) ∥ L 2 ( Ω ) . (74)
Lower bound. From (C.14) and the boundedness of B B B ,
∥ r ( v ) ∥ L 2 ( Ω ) = ∥ B e ( v ) ∥ L 2 ( Ω ) ≤ ∥ B ∥ L ( L 2 ( Ω ) ) ∥ e ( v ) ∥ L 2 ( Ω ) . (75) \lVert r(v)\rVert_{L^2(\Omega)} = \lVert Be(v)\rVert_{L^2(\Omega)} \leq\lVert B\rVert_{\mathcal{L}(L^2(\Omega))}\lVert e(v)\rVert_{L^2(\Omega)}. \tag*{(75)} ∥ r ( v ) ∥ L 2 ( Ω ) = ∥ B e ( v ) ∥ L 2 ( Ω ) ≤ ∥ B ∥ L ( L 2 ( Ω )) ∥ e ( v ) ∥ L 2 ( Ω ) . (75)
Since B B B is invertible, ∥ B ∥ L ( L 2 ( Ω ) ) > 0 \lVert B\rVert_{\mathcal{L}(L^2(\Omega))} > 0 ∥ B ∥ L ( L 2 ( Ω )) > 0 ; dividing yields
∥ B ∥ L ( L 2 ( Ω ) ) − 1 ∥ r ( v ) ∥ L 2 ( Ω ) ≤ ∥ e ( v ) ∥ L 2 ( Ω ) . (76) \lVert B\rVert_{\mathcal{L}(L^2(\Omega))}^{-1}\lVert r(v)\rVert_{L^2(\Omega)} \leq\lVert e(v)\rVert_{L^2(\Omega)}. \tag*{(76)} ∥ B ∥ L ( L 2 ( Ω )) − 1 ∥ r ( v ) ∥ L 2 ( Ω ) ≤ ∥ e ( v ) ∥ L 2 ( Ω ) . (76)
Combining the two inequalities, dividing by ∥ u ∥ L 2 ( Ω ) \lVert u\rVert_{L^2(\Omega)} ∥ u ∥ L 2 ( Ω ) , and using ∥ r ( v ) ∥ L 2 ( Ω ) = 2 J ( v ) \lVert r(v)\rVert_{L^2(\Omega)} = \sqrt{2J(v)} ∥ r ( v ) ∥ L 2 ( Ω ) = 2 J ( v ) gives (3.2) with
c 1 = 2 ∥ B ∥ L ( L 2 ( Ω ) ) ∥ u ∥ L 2 ( Ω ) , c 2 = 2 ∥ B − 1 ∥ L ( L 2 ( Ω ) ) ∥ u ∥ L 2 ( Ω ) . (77) c_1 = \frac{\sqrt{2}}{\lVert B\rVert_{\mathcal{L}(L^2(\Omega))}\lVert u\rVert_{L^2(\Omega)}}, \qquad c_2 = \frac{\sqrt{2}\lVert B^{-1}\rVert_{\mathcal{L}(L^2(\Omega))}}{\lVert u\rVert_{L^2(\Omega)}}. \tag*{(77)} c 1 = ∥ B ∥ L ( L 2 ( Ω )) ∥ u ∥ L 2 ( Ω ) 2 , c 2 = ∥ u ∥ L 2 ( Ω ) 2 ∥ B − 1 ∥ L ( L 2 ( Ω )) . (77)
Proof of Theorem 3.3
We define the space of one-hidden-layer networks as:
M ( σ ) = { f ( x ) = ∑ j = 1 N c j σ ( w j ⋅ x + b j ) : N ∈ N , c j ∈ C , w j ∈ R d , b j ∈ R } . (78) \mathcal{M}(\sigma) = \left\{ f(x) = \sum_{j=1}^{N} c_j \sigma(w_j \cdot x + b_j) : N \in\mathbb{N}, c_j \in\mathbb{C}, w_j \in\mathbb{R}^d, b_j \in\mathbb{R} \right\}. \tag*{(78)} M ( σ ) = { f ( x ) = j = 1 ∑ N c j σ ( w j ⋅ x + b j ) : N ∈ N , c j ∈ C , w j ∈ R d , b j ∈ R } . (78)
We start with the case of standard neural basis functions. Since M ( σ ) \mathcal{M}(\sigma) M ( σ ) is dense in L p ( Ω ) L^p(\Omega) L p ( Ω ) (refer, e.g., to [26 ] Corollary 3.19), the density of V \mathcal{V} V follows from the fact that M ( σ ) ⊂ V \mathcal{M}(\sigma) \subset\mathcal{V} M ( σ ) ⊂ V . Indeed, let f ∈ M ( σ ) f \in\mathcal{M}(\sigma) f ∈ M ( σ ) be an arbitrary function with N N N neurons, f ( x ) = ∑ j = 1 N c j σ ( w j ⋅ x + b j ) f(x) = \sum_{j=1}^{N} c_j \sigma(w_j \cdot x + b_j) f ( x ) = ∑ j = 1 N c j σ ( w j ⋅ x + b j ) . We define the architecture d = ( d 0 , 2 N ) d = (d_0, 2N) d = ( d 0 , 2 N ) . For each neuron j ∈ { 1 , … , N } j \in\{1,\ldots,N\} j ∈ { 1 , … , N } , we configure the parameters θ ∈ Θ d \theta\in\Theta_d θ ∈ Θ d such that the hidden layer features are identical for the pair ( 2 j − 1 , 2 j ) (2j-1,2j) ( 2 j − 1 , 2 j ) :
