Effective-One-Body Calibration and Extrapolation

Calibrated effective-one-body (EOB) waveforms are among the most accurate and useful models for compact binaries, but calibration does not turn sparse numerical-relativity coverage into a global guarantee. At the evidence cutoff of 2026-08-20, the SEOBNRv5 family is strongly validated inside and near important aligned-spin training regions and shows useful extrapolation in selected sectors. Accuracy remains conditional on mass ratio, spins, precession, eccentricity, mode content, detector band, signal-to-noise ratio, and the diagnostic used. The appropriate verdict is locally supported extrapolation, not universal validity outside the calibration set.

Helpful background. Calibration, Extrapolation, and Uncertainty explains the EOB construction, mismatch, calibration data, and the distinction between interpolation and extrapolation. This assessment asks what current comparisons actually establish for the SEOBNRv5 family.

Defining an extrapolation claim

A waveform claim must name a model and immutable software version, retained spherical-harmonic modes, intrinsic parameters $(q,\boldsymbol\chi_1,\boldsymbol\chi_2,e)$, and whether the point lies inside the calibration domain or outside its convex hull. It must also name the detector power spectral density, total mass, frequency interval, and error measure. A mismatch averaged over orientations can hide a poor weak mode; a small mismatch at moderate signal-to-noise ratio can still cause bias in an exceptionally loud signal.

Four uncertainty layers should remain separate. Numerical-relativity truncation and extraction error limit the reference waveform. Calibration uncertainty describes imperfectly determined fit coefficients within one ansatz. Structural uncertainty covers omitted mode mixing, precessional dynamics, eccentricity, or other missing physics. Model-family disagreement samples some structural choices but is not a probability distribution over all possible errors.

Assessment at 2026-08-20	Meaning
Question status	Partially resolved
Evidence status	Supported
Confidence	Moderate: extensive local validation coexists with sparse truth data in difficult corners
Domain	SEOBNRv5-family quasi-circular aligned and precessing binaries in declared mode sets and detector bands; no global claim for near-extremal, high-$q$, eccentric, or very-high-SNR systems
What could change the conclusion	Withheld simulations or independent comparisons enter a previously extrapolative region

Entry project: measure model error before exploring a sparse corner

Choose waveform validation if you like numerical analysis and signal processing. You should be able to define a noise-weighted inner product, align complex waveforms, and distinguish parameter fitting from testing. Generating a handful of EOB waveforms and comparing public simulations is a workstation-scale starting point; new numerical-relativity simulations and broad injection inference need an experienced group and a separate resource plan.

Read Pompili et al. 2023, §§ IV–VI and Figs. 8–11, recording the calibration set and the exact unfaithfulness convention. Use the pySEOBNR developers, standard interface and introductory notebook to generate an aligned-spin SEOBNRv5HM waveform in a pinned environment. Begin with the historical 0.3.6 version used by this assessment or explicitly document another version. Select one simulation identified in the paper and obtain its data through the SXS Collaboration, waveform database and accuracy notes. Record simulation version, resolution, extraction method, and mode conventions; select a case with an available resolution comparison.

Deliver a reproducible comparison of dominant-mode phase, one weaker mode, and mismatch over a stated total-mass range. Change tapering and alignment separately, and display numerical-relativity uncertainty alongside model residuals. Classify the case against the actual calibration list: a training waveform is a software and analysis baseline, not a withheld validation.

A candidate extension, with novelty unchecked, is a small set of genuinely withheld cases near one calibration boundary, using the same frozen comparison protocol. Stop if conditioning changes exceed the model difference, the weak mode lies below numerical accuracy, or the supposed withheld case was used in calibration. Pivot to resolving that uncertainty rather than declaring extrapolation failure. Agreement between two fitted models cannot replace an independent comparator.

What present validation establishes

SEOBNRv5HM is not calibrated to a handful of handpicked cases. Pompili and collaborators describe calibration to 442 numerical-relativity simulations plus 13 perturbative waveforms, incorporate second-order self-force information, and report broad aligned-spin multipolar comparisons Pompili et al. 2023, §§ IV–VI and Figs. 8–11. This supports high fidelity over much of the tested domain. It does not make the sparse high-spin and high-mass-ratio corners dense.

