Variability of earthquake stress drop in a subduction setting, the Hikurangi Margin, New Zealand

We calculate stress drops for 176 earthquakes (M2.6-M6.6) from four sequences of earthquakes in New Zealand. Two sequences are within the subducting Pacific plate (2014 Eketahuna and 2005 Upper Hutt), one in the over-riding plate (2013 Cook Strait) and one involved reverse faulting at the subduction interface (2015 Pongaroa). We focus on obtaining precise and accurate measurements of corner frequency and stress drop for the best-recorded earthquakes. We use an empirical Green's function (EGF) approach, and require the EGF earthquakes to be highly correlated (cross-correlation >= 0.8) to their respective main shocks. In order to improve the quality, we also stack the spectral ratios and source time functions obtained from the best EGF. We perform a grid search for each individual ratio, and each stacked ratio to obtain quantitative uncertainty measurements, and restrict our analysis to the well-constrained corner frequency measurements. We are able to analyse both P and S waves independently and the high correlation between these measurements strengthens the reliability of our results. We find that there is significant real variability in corner frequency, and hence stress drop, within each sequence; the range of almost 2 orders of magnitude is larger than the uncertainties. The four sequences have overlapping stress drop ranges, and the variability within a sequence is larger than any between different sequences. There is no clear systematic difference in the populations analysed here with tectonic setting. We see no dependence of the stress drop values on depth, time, or magnitude after taking the frequency bandwidth limitations into consideration. Small-scale heterogeneity must therefore exert a more primary influence on earthquake stress drop than these larger scale factors. We confirm that when fitting individual spectral ratios, a corner frequency within a factor of three of the maximum signal frequency is likely to be underestimated. Stacked ratios are smoother and more reliable near the frequency limits. We find that only corner frequencies within about a factor of two of the maximum signal frequency are likely to be underestimated.