Bayesian Geometric Inversion for Wave Equation: Well-Posedness and Multiscale Multimodal Posterior Approximation
Abstract
Geometric full-waveform inversion aims to recover sharp subsurface interfaces from noisy waveform observations. We formulate this problem within a Bayesian level-set framework, where the unknown interface is represented as zero level set of a latent function endowed with a Gaussian prior. Although the level-set-to-medium-parameter map is discontinuous, we establish that the level-set-to-observation map is continuous almost surely with respect to the prior, thereby guaranteeing that the posterior measure is well-posed and locally Lipschitz continuous with respect to the data in the Hellinger metric. To enable tractable posterior computation under the severe nonlinearity of the inverse problem and the resulting multimodality, we develop a multiscale multimodal Laplace approximation. A frequency-continuation strategy incorporates waveform data progressively from low to high frequencies, identifying and propagating multiple candidate modes across successive scales. At the final frequency level, local Gaussian approximations are constructed around the retained modes and combined into a Gaussian mixture, whose weights are determined by least-squares fitting to the unnormalized posterior density. Numerical experiments demonstrate that the method accurately reconstructs anomalous interfaces while effectively capturing posterior multimodality and the local uncertainty associated with distinct plausible geometries.
Related
- Well-posedness for a class of n× n hyperbolic systems and shock stability for the viscous approximation
- Well-posedness for time-fractional evolution equations by the Yosida approximation
- On the well-posedness and efficient approximation for the classical Melan equation in suspension bridges
- WEAKLY VISCOUS NONLOCAL WAVE EQUATIONS WITH VARIABLE COEFFICIENTS: WELL-POSEDNESS, RUNGE APPROXIMATION, AND INVERSE PROBLEMS
- High-order smoothness and high-order collocation approximation for Volterra integral equation with multiscale and nonlinear exponent kernel
BibTeX
@misc{yan2026bayesian,
title = {{Bayesian Geometric Inversion for Wave Equation: Well-Posedness and Multiscale Multimodal Posterior Approximation}},
author = {Xiong-Bin Yan and Ming-Jia Pei and Guang-Hui Zheng and Ting Wei},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.26151},
eprint = {2610.26151},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.26151},
}
RIS
TY - UNPB
TI - Bayesian Geometric Inversion for Wave Equation: Well-Posedness and Multiscale Multimodal Posterior Approximation
AU - Xiong-Bin Yan
AU - Ming-Jia Pei
AU - Guang-Hui Zheng
AU - Ting Wei
PY - 2026
DA - 2026/10/19
PB - arXiv
DO - 10.48550/arxiv.2610.26151
UR - https://arxiv.org/abs/2610.26151
AB - Geometric full-waveform inversion aims to recover sharp subsurface interfaces from noisy waveform observations. We formulate this problem within a Bayesian level-set framework, where the unknown interface is represented as zero level set of a latent function endowed with a Gaussian prior. Although the level-set-to-medium-parameter map is discontinuous, we establish that the level-set-to-observation map is continuous almost surely with respect to the prior, thereby guaranteeing that the posterior measure is well-posed and locally Lipschitz continuous with respect to the data in the Hellinger metric. To enable tractable posterior computation under the severe nonlinearity of the inverse problem and the resulting multimodality, we develop a multiscale multimodal Laplace approximation. A frequency-continuation strategy incorporates waveform data progressively from low to high frequencies, identifying and propagating multiple candidate modes across successive scales. At the final frequency level, local Gaussian approximations are constructed around the retained modes and combined into a Gaussian mixture, whose weights are determined by least-squares fitting to the unnormalized posterior density. Numerical experiments demonstrate that the method accurately reconstructs anomalous interfaces while effectively capturing posterior multimodality and the local uncertainty associated with distinct plausible geometries.
ER -