h 2 j − 1 ( 1 ) ( x ; θ ) = σ ( w j ⋅ x + b j ) , h 2 j ( 1 ) ( x ; θ ) = σ ( w j ⋅ x + b j ) . (79) h_{2j-1}^{(1)}(x;\theta) = \sigma(w_j \cdot x + b_j), \qquad h_{2j}^{(1)}(x;\theta) = \sigma(w_j \cdot x + b_j). \tag*{(79)} h 2 j − 1 ( 1 ) ( x ; θ ) = σ ( w j ⋅ x + b j ) , h 2 j ( 1 ) ( x ; θ ) = σ ( w j ⋅ x + b j ) . (79)
The basis function for this pair is then:
φ 1 , j ( x ; θ ) = h 2 j − 1 ( 1 ) ( x ; θ ) + i h 2 j ( 1 ) ( x ; θ ) = ( 1 + i ) σ ( w j ⋅ x + b j ) . (80) \varphi_{1,j}(x;\theta) = h_{2j-1}^{(1)}(x;\theta) + i h_{2j}^{(1)}(x;\theta) = (1+i)\sigma(w_j \cdot x+b_j). \tag*{(80)} φ 1 , j ( x ; θ ) = h 2 j − 1 ( 1 ) ( x ; θ ) + i h 2 j ( 1 ) ( x ; θ ) = ( 1 + i ) σ ( w j ⋅ x + b j ) . (80)
Defining coefficients α j = c j 1 + i \alpha_j = \frac{c_j}{1+i} α j = 1 + i c j , we get:
∑ j = 1 N α j φ 1 , j ( x ; θ ) = ∑ j = 1 N α j ( 1 + i ) σ ( w j ⋅ x + b j ) = ∑ j = 1 N c j σ ( w j ⋅ x + b j ) = f ( x ) . (81) \sum_{j=1}^{N} \alpha_j\varphi_{1,j}(x;\theta) = \sum_{j=1}^{N} \alpha_j(1+i)\sigma(w_j \cdot x+b_j) = \sum_{j=1}^{N} c_j\sigma(w_j \cdot x+b_j) = f(x). \tag*{(81)} j = 1 ∑ N α j φ 1 , j ( x ; θ ) = j = 1 ∑ N α j ( 1 + i ) σ ( w j ⋅ x + b j ) = j = 1 ∑ N c j σ ( w j ⋅ x + b j ) = f ( x ) . (81)
Thus, M ( σ ) ⊆ V d ( θ ) ⊂ V \mathcal{M}(\sigma) \subseteq\mathcal{V}_d(\theta) \subset\mathcal{V} M ( σ ) ⊆ V d ( θ ) ⊂ V .
We now show that the space V w \mathcal{V}_w V w , generated by the plane-wave neural basis functions, is dense in L p ( Ω ) L^p(\Omega) L p ( Ω ) for all p ∈ [ 1 , ∞ ) p \in[1,\infty) p ∈ [ 1 , ∞ ) . Let f ∈ L p ( Ω ) f \in L^p(\Omega) f ∈ L p ( Ω ) and let ε > 0 \varepsilon> 0 ε > 0 . Since C ( Ω ‾ ) C(\overline{\Omega}) C ( Ω ) is dense in L p ( Ω ) L^p(\Omega) L p ( Ω ) , there exists a function g ∈ C ( Ω ‾ ) g \in C(\overline{\Omega}) g ∈ C ( Ω ) such that: ∥ f − g ∥ L p ( Ω ) < ε 2 \lVert f-g\rVert_{L^p(\Omega)} < \frac{\varepsilon}{2} ∥ f − g ∥ L p ( Ω ) < 2 ε . For a given wavenumber k > 0 k>0 k > 0 and direction vector d ∈ S 1 d \in S^1 d ∈ S 1 , we define the function g ~ ( x ) = g ( x ) e − i k d ⋅ x ∈ C ( Ω ‾ ) \widetilde{g}(x) = g(x)e^{-ikd\cdot x} \in C(\overline{\Omega}) g ( x ) = g ( x ) e − ik d ⋅ x ∈ C ( Ω ) . Because M ( σ ) \mathcal{M}(\sigma) M ( σ ) is dense in C ( Ω ‾ ) C(\overline{\Omega}) C ( Ω ) , there exists φ ∈ V \varphi\in\mathcal{V} φ ∈ V such that: ∥ g ~ − φ ∥ L ∞ ( Ω ) < ε 2 ∣ Ω ∣ 1 / p \lVert\widetilde{g}-\varphi\rVert_{L^\infty(\Omega)} < \frac{\varepsilon}{2|\Omega|^{1/p}} ∥ g − φ ∥ L ∞ ( Ω ) < 2∣Ω ∣ 1/ p ε , where ∣ Ω ∣ |\Omega| ∣Ω∣ denotes the Lebesgue measure of the domain. Let h ( x ) = φ ( x ) e i k d ⋅ x h(x) = \varphi(x)e^{ikd\cdot x} h ( x ) = φ ( x ) e ik d ⋅ x . By definition, h ∈ V w h \in\mathcal{V}_w h ∈ V w . Notice that ∣ e i k d ⋅ x ∣ = 1 \lvert e^{ikd\cdot x}\rvert= 1 ∣ e ik d ⋅ x ∣ = 1 . Consequently:
