Born-Series-Inspired Residual Metric for Learning-based Preconditioners

Juntao Wang, Xinliang Liu, Jiwei Jia · 2026

Read Born-Series-Inspired Residual Metric for Learning-based Preconditioners by Juntao Wang, Xinliang Liu, Jiwei Jia (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-18527.txt.

Introduction

Wang, Jia, and Liu propose Neural Preconditioned Born Series for learning iterative neural PDE solvers. High-frequency waves in varied media challenge both classic iterative and neural solvers. A simple reference preconditioner feeds a network that predicts solution updates from preconditioned residuals. Training cuts the same preconditioned residual, giving a metric-matched objective. On heterogeneous Helmholtz problems, iteration counts fall, with larger gains at higher wavenumbers and the lowest wall-clock time in Krylov tests.

Catalog record

CREATORJuntao Wang, Xinliang Liu, Jiwei Jia
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE56070
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.18527

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Born-Series-Inspired Residual Metric for Learning-based Preconditioners Juntao Wang, Xinliang Liu, Jiwei Jia Abstract Loss functions for learning-based PDE preconditioners implicitly choose a metric in which residuals are matched, yet most approaches still optimize an unpreconditioned Euclidean residual norm. For indefinite operators such as the high-frequency Helmholtz equation, this default metric can make both learning and iterative correction overly sensitive to near-resonant spectral components, while classical preconditioning succeeds precisely by reshaping the residual geometry. We show that the Born Series and shifted-Laplacian left preconditioning are linked by the identity I − G η ​ V η = G η ​ A = L η − 1 ​ A , I-G_{\eta}V_{\eta}=G_{\eta}A=L_{\eta}^{-1}A, which turns the reference Green operator G η G_{\eta} into a natural Riesz-map residual metric R η = G η ∗ ​ G η R_{\eta}=G_{\eta}^{\ast}G_{\eta} and suggests measuring the physical residual via ‖ r ‖ R η = ‖ G η ​ r ‖ 2 . \|r\|_{R_{\eta}}=\|G_{\eta}r\|_{2}. Building on this viewpoint, we propose a Neural Preconditioned Born Series (NPBS) iteration that replaces the scalar CBS relaxation with a residual-driven neural operator, together with a metric-matched Born-series-inspired loss ℒ bs R η \mathcal{L}_{\mathrm{bs}}^{R_{\eta}} . The framework is architecture-agnostic and supports fast 𝒪 ⁡ ( N ​ log ⁡ N ) \mathcal{O}(N\log N) evaluation via FFT/DST/DCT. Numerical experiments on heterogeneous Helmholtz problems demonstrate the effectiveness of our method, and its advantage becomes more pronounced as the systems grow more ill-conditioned; we then extend the framework to other PDE classes, including convection–diffusion–reaction equations and linearized Newton systems for nonlinear PDEs, where it also yields substantial iteration reductions. Keywords Born Series \cdot Neural preconditioner \cdot Residual metric 1 Introduction The heterogeneous Helmholtz equation remains challenging for both classical and learning-based solvers in high-wavenumber and high-contrast regimes. Standard discretizations suffer from pollution errors 2 ; multilevel methods face coarse-grid dispersion mismatch 32 ; and indefiniteness degrades relaxation and coarse correction 13 . A variety of classical remedies: shifted-Laplacian preconditioning 12 , wave-ray multigrid 5 ; 23 , convergent Born series (CBS) 25 , and domain decomposition 9 ; 18 —have improved robustness, yet efficiency in strongly heterogeneous media is still not fully resolved. Learning-based PDE solvers have progressed from PINNs 28 to neural operators 1 ; 24 ; 21 ; 20 ; 27 ; 15 ; 7 ; 16 ; 8 . For Helmholtz, an increasingly effective paradigm is to embed learned modules inside iterative solvers as residual-to-correction maps rather than rely on end-to-end prediction 19 ; 29 ; 36 ; 10 ; 34 ; 31 . In its simplest form, a learned operator \mathcal{M}_{\theta} updates \bm{u}^{m+1}=\bm{u}^{m}+\mathcal{M}_{\theta}(\bm{f}-A\bm{u}^{m}) and training commonly minimizes an unpreconditioned Euclidean residual mismatch such as \|A\mathcal{M}_{\theta}(\bm{r})-\bm{r}\|_{2} . While this choice is natural, it implicitly fixes a residual metric and therefore determines which spectral components dominate optimization. For indefinite operators, Euclidean residual exhibit particularly stiff landscapes around near-resonant modes 22 ; 26 , motivating objective designs that reflect the geometry induced by effective preconditioning. To address this challenge, we observe that the Born Series (integral-equation) and shifted-Laplacian left preconditioning share the same algebraic core. Let L_{\eta}:=-\Delta-(k_{0}^{2}+i\eta) be a damped reference operator, G_{\eta}:=L_{\eta}^{-1} its inverse, and define V_{\eta}:=L_{\eta}-A (details in Section 2 ). Then the identity I-G_{\eta}V_{\eta}\;=\;G_{\eta}A\;=\;L_{\eta}^{-1}A shows that the integral residual in Born Series and the shifted-Laplacian left-preconditioned residual are equivalent representations. This equivalence suggests measuring residuals in the preconditioner-induced norm \|\bm{r}\|_{R_{\eta}}:=\|G_{\eta}\bm{r}\|_{2},\qquad R_{\eta}:=G_{\eta}^{\ast}G_{\eta}, i.e., a Riesz-map metric that inherits the conditioning mechanism of shifted-Laplacian preconditioning. Guided by these viewpoints, we