Differences from LAL
Use this page when substituting DiffGW for a LAL waveform or interpreting a reference discrepancy. Shared approximant names do not guarantee identical physics, calibration, conventions or numerical behavior. DiffGW evaluates its own tensor implementations; LAL is an optional comparison reference.
For availability across packages, use the dated package comparison. For DiffGW's inputs and options, use the waveform catalog and parameter conventions. The validation overview explains what reference agreement, frontend consistency and physical validation establish.
Implications for an analysis
This summary separates intentional modeling choices from unresolved agreement. Its evidence comes from the linked configurations and reports. Reference versions, DiffGW revisions and measurement dates belong to those individual evidence records.
| Model or path | Nature of the difference | What to do before using it |
|---|---|---|
| TaylorF2 | The dated FP64 campaign misses its strict direct-error gate, despite very small mismatch. The electric-tidal oracle agrees for nonspinning binaries, but a spinning binary with nonzero deformabilities shows an additional response in the LAL call even at tidal PN order zero. | Inspect both metrics and thresholds; treat spinning-tidal LAL substitution as unqualified. |
| IMRPhenomD / XAS | Agreement is established for the campaign's configurations, with explicit phase and time conventions. | Match options and grids; do not extrapolate the campaign to untested parameters or derivatives. |
| IMRPhenomXP | Agreement depends on precession and reference-phase options; aligned and small-spin residuals remain. | Use the recorded option set below and check your configuration independently. |
| IMRPhenomXHM / XPHM, including SpinTaylor / PNR | Independent higher-mode calibration is intentional. Numerical precession and attachment introduce additional differences. | Treat the carrier and angle prescriptions as scientific model choices; inspect the component checks and their limits below. |
| IMRPhenomT / THM / TP / TPHM | Sampling and epoch conventions differ; comparisons require matching precession and remnant-spin options. Numerical-precession residuals remain unresolved. | Use the model guide and reference report; the strongly spinning 10:1 case is a known concern. |
| TaylorF2_HM | Leading-order higher modes extend the dominant-mode reference. Paper/scalar checks cover all eight; pinned LAL checks the angular basis and only the (2,2) waveform. |
Do not describe angular agreement or the dominant-mode case as exact all-eight external waveform parity. |
| SEOBNR_ROM | The selected data asset defines the surrogate. | Match and record the asset before interpreting a reference comparison. |
Model-specific differences
DiffGW implements waveform physics in its own tensor backends. The following details explain how its coefficients, conventions, calibration and numerical methods relate to LAL, and summarize the available reference evidence.
| Model or family | DiffGW implementation | Reference evidence |
|---|---|---|
| IMRPhenomD / IMRPhenomXAS | Published models with explicit coefficient precision, reference-phase and time-origin conventions. XAS uses the author-supplied symbolic calibration deposit identified in the provenance ledger. | Both DiffGW frontends pass the strict LAL gate for all measured D/XAS configurations in the FP64 campaign. |
| IMRPhenomXP | The clean MSA version-223, final_spin_mod=3 route uses a paper-derived physical opening on one-spin-aligned rows. Valid unequal-mass generic two-spin rows also use the clean source chart at nonzero phi0; the observer-azimuth map is derived from the published frame conventions. Convention 6 applies its quadrupole output phase to clean convention 1. |
Frozen generic convention-1 cases agree closely. The declared generic nonzero-phi0 fixture passes the unchanged strict public-binary gate for conventions 0/1/6. Equal-mass and near-aligned v223 boundary comparisons still fail; the physical aligned convention-6 waveform has a measured residual of (10^{-5})–(5\times10^{-5}). See the boundary qualification and boundary analysis. |
| IMRPhenomXHM / IMRPhenomXPHM | Five modes: (2,2), (2,1), (3,3), (3,2) and (4,4). The four higher modes use independent SXS-based ringdown fits and native Kerr QNM calculations; XPHM adds the paper-backed rotation layer to this declared carrier dependency. |
The calibration differs from LAL's. A 2026-09-25 binary-oracle receipt records four frozen public-binary diagnostics; every strict catalog and holdout gate remains failed. Executable paper-derived math/reference gates are maintained separately in the test suite. |
| TaylorF2_HM | Literature-based frequency-domain PN extension with eight leading-order modes through l=4: (2,2), (2,1), (3,1), (3,2), (3,3), (4,2), (4,3), (4,4). |
Published equations and an independent scalar kernel cover all eight. Pinned LAL externally checks the angular basis, but its TaylorF2 waveform API supplies only the dominant carrier. Exact all-eight external waveform parity, higher-PN completeness, spin-amplitude completeness, and physical accuracy remain unresolved in the manifest. |
| SEOBNR_ROM | Data-backed reduced-order evaluation; the chosen surrogate and calibration asset define the model. Official SEOBNRv4ROM data are external. | Validation applies to the selected asset. See external data for acquisition and third-party notices for unresolved redistribution terms. |
The EOB peak locator solves a natural cubic spline's derivative on accepted samples, selecting an interior maximum and rejecting unresolved or tied maxima. The selected root is locally differentiable, while interval and maximum selection remain piecewise; continuous attachment does not remove integration-step dependence. JAX's non-symmetric Kerr angular eigensolver uses a host CPU callback.
Numerical precession: SpinTaylor and PNR
IMRPhenomXPHM-SpinTaylor (prec_version=320) integrates spin-precession
equations with fixed-grid RK4 in velocity space. IMRPhenomXPNR
(prec_version=330, final_spin_mod=2) is presently a provisional composition
of partially NR-tuned PNR angles with DiffGW's native symmetric XHM
coprecessing carrier. It is not the full 2025 XPNR prescription: transition-spin
mapping, published angle attachment/continuation, XHM-CP coefficient
deformations, effective ringdown frequencies and dominant-mode asymmetry remain
open. The exact component evidence and reduction relationships are in the
XPNR ledger.
The SpinTaylor tests
and PNR tests
use optimized plus-polarization overlaps with model-specific tolerances.
Their scope differs from the direct, both-polarization gate in the
FP64 campaign; these variants
were not timed in that campaign. The XPNR value 0.99 is an acceptance
threshold, not an observed lower bound and not a calibration-domain boundary.
Native IMRPhenomT family
T, THM, TP and TPHM have native time-domain evaluation in both frontends. The precessing variants offer NNLO (default), MSA and numerical prescriptions; reference comparisons must match angle and remnant-spin options. Fixed-grid first derivatives are supported, with restrictions on frequency inversion and automatic duration selection described in the model guide.
The reference-validation report records author checks, LAL regression criteria, measured residuals and diagnostic experiments. Numerical precession retains unresolved reference differences, including the strongly spinning 10:1 case; agreement in the documented cases does not establish accuracy throughout parameter space or independent physical accuracy.
Reproducing model checks
DiffGW's catalog harness retains model-specific references and diagnostic failures:
PYTHONPATH=src python tests/parity_harness.py --scope=analytic
PYTHONPATH=src python tests/parity_harness.py --scope=catalog
Each result records its reference identity, metrics and verification scope. The benchmark protocol sets out how those checks qualify performance results, and the package benchmark report provides timing receipts and reproduction commands.