∥ g − h ∥ L p ( Ω ) p = ∫ Ω ∣ g ~ ( x ) − φ ( x ) ∣ p d x ≤ ∣ Ω ∣ ∥ g ~ − φ ∥ L ∞ ( Ω ) p < ∣ Ω ∣ ( ε 2 ∣ Ω ∣ 1 / p ) p = ( ε 2 ) p . (82) \lVert g-h\rVert_{L^p(\Omega)}^p = \int_{\Omega} \lvert\widetilde{g}(x)-\varphi(x)\rvert^p\,dx \leq|\Omega|\lVert\widetilde{g}-\varphi\rVert_{L^\infty(\Omega)}^p < |\Omega|\left(\frac{\varepsilon}{2|\Omega|^{1/p}}\right)^p = \left(\frac{\varepsilon}{2}\right)^p. \tag*{(82)} ∥ g − h ∥ L p ( Ω ) p = ∫ Ω ∣ g ( x ) − φ ( x ) ∣ p d x ≤ ∣Ω∣ ∥ g − φ ∥ L ∞ ( Ω ) p < ∣Ω∣ ( 2∣Ω ∣ 1/ p ε ) p = ( 2 ε ) p . (82)
Finally, applying the triangle inequality, we get:
∥ f − h ∥ L p ( Ω ) ≤ ∥ f − g ∥ L p ( Ω ) + ∥ g − h ∥ L p ( Ω ) < ε . (83) \lVert f-h\rVert_{L^p(\Omega)} \leq\lVert f-g\rVert_{L^p(\Omega)} + \lVert g-h\rVert_{L^p(\Omega)} < \varepsilon. \tag*{(83)} ∥ f − h ∥ L p ( Ω ) ≤ ∥ f − g ∥ L p ( Ω ) + ∥ g − h ∥ L p ( Ω ) < ε . (83)
Thus, V w \mathcal{V}_w V w is dense in L p ( Ω ) L^p(\Omega) L p ( Ω ) .
Proof of Proposition 3.4
Fix μ > 0 \mu> 0 μ > 0 and θ ∈ R p \theta\in\mathbb{R}^{p} θ ∈ R p . The operator ( B Φ θ ) ∗ B Φ θ + μ I (\mathcal{B}\Phi_{\theta})^{*}\mathcal{B}\Phi_{\theta}+\mu I ( B Φ θ ) ∗ B Φ θ + μ I is positive definite and in particular injective. Since it acts on the finite-dimensional space C N \mathbb{C}^{N} C N , it is invertible, and therefore the map c μ ( θ ) = ( ( B Φ θ ) ∗ B Φ θ + μ I ) − 1 ( B Φ θ ) ∗ f c_{\mu}(\theta)=((\mathcal{B}\Phi_{\theta})^{*}\mathcal{B}\Phi_{\theta}+\mu I)^{-1}(\mathcal{B}\Phi_{\theta})^{*}f c μ ( θ ) = (( B Φ θ ) ∗ B Φ θ + μ I ) − 1 ( B Φ θ ) ∗ f is well defined. Since ( B Φ θ ) ∗ B Φ θ : C N → C N (\mathcal{B}\Phi_{\theta})^{*}\mathcal{B}\Phi_{\theta}:\mathbb{C}^{N}\to\mathbb{C}^{N} ( B Φ θ ) ∗ B Φ θ : C N → C N , it is identified with a matrix in C N × N \mathbb{C}^{N\times N} C N × N , denoted again by ( B Φ θ ) ∗ B Φ θ (\mathcal{B}\Phi_{\theta})^{*}\mathcal{B}\Phi_{\theta} ( B Φ θ ) ∗ B Φ θ . By definition of B Φ θ \mathcal{B}\Phi_{\theta} B Φ θ , for any c ∈ C N c\in\mathbb{C}^{N} c ∈ C N and g ∈ L 2 ( Ω ) g\in L^{2}(\Omega) g ∈ L 2 ( Ω ) ,
⟨ B Φ θ c , g ⟩ L 2 ( Ω ) = ∑ j = 1 N c j ⟨ B φ j ( ⋅ ; θ ) , g ⟩ L 2 ( Ω ) . (84) \langle\mathcal{B}\Phi_{\theta}c,g\rangle_{L^{2}(\Omega)}=\sum_{j=1}^{N}c_{j}\langle\mathcal{B}\varphi_{j}(\cdot\,;\theta),g\rangle_{L^{2}(\Omega)}. \tag*{(84)} ⟨ B Φ θ c , g ⟩ L 2 ( Ω ) = j = 1 ∑ N c j ⟨ B φ j ( ⋅ ; θ ) , g ⟩ L 2 ( Ω ) . (84)
Since ⟨ B Φ θ c , g ⟩ L 2 ( Ω ) = ⟨ c , ( B Φ θ ) ∗ g ⟩ C N \langle\mathcal{B}\Phi_{\theta}c,g\rangle_{L^{2}(\Omega)}=\langle c,(\mathcal{B}\Phi_{\theta})^{*}g\rangle_{\mathbb{C}^{N}} ⟨ B Φ θ c , g ⟩ L 2 ( Ω ) = ⟨ c , ( B Φ θ ) ∗ g ⟩ C N for all c ∈ C N c\in\mathbb{C}^{N} c ∈ C N , the j j j -th component of ( B Φ θ ) ∗ g (\mathcal{B}\Phi_{\theta})^{*}g ( B Φ θ ) ∗ g is
( ( B Φ θ ) ∗ g ) j = ⟨ g , B φ j ( ⋅ ; θ ) ⟩ L 2 ( Ω ) , j = 1 , … , N . (85) ((\mathcal{B}\Phi_{\theta})^{*}g)_{j}=\langle g,\mathcal{B}\varphi_{j}(\cdot\,;\theta)\rangle_{L^{2}(\Omega)},\qquad j=1,\ldots,N. \tag*{(85)} (( B Φ θ ) ∗ g ) j = ⟨ g , B φ j ( ⋅ ; θ ) ⟩ L 2 ( Ω ) , j = 1 , … , N . (85)
