Caroline Beghein 

Radial seismic anisotropy models are traditionally obtained using empirical constraints based on laboratory experiments and petrological considerations. We tested the hypothesis that such petrological constraints affect the uppermost mantle models of S wave anisotropy using a statistical approach. In addition, we were able to determine which model features are constrained by the data and which are dominated by the prior. We focused on large-scale models and found that the most likely models obtained in both cases are highly correlated. This demonstrates that for the best data-fitting solution, the geometry of uppermost mantle radial anisotropy is not strongly affected by prior petrological constraints. The amplitude of the anomalies, however, can change significantly: The best data-fitting model obtained without petrological constraints displays stronger amplitudes than the one obtained with prior. This could become an issue when quantitatively interpreting seismic anisotropy models, and thus emphasizes the importance of accurately accounting for parameter uncertainties and trade-offs, and of understanding whether the seismic data or the prior constraints the model. We showed that model uncertainties are strongly affected by the prior as the relative RMS uncertainties were reduced by a factor of 2. In addition, we showed that while the model distributions are not necessarily Gaussian a priori, imposing petrological constraints can force the distributions to be narrower and more Gaussian-like, as expected from inverse theory. Finally, we demonstrated that the age dependence of seismic wave velocities is robust and independent of prior constraints. A similar age signal exists for anisotropy, but with larger uncertainties without prior constraints.