Work

Factorized turnover scores on the corrected NANOGrav 15-yr spectrum: injection-response gain calibration and cross-correction stability

Year: 2026
Type: preprint
Abstract: Factorized, spectrum-space profile log-likelihoods are increasingly used to compare low-frequency turnover templates against pulsar-timing-array free spectra, and their scores are quoted alongside Bayesian evidence ratios although the two objects are not commensurable. On the corrected NANOGrav 15-year free-spectrum posteriors (v2.0.0, released after a parallel-tempering bug), we calibrate this mismatch empirically. An injection-response test, which shifts the per-bin posterior densities by a controlled multiple of the template contrast and re-profiles the likelihood, shows that the factorized score is an amplitude amplifier rather than an evidence: one unit of injected contrast registers as ~+6 in Delta log L for the 2.2 nHz window template families, with the gain set by the posterior widths. The observed turnover contrast of Delta log L = +3.32 ± 0.25 therefore corresponds to ~0.8 template units of actual contrast, and cannot be compared with the collaboration's full-likelihood Bayes factor of 1.46 ± 0.02. Independently, evaluating the same templates on the pre-correction (v1) posteriors shifts the contrasts by no more than 0.7 in Delta log L while shifting the posterior widths by ~40, so the turnover preference survives the sampler correction: it is a property of the likelihood construction, not of the bug or its fix. We name the practice of reading factorized scores as evidences an amplification read, give a two-step detection protocol (injection-response calibration; cross-version contrast check), and release the pipeline so the cross-check is a matter of configuration.
Author: Liu, Wei
Language: English
Open Access status: green
OpenAlex ID: W7235677091
Cites:
Cited by:
Related to: 10