The precessing model provides a different test. Ramos-Buades and collaborators compare against 1,543 precessing numerical-relativity simulations even though the precessing construction inherits rather than independently refits the aligned-spin calibration Ramos-Buades et al. 2023, construction, NR comparison, and conclusions. The breadth of this test is persuasive evidence that the reduction and twisting strategy work in many configurations. Antisymmetric modes and highly generic precession remain structurally more delicate.

Calibration uncertainty can now be carried into inference. Pompili, Buonanno, and Pürrer treat calibration coefficients probabilistically and show in injections that marginalizing them can reduce bias for very loud “golden” binaries Pompili, Buonanno, and Pürrer 2025, §§ II–IV and injection figures. This is an important correction to deterministic best-fit practice. It quantifies uncertainty within the calibrated model; it cannot account for a missing physical degree of freedom or a wrong functional form.

Results and open questions at 2026-08-20

Claim or evidence	What it supports	What it does not establish
442 NR plus 13 perturbative calibration waveforms	Strong aligned-spin training and validation base	Uniform density in every parameter-space corner
1,543 precessing NR comparisons	Broad empirical performance of SEOBNRv5PHM	Independent generic-precession calibration or all antisymmetric content
Probabilistic calibration study	Propagation of coefficient uncertainty and reduced injection bias	Structural-model completeness
EOB–2GSF–NR $Q_\omega$ comparison	A useful invariant challenge in an intermediate mass-ratio regime	Direct test of the SEOBNRv5 family or generic spin
pySEOBNR 0.3.6 release	An executable, pinnable software boundary	Accuracy where no truth waveform exists
GWTC-5 stable-release-9 revisions	Real catalog sensitivity to calibration reweighting details	A controlled waveform-accuracy experiment

The large comparison counts deserve weight, but count is not coverage. Many simulations cluster in tractable regions, and correlations among nearby points mean that one withheld corner can be more informative than many neighboring training cases.

Strongest competing interpretations

The practical interpretation is that current EOB models are sufficiently accurate for most present ground-based analyses. This is supported by broad NR comparisons and by successful catalog use. Stable release 9 of the GWTC-5 parameter-estimation products reports calibration-reweighting revisions for many RIFT SEOBNRv5PHM results, while describing most changes as statistically insignificant GWOSC 2026, GWTC-5.0 parameter-estimation and release-note sections. Mature models can therefore be adequate without being immutable.

The stricter interpretation asks whether modelling error is smaller than statistical uncertainty for a named source and detector. That threshold tightens with signal-to-noise ratio and with LISA or next-generation ground detectors. The LISA Waveform Working Group emphasizes that high-SNR science requires accuracy beyond current-catalog adequacy LISA Waveform Working Group 2025, accuracy-requirement and comparable-binary sections. Both interpretations are valid in their domains; conflict arises only when “works for most current events” is presented as “validated everywhere.”

Counterexamples, null results, and failed routes

The primary model paper itself reports tails of larger error for weak higher modes, including configurations with amplitude minima and approximate mode mixing. These do not invalidate the dominant-mode waveform, but they narrow claims about multipolar accuracy. A small total mismatch can give very little protection to a weak mode used for orientation, precession, or strong-field tests.

Albertini and collaborators compare an independent EOB family with second-order self-force and numerical relativity using $Q_\omega$, dephasing, and endpoint diagnostics Albertini et al. 2022, §§ III–V. They find a useful intermediate-mass-ratio overlap alongside deterioration from naive extrapolation and transition-to-plunge effects. Because the model is TEOBResumS rather than SEOBNRv5, this is not a direct failure of SEOBNRv5; it is strong adverse evidence against treating calibration plus resummation as a universal extrapolation certificate.

Accuracy is an observable-dependent statement

Model evaluation should begin with unfaithfulness over a mass and frequency grid, but it should not end there. Mode-by-mode phase and amplitude residuals can reveal compensating errors. $Q_\omega=\omega^2/\dot\omega$ reduces alignment ambiguity and is valuable for inspiral phasing. Injection studies test whether waveform differences shift physical posteriors. Each diagnostic answers a different question.

