Skip to content

Consistency

The ripple (JAX) backend is an alternative implementation of the same waveform models LAL provides. This page tracks their agreement so the JAX/GPU path can be trusted against the LAL baseline.

For every approximant, the frequency-domain overlap between the ripple and LAL waveforms is computed across several parameter sets on a shared frequency grid and f_ref. The overlap uses Re(⟨a,b⟩) with no time or phase maximization — the implementations must agree exactly, so any residual discrepancy lowers the overlap rather than being optimized away. Results are reported as log₁₀ overlap loss (log₁₀(1 − overlap)); more negative is better, and ≈ −15 is machine precision.

The chart plots both the worst-case and median loss per waveform model, with models sorted so the best (most negative worst case) is on the left. The gwmock-signal version is shown in the tooltip and the table.

TaylorF2 is omitted here (covered in the gwmock-signal test suite). This is a numerical property of the waveforms, independent of the hardware that computed it.

Reproduced in CI

Because the overlap is deterministic and hardware-independent, CI re-runs this suite against the same toolchain and asserts every committed record reproduces (gwmock-benchmark signal verify-consistency). These numbers are therefore not trust-based — a fabricated overlap cannot survive an independent recomputation.

Approximantgwmock-signalf_min (Hz)worst overlapworst log₁₀ lossmedian log₁₀ loss
IMRPhenomD0.9.020.01.000000-7.48-7.87
IMRPhenomD_NRTidalv20.9.040.01.000000-10.23-10.36
IMRPhenomHM0.9.020.01.000000-7.48-7.88
IMRPhenomPv20.9.020.01.000000-6.71-6.76
IMRPhenomXAS0.9.020.01.000000-14.03-14.04
IMRPhenomXAS_NRTidalv30.9.040.01.000000-7.92-8.04
IMRPhenomXHM0.9.020.01.000000-9.35-9.76
IMRPhenomXP0.9.020.01.000000-14.14-14.27
IMRPhenomXPHM0.9.020.01.000000-7.65-7.91

Reproduce / contribute

uv run --extra signal gwmock-benchmark signal consistency -o data/signal/consistency

Then open a pull request adding the data files. See Contribute a benchmark.