- Stability of Gibbs Posteriors from the Wasserstein Loss for Bayesian Full Waveform Inversion(arXiv)
Writer : : Matthew M. Dunlop, Yunan Yang
Summary : Lately, the Wasserstein loss perform has been confirmed to be efficient when utilized to deterministic full-waveform inversion (FWI) issues. We take into account the applying of this loss perform in Bayesian FWI in order that the uncertainty may be captured within the resolution. Different loss capabilities which might be generally utilized in observe are additionally thought of for comparability. Existence and stability of the ensuing Gibbs posteriors are proven on perform area beneath weak assumptions on the prior and mannequin. Specifically, the distribution arising from the Wasserstein loss is proven to be fairly secure with respect to high-frequency noise within the information. We then illustrate the distinction between the ensuing distributions numerically, utilizing Laplace approximations to estimate the unknown velocity area and uncertainty related to the estimates.