The public executable boundary at the cutoff is pySEOBNR 0.3.6, released 2026-02-02; live documentation may describe later development code and should not silently replace the stable wheel pySEOBNR release history and documentation, checked 2026-08-20. Reproducibility also requires calibration data, dependency versions, conditioning, mode selection, and numerical-relativity identifiers. A model-name string alone is insufficient.

What would change the assessment

The strongest positive update would be a preregistered set of high-resolution simulations deliberately placed outside the current calibration hull—high mass ratio, near-extremal spin, strong precession, or eccentricity—kept hidden from every fitting team until predictions are frozen. Identical conditioning should compare SEOBNRv5, an independent EOB family, a phenomenological model, and a surrogate. Results should include NR convergence, mode residuals, detector-weighted mismatch across total mass, and parameter bias across realistic signal-to-noise ratios.

A coherent failure across independent simulations and diagnostics would lower support in that region. Conversely, agreement smaller than independently estimated NR uncertainty would expand the validated domain. Adding more calibration cases without a withheld challenge would improve the model but provide weaker extrapolation evidence.

Research checklist

• Pin pySEOBNR 0.3.6 or another immutable release, its dependencies, calibration files, and model/mode names.
• Mark every test point as training, interpolation, near extrapolation, or remote extrapolation using the actual calibration set.
• Record numerical-relativity simulation ID, resolution, extraction radius/order, extrapolation, and estimated truth error.
• Specify PSD, total-mass and frequency grids, tapering, alignment, sky/orientation averaging, and mismatch convention.
• Examine phase and amplitude by mode, $Q_\omega$, and detector-weighted mismatch rather than one scalar alone.
• Inject the worst cases at several signal-to-noise ratios and compare posterior shifts with statistical widths.
• Compare at least two structurally distinct waveform families, without treating their spread as exhaustive uncertainty.
• No independent numerical rerun was performed for this assessment; these steps define a bounded reproduction and challenge protocol.

Source and data map

The two 2023 model papers are the primary calibration and precessing-validation sources. The 2025 calibration-uncertainty paper tests probabilistic propagation within that family. Albertini et al. supply an independent EOB–2GSF–NR diagnostic. The LISA review sets high-SNR requirements. PyPI and the official documentation pin executable pySEOBNR 0.3.6; GWOSC pins GWTC-5 stable-release-9 event products, including Zenodo records 20348005 and its linked Part 2.

Public software permits a bounded waveform comparison, but public code does not create missing truth waveforms in unsimulated corners. A defensible rerun must archive the environment, calibration-data hash, inputs, and analysis scripts. The catalog products document real inference use; they are not a substitute for withheld numerical-relativity validation.

References

• Albertini, A., Nagar, A., Albanesi, S., Bernuzzi, S., Damour, T., and Rettegno, P. “Comparing second-order gravitational self-force, numerical relativity and effective one body waveforms from inspiralling, quasi-circular and nonspinning black hole binaries.” Physical Review D 106, 084061 (2022). Open preprint.
• LIGO Scientific Collaboration, Virgo Collaboration, and KAGRA Collaboration. “GWTC-5.0 open-data documentation and parameter-estimation products.” GWOSC, stable release 9 (2026). Official release.
• LISA Waveform Working Group. “Waveform modelling for the Laser Interferometer Space Antenna.” Living Reviews in Relativity (2025). DOI.
• Pompili, L., Buonanno, A., and Pürrer, M. “Accounting for numerical-relativity calibration uncertainty in gravitational-wave modeling and inference.” Physical Review D 112, 044062 (2025). DOI; open preprint v2.
• Pompili, L., et al. “Laying the foundation of the effective-one-body waveform models SEOBNRv5: improved accuracy and efficiency for spinning non-precessing binary black holes.” Physical Review D 108, 124035 (2023). Open preprint v2.
• pySEOBNR developers. “pySEOBNR 0.3.6.” Python package, released 2 February 2026. PyPI; documentation.
• Ramos-Buades, A., et al. “SEOBNRv5PHM: Next generation of accurate and efficient multipolar precessing-spin effective-one-body waveforms for binary black holes.” Physical Review D 108, 124037 (2023). Open preprint v2.
• SXS Collaboration. “SXS Gravitational Waveform Database.” Public waveform catalog, data format, and accuracy documentation. Official database.