The Community Stress-Drop Validation Study—Part I: Source, Propagation, and Site Decomposition of Fourier Spectra

As part of the community stress-drop validation study initiative, we apply a spectral
decomposition approach to isolate the source spectra of 556 events occurred during
the 2019 Ridgecrest sequence (Southern California). We perform multiple decompositions by introducing alternative choices for some processing and model assumptions,
namely: three different S-wave window durations (i.e., 5 s, 20 s, and variable between 5
and 20 s); two attenuation models that account differently for depth dependencies; and
two different site amplification constraints applied to restore uniqueness of the solution. Seismic moment and corner frequency are estimated for the Brune and Boatwright
source models, and an extensive archive including source spectra, site amplifications,
attenuation models, and tables with source parameters is disseminated as the main
product of the present study. We also compare different approaches to measure the
precision of the parameters expressed in terms of 95% confidence intervals (CIs).
The CIs estimated from the asymptotic standard errors and from Monte Carlo resampling of the residual distribution show an almost one-to-one correspondence; the
approach based on model selection by setting a threshold for misfit chosen with an
F-ratio test is conservative compared to the approach based on the asymptotic standard
errors. The uncertainty analysis is completed in the companion article in which the outcomes from this work are used to compare epistemic uncertainty with precision of the
source parameters.

In part 1, we run multiple GIT decomposition for different choices of model assumptions, namely three different window duration for Fourier calculation, two different parametrization of the attenuation, two different site constraints. We also considered different source models (Brune, Boatwright, Brune with kappa_source) and different approaches to estimate uncertainties of source parameters (i.e., considering the covariance matrix, Monte Carlo sampling of the residual distribution, model selection with threshold based on F-test).