Taking g = B Φ θ e i g=\mathcal{B}\Phi_{\theta}e_{i} g = B Φ θ e i , where e i e_{i} e i is the i i i -th canonical basis vector of C N \mathbb{C}^{N} C N , gives the ( i , j ) (i,j) ( i , j ) -entry of ( B Φ θ ) ∗ B Φ θ (\mathcal{B}\Phi_{\theta})^{*}\mathcal{B}\Phi_{\theta} ( B Φ θ ) ∗ B Φ θ :
( ( B Φ θ ) ∗ B Φ θ ) i j = ⟨ B φ i ( ⋅ ; θ ) , B φ j ( ⋅ ; θ ) ⟩ L 2 ( Ω ) , i , j = 1 , … , N . (86) ((\mathcal{B}\Phi_{\theta})^{*}\mathcal{B}\Phi_{\theta})_{ij}=\langle\mathcal{B}\varphi_{i}(\cdot\,;\theta),\mathcal{B}\varphi_{j}(\cdot\,;\theta)\rangle_{L^{2}(\Omega)},\qquad i,j=1,\ldots,N. \tag*{(86)} (( B Φ θ ) ∗ B Φ θ ) ij = ⟨ B φ i ( ⋅ ; θ ) , B φ j ( ⋅ ; θ ) ⟩ L 2 ( Ω ) , i , j = 1 , … , N . (86)
Taking g = f g=f g = f gives the i i i -th entry of ( B Φ θ ) ∗ f (\mathcal{B}\Phi_{\theta})^{*}f ( B Φ θ ) ∗ f :
( ( B Φ θ ) ∗ f ) i = ⟨ f , B φ i ( ⋅ ; θ ) ⟩ L 2 ( Ω ) , i = 1 , … , N . (87) ((\mathcal{B}\Phi_{\theta})^{*}f)_{i}=\langle f,\mathcal{B}\varphi_{i}(\cdot\,;\theta)\rangle_{L^{2}(\Omega)},\qquad i=1,\ldots,N. \tag*{(87)} (( B Φ θ ) ∗ f ) i = ⟨ f , B φ i ( ⋅ ; θ ) ⟩ L 2 ( Ω ) , i = 1 , … , N . (87)
Since each φ j ( ⋅ ; θ ) \varphi_{j}(\cdot\,;\theta) φ j ( ⋅ ; θ ) is continuously differentiable in θ \theta θ by assumption, and B \mathcal{B} B is a bounded linear operator independent of θ \theta θ , the maps θ ↦ B φ j ( ⋅ ; θ ) \theta\mapsto\mathcal{B}\varphi_{j}(\cdot\,;\theta) θ ↦ B φ j ( ⋅ ; θ ) are continuously differentiable. Therefore, by (C.27) and (C.28), the entries of ( B Φ θ ) ∗ B Φ θ (\mathcal{B}\Phi_{\theta})^{*}\mathcal{B}\Phi_{\theta} ( B Φ θ ) ∗ B Φ θ and ( B Φ θ ) ∗ f (\mathcal{B}\Phi_{\theta})^{*}f ( B Φ θ ) ∗ f are continuously differentiable functions of θ \theta θ , and the map θ ↦ ( B Φ θ ) ∗ B Φ θ + μ I \theta\mapsto(\mathcal{B}\Phi_{\theta})^{*}\mathcal{B}\Phi_{\theta}+\mu I θ ↦ ( B Φ θ ) ∗ B Φ θ + μ I is continuously differentiable for all θ ∈ R p \theta\in\mathbb{R}^{p} θ ∈ R p . Moreover, the matrix inversion map M ↦ M − 1 M\mapsto M^{-1} M ↦ M − 1 is continuously differentiable on the open set of invertible matrices in C N × N \mathbb{C}^{N\times N} C N × N , and since ( B Φ θ ) ∗ B Φ θ + μ I (\mathcal{B}\Phi_{\theta})^{*}\mathcal{B}\Phi_{\theta}+\mu I ( B Φ θ ) ∗ B Φ θ + μ I is invertible for every θ ∈ R p \theta\in\mathbb{R}^{p} θ ∈ R p , the map θ ↦ ( ( B Φ θ ) ∗ B Φ θ + μ I ) − 1 \theta\mapsto((\mathcal{B}\Phi_{\theta})^{*}\mathcal{B}\Phi_{\theta}+\mu I)^{-1} θ ↦ (( B Φ θ ) ∗ B Φ θ + μ I ) − 1 is continuously differentiable by the chain rule. Therefore c μ ( θ ) = ( ( B Φ θ ) ∗ B Φ θ + μ I ) − 1 ( B Φ θ ) ∗ f c_{\mu}(\theta)=((\mathcal{B}\Phi_{\theta})^{*}\mathcal{B}\Phi_{\theta}+\mu I)^{-1}(\mathcal{B}\Phi_{\theta})^{*}f c μ ( θ ) = (( B Φ θ ) ∗ B Φ θ + μ I ) − 1 ( B Φ θ ) ∗ f is continuously differentiable with respect to θ \theta θ , as a product of continuously differentiable maps.