propose a Neural Preconditioned Born Series (NPBS) iteration that generalizes CBS by replacing its scalar correction with a residual-driven neural operator acting in preconditioned coordinates. Related Born-series-inspired works such as the learned Born series (LBS) 30 and the neural Born series operator (NBSO) 35 are fixed-depth, CBS-inspired unrolled forward solvers trained to predict Helmholtz wavefields. By contrast, we learn a residual-to-correction map inside an iterative solver. To optimize NPBS, we construct two complementary training objectives from the original system A\bm{u}=\bm{f} and its left-preconditioned counterpart G_{\eta}A\bm{u}=G_{\eta}\bm{f} . Accordingly, we obtain two residual metrics and their associated losses: the Euclidean-metric residual loss \mathcal{L}_{\mathrm{bs}}^{\ell_{2}} and the preconditioned-metric Born Series-inspired loss \mathcal{L}_{\mathrm{bs}}^{R_{\eta}} (Section 3.3 ). We develop our framework primarily based on the Helmholtz equation, and subsequently extend the methodology to other PDE classes, including convection–diffusion–reaction and linearized Newton systems. Our contributions are summarized as follows: 1. Neural Preconditioned Born Series (NPBS). We propose NPBS, a learned correction iteration in preconditioned residual coordinates that generalizes the Convergent Born Series (CBS) and is compatible with multiple neural operator architectures. 2. Residual-metric viewpoint and Born-series-inspired objectives. Leveraging the identity that connects Born Series residuals with shifted-Laplacian left preconditioning, we derive the preconditioner-induced residual metric R_{\eta}=G_{\eta}^{\ast}G_{\eta} and develop two training objectives, \mathcal{L}_{\mathrm{bs}}^{\ell_{2}} and \mathcal{L}_{\mathrm{bs}}^{R_{\eta}} , with consistent formulations for Helmholtz, CDR, and linearized Newton systems. 3. Robustness to ill-conditioning and algorithmic generalization. Experiments demonstrate that the convergence advantage of NPBS over direct baselines widens significantly as the underlying systems become increasingly ill-conditioned (e.g., higher wavenumbers or highly discontinuous media). Furthermore, the methodology generalizes effectively to other PDEs, delivering substantial iteration reductions on convection–diffusion–reaction equations and Newton linearizations of nonlinear problems. The paper is organized as follows: Section 2 derives the Born/integral equation and establishes its equivalence to shifted-Laplacian left-preconditioning; Section 3.2 introduces NPBS; Section 3.3 presents the Born-series-inspired objectives; and Section 4 reports the experiments. 2 From Born Series to Preconditioned Coordinates We begin from the classical Born series/Lippmann–Schwinger integral formulation and analyze its algebraic relation to shifted-Laplacian left-preconditioning. The key result (Proposition 2.1 ) shows that the integral residual and the shifted-Laplacian left-preconditioned residual are mathematically equivalent. This identity motivates both the Neural Preconditioned Born Series (NPBS) solver (Section 3 ) and the preconditioned training metric (Section 3.3 ). 2.1 Operator Splitting and Integral Equation We consider the Helmholtz equation for wave scattering in an unbounded heterogeneous medium, subject to the Sommerfeld radiation condition: \displaystyle\mathcal{L}u(x):=-\Delta u(x)-k(x)^{2}u(x) \displaystyle=f(x),\quad x\in\mathbb{R}^{d}, (1) \displaystyle\lim_{|x|\to\infty}|x|^{\frac{d-1}{2}}\left(\frac{\partial u}{\partial|x|}-ik_{0}u\right) \displaystyle=0, where u is the wavefield, f the source, d the spatial dimension, and k(x)=\omega/c(x) the spatially varying wavenumber ( \omega : angular frequency, c(x) : speed of sound). The constant k_{0} is the background wavenumber ( k(x)\to k_{0} as |x|\to\infty ). For computation, the unbounded domain is truncated to a bounded domain \Omega , and absorbing boundary layers—sponge layers 17 or PML 3 ; 31 —are incorporated so that the resulting bounded-domain operator (still denoted \mathcal{L} ) suppresses spurious reflections. After discretization (finite differences or FFT-based pseudospectral methods), we obtain the linear system A\bm{u}=\bm{f}, where A is the discrete Helmholtz operator (including absorbing-layer terms), \bm{u} the discrete unknown, and \bm{f} the discrete source. Introduce a damped reference operator \mathcal{L}_{\eta}:=-\Delta-(k_{0}^{2}+i\eta),\qquad\eta>0, (2) used only for preconditioning/training coordinates. The reference is chosen to be transform-diagonalizable (periodic FFT or Dirichlet/Neumann DST/DCT) for fast application; any mismatch between this reference and the physical operator, including absorbing-layer terms, is absorbed into the perturbation \mathcal{V}_{\eta}:=\mathcal{L}_{\eta}-\mathcal{L}. Let g_{\eta} be the Green kernel for the continuous reference operator \mathcal{L}_{\eta} under the chosen boundary condition on \Omega . Define the associated continuous Green operator (\mathcal{G}_{\eta}q)(x):=\int_{\Omega}g_{\eta}(x,y)\,q(y)\,\mathrm{d}y,\qquad q\in L^{2}(\Omega), (3) so that \mathcal{G}_{\eta}=\mathcal{L}_{\eta}^{-1} on the admissible function class. Discrete realization. In the discrete setting, let L_{\eta} denote the matrix obtained by discretizing \mathcal{L}_{\eta} (e.g., via finite differences or Fourier pseudospectral

Read the complete text