GW-JAX-Team / ripple / 26692143586

88%

main: 88%

push

github

Fix fM_CUT frequency-cutoff off-by-one in IMRPhenomD and IMRPhenomD_NRTidalv2

The amplitude and phase were zeroed at the grid point one step below the
intended fM_CUT/M_s boundary (argmin()-1), while LAL zeroes from the floor grid point. This produced a single-bin mismatch (ripple=0, LAL=nonzero) that was responsible for all of IMRPhenomD's above-machine-precision overlap loss.

Replace the argmin-based snapping with floor(fM_CUT/M_s/df)*df, which gives the correct floor grid point using only scalar arithmetic (no array ops, no lax.cond). IMRPhenomD OL drops from 5.3e-13 to machine precision (max 3.6e-16, 1000-sample GPU run); threshold tightened from 1e-12 to 1e-15.

3 of 3 new or added lines in 2 files covered. (100.0%)

4618 of 5134 relevant lines covered (89.95%)

0.9 hits per line