Skip to content

IMRPhenomXHM (3,2) four-target spheroidal ringdown phase solve

Scope and status

xhm.phase.32.spheroidal-four-target-solve.v1 is the conditional algebra of García-Quirós et al.'s Eq. (48): solve the four coefficients in the published Eq. (46) phase derivative from caller-supplied derivative targets. This does not supply or qualify the target fits, the production calibration, or a complete (3,2) mode.

Freeze CPU FP64 coefficient and collocation-residual tolerances rtol=2e-11, atol=5e-10, and full input-Jacobian tolerances rtol=2e-9, atol=5e-7 before comparing any native helper. The independent scalar reference uses 70 decimal digits. Torch eager and JAX JIT must pass on the same rows. A full-rank, reasonably conditioned scaled matrix is a precondition, not a claim about all physical parameters.

Inputs, conventions, and options

Inputs are (r22,r32,d32,G1,G2,G3,G4), with positive dimensionless Mf QNM frequencies and four phase-derivative targets in Table 9 order. The returned coefficients are (alpha2,alpha4,alpha_lambda,dphi_RD). No phase integration constant is solved by Eq. (48). The frozen rows are:

Row r22,r32,d32 G1,G2,G3,G4
1 .080,.095,.012 42,39,35,31
2 .082,.110,.014 -10,-9,-7,-6
3 .070,.087,.010 15,16,18,20

These are synthetic supplied targets, not author-fit evaluations. Row-wise Jacobians cover all seven inputs; derivatives across rows must vanish. A coincident-node case r22=r32-1.5*d32 must fail explicitly in the scalar reference and Torch eager helper; JAX JIT accepts only prevalidated domain inputs. A nearby distinct-node case is diagnostic for conditioning, not a production claim.

Mathematics and components

García-Quirós et al., arXiv:2001.10914v2, Eqs. (46) and (48), Table 9 specify

f = (r22, r32-1.5*d32, r32-0.5*d32, r32+0.5*d32)
G_i = alpha2/f_i² + alpha4/f_i⁴
      + alpha_lambda*d32/[d32²+(f_i-r32)²] + dphi_RD.

To expose numerical rank without changing this equation, let x_i=f_i/r32 and delta=d32/r32. The dimensionless matrix has row B_i=(x_i^-2, x_i^-4, delta/[delta²+(x_i-1)²], 1) and solves B b=G, where b=(alpha2/r32²,alpha4/r32⁴,alpha_lambda/r32,dphi_RD). Recover the printed coefficients by multiplying b by (r32²,r32⁴,r32,1). The gate reports rank(B) and spectral condition number kappa_2(B)=sigma_max/sigma_min on every frozen row. Positive, distinct nodes and nonzero determinant are required; closeness to a coincident node can still make the solve poorly conditioned.

Implementation mapping

The independent 70-digit solve belongs in diffgw_reference.imrphenomxhm_transition_mixing. The existing torchwave.approximants.xhm_phase_clean offers a standalone four-target helper; a JAX counterpart may be checked against the scalar reference. Neither standalone helper is called by the production waveform. The production Torch and JAX branches currently use targets at (r22,r32,r32+0.5*d32) and fix alpha2_32=alpha2_22, so agreement with the conditional four-target solve cannot qualify that branch.

Evidence and limitations

The source archive for arXiv:2001.10914v2 has SHA-256 cc05e7dbff7bf0c42a10b7da1f4840eefc55b252168d6bfb085155a7bdc5318e. Its author phase-fit notebook (SHA-256 48cc263ab9f2837220cd11fbfd99205562b8863af077074f46d730e35382108b) contains expressions for all four paper-era targets. The ancillary .m directory instead exports dphiS_p1.m, dphiS_p4.m, and an alternate dphiS_fring32.m; its 3Dfit_p1_lm_32.m through p4 are intermediate-phase targets. The arXiv non-exclusive distribution license does not establish MIT redistribution permission for the ancillary fits. This component embeds no fitted expressions or values.

The tests can establish Eq. (48) linear-algebra and differentiation fidelity for supplied targets only. They cannot establish that production has the four required fit surfaces, a release-specific mapping from paper-era fits, or a physically validated (3,2) ringdown phase.

Reproduce the checks

PYTHONPATH=src python -m pytest -q \
  tests/test_xhm_32_four_target_ringdown_phase_math_reference.py

The test needs mpmath, numpy, torch, jax, and pytest. It imports no ancillary fit asset. Changing node order, coefficient basis, or calibration requires a new component version.