Skip to content

Reference-Comparison Notes¤

This unlisted page records implementation details behind reference-specific limits. It intentionally excludes run outputs and historical calibration data. For the enforcing thresholds, see Reference Comparisons and Limits.

Frequency-domain comparisons¤

The frequency-domain overlap is a raw same-grid comparison at identical inputs. It is not maximised over time or phase.

Time origin and tc¤

LAL's frequency-domain models take no coalescence time. Each places a model-specific feature at t = 0, and the choice differs between models: IMRPhenomD uses the exact phase derivative at the peak-amplitude frequency, the IMRPhenomX family a fitted peak-time shift (linb), the IMRPhenomPv2 models a spline estimate at the ringdown or merger frequency, and IMRPhenomNSBH the merger-ringdown phase derivative at the requested upper frequency f_max. ripple takes f_max from the model's f_max argument, defaulting to LAL's own default Mf = 0.2; the LAL comparison sets it to the grid's upper edge. ripple reproduces each model's time origin, except for the two IMRPhenomPv2 models described below. The tc entry of the gen_* parameter arrays multiplies the strain by \(e^{-2\pi i f\,t_c}\), which delays the signal by tc under LAL's Fourier convention \(\tilde h(f) = \int h(t)\,e^{-2\pi i f t}\,dt\); the waveform classes fix tc = 0.

IMRPhenomPv2¤

Both LAL and ripple place the ringdown frequency f_RD at t = 0 using \(t_0 = \frac{1}{2\pi}\,\frac{d\Phi}{df}\) at f_RD. LAL estimates the derivative from a natural cubic spline through 10 phase samples on [0.8, 1.2] f_RD (clipped at Mf = 0.2); ripple evaluates it exactly with JAX autodiff. The spline derivative converges to the autodiff value as its grid is refined, so the difference is LAL's interpolation error rather than a modelling choice. The two waveforms differ by a parameter-dependent time shift of order microseconds, which appears as a linear phase ramp and accounts for the whole overlap loss: substituting LAL's spline time shift into ripple reproduces LAL at the float64 floor. The shift is applied identically to both polarizations after the twist-up, so it is equivalent to a shift of the coalescence time. At zero in-plane spin, both implementations remain continuous. IMRPhenomPv2 is nevertheless not required to reduce to IMRPhenomD: its LAL convention swaps the component masses and spins, changing their assignment in asymmetric phase terms.

IMRPhenomPv2_NRTidalv2¤

NRTidalv2 is applied to the co-precessing IMRPhenomD waveform before the precession twist-up, as in LALSimIMRPhenomP.c. Two LAL details must be reproduced: the quadrupole-monopole terms (dQuadMon) that LAL feeds into the IMRPhenomD inspiral phase (shared with IMRPhenomD_NRTidalv2), and the NRTidalv2 merger frequency computed with the mass ratio >= 1. The time origin follows IMRPhenomPv2 at the NRTidalv2 merger frequency: ripple takes the exact derivative of the co-precessing phase, tidal terms included, where LAL uses the same 10-point spline on [0.8, 1.2] f_merger. The resulting microsecond-scale time shift is the only difference from LAL.

IMRPhenomXHM¤

The phase of the (3, 2) mode is sensitive to spheroidal-to-spherical mixing near ringdown. The intermediate phase fit is constrained by the phase derivative at the transition, so small differences there can shift the fitted phase. Investigate a discrepancy in this region as a mode-mixing or phase-derivative issue rather than an amplitude discrepancy.

IMRPhenomXP and IMRPhenomXPHM¤

The MSA precession correction is ill-conditioned near angular-momentum resonances, where terms involving d0 + d2 + d4 suffer cancellation. Small float64 differences in spin-evolution coefficients can then be amplified in the precession angles, particularly at high inclination. IMRPhenomXPHM also inherits the (3, 2) mixing sensitivity through its XHM co-precessing seed. The LAL reference explicitly requests PhenomXPrecVersion=222, so an MSA-initialization failure is surfaced rather than silently selecting a different precession approximation. The BBH prior lands near this ill-conditioning rarely, so both models agree with LAL at the float64 floor.

IMRPhenomXP_NRTidalv3¤

LAL builds this approximant through its XLALSimIMRPhenomXPHM code path, restricted to the (2, ±2) modes. Multibanding is on by default there and must be disabled for the comparison (PhenomXPHMThresholdMband = PhenomXHMThresholdMband = 0, see tests/cross_validation/reference/lal.py); otherwise LAL disagrees with its own multibanding-off output for edge-on samples. The twist cutoff matches LAL bin-for-bin: Mf <= (fCutDef/M_sec)*M_sec, inclusive, with fCutDef in {0.3, 0.33}. The one corner case is chiEff > 0.99, where ripple's co-precessing amplitude is separately zeroed at Mf = 0.3 (fM_CUT in IMRPhenomXAS.py) so Mf in (0.3, 0.33] stays zero while LAL keeps it; it is unreachable within the BNS test prior. The dominant source of overlap loss is the MSA spin-evolution cubic's near-alignment ill-conditioning, which the BNS prior's low spins hit far more often than the BBH prior does. See tests/cross_validation/msa_precession_instability.md for the mechanism and why it shows up as an edge-on-amplified hc-only error that the SNR-weighted combined metric absorbs.

Continuous-wave comparisons¤

ExactPulsarSignal uses direct LALPulsar building blocks because CWMakeFakeData always includes the Einstein and Shapiro delays that the model intentionally omits. CWMakeFakeData instead provides the end-to-end comparison for the full pulsar models, including detector response, barycentering, and orbital modulation where applicable. That reference path interpolates propagation delays and detector response. A delay error produces a phase error that grows with frequency, so its normalized mismatch has a frequency-squared bound. Detector-response interpolation has a separate amplitude effect, which is covered by the relative-norm bound. The bounds account for those reference approximations; ripple should not reproduce them.