Proof of Proposition 3.5
Define g : R p → R p × C N g:\mathbb{R}^{p}\to\mathbb{R}^{p}\times\mathbb{C}^{N} g : R p → R p × C N by g ( θ ) : = ( θ , c μ ( θ ) ) g(\theta):=(\theta,c_{\mu}(\theta)) g ( θ ) := ( θ , c μ ( θ )) , then J μ = J μ ∘ g \mathcal{J}_{\mu}=\mathcal{J}_{\mu}\circ g J μ = J μ ∘ g . By Proposition 3.4, c μ c_{\mu} c μ is continuously differentiable, and therefore g g g is continuously differentiable. Since J μ \mathcal{J}_{\mu} J μ is also differentiable, the chain rule gives
Jac J μ ( θ ) = Jac J μ ( g ( θ ) ) Jac g ( θ ) . (88) \operatorname{Jac}\mathcal{J}_{\mu}(\theta)=\operatorname{Jac}\mathcal{J}_{\mu}(g(\theta))\operatorname{Jac}g(\theta). \tag*{(88)} Jac J μ ( θ ) = Jac J μ ( g ( θ )) Jac g ( θ ) . (88)
The Jacobian of g g g is
Jac g ( θ ) = ( I p Jac c μ ( θ ) ) ∈ R ( p + N ) × p , (89) \operatorname{Jac}g(\theta)=\begin{pmatrix} I_{p}\\ \operatorname{Jac}c_{\mu}(\theta)\end{pmatrix}\in\mathbb{R}^{(p+N)\times p}, \tag*{(89)} Jac g ( θ ) = ( I p Jac c μ ( θ ) ) ∈ R ( p + N ) × p , (89)
and the Jacobian of J μ \mathcal{J}_{\mu} J μ at g ( θ ) g(\theta) g ( θ ) decomposes as
Jac J μ ( g ( θ ) ) = ( ∇ θ J μ ( θ , c μ ( θ ) ) T ∇ c J μ ( θ , c μ ( θ ) ) T ) ∈ R 1 × ( p + N ) . (90) \operatorname{Jac}\mathcal{J}_{\mu}(g(\theta))=\left(\nabla_{\theta}\mathcal{J}_{\mu}(\theta,c_{\mu}(\theta))^{\mathsf{T}}\ \ \nabla_{c}\mathcal{J}_{\mu}(\theta,c_{\mu}(\theta))^{\mathsf{T}}\right)\in\mathbb{R}^{1\times(p+N)}. \tag*{(90)} Jac J μ ( g ( θ )) = ( ∇ θ J μ ( θ , c μ ( θ ) ) T ∇ c J μ ( θ , c μ ( θ ) ) T ) ∈ R 1 × ( p + N ) . (90)
Substitution into (C.29) and multiplication of the two block matrices yields
∇ J μ ( θ ) = ∇ θ J μ ( θ , c μ ( θ ) ) + Jac c μ ( θ ) T ∇ c J μ ( θ , c μ ( θ ) ) . (91) \nabla\mathcal{J}_{\mu}(\theta)=\nabla_{\theta}\mathcal{J}_{\mu}(\theta,c_{\mu}(\theta))+\operatorname{Jac}c_{\mu}(\theta)^{\mathsf{T}}\nabla_{c}\mathcal{J}_{\mu}(\theta,c_{\mu}(\theta)). \tag*{(91)} ∇ J μ ( θ ) = ∇ θ J μ ( θ , c μ ( θ )) + Jac c μ ( θ ) T ∇ c J μ ( θ , c μ ( θ )) . (91)
Since c μ ( θ ) c_{\mu}(\theta) c μ ( θ ) minimizes c ↦ J μ ( θ , c ) c\mapsto\mathcal{J}_{\mu}(\theta,c) c ↦ J μ ( θ , c ) , the first-order optimality condition gives ∇ c J μ ( θ , c μ ( θ ) ) = 0 \nabla_{c}\mathcal{J}_{\mu}(\theta,c_{\mu}(\theta))=0 ∇ c J μ ( θ , c μ ( θ )) = 0 , and (C.32) reduces to ∇ J μ ( θ ) = ∇ θ J μ ( θ , c μ ( θ ) ) \nabla\mathcal{J}_{\mu}(\theta)=\nabla_{\theta}\mathcal{J}_{\mu}(\theta,c_{\mu}(\theta)) ∇ J μ ( θ ) = ∇ θ J μ ( θ , c μ ( θ )) . Since the regularization term μ 2 ∥ c ∥ C N 2 \frac{\mu}{2}\lVert c\rVert_{\mathbb{C}^{N}}^{2} 2 μ ∥ c ∥ C N 2 does not depend on θ \theta θ , ∇ θ J μ = ∇ θ J \nabla_{\theta}\mathcal{J}_{\mu}=\nabla_{\theta}\mathcal{J} ∇ θ J μ = ∇ θ J , and (3.19) follows.
Regularization of the free-space Green’s function singularity
The free-space Green’s function g 0 g_{0} g 0 for the Helmholtz equation has a logarithmic singularity at the origin. When the integral operator G \mathcal{G} G is discretized on a uniform grid, the value g 0 ( 0 ) g_{0}(0) g 0 ( 0 ) must be approximated by a suitable regularized quantity. An efficient strategy is to replace the pointwise value g 0 ( 0 ) g_{0}(0) g 0 ( 0 ) by the average of g 0 g_{0} g 0 over a small domain D D D centered at the origin with area equal to the grid cell area Δ x Δ z \Delta x\,\Delta z Δ x Δ z :
g 0 ( 0 ) ≈ 1 ∣ D ∣ ∫ D g 0 ( x ) d A . (92) g_{0}(0)\approx\frac{1}{\lvert D\rvert}\int_{D}g_{0}(x)\,\mathrm{d}A. \tag*{(92)} g 0 ( 0 ) ≈ ∣ D ∣ 1 ∫ D g 0 ( x ) d A . (92)
We present two approaches.
