Window-Width Scanning Detects Genuine Aftershock-Sequence Non-Stationarity in Earthquake Catalogs — But Confers No Discriminative Precursor Power

Bridge Paper T (10.5281/zenodo.21484429) applied PL-27/PL-28 to earthquake catalogs and, aftertracing two apparent leads to specific confounds (a foreshock-composition artifact and a single-baseline-window bias), reported a clean negative result at a single, fixed pre-event window width.This paper extends that investigation along an axis Bridge Paper T did not test: instead of holdingthe window width fixed and sliding it chronologically, we fix the window’s END at a target eventand vary how far back it reaches (a window-WIDTH scan). Applied to the JMA catalog’s magnitudesequence around the Kumamoto M7.3 (2016) and Tohoku M9.0 (2011) mainshocks, this reveals awidth range (roughly 2,000-3,000 preceding events) at which both rank and robust normalizationagree strongly (z = 5-10, versus an empirical background distribution built from up to 40 same-width windows spanning each catalog) ‒ the first result in this investigation series that survivedthe same scrutiny that debunked every earlier lead: it is not explained by a few extrememagnitude values (removing all M≥5 events leaves it largely intact), and it reproducesindependently in a third, entirely unrelated M6-M7 cluster elsewhere in the Tohoku catalog. Wethen tested, as directly as the data allow, whether this is a usable precursor signal, and found it isnot. A window positioned around an M7.1 event that wasnot followed by anything larger for thenext ~3,000 events shows the same elevated signal as windows that were. Formalizing this with aproperly powered sample ‒ all 31 independent M≥6.0 clusters in the Tohoku region catalog, splitinto escalating (n=5) and non-escalating (n=26) groups by whether a larger event followed within3,000 events ‒ a Mann-Whitney U test finds no discriminative difference (rank p=0.42, with theescalating group’s mean actuallylower; robust p=0.26). And the original width-scan result itself,while the most extreme value obtainable given the catalog’s natural background-pool size(p=0.024, exceeding all 40 available background windows), does not survive Bonferroni correctionfor the 10 widths originally scanned (alpha=0.005). We conclude the diagnostic detects real,reproducible, order-dependent structure ‒ consistent with the well-established non-stationarity ofactive aftershock sequences (Omori-Utsu decay, ETAS-style clustering) ‒ but this structure existswhether or not a sequence goes on to produce a larger event, and is therefore not informativeabout which sequences w. ※ The License for codes is AGPL-3.0-or-later