Bayesian Inversion with an Improved Proposal Distribution

عنوان دوره: بیستمین کنفرانس ژئوفیزیک ایران
کد مقاله : 1896-NIGS
نویسندگان
1Institute of Geophysics, University of Tehran
2موسسه ژئوفیزیک دانشگاه تهران
چکیده
Geophysical inverse problems are generally formulated as an ill-posed non-linear optimization problem commonly solved through deterministic gradient-based approaches. Using these methods, despite fast convergence properties, may lead to local minima as well as impend accurate uncertainty analysis. On the contrary, formulating a geophysical inverse problem in a probabilistic framework and solving it by constructing the multi-dimensional posterior probability density (PPD) allow for complete sampling of the parameter space and uncertainty quantification. The PDD is numerically characterized using Markov Chain Monte Carlo (MCMC) approaches. However, the convergence of the MCMC algorithm (i.e. sampling efficiency) toward the target stationary distribution highly depends upon the choice of the proposal distribution. In this paper, we develop an efficient proposal distribution based on perturbing the model parameters through an eigenvalue decomposition of the model covariance matrix in a principal component space. The proposed strategy is illustrated for the inversion of synthetic and real geo-electrical data sets. The numerical experiments demonstrate that the presented proposal distribution takes advantage of the benefits from an accelerated convergence and mixing rate compared to the conventional Gaussian proposal distribution.
کلیدواژه ها
 
Title
Bayesian Inversion with an Improved Proposal Distribution
Authors
Reza Ghanati, Zahra Tafaghod Khabaz
Abstract
Geophysical inverse problems are generally formulated as an ill-posed non-linear optimization problem commonly solved through deterministic gradient-based approaches. Using these methods, despite fast convergence properties, may lead to local minima as well as impend accurate uncertainty analysis. On the contrary, formulating a geophysical inverse problem in a probabilistic framework and solving it by constructing the multi-dimensional posterior probability density (PPD) allow for complete sampling of the parameter space and uncertainty quantification. The PDD is numerically characterized using Markov Chain Monte Carlo (MCMC) approaches. However, the convergence of the MCMC algorithm (i.e. sampling efficiency) toward the target stationary distribution highly depends upon the choice of the proposal distribution. In this paper, we develop an efficient proposal distribution based on perturbing the model parameters through an eigenvalue decomposition of the model covariance matrix in a principal component space. The proposed strategy is illustrated for the inversion of synthetic and real geo-electrical data sets. The numerical experiments demonstrate that the presented proposal distribution takes advantage of the benefits from an accelerated convergence and mixing rate compared to the conventional Gaussian proposal distribution.
Keywords
MCMC, Principal component space, Proposal distribution