Circular-cell regularization. The first choice is the disk D = B ( 0 , h ) D=B(0,h) D = B ( 0 , h ) , where h h h is chosen so that π h 2 = Δ x Δ z \pi h^2=\Delta x\,\Delta z π h 2 = Δ x Δ z , that is,
h = Δ x Δ z π . (93) h=\sqrt{\frac{\Delta x\,\Delta z}{\pi}}. \tag*{(93)} h = π Δ x Δ z . (93)
Substitution into (D.1) and use of the local expansion of g 0 g_0 g 0 near the origin yield
g 0 ( 0 ) ≈ 1 2 π ( log h + log k 0 2 + γ − 1 2 ) + i 4 , (94) g_0(0)\approx\frac{1}{2\pi}\left(\log h+\log\frac{k_0}{2}+\gamma-\frac{1}{2}\right)+\frac{i}{4}, \tag*{(94)} g 0 ( 0 ) ≈ 2 π 1 ( log h + log 2 k 0 + γ − 2 1 ) + 4 i , (94)
where γ \gamma γ denotes the Euler-Mascheroni constant.
Rectangular-cell regularization. The second choice is to take D D D as the actual grid cell centered at the origin,
D = [ − Δ x 2 , Δ x 2 ] × [ − Δ z 2 , Δ z 2 ] , ∣ D ∣ = Δ x Δ z . (95) D=\left[-\frac{\Delta x}{2},\frac{\Delta x}{2}\right]\times\left[-\frac{\Delta z}{2},\frac{\Delta z}{2}\right],\qquad|D|=\Delta x\,\Delta z. \tag*{(95)} D = [ − 2 Δ x , 2 Δ x ] × [ − 2 Δ z , 2 Δ z ] , ∣ D ∣ = Δ x Δ z . (95)
Substitution into (D.1) and use of the local expansion of g 0 g_0 g 0 give
g 0 ( 0 ) ≈ I + 1 2 π ( log k 0 2 + γ ) + i 4 , (96) g_0(0)\approx I+\frac{1}{2\pi}\left(\log\frac{k_0}{2}+\gamma\right)+\frac{i}{4}, \tag*{(96)} g 0 ( 0 ) ≈ I + 2 π 1 ( log 2 k 0 + γ ) + 4 i , (96)
where I I I is the logarithmic average over the rectangular cell,
I = 1 Δ x Δ z ∫ D 1 2 π log ∥ x ∥ d A . (97) I=\frac{1}{\Delta x\,\Delta z}\int_D\frac{1}{2\pi}\log\lVert x\rVert\,\mathrm{d}A. \tag*{(97)} I = Δ x Δ z 1 ∫ D 2 π 1 log ∥ x ∥ d A . (97)
This integral admits the closed-form expression
I = 1 4 π ( log ( Δ x 2 + Δ z 2 ) − 2 log 2 − 3 + Δ x Δ z arctan Δ z Δ x + Δ z Δ x arctan Δ x Δ z ) . (98) I=\frac{1}{4\pi}\left(\log(\Delta x^2+\Delta z^2)-2\log2-3+\frac{\Delta x}{\Delta z}\arctan\frac{\Delta z}{\Delta x}+\frac{\Delta z}{\Delta x}\arctan\frac{\Delta x}{\Delta z}\right). \tag*{(98)} I = 4 π 1 ( log ( Δ x 2 + Δ z 2 ) − 2 log 2 − 3 + Δ z Δ x arctan Δ x Δ z + Δ x Δ z arctan Δ z Δ x ) . (98)
Computational complexity of Algorithm 1
We summarize the per-iteration cost of Algorithm 1. At each outer iteration the parameters θ \theta θ are fixed, the discrete operator B θ ∈ C N q × N B_\theta\in\mathbb{C}^{N_q\times N} B θ ∈ C N q × N is assembled, and the discrete normal equations (3.23) are solved for the coefficient vector before the Adam update of θ \theta θ . Throughout, N q N_q N q denotes the number of quadrature points, N N N the number of basis functions, and p p p the number of network parameters. The assembly of B θ B_\theta B θ costs
O ( N q p + N N q 2 ) (direct) , O ( N q p + N N q log N q ) (FFT) , (99) O(N_qp+NN_q^2)\ \text{(direct)},\qquad O(N_qp+NN_q\log N_q)\ \text{(FFT)}, \tag*{(99)} O ( N q p + N N q 2 ) (direct) , O ( N q p + N N q log N q ) (FFT) , (99)
and the full iteration cost using FFT
O ( N q p + N N q log N q + N 2 N q + N 3 ) . (100) O(N_qp+NN_q\log N_q+N^2N_q+N^3). \tag*{(100)} O ( N q p + N N q log N q + N 2 N q + N 3 ) . (100)
Table 7. Per-iteration computational cost and memory of the main operations in Algorithm 1, using the FFT realization of the Green operator. Here N q N_q N q is the number of quadrature points, N N N the number of basis functions, and p p p the number of network parameters.
