IMRPhenomHM plus/cross pinned-binary holdout — 27 September 2026

Frozen comparison contract

The maintained tests/test_imrphenomhm_lal_polarizations_binary.py compares complex plus and cross from public Torch/JAX IMRPhenomHM with LALSuite 7.26.15 SimInspiralChooseFDWaveform for the same aligned-spin D-carrier model. The test, cases, and bounds in this section were committed before running these holdouts. Only the public LAL binary is queried; no LAL source or coefficient is copied.

London et al., Eq. (1) state the spin-weighted-harmonic polarization sum. García-Quirós et al., Sec. II fix the positive-frequency mode-pair convention. The separate projection specification derives and checks the six angular coefficients. The public LAL waveform API names the inclination and reference phase arguments.

One non-holdout exploratory system (60,20 M_sun, spins 0.5,-0.3, f_ref=31 Hz) established that LAL phiRef corresponds to native phi0 + phic for this API/model; the opposite sign had order-one complex residuals. This fixed convention was selected before the new cases below. LAL has no tc argument in this call, so its output is multiplied by the analytic Fourier shift exp(-2πiftc); no shift is estimated from the data. No per-case, polarization-specific, phase, time, or amplitude alignment is allowed. Native settings use default D carrier and all six modes.

Unseen system	`(m1,m2)` \(M_\odot\)	`(chi1,chi2)`	`f_ref` Hz	Distance Mpc	Inclination	`phi0`, `phic` rad	`tc` s
Mixed generic	`(42,23)`	`(0.31,-0.18)`	31	610	0.73	`(0.21,-0.08)`	0.011
Asymmetric generic	`(55,17)`	`(-0.32,0.41)`	27	930	1.29	`(-0.28,0.37)`	-0.007
Positive-spin backside	`(78,31)`	`(0.52,0.07)`	37	740	2.15	`(0.43,-0.19)`	0.019

LAL uses deltaF=1 Hz, f_min=20 Hz, f_max=512 Hz; both native frontends sample 20–400 Hz inclusive, 381 bins, CPU FP64. The unnormalized complex plus/cross retain absolute distance and polarization scale. For each polarization separately, the predeclared bounds are complex relative L2 = ||native−binary||₂/||binary||₂ ≤ 10⁻³ across all bins and maximum pointwise complex relative residual ≤ 10⁻² where the binary amplitude is at least 1% of that polarization's peak. No overlap maximization is used.

Result and scope

The first held-out run passed 6/6 tests, with no skips or threshold changes. Each cell below is full-band complex L2 / strongest supported-bin pointwise complex residual. The strongest supported-bin test covers 299–365 bins, depending on case and polarization.

System	Polarization	Torch	JAX
Mixed generic	plus	`3.94862954e-8 / 1.06454329e-6`	`3.94863559e-8 / 1.06454248e-6`
Mixed generic	cross	`4.04219292e-8 / 1.07391291e-6`	`4.04219975e-8 / 1.07391214e-6`
Asymmetric generic	plus	`4.26317697e-7 / 8.98561674e-6`	`4.26318119e-7 / 8.98562140e-6`
Asymmetric generic	cross	`3.94396972e-7 / 9.89964493e-6`	`3.94397364e-7 / 9.89964899e-6`
Positive-spin backside	plus	`2.87010133e-6 / 7.82657707e-5`	`2.87010132e-6 / 7.82657697e-5`
Positive-spin backside	cross	`2.76494833e-6 / 5.04790240e-5`	`2.76494832e-6 / 5.04790230e-5`

The run used host len, cwd /home/xangma/tmp/diffgw-hm-polarization-holdout-20260927, Python 3.12.3, LALSuite 7.26.15, Torch 2.14.0+cpu, JAX/JAXlib 0.11.2, NumPy 2.5.3, and pytest 9.1.1. Protocol commit 1a83e2c predates the held-out run. The exact command was:

cd /home/xangma/tmp/diffgw-hm-polarization-holdout-20260927
env JAX_ENABLE_X64=1 JAX_PLATFORMS=cpu DIFFGW_REQUIRE_LAL_BINARY=1 \
  OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 PYTHONPATH=src:tests \
  /home/xangma/.cache/diffgw-taylorf2-hm-venv/bin/python -m pytest -q -s \
  tests/test_imrphenomhm_lal_polarizations_binary.py -ra

The complete output is hm_polarization_binary.log in that cwd, with 6 passed in 17.61s. The launcher PID was 4081252, Python PID 4081253; the job is complete, so no next check or stop action remains. During execution, ssh len 'kill 4081253' was the stop command.

This is a sampled same-model public-binary polarization gate, distinct from the existing six-mode binary holdout and source projection gate. A pass will not establish exact paper fit provenance, derivatives, all parameter/frequency limits, negative-frequency reconstruction, or physical accuracy against NR.