, D.Turcotte 5

It is well established that earthquakes are correlated over distances greatly exceeding their source dimension. Recent studies hypothesize for important associated phenomenon: The area over which earthquake activity is correlated varies in time and might grow prior to a large earthquake. This hypothesis is supported by a wealth of observations, computer simulation, and has theoretical interpretations. Several measures of earthquake correla-tion lengths were recently suggested by dierent authors. Here we analyze one of these measures, (x; t), based on single-link cluster analysis of epicenters. Previous studies have shown the growth of prior to nine large earthquakes in California during 1945-2000. In this paper we study whether the reported growth of the correlation length (x; t) can be used for earthquake prediction. Our results show that reasonable retrospective prediction of large earthquakes (M 6:5) in California can be achieved by using the increase of as a signal for the approach of a large earthquake. Extensive variations of numerical param-eters demonstrate the stability of this prediction method. Additionally, we compare the distributions of (x; t) close and distant in time and space to large earthquakes and nd a systematic shift re ecting the increase of the correlation length prior to large earthquakes