Other implementation details
Enforcing compact support of m m m via damping taper
The Lippmann-Schwinger formulation requires supp ( δ m ) ⊂ Ω \operatorname{supp}(\delta m) \subset\Omega supp ( δ m ) ⊂ Ω . In practice, the physical velocity model is defined on a finite grid and the contrast δ m \delta m δ m may be nonzero up to and including the boundary of the computational domain. We enforce the compact support condition numerically by applying a separable raised-cosine (Hann) taper that smoothly drives δ m \delta m δ m to zero in a prescribed width border region (refer to Figure 3 ).
Let the computational grid have n x n_x n x points in the horizontal direction and n z n_z n z points in the vertical direction. We reserve a border of w x w_x w x cells on each side in x x x and w z w_z w z cells on each side in z z z , defining an interior region of size ( n z − 2 w z ) × ( n x − 2 w x ) (n_z-2w_z)\times(n_x-2w_x) ( n z − 2 w z ) × ( n x − 2 w x ) . The taper is constructed as the outer product of two one-dimensional windows,
τ k ℓ = ψ z ( k ) ψ x ( ℓ ) , k = 1 , … , n z , ℓ = 1 , … , n x , (101) \tau_{k\ell}=\psi_z(k)\,\psi_x(\ell),\qquad k=1,\ldots,n_z,\quad\ell=1,\ldots,n_x, \tag*{(101)} τ k ℓ = ψ z ( k ) ψ x ( ℓ ) , k = 1 , … , n z , ℓ = 1 , … , n x , (101)
where each factor ψ \psi ψ is a raised-cosine ramp: taking the x x x -direction as representative,
ψ x ( ℓ ) = { 1 2 ( 1 − cos ( π ( ℓ − 1 ) w x ) ) , ℓ = 1 , … , w x , 1 , ℓ = w x + 1 , … , n x − w x , 1 2 ( 1 − cos ( π ( n x − ℓ ) w x ) ) , ℓ = n x − w x + 1 , … , n x , (102) \psi_x(\ell)=
\begin{cases}
\frac{1}{2}\left(1-\cos\left(\frac{\pi(\ell-1)}{w_x}\right)\right), & \ell=1,\ldots,w_x,\\
1, & \ell=w_x+1,\ldots,n_x-w_x,\\
\frac{1}{2}\left(1-\cos\left(\frac{\pi(n_x-\ell)}{w_x}\right)\right), & \ell=n_x-w_x+1,\ldots,n_x,
\end{cases} \tag*{(102)} ψ x ( ℓ ) = ⎩ ⎨ ⎧ 2 1 ( 1 − cos ( w x π ( ℓ − 1 ) ) ) , 1 , 2 1 ( 1 − cos ( w x π ( n x − ℓ ) ) ) , ℓ = 1 , … , w x , ℓ = w x + 1 , … , n x − w x , ℓ = n x − w x + 1 , … , n x , (102)
and ψ z ( k ) \psi_z(k) ψ z ( k ) is defined analogously. The window equals one throughout the interior and ramps smoothly to exactly zero at the grid boundary, so that τ k ℓ = 0 \tau_{k\ell}=0 τ k ℓ = 0 whenever k ∈ { 1 , n z } k\in\{1,n_z\} k ∈ { 1 , n z } or ℓ ∈ { 1 , n x } \ell\in\{1,n_x\} ℓ ∈ { 1 , n x } . The tapered contrast is then
δ m ~ ( x k ℓ ) = τ k ℓ δ m ( x k ℓ ) , (103) \widetilde{\delta m}(x_{k\ell})=\tau_{k\ell}\,\delta m(x_{k\ell}), \tag*{(103)} δ m ( x k ℓ ) = τ k ℓ δ m ( x k ℓ ) , (103)
which satisfies δ m ~ ≡ 0 \widetilde{\delta m}\equiv0 δ m ≡ 0 on the border cells and hence supp ( δ m ~ ) ⊂ Ω f u l l \operatorname{supp}(\widetilde{\delta m})\subset\Omega_{\mathrm{full}} supp ( δ m ) ⊂ Ω full , as required. This replaces the original contrast δ m \delta m δ m in the weighted basis matrix (36) and therefore in the Green operator (37) .
FFT-accelerated evaluation
The matrix product (37) consists of N N N independent Toeplitz matrix-vector products (38) . A direct evaluation of each product costs O ( N q 2 ) O(N_q^2) O ( N q 2 ) , giving a total cost of O ( N N q 2 ) O(NN_q^2) O ( N N q 2 ) . this can be reduced to O ( N N q log N q ) O(NN_q\log N_q) O ( N N q log N q ) by embedding the Toeplitz product into a circulant one and applying the convolution theorem, implementing the discrete Fourier transform and its inverse using an FFT algorithm.
