Fully data-driven time-delay interferometry with time-varying delays

TL;DR

This paper introduces an advanced, data-driven method called aPCI for canceling laser frequency noise in space-based gravitational-wave data, which adapts to time-varying delays without relying on detailed noise models.

Contribution

It generalizes the aPCI approach to handle slowly varying noise covariance, achieving noise mitigation comparable to second-generation TDI without detailed modeling.

Findings

aPCI effectively mitigates laser frequency noise below other noise levels

Sensitivity of aPCI matches second-generation TDI within 2% error

Method adapts to time-varying delays in LISA-like data

Abstract

Raw space-based gravitational-wave data like LISA's phase measurements are dominated by laser frequency noise. The standard technique to make this data usable for science is time-delay interferometry (TDI), which cancels laser noise terms by forming suitable combinations of delayed measurements. We recently introduced the basic concepts of an alternative approach which, unlike TDI, does not rely on independent knowledge of temporal correlations in the dominant noise. Instead, our automated Principal Component Interferometry (aPCI) processing only assumes that one can produce some linear combinations of the temporally nearby regularly spaced phase measurements, which cancel the laser noise. Then we let the data reveal those combinations. Our previous work relies on the simplifying additional assumption that the filters which lead to the laser-noise-free data streams are time-independent. In LISA, however, these filters will vary as the constellation armlengths evolve. Here, we discuss a generalization of the basic aPCI concept compatible with data dominated by a still unmodeled but slowly varying noise covariance. Despite its independence on any model, aPCI successfully mitigates laser frequency noise below the other noises' level, and its sensitivity to gravitational waves is the same as the state-of-the-art second-generation TDI, up to a 2\% error.