Full waveform inversion reconstructs the Earth's interior from seismic data. The inverse problem is nonlinear — small changes in subsurface velocity structure can produce large, non-proportional changes in observed waveforms. Bayesian inversion quantifies the uncertainty: not just the best model, but the full family of models consistent with the data.
The posterior distribution can be estimated two ways. The nonlinear method explores the full probability landscape, capturing the non-Gaussian, multimodal, asymmetric character of the uncertainty. The linearized method approximates the posterior as a Gaussian around the best-fit model. This is faster, cheaper, and wrong.
Both methods recover the posterior mean accurately (arXiv:2603.11711). The best-fit model is the same either way. But the uncertainty structures diverge, especially at layer interfaces — exactly where the geological interpretation matters most. The linearized uncertainties fit the observed data poorly and produce biased estimates of geological body volumes. The mean is correct; the error bars are fiction.
The structural observation: the point estimate survives linearization but the uncertainty does not. The best model is a robust quantity — it lives near the peak of the posterior regardless of how you approximate the distribution's shape. But the width, asymmetry, and multimodal structure of the posterior are nonlinear features that the Gaussian approximation cannot represent. At layer boundaries, where the physics is most nonlinear — small velocity changes produce large waveform changes — the linearized uncertainty is most wrong. The linearization is accurate where it matters least and inaccurate where it matters most.
Note (2026-11-11): checked against the full text of the paper (arXiv 2603.11711, linearised against nonlinear uncertainty estimates in Bayesian full waveform inversion). The finding that both methods recover the posterior mean accurately while their uncertainty structures differ, especially around layer interfaces, with biased estimates of geological volumes from the linearised one, is as in the paper. Three points are framed more strongly. (1) The paper's tests are two-dimensional acoustic problems (a simple layered model in a near-ideal setting, then a more realistic example), and its nonlinear estimates are themselves approximations: a variational method that optimizes a transformed Gaussian, compared with Stein variational gradient descent. The essay speaks of the nonlinear method as exploring the full probability landscape. (2) The essay says the linearised method is 'wrong' and 'the error bars are fiction'; the paper says the linearised uncertainties are 'significantly less accurate' and give far poorer fits to the observed waveform data and biased volume estimates. (3) The essay's statement that the linearisation is 'accurate where it matters least and inaccurate where it matters most' is its own summary of the interface result; the paper's conclusion is that fully nonlinear inversion is warranted when accurate uncertainty on parameters or on interpreted meta-properties is important.