Since g 0 g_0 g 0 is Toeplitz with 2 N q − 1 2N_q-1 2 N q − 1 distinct entries t − ( N q − 1 ) , … , t N q − 1 t_{-(N_q-1)},\ldots,t_{N_q-1} t − ( N q − 1 ) , … , t N q − 1 , we set M = 2 N q − 1 M=2N_q-1 M = 2 N q − 1 and pad both the kernel and each column of F θ F_\theta F θ to length M M M . Define the padded kernel g ~ ∈ C M \widetilde{g}\in\mathbb{C}^{M} g ∈ C M by
g ~ r = { t r − 1 , r = 1 , … , N q , t r − 1 − M , r = N q + 1 , … , M , (104) \widetilde{g}_r=
\begin{cases}
t_{r-1}, & r=1,\ldots,N_q,\\
t_{r-1-M}, & r=N_q+1,\ldots,M,
\end{cases} \tag*{(104)} g r = { t r − 1 , t r − 1 − M , r = 1 , … , N q , r = N q + 1 , … , M , (104)
which packs all 2 N q − 1 2N_q-1 2 N q − 1 values into a single period of length M M M : the non-negative lags t 0 , … , t N q − 1 t_0,\ldots,t_{N_q-1} t 0 , … , t N q − 1 occupy the first N q N_q N q entries and the negative lags t − ( N q − 1 ) , … , t − 1 t_{-(N_q-1)},\ldots,t_{-1} t − ( N q − 1 ) , … , t − 1 are placed at the end. This vector depends only on the kernel g 0 g_0 g 0 and is the same for every column j j j . For each j = 1 , … , N j=1,\ldots,N j = 1 , … , N , define the padded column f ~ j ∈ C M \widetilde{f}_j\in\mathbb{C}^{M} f j ∈ C M by
( f ~ j ) r = { ( F θ ) r , j , r = 1 , … , N q , 0 , r = N q + 1 , … , M . (105) (\widetilde{f}_j)_r=
\begin{cases}
(F_\theta)_{r,j}, & r=1,\ldots,N_q,\\
0, & r=N_q+1,\ldots,M.
\end{cases} \tag*{(105)} ( f j ) r = { ( F θ ) r , j , 0 , r = 1 , … , N q , r = N q + 1 , … , M . (105)
The circular convolution of length M M M is
( g ~ ∗ f ~ j ) k = ∑ r = 1 M g ~ ( k − r ) m o d M + 1 ( f ~ j ) r , k = 1 , … , M , (106) (\widetilde{g}\ast\widetilde{f}_j)_k=\sum_{r=1}^{M}\widetilde{g}_{(k-r)\bmod M+1}(\widetilde{f}_j)_r,\qquad k=1,\ldots,M, \tag*{(106)} ( g ∗ f j ) k = r = 1 ∑ M g ( k − r ) mod M + 1 ( f j ) r , k = 1 , … , M , (106)
which defines a circulant matrix-vector product of size M M M . For k ∈ { 1 , … , N q } k\in\{1,\ldots,N_q\} k ∈ { 1 , … , N q } , the zero-padding kills all terms with r > N q r>N_q r > N q , and the construction of g ~ \widetilde{g} g ensures that the circulant index resolves to t k − r t_{k-r} t k − r for every r ∈ { 1 , … , N q } r\in\{1,\ldots,N_q\} r ∈ { 1 , … , N q } , so
( g ~ ∗ f ~ j ) k = ∑ ℓ = 1 N q t k − ℓ ( F θ ) ℓ , j = ( g 0 F θ ) k , j = ( G Φ θ ) k , j . (107) (\widetilde{g}\ast\widetilde{f}_j)_k=\sum_{\ell=1}^{N_q}t_{k-\ell}(F_\theta)_{\ell,j}=(g_0F_\theta)_{k,j}=(G\Phi_\theta)_{k,j}. \tag*{(107)} ( g ∗ f j ) k = ℓ = 1 ∑ N q t k − ℓ ( F θ ) ℓ , j = ( g 0 F θ ) k , j = ( G Φ θ ) k , j . (107)
That is, the first N q N_q N q entries of the length-M M M circular convolution reproduce the Toeplitz product (38) exactly, with no wraparound. By the convolution theorem
g ~ ⊛ f ~ j = F M − 1 ( F M [ g ~ ] ⊙ F M [ f ~ j ] ) , (108) \tilde{g} \circledast\tilde{f}_j = \mathcal{F}_M^{-1}\left(\mathcal{F}_M[\tilde{g}] \odot\mathcal{F}_M[\tilde{f}_j]\right), \tag*{(108)} g ~ ⊛ f ~ j = F M − 1 ( F M [ g ~ ] ⊙ F M [ f ~ j ] ) , (108)
where F M \mathcal{F}_M F M denotes the DFT of length M M M and ⊙ \odot ⊙ the pointwise product. Combining the two, each column of G Φ θ G\Phi_\theta G Φ θ is recovered by extracting the first N q N_q N q entries,
( G Φ θ ) ⋅ , j = [ F M − 1 ( F M [ g ~ ] ⊙ F M [ f ~ j ] ) ] 1 : N q , j = 1 , … , N , (109) (G\Phi_\theta)_{\cdot,j} = \left[\mathcal{F}_M^{-1}\left(\mathcal{F}_M[\tilde{g}] \odot\mathcal{F}_M[\tilde{f}_j]\right)\right]_{1:N_q}, \qquad j = 1,\ldots,N, \tag*{(109)} ( G Φ θ ) ⋅ , j = [ F M − 1 ( F M [ g ~ ] ⊙ F M [ f ~ j ] ) ] 1 : N q , j = 1 , … , N , (109)
where [ ⋅ ] 1 : N q [\cdot]_{1:N_q} [ ⋅ ] 1 : N q denotes extraction of the first N q N_q N q entries. The transform F M [ g ~ ] \mathcal{F}_M[\tilde{g}] F M [ g ~ ] is computed once at cost O ( M log M ) O(M\log M) O ( M log M ) , and each of the N N N columns requires one forward FFT, one pointwise product, and one inverse FFT, giving a total cost of O ( N M log M ) = O ( N N q log N q ) O(NM\log M) = O(NN_q\log N_q) O ( N M log M ) = O ( N N q log N q ) , compared with O ( N N q 2 ) O(NN_q^2) O ( N N q 2 ) for the direct Toeplitz products.