It is much more difficult to estimate the effect of varying ray configurations
between different time periods on the model variations. To
assess this, we have performed a series of synthetic tests with realistic
patterns and checkerboards. The synthetic tests were performed in a
way that reproduces the procedures for real data processing, as far as
possible. The travel times are computed for sources and receivers corresponding
to the real observations in different years using 3D ray tracing
by the bending algorithm. The travel times are then perturbed
with noise with a predefined RMS for P and S data (0.1 and 0.15 s,
respectively) that enables similar variance reductions for synthetic
data as for observed data. The noise was produced by a random number
generator with the statistical distribution corresponding to the residuals
in real earthquake catalogs. After computing the synthetic travel
times, we “forget” the hypocenters, origin times and the velocity distribution.
The reconstruction of the synthetic model is performed in the
same way as the real data processing including the estimation of the
initial source location using the starting 1D model. We use the same
values of free parameters (damping, smoothing, grid spacing etc.) as
in the real case.