EMRI frequency tracks skip waveform generation: harmonic pairs identified 98% of the time among physical candidates, parameter volume cut by up to four more orders
Synopsis
Extreme mass ratio inspirals (EMRIs) have long, multi-harmonic waveforms that make search and parameter estimation expensive. This work trains simulation-based inference networks directly on frequency tracks from a time-frequency search: harmonic pairs are identified correctly in 98% of cases among physically plausible candidates, and sequential inference shrinks the intrinsic-parameter prior volume, with a detection-statistic refinement cutting up to four further orders of magnitude so heavily masked tracks approach the precision of the least-masked cases.
Interpretation
The output of an EMRI search, one or more frequency-evolution tracks, can carry two jobs: deciding which harmonic a track corresponds to, and constraining the intrinsic parameters. Earlier approaches either relied on expensive matched filtering of long waveforms or examined only a few systems to see how much information frequency tracks carry, without a systematic assessment. On an injection set of 10 masking levels with 100 harmonic-track pairs each, the true harmonic pair ranks in the top two, and 98% of cases are identified outright when restricted to physically plausible candidates.
The method combines two candidate tracks under different harmonic labels to recover the azimuthal and radial fundamental frequencies, then passes them to three inference networks. The three networks see only the azimuthal fundamental, only the radial fundamental, and both together; with correct labels they should point to the same parameter region, while wrong labels drive the composite network's sampling acceptance rate very low. A classifier trained on four features, the three posterior medians and the composite acceptance rate, ranks 25000 label assignments under cross-validation.
Sequential inference starts from a broad prior and truncates it round by round to shrink the parameter volume, then a detection statistic picks the highest-scoring samples inside the reduced region. Sequential inference alone stalls under heavy masking, because most parameter information sits in the fast frequency evolution near plunge. For the most heavily masked reference source, adding detection-statistic selection after sequential inference lowers the prior volume ratio by four more orders of magnitude, reaching a final level comparable to the least-masked case.
Perspective
The demonstrated scope is equatorial, prograde eccentric EMRIs using frequency evolution only, without amplitude or phase information; training data consist of two-year, 1262-point fundamental-frequency tracks with noise injected from the residual distribution of an existing semi-coherent search pipeline. The intended readers are researchers working on LISA data analysis and gravitational-wave parameter estimation.
Harmonic identification reaches 98% among physically plausible candidates, but that figure depends on the chosen harmonic label library; when misclassified, the true label still ranks near the top, yet whether sequential inference keeps the same parameter region under a wrong label has not been tested. Detection-statistic selection requires choosing one harmonic track, and at higher eccentricity the (2,0) mode may not be the best choice, introducing bias.