- Accepted Paper
Unified matched-filter and excess-power pipeline for gravitational-wave detector characterization
Phys. Rev. D - Accepted 13 November, 2026
DOI: https://doi.org/10.1103/kw6s-517g
Phys. Rev. D - Accepted 13 November, 2026
DOI: https://doi.org/10.1103/kw6s-517g
We present sgnax, an open-source pipeline for gravitational-wave detector characterization that delivers both matched-filter and excess-power transient triggers from a single streaming dataflow graph. Built on the Stream Graph Navigator (sgn) framework, sgnax unifies the multi-rate sine-Gaussian matched-filter search of snax and the multi-resolution Q-transform excess-power search of omicron in one Python-native package with a shared configuration and output schema. Strain and auxiliary channels from a single interferometer can be analyzed as parallel branches of one graph; because the branches share one data-read and whitening front end, the per-channel processing cost falls as channels are added—by a factor of 2.8 between one and 32 channels for the Q-transform—and a full day of 16 kHz strain is analyzed in 14 minutes of wall-clock time. In a 512 s archival-data benchmark, four processes sharing one NVIDIA V100 analyze 792 production auxiliary channels with the Q-transform at 1.11× real-time streaming throughput. Selectable data sources span offline frame caches, shared-memory buffers, and the arrakis distribution service, with one configuration serving offline and online operation. The streaming matched filter carries no per-chunk buffering and delivers triggers at a measured end-to-end latency of about five seconds, while the Q-transform reaches low-latency operation within tens of seconds of acquisition, well ahead of the cadence at which production omicron deployments deliver triggers in practice. Injection campaigns validate both modes, recovering 99.4% of recoverable sine-Gaussian injections with parameters within the expected template mismatch, and broadband white-noise-burst recovery consistent with the established event-trigger generators. On 24 hours of archival LIGO strain, sgnax reproduces the production omicron trigger population to 0.1% in total trigger count, with per-trigger SNRs matched at the percent level in every frequency band and the residual differences localized to a documented difference in whitening adaptation during noise transients; on production auxiliary channels, its dense feature streams recover the snax loud-feature population with 80% per-bin coincidence, a comparison limited less by the reimplementation than by the production reference itself. Indeed, the injection-calibrated reimplementation uncovers a multiband amplitude error in the production snax deployment that inflates its reported SNRs below 25.6 Hz by factors of √2–2, for which we give the mechanism and the correction.
If the author has provided any supplemental materials with this article they will be available upon publication of the version of record.