We are aware of the risks when interpreting temporal variations in
tomographic models. In particular, study of the posterior residuals
after the inversion shows that the dataset used in this study is rather
noisy. Given the short source-receiver distances in this study, even
strong heterogeneities hardly provide large relevant residuals. Thus,
noise is a considerable component of the data in addition to signal,
which explains the moderate variance reduction in some years
(see Fig. 2C). The influence of random noise on the final result can be
estimated using the “odd/even test” which consists in independent
inversions of two data subsets separated using odd (1, 3, 5…) and
even (2, 4, 6…) numbers of events. The inversion procedure is identical
to the one used to derive the main results. The results of the
reconstructed Vp/Vs ratio for the two subsets in different years from
2003 to 2006 are presented in the vertical section in Fig. 3 (results in
other years are similar and not presented to save the space).